What is with the negativity in these comments? This is a huge, huge surface area that touches a large percentage of white collar work. Even just basic automation/scaffolding of spreadsheets would be a big productivity boost for many employees.
My wife works in insurance operations - everyone she manages from the top down lives in Excel. For line employees a large percentage of their job is something like "Look at this internal system, export the data to excel, combine it with some other internal system, do some basic interpretation, verify it, make a recommendation". Computer Use + Excel Use isn't there yet...but these jobs are going to be the first on the chopping block as these integrations mature. No offense to these people but Sonnet 4.5 is already at the level where it would be able to replicate or beat the level of analysis they typically provide.
Having wrangled many spreadsheets personally, and worked with CFOs who use them to run small-ish businesses, and all the way up to one of top 3 brokerage houses world-wide using them to model complex fixed income instruments... this is a disaster waiting to happen.
Spreadsheet UI is already a nightmare. The formula editing and relationship visioning is not there at all. Mistakes are rampant in spreadsheets, even my own carefully curated ones.
Claude is not going to improve this. It is going to make it far, far worse with subtle and not so subtle hallucinations happening left and right.
The key is really this - all LLMs that I know of rely on entropy and randomness to emulate human creativity. This works pretty well for pretty pictures and creating fan fiction or emulating someone's voice.
It is not a basis for getting correct spreadsheets that show what you want to show. I don't want my spreadsheet correctness to start from a random seed. I want it to spring from first principles.
In my opinion the biggest use case for spread sheet with LLM is to ask them to build python scripts to do what ever manipulations you want to do with the data. Once people learn to do this workplace productivity would increase greatly I have been using LLM for years now to write python scripts that automate different repeatable tasks. Want a pdf of this data to be overlayed on this file create a python script with an LLM. Want the data exported out of this to be formated and tallied create a script for that.
How will people without Python knowledge know that the script is 100% correct? You can say "Well they shouldn't use it for mission critical stuff" or "Yeah that's not a use case, it could be useful for qualitative analysis" etc., but you bet they will use it for everything. People use ChatGPT as a search engine and a therapist, which tells us enough
Yesterday I had to pass a bunch of data to finance as the person that does so had left the company. But they wanted me to basically group by a few columns, so instead of spending an hour on this in excel, I created 3 rows of fake data, gave it to the llm, it created a Python script which I ran against the dataset. After manual verification of the results, it could be submitted to finance.
My first job out of uni was building a spreadsheet infra as code version control system after a Windows update made an eight year old spreadsheet go haywire and lose $10m in a afternoon.
Compared to what? Granted, Excel incidents are probably underreported and might produce "silent" consequential losses. But compared to that, for enterprise or custom software in general we have pretty scary estimates of the damages. Like Y2K (between 300-600bn) and the UK Postal Office thing (~1bn).
Excel spreadsheets ARE custom software, with custom requirements, calculations, and algorithms. They're just not typically written by programmers, have no version control or rollback abilities, are not audited, are not debuggable, and are typically not run through QA or QC.
Agreed. The tradition has been continued by workflow engines, low code tools, platforms like Salesforce and lately AI-builders. The issue is generally not that these are bad, but because they don't _feel_ like software development everyone is comfortable skipping steps of the development process.
To be fair, I've seen shops which actually apply good engineering practices to Excel sheets too. Just definitely not a majority...
Thing is, they are also the common workaround solution for savy office workers that don't want to wait for the IT department if it exists, or some outsourced consultancy, to finally deliver something that only does half the job they need.
So far no one has managed to deliver an alternative to spreedsheets that fix this issue, doesn't matter if we can do much better in Python, Java, C# whatever, if it is always over budget and only covers half of the work.
I know, I have taken part in such project, and it run over budget because there was always that little workflow super easy to do in Excel and they would refuse to adopt the tool if it didn't cover that workflow as well.
exactly. And Claude and other code assistants are more of the same, allowing non-programmers[1] to write code for their needs. And that's a good thing overall.
[1] well, people that don't consider themselves programmers.
I'll add to this - if you work on a software project to port an excel spreadsheet to real software that has all those properties, if the spreadsheet is sophisticated enough to warrant the process, the creators won't be able to remember enough details abut how they created it to tell you the requirements necessary to produce the software. You may do all the calculations right, and because they've always had a rounding error that they've worked around somewhere else, your software shows calculations that have driven business decisions for decades were always wrong, and the business will insist that the new software is wrong instead of owning some mistake. It's never pretty, and it always governs something extremely important.
Now, if we could give that excel file to an llm and it creates a design document that explains everything is does, then that would be a great use of an LLM.
Yeah, that's what OP said. Now add a bunch of random hallucinations hidden inside formulas inside cells.
If they really have a good spreadsheet solution they've either fixed the spreadsheet UI issues or the LLM hallucination issues or both. My guess is neither.
I don't think tools like Claude are there yet, but I already trust GPT-5 Pro to be more diligent about catching bugs in software than me, even when I am trying to be very careful. I expect even just using these tools to help review existing Excel spreadsheets could lead to a significant boost in quality if software is any guide (and Excel spreadsheets seem even worse than software when it comes to errors).
That said, Claude is still quite behind GPT-5 in its ability to review code, and so I'm not sure how much to expect from Sonnet 4.5 in this new domain. OpenAI could probably do better.
> That said, Claude is still quite behind GPT-5 in its ability to review code, and so I'm not sure how much to expect from Sonnet 4.5 in this new domain. OpenAI could probably do better.
It’s always interesting to see others opinions as it’s still so variable and “vibe” based. Personally, for my use, the idea that any GPT-5 model is superior to Claude just doesn’t resonate - and I use both regularly for similar tasks.
I also find the subjective nature of these models interesting, but in this case the difference in my experiences between Sonnet 4.5 and GPT-5 Codex, and especially GPT-5 Pro, for code review is pretty stark. GPT-5 is consistently much better at hard logic problems, which code review often involves.
I have had GPT-5 point out dozens of complex bugs to me. Often in these cases I will try to see if other models can spot the same problems, and Gemini has occasionally but the Claude models never have (using Opus 4, 4.1, and Sonnet 4.5). These are bugs like complex race conditions or deadlocks that involve complex interactions between different parts of the codebase. GPT-5 and Gemini can spot these types of bugs with a decent accuracy, while I’ve never had Claude point out a bug like this.
If you haven’t tried it, I would try the codex /review feature and compare its results to asking Sonnet to do a review. For me, the difference is very clear for code review. For actual coding tasks, both models are much more varied, but for code review I’ve never had an instance where Claude pointed out a serious bug that GPT-5 missed. And I use these tools for code review all the time.
Try the TypeScript codex CLI with the gpt-5-codex model with reasoning always set to high, or GPT-5 Pro with max reasoning. Both are currently undeniably better than Claude Opus 4.1 or Sonnet 4.5 (max reasoning or otherwise) for all code-related tasks. Much slower but more reliable and more intelligent.
I've been a Claude Code fanboy for many months but OpenAI simply won this leg of the race, for now.
Same. I switched from sonnet 4 when it was out to codex. Went back to try sonnet 4.5 and it really hates to work for longer than like 5 minutes at a time
Codex meanwhile seems to be smarter and plugs away at a massive todo list for like 2 hours
Yeah, it's like that commercial for OpenAI (or was it Gemini?) where the guy says it lets the tool work on it's complex financial spreadsheets, goes for a walk with a dog, gets back and it is done with "like 98% accuracy". I cannot imagine what the 2% margin of error looks like for a company that moves around hundreds of billions of dollars...
Having AI create the spreadsheet you want is totally possible, just like generating bash scripts works well. But to get good results, there needs to be some documentation describing all the hidden relationships and nasty workarounds first.
Don't try to make LLMs generate results or numbers, that's bound to fail in any case. But they're okay to generate a starting point for automations (like Excel sheets with lots of formulas and macros), given they get access to the same context we have in our heads.
I like this take. There seems to be an over-focus on 'one-shot' results, but I've found that even the free tools are a significant productivity booster when you focus on generating smaller pieces of code that you can verify. Maybe I'm behind the power curve since I'm not leveraging the full capability of the advanced LLM's, but if the argument is disaster is right around the corner due to potential hallucinations, I think we should consider that you still have to check your work for mission critical systems. That said, I don't really build mission critical systems - I just work in Aerospace Engineering and like building small time saving scripts / macros for other engineers to use. For this use, free LLMs even have been huge for me. Maybe I'm in a very small minority, but I do use Excel & Python nearly every day.
I tend to agree that dropping the tool as it is into untrained hands is going to be catastrophic.
I’ve had similar professional experiences as you and have been experimenting with Claude Code. I’ve found I really need to know what I’m doing and the detail in order to make effective (safe) use out of it. And that’s been a learning curve.
The one area I hope/think it’s closest to (given comments above) is potentially as a “checker” or validator.
But even then I’d consider the extent to which it leaks data, steers me the wrong way, or misses something.
The other case may be mocking up a simple financial model for a test / to bounce ideas around. But without very detailed manual review (as a mitigating check), I wouldn’t trust it.
So yeah… that’s the experience of someone who maybe bridges these worlds somewhat… And I think many out there see the tough (detailed) road ahead, while these companies are racing to monetize.
You would hope so. But how many companies have actually changed their IT policy of outsourcing everything to Tata Consultancy Services (or similar) where a sweaty office in Mumbai full of people who don't give a shit run critical infrastructure?
Jaguar Landrover had production stopped for over a month I think, and 100+ million impact to their business (including a trail of smaller suppliers put near bankruptcy). I'd bet Tata are still there and embedded even further in 5 years.
If AI provides some day-to-day running cost reduction that looks good on quarterly financial statements it will be fully embraced, despite the odd "act of god".
My take is more optimistic. This could be an off ramp to stop putting critical business workflows in spreadsheets. If people start to learn that general purpose programming languages are actually easier than Excel (and with LLMs, there is no barrier), then maybe more robust workflows and automation will be the norm.
I think the world would be a lot better off if excel weren’t in it. For example, I work at business with 50K+ employees where project management is done in a hellish spreadsheet literally one guy in Australia understands. Data entry errors can be anywhere and are incomprehensible. 3 or 4 versions are floating around to support old projects. A CRUD app with a web front end would solve it all. Yet it persists because Excel is erroneously seen as accessible whereas Rails, Django, or literally anything else is witchcraft.
> all LLMs that I know of rely on entropy and randomness to emulate human creativity
Those are tuneable parameters. Turn down the temperature and top_p if you don't want the creativity.
> Claude is not going to improve this.
We can measure models vs humans and figure this out.
To your own point, humans already make "rampant" mistakes. With models, we can scale inference time compute to catch and eliminate mistakes, for example: run 6x independent validators using different methodologies.
One-shot financial models are a bad idea, but properly designed systems can probably match or beat humans pretty quickly.
Not the parent poster, but this is pretty much the foundation of LLMs. They are by their nature probabilistic, not deterministic. This is precisely what the parent is referring to.
All processes in reality, everywhere, are probablistic. The entire reason "engineering" is not the same as theoretical mathematics is about managing these probabilities to an acceptable level for the task you're trying to perform. You are getting a "probablistic" output from a human too. Human beings are not guaranteeing theoretically optimal excel output when they send their boss Final_Final_v2.xlsx. You are using your mental model of their capabilities to inform how much you trust the result.
Building a process to get a similar confidence in LLM output is part of the game.
Yup. It becomes clearer to me when I think about the existing validators. Can these be improved, for sure.
It’s when people make the leaps to the multi-year endgame and in their effort to monetise by building overconfidence in the product where I see the inherent conflict.
It’s going to be a slog… the detailed implementations. And if anyone is a bit more realistic about managing expectations I think Anthropic is doing it a little better.
> All processes in reality, everywhere, are probablistic.
If we want to go in philosophy then sure, you're correct, but this not what we're saying.
For example, an LLM is capable (and it's highly plausible for it to do so) of creating a reference to a non-existent source. Humans generally don't do that when their goal is clear and aligned (hence deterministic).
> Building a process to get a similar confidence in LLM output is part of the game.
Which is precisely my point. LLMs are supposed to be better than humans. We're (currently) shoehorning the technology.
> Humans generally don't do that when their goal is clear and aligned (hence deterministic).
Look at the language you're using here. Humans "generally" make less of these kinds of errors. "Generally". That is literally an assessment of likelihood. It is completely possible for me to hire someone so stupid that they create a reference to a non-existent source. It's completely possible for my high IQ genius employee who is correct 99.99% of the time to have an off-day and accidentally fat finger something. It happens. Perhaps it happens at 1/100th of the rate that an LLM would do it. But that is simply an input to the model of the process or system I'm trying to build that I need to account for.
Not OP but using LLMs in any professional setting, like programming, editing or writing technical specifications, OP is correct.
Without extensive promoting and injectimg my own knowledge and experience, LLMs generate absolute unusable garbage (on average). Anyone who disagrees very likely is not someone who would produce good quality work by themselves (on average). That's not a clever quip; that's a very sad reality. SO MANY people cannot be bothered to learn anything if they can help it.
The triad of LLM dependencies in my view: initiation of tasks, experience based feedback, and consequence sink. They can do none of these, they all connect to the outer context which sits with the user, not the model.
You know what? This is also not unlike hiring a human, they need the hirer party tell them what to do, give feedback, and assume the outcomes.
It's all about context which is non-fungible and distributed, not related to intelligence but to the reason we need intelligence for.
> Anyone who disagrees very likely is not someone who would produce good quality work by themselves (on average).
So for those producing slop and not knowing any better (or not caring), AI just improved the speed at which they work! Sounds like a great investment for them!
For many mastering any given craft might not be the goal, but rather just pushing stuff out the door and paying bills. A case of mismatched incentives, one might say.
I would completely disagree. I use LLMs daily for coding. They are quite far from AGI and it does not appear they are replacing Senior or Staff Engineers any time soon. But they are incredible machines that are perfectly capable of performing some economically valuable tasks in a fraction of the time it would have taken a human. If you deny this your head is in the sand.
Capable, yeah, but not reliable, that's my point. They can one shot fantastic code, or they can one shot the code I then have to review and pull my hair out over for a week, because it's such crap (and the person who pushed it is my boss, for example, so I can't just tell him to try again).
It's not much better with planning either. The amount of time I spent planning, clarifying requirements, hand-holding implementation details always offset any potential savings.
To me, the case for LLMs is strongest not because LLMs are so unusually accurate and awesome, but because if human performance were put on trial in aggregate, it would be found wanting.
Humans already do a mediocre job of spreadsheets, so I don't think it is a given that Claude will make more mistakes than humans do.
But isn't this only fine as long someone who knows what they are doing has oversight and can fix issues when they arise and Claude gets stuck?
Once we all forget how to write SUM(A:A), will we just invent a new kind of spreadsheet once Claude gets stuck?
Or in other words; what's the end game here? LLMs clearly cannot be left alone to do anything properly, so what's the end game of making people not learn anything anymore?
Well the end game with AI is AGI of course. But realistically the best case scenario with LLM’s is having fewer people with the required knowledge, leveraging LLM’s to massively enhance productivity.
We’re already there to some degree. It is hard to put a number on my productivity gain, but as a small business owner with a growing software company it’s clear to me already that I can reduce developer hiring going forward.
When I read the skeptics I just have to conclude that they’re either poor at context building and/or work on messy, inconsistent and poorly documented projects.
My sense is that many weaker developers who can’t learn these tools simply won’t compete in the new environment. Those who can build well designed and documented projects with deep context easy for LLM’s to digest will thrive.
Why isn't there a single study that would back up your observations?
The only study with a representative experimental design that I know about is the METR study and it showed the opposite. Every study citing significant productivity improvements that I've seen is either:
- relying on self-assessments from developers about how much time they think they saved, or
- using useless metrics like lines of code produced or PRs opened, or
- timing developers on toy programming assignments like implementing a basic HTTP server that aren't representative of the real world.
Why is it that any time I ask people to provide examples of high quality software projects that were predominantly LLM-generated (with video evidence to document the process and allow us to judge the velocity), nobody ever answers the call? Would you like to change that?
My sense is that weaker developers and especially weaker leaders are easily impressed and fascinated by substandard results :)
Okay, and now you give those mediocre humans a tool hat is both great and terrible. The problem is, unless you know your way around very well, they won't know which is which.
Since my company uses Excel a lot, and I know the basics but don't want to become an expert, I use LLMs to ask intermediate questions, too hard to answer with the few formulas I know, not too hard for a short solution path.
I have great success and definitely like what I can get with the Excel/LLM combo. But if my colleagues used it the same way, they would not get my good results, which is not their fault, they are not IT but specialists, e.g. for logistics. The best use of LLMs is if you could already do the job without them, but it saves you time to ask them and then check if the result is actually acceptable.
Sometimes I abandon the LLM session, because sometimes, and it's not always easy to predict, fixing the broken result would take more effort than just doing it the old way myself.
A big problem is that the LLMs are so darn confident and always present a result. For example, I point it to a problem, it "thinks", and then it gives me new code and very confidently summarizes what the problem was, correctly, that it now for sure fixed the problem. Only that when I actually try the result has gotten worse than before. At that point I never try to get back to a working solution by continuing to try to "talk" to the AI, I just delete that session and do another, non-AI approach.
But non-experts, and people who are very busy and just want to get some result to forward to someone waiting for it as quickly as possible will be tempted to accept the nice looking and confidently presented "solution" as-is. And you may not find a problem until half a year later somebody finds that prepayments, pro forma bills and the final invoices don't quite match in hard to follow ways.
Not that these things don't happen now already, but adding a tool with erratic results might increase problems, depending on actual implementation of the process. Which most likely won't be well thought out, many will just cram in the new tool and think it works when it doesn't implode right away, and the first results, produced when people still pay a lot of attention and are careful, all look good.
I am in awe of the accomplishments of this new tool, but it is way overhyped IMHO, still far too unpolished and random. Forcing all kinds of processes and people to use it is not a good match, I think.
This is a great point. LLMs make good developers better, but they make bad developers even worse. LLMs multiply instead of add value. So if you're a good developer, who is careful, pays attention, watches out for trouble, and is constantly reviewing and steering, the LLM is multiplying by a positive number and will make you better. However, if you're a mediocre/bad developer, who is not careful, who lacks attention to detail, and just barely gets things to compile / run, then the LLM is multiplying by a negative number and will make your output even worse.
Or you could, you know, read the article before commenting to see the limited scope of this integration?
Anyway, Google has already integrated Gemini into Sheets, and recently added direct spreadsheet editing capability so your comment was disproven before you even wrote it
> The key is really this - all LLMs that I know of rely on entropy and randomness to emulate human creativity. This works pretty well for pretty pictures and creating fan fiction or emulating someone's voice.
I think you need to turn down the temperature a little bit. This could be a beneficial change.
I don't trust LLMs to do the kind of precise deterministic work you need in a spreadsheet.
It's one thing to fudge the language in a report summary, it can be subjective, however numbers are not subjective. It's widely known LLMs are terrible at even basic maths.
Even Google's own AI summary admits it which I was surprised at, marketing won't be happy.
Yes, it is true that LLMs are often bad at math because they don't "understand" it as a logical system but rather process it as text, relying on pattern recognition from their training data.
Seems like you're very confused about what this work typically entails. The job of these employees is not mental arithmatic. It's closer to:
- Log in to the internal system that handles customer policies
- Find all policies that were bound in the last 30 days
- Log in to the internal system that manages customer payments
- Verify that for all policies bound, there exists a corresponding payment that roughly matches the premium.
- Flag any divergences above X% for accounting/finance to follow up on.
Practically this involves munging a few CSVs, maybe typing in a few things, setting up some XLOOKUPs, IF formulas, conditional formatting, etc.
Will AI replace the entire job? No...but that's not the goal. Does it have to be perfect? Also no...the existing employees performing this work are also not perfect, and in fact sometimes their accuracy is quite poor.
Actually, yes. This kind of management reporting is either (1) going to end up in the books and records of the company - big trouble if things have to be restated in the future or (2) support important decisions by leadership — who will be very much less than happy if analysis turns out to have been wrong.
A lot of what ties up the time of business analysts is ticking and tying everything to ensure that mistakes are not made and that analytics and interpretations are consistent from one period to the next. The math and queries are simple - the details and correctness are hard.
Is this not belligerently ignoring the fact that this work is already done imperfectly? I can’t tell you how many serious errors I’ve caught in just a short time of automating the generation of complex spreadsheets from financial data. All of them had already been checked by multiple analysts, and all of them contained serious errors (in different places!)
No belligerence intended! Yes, processes are faulty today even with maker-checker and other QA procedures. To me it seems the main value of LLMs in a spreadsheet-heavy process is acceleration - which is great! What is harder is quality assurance - like the example someone gave regarding deciding when and how to include or exclude certain tables, date ranges, calc, etc. Properly recording expert judgment and then consistently applying that judgement over time is key. I’m not sure that is the kind of thing LLMs are great at, even ignoring their stochastic nature. Let’s figure out how to get best use out of the new kit - and like everything else, focus on achieving continuously improving outcomes.
There’s actually different classes of errors though. There’s errors in the process itself versus errors that happen when performing the process.
For example, if I ask you to tabulate orders via a query but you forgot to include an entire table, this is a major error of process but the query itself actually is consistently error-free.
Reducing error and mistakes is very much modeling where error can happen. I never trust an LLM to interpret data from a spreadsheet because I cannot verify every individual result, but I am willing to ask an LLM to write a macro that tabulates the data because I can verify the algorithm and the macro result will always be consistent.
Using Claude to interpret the data directly for me is scary because those kinds of errors are neither verifiable nor consistent. At least with the “missing table” example, that error may make the analysis completely bunk but once it is corrected, it is always correct.
Speak for yourself and your own use cases. There are a huge diversity of workflows with which to apply automation in any medium to large business. They all have differing needs. Many excel workflows I'm personally familiar with already incoporate a "human review" step. Telling a business leader that they can now jump straight to that step, even if it requires 2x human review, with AI doing all of the most tediuous and low-stakes prework, is a clear win.
I'm taking a much weaker position than the respondent: LLMs are useful for many classes of problem that do not require zero shot perfect accuracy. They are useful in contexts where the cost of building scaffolding around them to get their accuracy to an acceptable level is less than the cost of hiring humans to do the same work to the same degree of accuracy.
>LLMs are useful for many classes of problem that do not require zero shot perfect accuracy. They are useful in contexts where the cost of building scaffolding around them to get their accuracy to an acceptable level is less than the cost of hiring humans to do the same work to the same degree of accuracy.
Well said. Concise and essentially inarguable, at least to the extent it means LLMs are here to stay in the business world whether anyone likes it or not (barring the unforeseen, e.g. regulation or another pressure).
Sometimes there can be an advantage in leading or lagging some aspects of internal accounting data for a time period. Basically sitting on credits or debits to some accounts for a period of weeks. The tacit knowledge to know when to sit on a transaction and when to action it is generally not written down in formal terms.
I'm not sure how these shenanigans will translate into an ai driven system.
That’s the kind of thing that can get a company into a lot of trouble with its auditors and shareholders. Not that I am offering accounting advice of course. And yeah, one can not “blame” and ai system or try to ai-wash any dodgy practices.
Checking someone elses spreadsheet is a fucking nightmare. If your company has extremely good standards it's less miserable because at least the formatting etc will be consistent...
The one thing LLMs should consistently do is ensure that formatting is correct. Which will help greatly in the checking process. But no, I generally don't trust them to do sensible things with basic formulation. Not a week ago GPT 5 got confused whether a plus or a minus was necessary in a basic question of "I'm 323 days old, when is my birthday?"
I think you have a misunderstanding of the types of things that LLMs are good at. Yes you're 100% right that they can't do math. Yet they're quite proficient at basic coding. Most Excel work is similar to basic coding so I think this is an area where they might actually be pretty well suited.
My concern would be more with how to check the work (ie, make sure that the formulas are correct and no columns are missed) because Excel hides all that. Unlike code, there's no easy way to generate the diff of a spreadsheet or rely on Git history. But that's different from the concerns that you have.
The model ought to be calling out to some sort of tool to do the math—effectively writing code, which it can do. I'm surprised the major LLM frontends aren't always doing this by now.
MS Office Tools menu has a "Spreadsheet Compare" application. It is quite good for diffing 2 spreadsheets. Of course it cannot catch logic errors, human or ML.
I've built spreadsheet diff tools on Google sheets multiple times. As the needs grows I think we will see diffs and commits and review tools reach customers
hey Collin! I am working on an AI agent on Google Sheets, I am curious if any of your designs are out in the public. We are trying to re-think how diffs should look like and want to make something nicer than what we currently have, so curious.
Hi! Nothing public nor generic enough to be a good building block. I found myself often frustrated by the tools that came out of the box but I believe better apis could make this slightly easier to solve.
The UX of spreadsheet diffs is a hard one to solve because of how weird the calculation loops are and how complicated the relationship between fields might be.
I've never tried to solve this for a real end user before in a generic way - all my past work here was for internal ability to audit changes and rollback catastrophes. I took a lot of shortcuts by knowing which cells are input data vs various steps of calculations -- maybe part of your ux is being able to define that on a sheet by sheet basis? Then you could show how different data (same formulas) changed outputs or how different formulas (same data) did differently?
Spreadsheets are basically weird app platforms at this point so you might not be able to create a single experience that is both deep and generic. On the other hand maybe treating it as an app is the unlock? Get your AI to noodle on what the whole thing is for, then show diff between before and after stable states (after all calculation loops stabilize or are killed) side by side with actual diffs of actual formulas? I feel like Id want to see a diff as a live final spreadsheet and be able to click on changed cells and see up the chain of their calculations to the ancestors that were modified.
Fun problem that sounds extremely complicated. Good luck distilling it!
So do it in basic code where numbering your line G53 instead of G$53 doesn't crash a mass transit network because somebody's algorithm forgot to order enough fuel this month.
Maybe LLMs will enable a new type of work in spreadsheets. Just like in coding we have PR reviews, with an LLM it should be possible to do a spreadsheet review. Ask the LLM to try to understand the intent and point out places where the spreadsheet deviates from the intent. Also ask the LLM to narrate the spreadsheet so it can be understood.
That first condition "try to understand the intent" is where it could go wrong. Maybe it thinks the spreadsheet aligns with the intent, but it misunderstood the intent.
LLMs are a lossy validation, and while they work sometimes, when they fail they usually do so 'silently'.
Maybe we need some kind of method, framework to develop intent. Most of things that go wrong in knowledge working are down to lack of common understanding of intent.
> The one thing LLMs should consistently do is ensure that formatting is correct.
In JavaScript (and I assume most other programming languages) this is the job of static analysis tools (like eslint, prettier, typescript, etc.). I’m not aware of any LLM based tools which performs static analysis with as good a results as the traditional tools. Is static analysis not a thing in the spreadsheet world? Are there the tools which do static analysis on spreadsheets subpar, or offer some disadvantage not seen in other programming languages? And if so, are LLMs any better?
Just use a normal static analysis tool and shove the result to an LLM. I believe Anthropic properly figured that agents are the key, in addition to models, contrary to OpenAI that is run by a psycho that only believes in training the bigger model.
Sysadmin of a small company. I get asked pretty often to help with a pivot table, vlookup, or just general excel functions (and smartsheet, these users LOVE smartsheet)
Last time, I gave claude an invoice and asked it to change one item on it, it did so nicely and gave me the new invoice. Good thing I noticed it had also changed the bank account number..
The more complicated the spreadsheet and the more dependencies it has, the greater the room for error. These are probabilistic machines. You can use them, I use them all the time for different things, but you need to treat them like employees you can't even trust to copy a bank account number correctly.
We’ve tried to gently use them to automate some of our report generation and PDF->Invoice workflows and it’s a nightmare of silent changes and absence of logic.. basic things like specifically telling it “debits need to match credits” and “balance sheets need to balance” that are ignored.
Yeah, asking llm to edit one specific thing in a large or complex document/ codebase is like those repeated "give me the exact same image" gifs. It's fundamentally a statistical model so the only thing we can be _certain_ of is that _it's not_. It might get the desired change 100% correct but it's only gonna get the entire document 99 5%
Something that Claude Sonnet does when you use it to code is write scripts to test whether or not something is working. If it does that for Excel (e.g. some form of verification) it should be fine.
Besides, using AI is an exercise in a "trust but verify" approach to getting work done. If you asked a junior to do the task you'd check their output. Same goes for AI.
The use cases for spreadsheets are much more diverse than that. In my experience, spreadsheets just as often used for calculation. Many of them do require high accuracy, rely on determinism, and necessitate the understanding of maths ranging from basic arithmetic to statistics and engineering formulas. Financial models, for example, must be built up from ground truth and need to always use the right formulas with the right inputs to generate meaningful outputs.
I have personally worked with spreadsheet based financial models that use 100k+ rows x dozens of columns and involve 1000s of formulas that transform those data into the desired outputs. There was very little tolerance for mistakes.
That said, humans, working in these use cases, make mistakes >0% of the time. The question I often have with the incorporation of AI into human workflows is, will we eventually come to accept a certain level of error from them in the way we do for humans?
I don’t see the issue so much as the deterministic precision of an LLM, but the lack of observability of spreadsheets. Just looking at two different spreadsheets, it’s impossible to see what changes were made. It’s not like programming where you can run a `git diff` to see what changes an LLM agent made to a source code file. Or even a word processing document where the text changes are clear.
Spreadsheets work because the user sees the results of complex interconnected values and calculations. For the user, that complexity is hidden away and left in the background. The user just sees the results.
This would be a nightmare for most users to validate what changes an LLM made to a spreadsheet. There could be fundamental changes to a formula that could easily be hidden.
For me, that the concern with spreadsheets and LLMs - which is just as much a concern with spreadsheets themselves. Try collaborating with someone on a spreadsheet for modeling and you’ll know how frustrating it can be to try and figure out what changes were made.
I like to use Claude Code to write deterministic computer programs for me, which then perform the actual work. It saves a lot of time.
I had a big backlog of "nice to have scripts" I wanted to write for years, but couldn't find the time and energy for. A couple of months after I started using Claude Code, most of them exist.
That’s great and the only legitimate use case here. I suspect Microsoft will not try to limit customers to just writing scripts and will instead allow and perhaps even encourage them to let the AI go ham on a bunch of raw data with no intermediary code that could be reviewed.
Most real-world spreadsheets I've worked with were fragile and sloppy, not precise and deterministic. Programmers always get shocked when they realize how many important things are built on extremely messy spreadsheets, and that people simply accept it. They rather just spend human hours correcting discrepancies than trying to build something maintainable.
Usually this is very hard because the tasks and the job often subtly shifts in somewhat unpredictable and unforeseen ways and there is no neat clean abstraction that you can just implement as an application. Too hererogeneous, too messy, too many exceptions. If you develop some clean elegant solution, next week there will be something that your shiny app doesn't allow and they'd have to submit a feature request or whatever.
In Excel, it's possible to just ad hoc adjust things and make it up as you go. It's not clean but very adaptable and flexible.
"I don't trust LLMs to do the kind of precise deterministic work" => I think LLM is not doing the precise arithmetic. It is the agent with lots of knowledge (skills) and tools. Precise deterministic work is done by tools (deterministic code). Skills brings domain knowledge and how to sequence a task. Agent executes it. LLM predicts the next token.
Sure, but this isn't requiring that the LLM do any math. The LLM is writing formulas and code to do the math. They are very good at that. And like any automated system you need to review the work.
Exactly, and if it can be done in a way that helps users better understand their own spreadsheets (which are often extremely complex codebases in a single file!) then this could be a huge use case for Claude.
Do you trust humans to be precise and deterministic, or even to be especially good at math?
This is talking about applying LLMs to formula creation and references, which they are actually pretty good at. Definitely not about replacing the spreadsheet's calculation engine.
Eh, yes. In theory. In practice, and this is what I have experienced personally, bosses seem to think that you now have interns so you should be able to do 5x the output.. guess what that means. No verification or rubber stamp.
>I don't trust LLMs to do the kind of precise deterministic work
not just in a spreadsheet, any kind of deterministic work at all.
find me a reliable way around this. i don't think there is one. mcp/functions are a band aid and not consistent enough when precision is important.
after almost three years of using LLMs, i have not found a single case where i didn't have to review its output, which takes as long or longer than doing it by hand.
ML/AI is not my domain, so my knowledge is not deep nor technical. this is just my experience. do we need a new architecture to solve these problems?
ML/AI is not my domain but you don’t have to get all that technical to understand that LLMs run on probability. We need a new architecture to solve these problems.
If only we had created some device that could perform deterministic calculations and then wrote software that made it easy for humans to use such calculations.
ok but humans are idiots, if only we could make some sort of Alternate Idiot, a non-human but every bit as generally stupid as humans are! This A.I would be able to do every stupid thing humans did with the device that performed deterministic calculations only many times faster!
Yes and when the AI did that all the stupid humans could accept its output without question. This would save the humans a lot of work and thought and personal responsibility for any mistakes! See also Israel’s Lavender for an exciting example of this in action.
They already doing that with AI, rejecting claims at higher numbers than before .
Privatized insurance will always find a way to pay out less if they could get away with it . It is just nature of having the trifecta of profit motive , socialized risk and light regulation .
If you think that insurance companies have "light regulation", I shudder to think of what "heavy regulation" would look like. (Source: I'm the CTO at an insurance company.)
Light did not mean to imply quantity of paperwork you have to do, rather are you allowed to do the things you want to do as a company.
More compliance or reporting requirements usually tend to favor the larger existing players who can afford to do it and that is also used to make the life difficult and reject more claims for the end user.
It is kind of thing that keeps you and me busy, major investors don't care about it all, the cost of the compliance or the lack is not more than a rounding number in the balance, the fines or penalties are puny and laughable.
The enormous profits year on year for decades now, the amount of consolidation allowed in the industry show that the industry is able to do mostly what they want pretty much, that is what I meant by light regulation.
I'm not sure we're looking at the same industry. Overall, insurance company profit margins are in the single digits, usually low single digits - and in many segments, they're frequently not profitable at all. To take one example, 2024 was the first profitable year for homeowners insurance companies since 2019, and even then, the segment's entire profit margin was 0.3% (not 3% - 0.3%).
So it's weird to see folks on a tech site talking about how enormous all the profits are in health insurance, and citations with numbers would be helpful to the discussion.
I worked in insurance-related tech for some time, and the providers (hospitals, large physician groups) and employers who actually pay for insurance have signficant market power in most regions, limiting what insurers can charge.
I agree, and I can see where it comes from (at least at the state level).
The cycle is: bad trend happens that has deep root causes (let's say PE buying rural hospitals because of reduced Medicaid/Medicare reimbursements); legislators (rightfully) say "this shouldn't happen", but don't have the ability to address the deep root causes so they simply regulate healthcare M&As – now you have a bandaid on a problem that's going to pop up elsewhere.
I mean even in the simple stuff like denying payment for healthcare that should have been covered. CMS will come by and out a handful of cases, out of millions, every few years.
So obviously the company that prioritizes accuracy of coverage decisions by spending money on extra labor to audit itself is wasting money. Which means insureds have to waste more time getting the payment for healthcare they need.
I should have been clearer - profit maximization above all else as long it is mostly legal. Neither profit or profit maximization at all cost is nature of everything .
There are many other entity types from unions[1], cooperatives , public sector companies , quasi government entities, PBC, non profits that all offer insurance and can occasionally do it well.
We even have some in the US and don’t think it is communism even - like the FDIC or things like social security/ unemployment insurance.
At some level government and taxation itself is nothing but insurance ? We agree to paying taxes to mitigate against variety of risks including foreign invasion or smaller things like getting robbed on the street.
[1] Historically worker collectives or unions self-organized to socialize the risks of both major work ending injuries or death.
Ancient to modern armies operate on because of this insurance the two ingredients that made them not mercenaries - a form of long term insurance benefit (education, pension, land etc) or family members in the event of death and sovereign immunity for their actions.
That would be illegal though, the goal is do this legally after all.
We also have to remember all claims aren't equal. i.e. some claims end up being way costlier than others. You can achieve similar % margin outcomes by putting a ton of friction like, preconditions, multiple appeals processes and prior authorization for prior authorization, reviews by administrative doctors who have no expertise in the field being reviewed don't have to disclose their identity and so and on.
While U.S. system is most extreme or evolved, it is not unique, it is what you get when you end up privatize insurance any country with private insurance has some lighter version of this and is on the same journey .
Not that public health system or insurance a la NHS in UK or like Germany work, they are underfunded, mismanaged with long times in months to see a specialist and so on.
We have to choose our poison - unless you are rich of course, then the U.S. system is by far the best, people travel to the U.S. to get the kind of care that is not possible anywhere else.
> While U.S. system is most extreme or evolved, it is not unique, it is what you get when you end up privatize insurance any country with private insurance has some lighter version of this and is on the same journey .
I disagree with the statement that healthcare insurance is predominantly privatized in the US: Medicare and Medicaid, at least in 2023, outspent private plans for healthcare spending by about ~10% [1]; this is before accounting for government subsidies for private plans. And boy, does America have a very unique relationship with these programs.
It is more nuanced, for example Medicare Advantage(Part C) is paid by Medicare money but it is profitable private operators who provide the plans and service it a fast growing part of Medicare .
John Oliver had an excellent segment coincidentally yesterday on this topic.
While the government pays for it, it is not managed or run by them so how to classify the program as public or private ?
This is a great application of this quote. Insurance providers have 0 incentive to make their AI "good" at processing claims, in fact it's easy to see how "bad" AI can lead to a justification to deny more claims.
The question is how you define good. They surely want the Ai to be good in the sense that it rejects all claims that they think can get away with rejecting. But it should not reject those where rejection likely results in litigation and losing and having to pay damages.
Yes it's surprising to see so much cynicism for something that has a real possibility of making so many people so much more productive.
My mental model of the average excel user is of someone who doesn't care about excel, but cares about their business. If Claude can help them use excel and learn about their business faster, then this should make the world more productive and we all get richer.
Claude can make mistakes, but it's not clear to me why people think that the ratio of results to mistakes will get worse here.
I think there are many possible reasons why this could not work out, but many of the comments here just seem like unfounded cynicism.
HN has a base of strong anti-AI bias, I assume is partially motivated by insecurity over being replaced, losing their jobs or having missed the boat on the AI.
I use AI every day. Without oversight, it does not work well.
If it doesn't work well, I will do it myself, because I care that things are done well.
None of this is me being scared of being replaced; quite the opposite. I'm one of the last generations of programmers who learned how to program and can debug and fix the mess your LLM leaves behind when you forgot to add "make sure it's a clean design and works" to the prompt.
Okay, that's maybe hyperbole, but sadly only a little bit. LLMs make me better at my job, they don't replace me.
Based on the comments here, it's surprisingly anything in society works at all. I didn't realize the bar was "everything perfect every time, perfectly flexible and adaptable". What a joy some of these folks must be to work with, answering every new technology with endless reasons why it's worthless and will never work.
I think perhaps you underestimate how antithetical the current batch of LLM AI's is to what most programmers strive for every day, and what we want from our tools. Its not about losing our job, its about "correctness". (or as said below - deterministic)
In a lot of jobs, particularly in creative industries, or marketing, media and writing, the definition of a job well done is a fairly grey area. I think AI will be mostly disruptive in these areas.
But in programming there is a hard minimum of quality. Given a set of inputs, does the program return the correct answer or not? When you ask it what 2+2, do you get 4?
When you ask AI anything, it might be right 50% of the time, or 70% of the time, but you can't blindly trust the answer. A lot of us just find that not very useful.
I am a SWE myself and use LLMs to write ~100% of my code. That does not mean I fire and forget multiplexed codex instances. Many times I step through and approve every edit. Even if it was nothing but a glorified stenographer - there are substantial time savings in being able to prototype and validate ideas quickly.
> But in programming there is a hard minimum of quality. Given a set of inputs, does the program return the correct answer or not? When you ask it what 2+2, do you get 4?
Whether something works or not matters less than whether someone will pay for it.
HN constantly points out the flaws, gaps, and failings of AI. But the same is true of any technology discussed on HN. You could describe HN as having an anti-technology bias, because HN complains about the failings of tech all day every day.
Quite the opposite, actually. You can always find five stories on the front page about some AI product or feature. Meanwhile, you have people like yourself who convince themselves that any pushback is done by people who just don't see the true value of it yet and that they're about to miss out!! Some kind of attempt at spreading FOMO, I guess.
If anything, HN, has a pro-AI bias. I don't know of any other medium where discussions about AI consistently get this much frontpage time, this amount of discussion, and this many people reporting positive experiences with it. It's definitely true that HN isn't the raging pro-AI hypetrain it was two years ago, but that shouldn't be mistaken for "strong anti-AI bias".
Outside of HN I am seeing, at best, an ambivalent reaction: plenty of people are interested, almost everyone tried it, very few people genuinely like it. They are happy to use it when it is convenient, but couldn't care less if it disappeared tomorrow.
There's also a small but vocal group which absolutely hates AI and will actively boycott any creative-related company stupid enough to admit to using it, but that crowd doesn't really seem to hang out on HN.
I really don’t think this is accurate. I think the median opinion here is to be suspicious of claims made about AI, and I don’t think that’s necessarily a bad thing. But I also regularly see posts talking about AI positively (e.g. simonw), or talking about it negatively. I think this is a good thing, it is nice to have a diversity of opinions on a technology. It's a feature, not a bug.
HN has an obsession with quality too, which has merit, but is often economically irrelevant.
When US-East-1 failed, lots of people talked about how the lesson was cloud agnosticism and multi cloud architecture. The practical economic lesson for most is that if US-East-1 fails, nobody will get mad at you. Cloud failure is viewed as an act of god.
Anti-AI bias is motivated by the waste of natural resources due to a handful of non-technical douchebag tech bros.
Everything isn't about money, I know that status and power are all you ai narcissists dream about. But you'll never be Bill Gates, nor will you be Elon Musk.
Once ai has gone the way of "Web3", "NFTs", "blockchain", "3D tvs", etc; You'll find a new grift to latch your life savings onto.
Non-reproducability is the biggest issue here. You deliver a report in 5 minutes to CFO, he comes back after lunch, gives you updated data to adjust a bit of a report and 5 minutes later gets a new report that has some non related to update number changed and asks why? what do you do?
The vast majority of people in business and science are using spreadsheets for complex algorithmic things they weren't really designed for, and we find a metric fuckton of errors in the sheets when you actually bother looking auditing them, mistakes which are not at all obvious without troubleshooting by... manually checking each and every cell & cell relation, peering through parentheses, following references. It's a nightmare to troubleshoot.
LLMs specialize in making up plausible things with a minimum of human effort, but their downside is that they're very good at making up plausible things which are covertly erroneous. It's a nightmare to troubleshoot.
There is already an abject inability to provision the labor to verify Excel reasoning when it's composed by humans.
I'm dead certain that Claude will be able to produce plausibly correct spreadsheets. How important is accuracy to you? How life-critical is the end result? What are your odds, with the current auditing workflow?
Okay! Now! Half of the users just got laid off because management thinks Claude is Good Enough. How about now?
Not really, they're only as good as their context and they do miss and forget important things. It doesn't matter how often, because they do, and they will tell you with 100% confidence and with every synonym of "sure" that they caught it all. That's the issue.
I am very confident that these tools are better than the median programmer at code review now. They are certainly much more diligent. An actually useful standard to compare them to is human review, and for technical problems, they definitely pass it. That said, they’re still not great at giving design feedback.
But GPT-5 Pro, and to a certain extent GPT-5 Codex, can spot complex bugs like race conditions, or subtly incorrect logic like memory misuse in C, remarkably well. It is a shame GPT-5 Pro is locked behind a $200/month subscription, which means most people do not understand just how good the frontier models are at this type of task now.
Yeah, this could be a pretty big deal. Not everyone is an excel expert, but nearly everyone finds themselves having to work with data in excel at some time or other.
A lot of us have seen the effects of AI tools in the hands of people who don't understand how or why to use the tools. I've already seen AI use/misuse get two people fired. One was a line-of-business employee who relied on output without ever checking it, got herself into a pretty deep hole in 3 weeks. Another was a C suite person who tried to run an AI tool development project and wasted double their salary in 3 months, nothing to show for it but the bill, fired.
In both cases the person did not understand the limits of the tools and kept replacing facts with their desires and their own misunderstanding of AI. The C suite person even tried to tell a vendor they were wrong about their own product because "I found out from AI".
AI right now is fireworks. It's great when you know how to use it, but if you half-ass it you'll blow your fingers off very easily.
> but these jobs are going to be the first on the chopping block as these integrations mature.
I'm not even sure that has to be true anymore. From my admittedly superficial impression of the page, this appears to be a tool for building tools. There are plenty of organizations that are resource constrained, that are doing things the way they have always done thing in Excel, simply because they cannot allocate someone to modify what is already in place to better suit their current needs. For them, this is more of a quality of life and quality of out improvement. This is not like traditional software development, where organizations are far more likely to purchase a product or service to do a job (and where the vendors of those products and services are going to do their best to eliminate developers).
It is bad in a very specific sense, but I did not see any other comments express the bad parts instead of focusing merely on the accuracy part ( which is an issue, but not the issue ):
- this opens up ridiculous flood of data that would otherwise be semi-private to one company providing this service
- this works well small data sets, but will choke on ones it will need to divvy up into chunks inviting interesting ( and yet unknown ) errors
There is a real benefit to being able to 'talk to data', but anyone who has seen corporate culture up close and personal knows exactly where it will end.
edit: an i saying all this as as person, who actually likes llms.
Probably because many people here are software developers, and wrapping spreadsheets in deterministic logic and a consistent UI covers... most software use cases.
What does scaffolding of spreadsheets mean? I see the term scaffolding frequently in the context of AI-related articles and not familiar with this method and I’m hesitant to ask an LLM.
Scaffolding typically just refers to a larger state machine style control flow governing an agent's behavior and the suite of external tools it has access to.
Anthropic now has all your company's data, and all you saved was the cost of one human minus however much they charge for this. The good news is it can't have your data again! So starting from the 163rd-165th person you fire, you start to see a good return and all you've sacrificed is exactitude, precision, judgement, customer service and a little bit of public perception!
in the short run. In the long run, productivity gains benefit* all of us (in a functional market economy).
*material benefit. In terms of spirit and purpose, the older I get the more I think maybe the Amish are on to something. Work gives our lives purpose, and the closer the work is to our core needs, the better it feels. Labor saving so that most of us are just entertaining each other on social networks may lead to a worse society (but hey, our material needs are met!)
My theory: a lot of software we build is the supposed solve for a 'crappy spreadsheet'. a) that isnt' much of a moat, b) you're watching generalization of software happen in real time.
Crappy spreadsheet is just the codification of business processes. Those are inherently messy and there's lots of assumptions, lots of edge cases. That's why spreadsheets tend towards crappy on a long enough timeline. It's a fundamentally messy problem.
Spreadsheets are an abstraction over a messy reality, lossy. They were already generalizing reality.
Now we generalize the generalization. It is this lossy reality that people are worried about with AI in HN.
The biggest problem with spreadsheets is that they tend to be accounts for the accumulation of technical debt, which is an area that AI tools are not yet very good at retiring, but very good at making additional withdrawals from.
I have to admit that my first thought was “April’s fool”. But you are right. It makes a lot of sense (if they can get it to work well). Not only is Excel the world’s biggest “programming language”. It’s probably also one of the most unintuitive ways to program.
The issue isn’t in creating a new monstrosity in excel.
The issue is the poor SoB who has to spelunk through the damn thing to figure out what it does.
Excel is the sweet spot of just enough to be useful, capable enough to be extensible, yet gated enough to ensure everyone doesn’t auto run foreign macros (or whatever horror is more appropriate).
In the simplest terms - it’s not excel, it’s the business logic. If an excel file works, it’s because theres someone who “gets” it in the firm.
I used to live in Excel too. I've trudged through plenty of awful worksheets. The output I've seen from AI is actually more neatly organized than most of what I used to receive in outlook. Most of that wasn't hyper-sophisticated cap table analyses. It was analysis from a Jr Analyst or line employee trying to combine a few different data sources to get some signal on how XYZ function of the business was performing. AI automation is perfectly suitable for this.
Neat formatting didn't save any model from having the wrong formula pasted in.
Being neat was never a substitute for being well rested, or sufficiently caffeinated.
Have you seen how AI functions in the hands of someone who isn't a domain expert? I've used it for things I had no idea about, like Astro+ web dev. User ignorance was magnified spectacularly.
This is going to have Jr Analysts dumping well formatted junk in email boxes within a month.
> offense to these people but Sonnet 4.5 is already at the level where it would be able to replicate or beat the level of analysis they typically provide.
No offense, but this is pure fantasy. The level of analysis they typically provide doesn't suffer from the same high baseline level of completely made up numbers of your favorite LLM.
Honestly as a dev I hate Excel its a whole mess I dont understand. I will gladly use Claude for Excel. It will understand the business needs from the data more than I a mere developer just trying to get back to regular developer work.
> these jobs are going to be the first on the chopping block as these integrations mature.
Those two things are maybe related? So many of my friends don't enjoy the same privileges as I do, and have a more tenuous connection to being gainfully employed.
Some people - normal people - understand the difference between the holistic experience of a mathematically informed opinion and an actual model.
It's just that normal people always wanted the holistic experience of an answer. Hardly anyone wants a right answer. They have an answer in their heads, and they want a defensible journey to that answer. That is the purpose of Excel in 95% of places it is used.
Lately people have been calling this "syncophancy." This was always the problem. Sycophancy is the product.
It seems like to me the answer is moreso "People on HN are so far removed from the real use cases for this kind of automation they simply have no idea what they're talking about".
It's actually really cool. I will say that "spreadsheets" remain a bandaid over dysfunctional UIs, processes, etc and engineering spends a lot of time enabling these bandaids vs someone just saying "I need to see number X" and not "a BI analytics data in a realtime spreadsheet!", etc.
I second this. Spreadsheets are the primary tool used for 15% of the U.S. economy. Productivity improvements will affect hundreds of millions of users globally. Each increment in progress is a massive time save and value add.
The criticisms broadly fall between "spreadsheets are bad" and "AI will cause more trouble than it solves".
This release is a dot in a trend towards everyone having a Goldman-Sachs level analyst at their disposal 24/7. This is a huge deal for the average person or business. Our expectation (disclaimer: I work in this space) is that spreadsheet intelligence will soon be a solved problem. The "harder" problem is the instruction set and human <> machine prompting.
For the "spreadsheets are bad" crowd -- sure, they have problems, but users have spoken and they are the preferred interface for analysis, project management and lightweight database work globally. All solutions to "the spreadsheet problem" come with their own UX and usability tradeoffs, so it'a a balance.
Congrats to the Claude team and looking forward to the next release!
> Each increment in progress is a massive time save and value add.
Based on the history of digitalization of businesses from the 1980s onwards, the spreadsheets will just balloon in number and size and there will be more rules and more procedures and more forms and reports to file until the efficiency gains are neutralized (or almost neutralized).
I'm a co-founder of Calcapp, an app builder for formula-driven apps using Excel-like formulas. I spent a couple of days using Claude Code to build 20 new templates for us, and I was blown away. It was able to one-shot most apps, generating competent, intricate apps from having looked at a sample JSON file I put together. I briefly told it about extensions we had made to Excel functions (including lambdas for FILTER, named sort type enums for XMATCH, etc), and it picked those up immediately.
At one point, it generated a verbose formula and mentioned, off-handedly, that it would have been prettier had Calcapp supported LET. "It does!", I replied, "and as an extension, you can use := instead of , to separate names and values!") and it promptly rewrote it using our extended syntax, producing a sleek formula.
These templates were for various verticals, like real estate, financial planning and retail, and I would have been hard-pressed to produce them without Claude's domain knowledge. And I did it in a weekend! Well, "we" did it in a weekend.
So this development doesn't really surprise me. I'm sure that Claude will be right at home in Excel, and I have already thought about how great it would be if Claude Code found a permanent home in our app designer. I'm concerned about the cost, though, so I'm holding off for now. But it does seem unfair that I get to use Claude to write apps with Calcapp, while our customers don't get that privilege.
Seems everyone is speculating features instead of just reading TFA which does in fact list features:
- Get answers about any cell in seconds:
Navigate complex models instantly. Ask Claude about specific formulas, entire worksheets, or calculation flows across tabs. Every explanation includes cell-level citations so you can verify the logic.
- Test scenarios without breaking formulas:
Update assumptions across your entire model while preserving all dependencies. Test different scenarios quickly—Claude highlights every change with explanations for full transparency.
- Debug and fix errors:
Trace #REF!, #VALUE!, and circular reference errors to their source in seconds. Claude explains what went wrong and how to fix it without disrupting the rest of your model.
- Build models or fill existing templates:
Create draft financial models from scratch based on your requirements. Or populate existing templates with fresh data while maintaining all formulas and structure.
Nah, the greatest trick the devil ever pulled was convincing the world that Machine Learning is a legitimate field of study, and not just thinly veiled demon summoning.
I feel similarly about MS Word. It can actually produce decent documents if you learn how to use it, in particularly if you use styles consistently and never, ever touch the bold, italic, colours etc. (outside of defining said styles, although the defaults are probably all most people need). Unfortunately I think the appeal of Word is that you don't have to learn this and it will just do what you want. Is AI the panacea that will both do what you want and give you the right answers every time?
Also people complaining about AI inaccuracy are just technical people that like precision. The vast majority of the world is people who dont give a damn about accuracy or even correctness. They just want to appear as if not completely useless to people that could potentially affect their salary
I can pretty reliably guess that approximately 100% of all companies in the world use excel tables for financial data and for processes. Ok, this was a joke. It's actually 99.99% of all companies. One would think that financial data, inventory and stuff like that should be damn precise. No?
How precise do they really need to be? If there's 3 of a widget on the shelf in the factory, and the factory uses 1000 per day, is it crucial to know that there's 3 of them, and not 0 or 50? Either way, the factory ain't running today or tomorrow or until more of those things come in. Similarly, what's $3 missing from an internal spreadsheet when the company costs $5,000 an hour to operate (or $10 million a year). Obviously errors accumulate so the books need to be reconciled, but apl that stuff only need to be sufficiently directionally accurate with enough precision. If precision is free, then sure, but if a good enough job is cheaper? We all make that call every day.
If you have 2000 hectares of land you need to buy the exact amount of seeds to sow them. If you buy less you are losing money, if you buy more it is useless and you are losing money. If you have trucks or other machinery in the company you need to report exact amount of fuel needed/used, or either they won't run or you lose money on machinery missing fuel. If you need to tax a company, it is pretty important if there are 100 tons of steel used or 1000 tons. Or if the company has 5 factories to be taxed or 15. Etc.
You are anthropomorphizing LLM programs, you assume that if a number in a spreadsheed is big, then program can somehow understand it that it is a big number and if it will make an error it will be a small order error like a human would make. Human process: "hmm, here is a calculation where we divide our imports by number of subsidiaries, let me estimate this in my head, ok, looks like 7320." (actual correct answer was 7340, bu human made a small, typical mistake in the math) LLM program process: it literally uses heat maps and randomization to arrive at each particular character in a row. So it may be 7340, or it may be 8745632, or 1320, or whatever. There is a comment here at a top, from another user, where he queried LLM to make a change of value in the document and it did it correctly. But at the same time it replaced bank account number with a different bank account number. Because to LLM it is the same - sixteen digit in the field, or another sixteen digits in a field, it is the same for LLM. Because it is not AI and doesn't "understand" what it does.
If you have 2000 hecatares land, there is no way you're buying the exact right amount of seeds. You overbuy seeds by as little as you can, but seeds get loaded via tractor bucket, which is fairly messy. You're going to lose a decent amount of seeds. Thus, a lb or kilo of seeds or < 1% in the scheme of things isn't even going to be noticed, much less cause the demise of your farm
For fuel, similarly, you're going to lose militers to evaporation on a hot day, so similarly, being off my ml isn't material.
If you tax a company, fine, sure, the company is going to want it to be right, but 1 or two tons in a 10,000 ton order is again, < 1%. There is some threshold below which precision is extra unnecessary work, though if you have problems with thieves and corruption, you're going to want additional precision that isn't necessary elsewhere.
As to where in my comment I'm anthropomorphizing LLMs, you're going to have to point out where I did that, as the word LLM doesn't appear anywhere in my comment, so it feels like you're projecting claims my comment does not make, as it is LLM neutral and merely point of that 100% exact precision doesn't come without a cost.
"just" technical people who like precision are the reason we are here, typing this, and why lots of parts of our world is pretty cool and comfortable. I wouldn't say that's useless and "just" some people when it clearly is generating unmistakable value
The best thing that can come from this is unit tests for Excel.
LLMs work best when they can call tools (edit the sheet) and test their results in a loop.
It's like the "value seek" thing Excel has had since forever; "adjust these values until this cell is X"
Excel doesn't have any way to verify that every formula in that 60k line sheet is correct and someone hasn't accidentally replaced one with a static number for example.
They can try, but doubt anyone serious will adopt it.
Tried integrating chatgpt into my finance job to see how far I can get. Mega jikes...millions of dollars of hallucinated mistakes.
Worse you don't have the same tight feedback loop you've got in programming that'll tell you when something is wrong. Compile errors, unit tests etc. You basically need to walk through everything it did to figure out what's real and what's hallucinations. Basically fails silently. If they roll that out at scale in the financial system...interesting times ahead.
Still presumably there is something around spreadsheets it'll be able to do - the spreadsheet equivalent of boilerplate code whatever that may be
I'm bad with spread sheets so maybe this is trivial but having an llm tell me how to connect my sheet to whatever data I'm using at the moment and it coming up with a link or sql query or both has allowed me to quickly pull in data where I'd normally eyeball it and move on or worst case do it partially manually if really important.
It's like one off scripts in a sense? I'm not doing complex formulas I just need to know how I can pull data into a sheet and then I'll bucketize or graph it myself.
Again probably because I'm not the most adept user but it has definitely been a positive use case for me.
Neither am I frankly. Finance stuff can get conceptually complicated even with simple addition & multiplication though. e.g. I deal with a lot of offshore stuff, so the average spreadsheet is a mix of currencies, jurisdictions and companies that are interlinked. I could probably talk you through it high level in an hour with a pen & paper, but the LLMs just can't see the forest for all the trees in the raw sheet.
Anthropic is in a weird place for me right now. They're growing fast , creating little projects that i'd love to try, but their customer service was so bad for me as a max subscriber that I set an ethical boundary for myself to avoid their services until such point that it appears that they care about their customers whatsoever.
I keep searching for a sign, but everyone I talk to has horror stories. It sucks as a technologist that just wants to play with the thing; oh well.
> I keep searching for a sign, but everyone I talk to has horror stories. It sucks as a technologist that just wants to play with the thing; oh well.
The reason that Claude Code doesn't have an IDE is because ~"we think the IDE will obsolete in a year, so it seemed like a waste of time to create one."
Noam Shazeer said on a Dwarkesh podcast that he stopped cleaning his garage, because a robot will be able to do it very soon.
If you are operating under the beliefs these folks have, then things like IDEs, cleaning up, and customer service are stupid annoyances that will become obsolete very soon.
To be clear, I have huge respect for everyone mentioned above, especially Noam.
> Noam Shazeer said on a Dwarkesh podcast that he stopped cleaning his garage, because a robot will be able to do it very soon.
We all come up with excuses for why we haven't done a chore, but some of us need to sound a bit more plausible to other members of the household than that.
It would get about the same reaction as "I'm not going to wash the dishes tonight, the rapture is tomorrow."
I want to make it very clear that this was a lighthearted response from Noam to the "AGI timeline" question.
Noam does not do a lot of interviews, and I really hope that stuff like my dumb comment does not prevent him from doing more in the future. We could all learn a lot from him. I am not sure that everyone understands everything that this man has given us.
I would do it for free, just to answer the question of what does a genius of his caliber have in his garage? Probably the same stuff most people do, but it would still be interesting.
I don’t think the point was about having a clean space, it was in response to a question along the lines of: when do you think we will achieve AGI?
There is this homogenization happening in AI. No matter what their original mission was, all the AI companies are now building AI-powered gimmicks hoping to stumble upon something profitable. The investors are waiting...
Customer service at B2C companies can only go downhill or stay level. See Google, Apple, Microsoft etc. At B2B it maaaybe can improve, but only when a ten times bigger customer strongarms a company into doing it.
bad customer service comes from low priority.
I think anthropic prioritize new growth point over small number of customer’s feedback, that’s why they publish new product, features so frequently, there are so much possible potential opportunities for them to focus
Best way to think of it is this: Right now you are not the customer. Investors are.
The money people pay in monthly fees to Anthropic for even the top Max sub likely doesn't come closer to covering the energy & infrastructure costs for running the system.
You can prove this to yourself by just trying to cost out what it takes to build the hardware capable of running a model of this size at this speed and running it locally. It's tens of thousands of dollars just to build the hardware, not even considering the energy bills.
So I imagine the goal right now is to pull in a mass audience and prove the model, to get people hooked, to get management and talent at software firms pushing these tools.
And I guess there's some in management and the investment community that thinks this will come with huge labour cost reductions but I think they may be dreaming.
... And then.. I guess... jack the price up? Or wait for Moore's Law?
So it's not a surprise to me they're not jumping to try and service individual subscribers who are paying probably a fraction of what it costs them to the run the service.
I dunno, I got sick of paying the price for Max and I now use the Claude Code tool but redirect it to DeepSeek's API and use their (inferior but still tolerable) model via API. It's probably 1/4 the cost for about 3/4 the product. It's actually amazing how much of the intelligence is built into the tool itself instead of just the model. It's often incredibly hard to tell the difference bertween DeepSeek output and what I got from Sonnet 4 or Sonnet 4.5
I've been playing around with local LLMs in Ollama, just for fun. I have an RTX 4080 Super, a Ryzen 5950X with 32 threads, and 64 GB of system memory. A very good computer, but decidedly consumer-level hardware.
I have primarily been using the 120b gpt-oss model. It's definitely worse than Claude and GPT-5, but not by, like, an order of magnitude or anything. It's also clearly better than ChatGPT was when it first came out. Text generates a bit slowly, but it's perfectly usable.
So it doesn't seem so unreasonable to me that costs could come down in a few years?
Every AI company right now (except Google Meta and Microsoft) has their valuations based on the expectation of a future monopoly on AGI. None of their business models today or in the foreseeable horizon are even positive let alone world-dominating. The continued funding rounds are all apparently based on expectation of becoming the sole player.
The continuing advancement of open source / open weights models keeps me from being a believer.
The thing really missing from multi-megabyte excel sheets of business critical carnage was a non-deterministic rewrite tool. It'll interact excitingly with the industry standard of no automated testing whatsoever.
I 100% believe generative AI can change a spreadsheet. Turn the xslx into text, mutate that, turn it back into an xslx, throw it away if it didn't parse at all. The result will look pretty similar to the original too, since spreadsheets are great at showing immediately local context and nothing else.
Also, we've done a pretty good job of training people that chatgpt works great, so there's good reason for them to expect claude for excel to work great too.
I'd really like the results of this to be considered negligence with non-survivable fines for the reckless stupidity, but more likely, it'll be seen as an act of god. Like all the other broken shit in the IT world.
This is going to be massive if it works as well as I suspect it might.
I think many software engineers overlook how many companies have huge (billion dollar) processes run through Excel.
It's much less about 'greenfield' new excel sheets and much more about fixing/improving existing ones. If it works as well as Claude Code works for code, then it will get pretty crazy adoption I suspect (unless Microsoft beats them to it).
> I think many software engineers overlook how many companies have huge (billion dollar) processes run through Excel.
So they can fire the two dudes that take care of it, lose 15 years of in house knowledge to save 200k a year and cry in a few months when their magic tool shits the bed ?
You think it's better for the company to have "two dudes" that are completely indispensable and whose work will be completely useless if they die / leave?
I think you're making an argument for LLMs, not against.
These two dudes can train the next generation, you know, like we've been doing since humans exist... instead of relying on some centralised point of failure somewhere thousands of km away which might or might not break your company whenever they decide to update something.
You're one of the people who saw nothing wrong with moving all our industries to asia right ? "It's cheaper so it's obviously better", if you don't think about any of the externalities and long term consequences sure...
If the company is half baked, those "two dudes" will become indispensable beyond belief. They are the ones that understand how Excel works far deeper, and paired with Claude for Excel they become far far more valuable.
At my org it more that these AI tools finally allow the employees to get through things at all. The deadlines are getting met for the first time, maybe ever. We can at last get to the projects that will make the company money instead of chasing ghosts from 2021. The burn down charts are warm now.
From the signup form mentioning Private Equity / Venture Capital, Hedge Fund, Investment Banking... this seems squarely aimed at financial modeling. Which is really, really cool.
I've worked alongside sell-side investment bankers in a prior startup, and so much of the work is in taking a messy set of statements from a company, understanding the underlying assumptions, and building, and rebuilding, and rebuilding, 3-statement models that not only adhere to standard conventions (perhaps best introed by https://www.wallstreetprep.com/knowledge/build-integrated-3-... ) but also are highly customized for different assumptions that can range from seasonality to sensitivity to creative deal structures.
It is quite common for people to pull many, many all-nighters to try to tweak these models in response to a senior banker or a client having an idea! And one might argue there are way too many similar-looking numbers to keep a human banker from "hallucinating," much less an LLM.
But fundamentally, a 3-statement model and all its build-sheets are a dependency graph with loosely connected human-readable labels, and that means you can write tools that let an LLM crawl that dependency graph in a reliable and semantically meaningful way. And that lets you build really cool things, really fast.
I'm of the opinion that giving small companies the ability to present their finances to investors, the same way Fortune 500 companies hire armies of bankers to do, is vital to a healthy economy, and to giving Main Street the best possible chance to succeed and grow. This is a massive step in the right direction.
Presenting false data to investors is fraud, doesn't matter how it was generated. In fact, humans are quite good at "generating plausible looking data", doesn't mean human generated spreadsheets are fraud.
On the other hand, presenting truthful data to investors is distinctly not fraud, and this again does not depend on the generation method.
Establishes that accountants who certify financials are liable if they are incorrect. In particular, if they have a reason to believe they might not be accurate and they certify anyway they are liable. And at this stage of development it’s pretty clear that you need to double check LLM generated numbers.
Obviously no clue if this would hold up with today’s court, but I also wasn’t making a legal statement before. I’m not a lawyer and I’m not trying to pretend to be one.
If humans "generate plausible looking data" despite any processes to ensure data quality they've likely engaged in willful fraud.
An LLM doing so needn't even be willful from the author's part. We're going to see issues with forecasts/slide decks full of inaccuracies that are hard to review.
I think my main point is just because an LLM can lie, doesn’t necessarily mean an LLM generated slide is fraud. It could very easily be correct and verified/certified by the accountant and not fraud. Just cuz the text was generated first by an LLM doesn’t mean fraud.
That being said, oh for sure this will lead to more incidental fraud (and deliberate fraud) and I’m sure it already has. Would be curious to see the prevalence of em-dash’s in 10k’s over the years.
Completely understand the sentiment, but it doesn't apply here, because what's being generated are formulas!
Standardized 3-statement models in Excel are designed to be auditable, with or without AI, because (to only slightly simplify) every cell is either a blue input (which must come from standard exports of the company's accounting books, other auditable inventory/CRM/etc. data, or a visible hardcoded constant), or a black formula that cannot have hardcoded values, and must be simple.
If every buyer can audit, with tools like this, that the formulas match the verbal semantics of the model, there's even less incentive than there is now to fudge the formula level. (And with Wall Street conventions, there's nowhere to hide a prompt injection, because you're supposed to keep every formula to only a few characters, and use breakout "build" rows that can themselves be visually audited.)
And sure, you could conceivably use any AI tool to generate a plausible list of numbers at the input level, but that was equally easy, and equally dependent on context to be fraudulent or not, ever since that famous Excel 1990 elevator commercial: https://www.youtube.com/watch?v=kOO31qFmi9A&t=61s
At the end of the day, the difference between "they want to see this growth, let's fudge it" and "they want to see this growth, let's calculate the exact metrics we need to hit to make that happen, and be transparent about how that's feasible" has always been a matter of trust, not technology.
Tech like this means that people who want to do things the right way can do it as quickly as people who wanted to play loose with the numbers, and that's an equalizer that's on the right side of history.
I'm not excited about having LLMs generate spreadsheets or formulas. But, I think LLMs could be particularly useful in helping me find inconsistent formulas or errors that are challenging to identify. Especially in larger, complex spreadsheets touched by multiple people over the course of months.
For once in my life, I actually had a delightful interaction with an LLM last week. I was changing some text in an Excel sheet in a very progromatic way that could have easily been done with the regex functions in Excel. But I'm not really great with regex, and it was only 15 or so cells, so I was content to just do it manually. After three or four cells, Copilot figured out what I was doing and suggested the rest of the changes for me.
This is what I want AI to do, not generate wrong answers and hallucinate girlfriends.
Thanks for reminding me to check if the REGEXEXTRACT, REGEXREPLACE, and REGEXTEST functions had landed for me yet. They have! Good, because sometime in 2027 the library providing RegEx in VBA will be yanked. https://youtu.be/pGH9LdgkJio
One approach is to produce read-only data in BI tools: users are free to export anything they want and make their own spreadsheets, but those are for their own use only. Reference data is produced every day by a central, controlled process and cannot in any circumstance be modified by the end user.
I have implemented this a couple of times and not only does it work well, it tends to be fairly well accepted. People need spreadsheets to work on them, but generally they kind of hate sending those around via email. Having a reference source of data is welcomed.
So cool, I hope they pull it off. So many people use Excel. Although, I always thought the power of AI in Excel would come from the ability to use AI _as_ a formula. For example, =PROMPT("Classify user feedback as positive, neutral or negative", A1). This would enable normal people (non-programmers) to fire off thousands of prompts at once and automate workflows like programmers do (disclaimer: I am the author of Cellm that does exactly this). Combined with Excel's built-in functions for deterministic work, Claude could really kill the whole copy-pasting data in and out of chat windows for bulk-processing data.
I can't wait until someone does this, then autofills 50k rows down, then gets a $50k bill for all the tokens.
Reminds me of when our CIO insisted on moving to the cloud (back when AWS was just getting started) and then was super pissed when he got a $60k bill because no one knew to shutdown their VMs when leaving for the day.
If someone is processing 50k rows, that means they found real value and the UX is working. That's the whole point.
Also, 50k rows wouldn't cost $50k. More like $100 with Sonnet 4.5 pricing and typical numbers of input/output tokens. Imagine the time needed to go through 50k rows manually and math doesn't really work for a horror story.
On the first glance this seems to be a very bad idea. But re-readig this:
> Get answers about any cell in seconds: Navigate complex models instantly. Ask Claude about specific formulas, entire worksheets, or calculation flows across tabs. Every explanation includes cell-level citations so you can verify the logic.
this might just be an excellent tool for refactoring Excel sheets into something more robust and maintainable. And making a bunch of suits redundant.
On a related note, has anyone found a good local LLM option for working with Excel files?
Here's my use case: I have a set of responses from a survey and want to perform sentiment analysis on them, classify them, etc. Ideally, I'd like to feed them one at a time to a local LLM with a prompt like: "Classify this survey response as positive, negative, or off-topic...etc".
If I dump the whole spreadsheet into ChatGPT, I found that because of the context window, it can get "lazy"; while with a local LLM, I could just literally prompt it one row at a time to accomplish my goal, even if it takes a little longer in terms of GPU and wall-clock time.
However, I can't find anything that works off the shelf like this. It seems like a prime use case for local models.
Don't know about excel, but for Google Sheets. You can ask chatgpt to write you a appsscript custom function e.g CALL_OPENAI. Then you can pass in variables into. =CALL_OPEN("Classify this survey response as positive, negative, or off-topic: "&A1)
When I tried the Gemini/AI formula it didn’t work very well, gpt-5 mini or nano are cheap and generally do what you want if you are asking something straightforward about a piece of content you give them. You can also give a json schema to make the results more deterministic.
Maybe this is how we get code versioning for Excel.
Git LFS for workbook + the following prompt
:
“Create a commit explains what has changed in the workbook since the last commit. Be brief, but explain the change in business terms as well as code change terms.”
I've worked at MULTIPLE million dollar firms whose entire business relies on 10 Excel workbooks that were created 30 years ago by a person who is either passed on or retired.
Give users who aren't intimately knowledgeable with their source material ai, and you're asking for trouble.
The undo function has a history limit.
The real issue is: at what point are we going to stop chasing efficiency and profit at the sake of humanity?
Claude and OpenAI are built on stretched truths, stolen creativity and what-if statements.
I’m decent at excel, but not amazing. I’ve tried again and again to use LLMs including Claude to solve specific, small, well defined problems in excel with a 0% success rate. My experience so far has been if I can’t do it LLMs can’t either.
If LLMs are a 6/10 right now at basic coding then they’re a 3/10 at excel from my experience.
IMO, a real solution here has to be hybrid, not full LLM, because these sheets can be massive and have very complicated structures. You want to be able to use the LLM to identify / map column headers, while using non-LLM tool calling to run Excel operations like SUMIFs or VLOOKUPs. One of the most important traits in these systems is consistency with slight variation in file layout, as so much Excel work involves consolidating / reconciling between reports made on a quarterly basis or produced by a variety of sources, with different reporting structures.
Disclosure: My company builds ingestion pipelines for large multi-tab Excel files, PDFs, and CSVs.
"Hey Claude, what's the total from this file?
> grep for headers
> "Ah, I see column 3 is the price value"
> SUM(C2:C) -> $2020
> "Great! I found your total!"
If you can find me an example of tech that can solve this at scale on large, diverse Excel formats, then I'll concede, but I haven't found something actually trustworthy for important data sets
I would love to learn more about their challenges as I have been working on an Excel AI add-in for quite some time and have followed Ask Rosie from almost their start.
That they now gone through the whole cycle worries me I‘m too slow as a solo building on the side in these fast paced times.
That seems to be true for any startup that offers a wrapper to existing AIs rather than an AI on their own. The lucky ones might be bought but many if not most of them will perish trying to compete with companies that actually create AI models and companies large enough to integrate their own wrappers.
I wonder if this will be more/less useful than what we have with AI in software development.
There's a lot less to understand than a whole codebase.
I don't do spreadsheets very often, but I can emphasize with tracking down "Trace #REF!, #VALUE!, and circular reference errors to their source in seconds." I once hit something like that, and I found it a lot harder to trace a typical compiler error.
I guess Claude maybe useful for finding errors in large Excel Workbooks.
May also help beginners to learn the more complex Excel functions (which are still pretty easy).
But if you are proficient at building Excel models I don't see any benefit.
Excel already has a superb very efficient UI for entering formulas, ranges, tables, data sources etc I'm sceptical that a different UI especially a text based one can improve on this.
I understand the sentiment about a skilled user not needing this, but I think having a little buddy that I can use to offload some menial tasks would be helpful for me to iterate through my models more efficiently; even if the AI is not perfect. As a highly skilled excel user, I admit the software has terrible ergonomics. It would be a productivity boon for me if an AI can help me stay focused on model design vs model implementation.
For some reason, I find that these tools are TERRIBLE at helping someone learn. I suspect because turning one on, results in turning the problem solving part of ones brain off.
Its obviously not the same experience for everyone. ( If you are one of those energized while working in a chat window, you might be in a minority - given what we see from the ongoing massacre of brains in education. )
Paraphrasing something I read here "people don't use ChatGPT to do learn more, they use it to study less".
Pretty cool idea — AI inside spreadsheets makes sense since most of our work already lives there. But I’m a bit cautious too — spreadsheets are messy enough, and adding probabilistic AI could make mistakes harder to spot. Useful if done right, risky if not.
Gemini already has its hooks in Google Sheets, and to be honest, I've found it very helpful in constructing semi-complicated Excel formulas.
Being able to select a few rows and then use plain language to describe what I want done is a time saver, even though I could probably muddle through the formulas if I needed to.
It is an entire agent loop. You can ask it to build a multi sheet analysis of your favorite stock and it will. We are seeing a lot of early adopters use it for financial modeling, research automation, and internal reporting tasks that used to take hours.
I have had the opposite experience. I've never had Gemini give me something useful in sheets, and I'm not asking for complicated things. Like "group this data by day" or "give me p50 and p90"
Last time I tried using Gemini in Google Sheets it hallucinated a bunch of fake data, then gave me a summary that included all that fake data. I'd given it a bunch of transaction data, and asked it to group the records into different categories for budgeting. When asking it to give the largest values in each category, all the values that came back were fake. I'm not sure I'd really trust it to touch a spreadsheet after that.
-stop using the free plan
-don't use gemini flash for these tasks
-learn how to do things over time and know that all ai models have improved significantly every few months
I forgot to add, you can try TabTabTab, without installing anything as well.
To see something much more powerful on Google Sheets than Gemini for free, you can add "try@tabtabtab.ai" to your sheet, and make a comment tagging "try@tabtabtab.ai" and see it in action.
took me four prompts to do generate a waterfall chart using d3 js because it didn't want to run it. obviously with real numbers and not generated data, you'd need to check the results thoroughly.
Interesting their X post mentions "pre-built Agent Skills" but it's not on the webpage. I wonder if they will give you the ability to edit/add/delete Skills, that would be phenomenal.
From their FAQ “Claude doesn’t have advanced Excel capabilities including pivot tables, conditional formatting, data validation, data tables, macros, and VBA. We’re actively working on these features.”
George Hotz said there's 5 tiers of AI systems, Tier 1 - Data centers, Tier 2 - fabs, Tier 3 - chip makers, Tier 4 - frontier labs, Tier 5 - Model wrappers. He said Tier 4 is going to eat all the value of Tier 5, and that Tier 5 is worthless. It's looking like that's going to be the case
That is a common refrain by people who have no domain expertise in anything outside of tech.
Spend a few years in an insurance company, a manufacturing plant, or a hospital, and then the assertion that the frontier labs will figure it out appears patently absurd. (After all, it takes humans years to understand just a part of these institutions, and they have good-functioning memory.)
This belief that tier 5 is useless is itself a tell of a vulnerability: the LLMs are advancing fastest in domain-expertise-free generalized technical knowledge; if you have no domain expertise outside of tech, you are most vulnerable to their march of capability, and it is those with domain expertise who will rely increasingly less on those who have nothing to offer but generalized technical knowledge.
yeah but if Anthropic/OpenAI dedicate resources to gaining domain expertise then any tier 5 is dead in the water. For example, they recently hired a bunch of finance professionals to make specialized models for financial modeling. Any startup in that space will be wiped out
I dont think the claim is exactly that tier 5 is useless more that tier 5 synergizes so well with tier 4 that all the popular tier 5 products will eventually be made by the tier 4 companies.
George Hotz says a lot of things. I think he's directionally correct but you could apply this argument to tech as a whole. Even outside of AI, there are plenty of niches where domain-specific solutions matter quite a bit but are too small for the big players to focus on.
Tier 5 requires domain expertise until we reach AGI or something very different from the latest LLMs.
I don’t think the frontier labs have the bandwidth or domain knowledge (or dare I say skills) to do tier 5 tasks well. Even their chat UIs leave a lot to be desired and that should be their core competency.
That OpenAI is now apparantly striving to become the next big app layer company could hint at George Hotz being right but only if the bets work out. I‘m glad that there is competition on the frontier labs tier.
Interesting. I found a reference to this in a tweet [1], and it looks to be a podcast. While I'm not extremely knowledgable. I'd put it like this: Tier 1 - fabs, Tier 2 - chip makers, Tier 3 - data centers, Tier 4 - frontier labs, Tier 5 - Model wrappers
However I would think more of elite data centers rather than commodity data centers. That's because I see Tier 4 being deeply involved in their data centers and thinking of buying the chips to feed their data centers. I wouldn't be so inclined to throw in my opinion immediately if I found an article showing this ordering of the tiers, but being a tweet of a podcast it might have just been a rough draft.
Copilot is getting better - I'm getting fewer of those than I used to - but it's still significantly more stupid than other agents, even when in theory it's using the same model.
As I was reading through the post, and the comments here, and pondering my own many hours with these tools, I was suddenly reminded of one of my favorite studio C sketches: An Unfortunate Fortune
I have just launched a product (easyanalytica.com) to create dashboards from spreadsheets, and Excel is on my to-do list of formats to be supported. However, I'm having second thoughts. Although, from the description, it seems like it would be more helpful on the modeling side rather than the presentation side. I guess I'll have to wait until it's publicly available
Anyone understand how this could work? My mental model for llm is predictive text but here how can it understand cell A1 which has a string is the “header” for all values under it? How does it learn to understand table data like that?
LLMs already understand table data. "Predictive text" is somewhat true but so reductive that it leads to that kind of misconception.
HN is going to mangle this but here's a quick table:
| Type of Horse | Average Height | Typical Color |
|----------------|----------------|-----------------|
| Arabian | 15 hh | Bay, Gray |
| Thoroughbred | 16 hh | Chestnut, Bay |
| Clydesdale | 17.5 hh | Bay with White |
| Shetland Pony | 10.5 hh | Black, Chestnut |
And after a prompt "pivot the table so rows are colors":
| Typical Color | Type of Horse | Average Height |
|----------------|----------------------------------------|-----------------------|
| Bay | Arabian, Thoroughbred, Clydesdale | 15 hh, 16 hh, 17.5 hh |
| Gray | Arabian | 15 hh |
| Chestnut | Thoroughbred, Shetland Pony | 16 hh, 10.5 hh |
| Bay with White | Clydesdale | 17.5 hh |
| Black | Shetland Pony | 10.5 hh |
> Anyone understand how this could work? My mental model for llm is predictive text but here how can it understand cell A1 which has a string is the “header” for all values under it? How does it learn to understand table data like that?
Modern Excel add-ins work in desktop Windows, macOS, and web. They're just a bit of XML that Excel looks at to call a whatever web endpoint is defined in the XML.
It’s interesting to me that this page talks a lot about “debugging models” etc. I would’ve expected (from the title) this to be going after the average excel user, similar to how chatgpt went after every day people.
I would’ve expected “make a vlookup or pivot table that tells me x” or “make this data look good for a slide deck” to be easier problems to solve.
I think actually Anthropic themselves are having trouble with imagining how this could be used. Coders think like coders - they are imagining the primary use case being managing large Excel sheets that are like big programs. In reality most Excel worksheets are more like tiny, one-off programs. More like scripts than applications. AI is very very good at scripts.
The issue is that the average Excel user doesn’t quite have the skills to validate and double-check the Excel formulas that Claude would produce, and to correct them if needed. It would be similar to a non-programmer vibe-coding an app. And that’s really not what you want to happen for professionally used Excel sheets.
IMO that is exactly what people want. At my work everyone uses LLMs constantly and the trade off of not perfect information is known. People double check it, etc, but the information search is so much faster even if it finds the right confluence but misquotes it, it still sends me the link.
For easy spreadsheet stuff (which 80% of average white collars workers are doing when using excel) I’d imagine the same approach. Try to do what I want, and even if you’re half wrong the good 50% is still worth it and a better starting point.
Vibe coding an app is like vibe coding a “model in excel”. Sure you could try, but most people just need to vibe code a pivot table
As an inveterate Excel lover, I can just sense the blinding pain wafting off the legions of accountants, associates, seniors, and tech people who keep the machine spirits placated.
lies, damn lies, statistics, and then Excel deciding cell data types.
If it has a concept of data sources and can digest them, sure.
Anecdotally, most issues with Excel at my job are caused by data sources being renamed, moved or reformatted, by broken logins, or by insufficient access rights.
I'm excited to see what national disasters will be caused by auto-generated Excel sheets that nobody on the planet understands. A few selections from past HN threads to prime your imagination:
Especially combined with the dynamic array formulas that have recently been added (LET, LAMBDA etc). You can have much more going on within each cell now. Think whole temporary data structures. The "evaluate formula" dialog doesn't quite cut it anymore for debugging.
from my experience in the corporate world, i'd trust an excel generated / checked by an LLM more than i would one that has been organically grown over years in a big corporation where nobody ever checks or even can check anything because its one big growing pile of technical debt people just accept as working
How well does change tracking work in Excel... how hard would it be to review LLM changes?
AFAIK there is no 'git for Excel to diff and undo', especially not built-in (aka 'for free' both cost-wise and add-ons/macros not allowed security-wise).
My limited experience has been that it is difficult to keep LLMs from changing random things besides what they're asked to change, which could cause big problems if unattackable in Excel.
I thought there was track changes on all office products. Most Office documents are zip files of XML files and assets, so I'd imagine it would be possible to rollback changes.
When I think how easy I can misclick to stuff up a spreadsheet I can't begin to imagine all the subtle ways LLMs will screw them up.
Unlike code where it's all on display, with all these formulas are hidden in each cell, you won't see the problem unless click on the cell so you'll have a hard time finding the cause.
If Claude is going to work on the underpinning technology of every business in the Capitalist world, We should let Claude loose on the COBOL code out there too, I can't imagine anything going wrong.
Yeah now tell the Auditors that the financial spreadsheet we have here has AI touching it left and right. "I did not cook the books I promise it is the AI that made our financials seem better than they actually are trust me bro!", said Joe from Accounting.
Yet more evidence of the bubble burst being imminent. If any of these companies really had some almost-AGI system internally, they wouldn’t be spending any effort making f’ing Excel plugins. Or at the very least, they’d be writing their own Excel because AI is so amazing at coding, right?
You make a great point. Where is all of the complex applications? They haven't been able to create than own office suite or word processor or really anything aside from a halloween matching game in js. You would think we would have some complex application they can point to but nothing.
Yes. I once interviewed a developer who’s previous job was maintaining the .NET application that used an Excel sheet as the brain for decisions about where to drill for oil on the sea floor. No one understood what was in the Excel sheet. It was built by a geologist who was long gone. The engineering team understood the inputs and outputs. That’s all they needed to know.
Years ago when I worked for an engineering consulting company we had to work with a similarly complex, opaque Excel spreadsheet from General Electric modeling the operation of a nuclear power plant in exacting detail.
Same deal there -- the original author was a genius and was the only person who knew how it was set up or how it worked.
I think you’re misunderstanding me. This might be something somewhat useful, I don’t know, and I’m not judging it based on that.
What I’m saying is that if you really believed we were 2, maybe 3 years tops from AGI or the singularity or whatever you would spend 0 effort serving what already seems to be a domain that is already served by 3rd parties that are already using your models! An excel wrapper for an LLM isn’t exactly cutting edge AI research.
They’re desperate to find something that someone will pay a meaningful amount of money for that even remotely justifies their valuation and continued investment.
The current valuations do not require AGI. They require products like this that will replace scores of people doing computer based grunt work. MSFT is worth $4 trillion off the back of enterprise productivity software, the AI labs just need some of that money.
If this works very well and reliable, it might not kill programming as such, but it might put a lot of small businesses who do custom software for other small businesses out of work.
Been working with Claude Code lately and been pretty impressed.
If this works as well could be a nice add on. Its probably a smart market to enter as Excel is essentially everywhere.
Just like Claude Code allows 1 dev to potentially do the work of 2 or 3, I could see this allowing 1 accountant or operations person to do the work of 2 or 3. Financial savings but human cost
If AI turns out to be the powerhouse it is claimed to be, AI's impact will be corporations replacing corporate dependencies upon 'Excel projects' created by self-taught assistants to department managers.
Cool but now companies POs will be like "you must add the Excel export for all the user data!" and when asked why, will basically be "so I can do this roundabout query of data for some number in a spreadsheet using AI (instead of just putting the number or chart directly in the product with a simple db call)"
Okay. But then you could say the same for a human, isn't your brain just a cloud of matter and electricity that just reacts to senses deterministically?
> isn't your brain just a cloud of matter and electricity that just reacts to senses deterministically?
LLMs are not deterministic.
I'd argue over the short term humans are more deterministic. I ask a human the same question multiple times and I get the same answer. I ask an LLM and each answer could be very different depending on its "temperature".
I mean - try clicking the CoPilot button and see what it can actually do. Last I checked, it told me it couldn't change any of the actual data itself, but it could give you suggestions. Low bar for excellence here.
What is with the negativity in these comments? This is a huge, huge surface area that touches a large percentage of white collar work. Even just basic automation/scaffolding of spreadsheets would be a big productivity boost for many employees.
My wife works in insurance operations - everyone she manages from the top down lives in Excel. For line employees a large percentage of their job is something like "Look at this internal system, export the data to excel, combine it with some other internal system, do some basic interpretation, verify it, make a recommendation". Computer Use + Excel Use isn't there yet...but these jobs are going to be the first on the chopping block as these integrations mature. No offense to these people but Sonnet 4.5 is already at the level where it would be able to replicate or beat the level of analysis they typically provide.
Having wrangled many spreadsheets personally, and worked with CFOs who use them to run small-ish businesses, and all the way up to one of top 3 brokerage houses world-wide using them to model complex fixed income instruments... this is a disaster waiting to happen.
Spreadsheet UI is already a nightmare. The formula editing and relationship visioning is not there at all. Mistakes are rampant in spreadsheets, even my own carefully curated ones.
Claude is not going to improve this. It is going to make it far, far worse with subtle and not so subtle hallucinations happening left and right.
The key is really this - all LLMs that I know of rely on entropy and randomness to emulate human creativity. This works pretty well for pretty pictures and creating fan fiction or emulating someone's voice.
It is not a basis for getting correct spreadsheets that show what you want to show. I don't want my spreadsheet correctness to start from a random seed. I want it to spring from first principles.
In my opinion the biggest use case for spread sheet with LLM is to ask them to build python scripts to do what ever manipulations you want to do with the data. Once people learn to do this workplace productivity would increase greatly I have been using LLM for years now to write python scripts that automate different repeatable tasks. Want a pdf of this data to be overlayed on this file create a python script with an LLM. Want the data exported out of this to be formated and tallied create a script for that.
How will people without Python knowledge know that the script is 100% correct? You can say "Well they shouldn't use it for mission critical stuff" or "Yeah that's not a use case, it could be useful for qualitative analysis" etc., but you bet they will use it for everything. People use ChatGPT as a search engine and a therapist, which tells us enough
If you have a mechanism that can prove arbitrary program correctness with 100% accuracy you’re sitting on something more valuable than LLMs.
Yesterday I had to pass a bunch of data to finance as the person that does so had left the company. But they wanted me to basically group by a few columns, so instead of spending an hour on this in excel, I created 3 rows of fake data, gave it to the llm, it created a Python script which I ran against the dataset. After manual verification of the results, it could be submitted to finance.
That’s exactly how it should be done if accuracy is important.
Congrats? But you are not likely a typical user.
My first job out of uni was building a spreadsheet infra as code version control system after a Windows update made an eight year old spreadsheet go haywire and lose $10m in a afternoon.
Spreadsheets are already a disaster.
Compared to what? Granted, Excel incidents are probably underreported and might produce "silent" consequential losses. But compared to that, for enterprise or custom software in general we have pretty scary estimates of the damages. Like Y2K (between 300-600bn) and the UK Postal Office thing (~1bn).
Excel spreadsheets ARE custom software, with custom requirements, calculations, and algorithms. They're just not typically written by programmers, have no version control or rollback abilities, are not audited, are not debuggable, and are typically not run through QA or QC.
Agreed. The tradition has been continued by workflow engines, low code tools, platforms like Salesforce and lately AI-builders. The issue is generally not that these are bad, but because they don't _feel_ like software development everyone is comfortable skipping steps of the development process.
To be fair, I've seen shops which actually apply good engineering practices to Excel sheets too. Just definitely not a majority...
If I could teach managers one lesson, it would be this one.
Thing is, they are also the common workaround solution for savy office workers that don't want to wait for the IT department if it exists, or some outsourced consultancy, to finally deliver something that only does half the job they need.
So far no one has managed to deliver an alternative to spreedsheets that fix this issue, doesn't matter if we can do much better in Python, Java, C# whatever, if it is always over budget and only covers half of the work.
I know, I have taken part in such project, and it run over budget because there was always that little workflow super easy to do in Excel and they would refuse to adopt the tool if it didn't cover that workflow as well.
exactly. And Claude and other code assistants are more of the same, allowing non-programmers[1] to write code for their needs. And that's a good thing overall.
[1] well, people that don't consider themselves programmers.
I'll add to this - if you work on a software project to port an excel spreadsheet to real software that has all those properties, if the spreadsheet is sophisticated enough to warrant the process, the creators won't be able to remember enough details abut how they created it to tell you the requirements necessary to produce the software. You may do all the calculations right, and because they've always had a rounding error that they've worked around somewhere else, your software shows calculations that have driven business decisions for decades were always wrong, and the business will insist that the new software is wrong instead of owning some mistake. It's never pretty, and it always governs something extremely important.
Now, if we could give that excel file to an llm and it creates a design document that explains everything is does, then that would be a great use of an LLM.
It's interesting that you mention disaster; there is at least one annual conference dedicated to "spreadsheet risk management".[1]
[1] https://eusprig.org/
> Spreadsheets are already a disaster.
Yeah, that's what OP said. Now add a bunch of random hallucinations hidden inside formulas inside cells.
If they really have a good spreadsheet solution they've either fixed the spreadsheet UI issues or the LLM hallucination issues or both. My guess is neither.
I know you probably can't share the details, but if you can I (and I'm sure all of us) would love to hear them
I don't think tools like Claude are there yet, but I already trust GPT-5 Pro to be more diligent about catching bugs in software than me, even when I am trying to be very careful. I expect even just using these tools to help review existing Excel spreadsheets could lead to a significant boost in quality if software is any guide (and Excel spreadsheets seem even worse than software when it comes to errors).
That said, Claude is still quite behind GPT-5 in its ability to review code, and so I'm not sure how much to expect from Sonnet 4.5 in this new domain. OpenAI could probably do better.
> That said, Claude is still quite behind GPT-5 in its ability to review code, and so I'm not sure how much to expect from Sonnet 4.5 in this new domain. OpenAI could probably do better.
It’s always interesting to see others opinions as it’s still so variable and “vibe” based. Personally, for my use, the idea that any GPT-5 model is superior to Claude just doesn’t resonate - and I use both regularly for similar tasks.
I also find the subjective nature of these models interesting, but in this case the difference in my experiences between Sonnet 4.5 and GPT-5 Codex, and especially GPT-5 Pro, for code review is pretty stark. GPT-5 is consistently much better at hard logic problems, which code review often involves.
I have had GPT-5 point out dozens of complex bugs to me. Often in these cases I will try to see if other models can spot the same problems, and Gemini has occasionally but the Claude models never have (using Opus 4, 4.1, and Sonnet 4.5). These are bugs like complex race conditions or deadlocks that involve complex interactions between different parts of the codebase. GPT-5 and Gemini can spot these types of bugs with a decent accuracy, while I’ve never had Claude point out a bug like this.
If you haven’t tried it, I would try the codex /review feature and compare its results to asking Sonnet to do a review. For me, the difference is very clear for code review. For actual coding tasks, both models are much more varied, but for code review I’ve never had an instance where Claude pointed out a serious bug that GPT-5 missed. And I use these tools for code review all the time.
Try the TypeScript codex CLI with the gpt-5-codex model with reasoning always set to high, or GPT-5 Pro with max reasoning. Both are currently undeniably better than Claude Opus 4.1 or Sonnet 4.5 (max reasoning or otherwise) for all code-related tasks. Much slower but more reliable and more intelligent.
I've been a Claude Code fanboy for many months but OpenAI simply won this leg of the race, for now.
Same. I switched from sonnet 4 when it was out to codex. Went back to try sonnet 4.5 and it really hates to work for longer than like 5 minutes at a time
Codex meanwhile seems to be smarter and plugs away at a massive todo list for like 2 hours
Yeah, it's like that commercial for OpenAI (or was it Gemini?) where the guy says it lets the tool work on it's complex financial spreadsheets, goes for a walk with a dog, gets back and it is done with "like 98% accuracy". I cannot imagine what the 2% margin of error looks like for a company that moves around hundreds of billions of dollars...
Having AI create the spreadsheet you want is totally possible, just like generating bash scripts works well. But to get good results, there needs to be some documentation describing all the hidden relationships and nasty workarounds first.
Don't try to make LLMs generate results or numbers, that's bound to fail in any case. But they're okay to generate a starting point for automations (like Excel sheets with lots of formulas and macros), given they get access to the same context we have in our heads.
I like this take. There seems to be an over-focus on 'one-shot' results, but I've found that even the free tools are a significant productivity booster when you focus on generating smaller pieces of code that you can verify. Maybe I'm behind the power curve since I'm not leveraging the full capability of the advanced LLM's, but if the argument is disaster is right around the corner due to potential hallucinations, I think we should consider that you still have to check your work for mission critical systems. That said, I don't really build mission critical systems - I just work in Aerospace Engineering and like building small time saving scripts / macros for other engineers to use. For this use, free LLMs even have been huge for me. Maybe I'm in a very small minority, but I do use Excel & Python nearly every day.
I tend to agree that dropping the tool as it is into untrained hands is going to be catastrophic.
I’ve had similar professional experiences as you and have been experimenting with Claude Code. I’ve found I really need to know what I’m doing and the detail in order to make effective (safe) use out of it. And that’s been a learning curve.
The one area I hope/think it’s closest to (given comments above) is potentially as a “checker” or validator.
But even then I’d consider the extent to which it leaks data, steers me the wrong way, or misses something.
The other case may be mocking up a simple financial model for a test / to bounce ideas around. But without very detailed manual review (as a mitigating check), I wouldn’t trust it.
So yeah… that’s the experience of someone who maybe bridges these worlds somewhat… And I think many out there see the tough (detailed) road ahead, while these companies are racing to monetize.
IMO people tend to over-trust both AI and Excel. Maybe this will recalibrate that after it leads to a catastrophic business failure or two.
You would hope so. But how many companies have actually changed their IT policy of outsourcing everything to Tata Consultancy Services (or similar) where a sweaty office in Mumbai full of people who don't give a shit run critical infrastructure?
Jaguar Landrover had production stopped for over a month I think, and 100+ million impact to their business (including a trail of smaller suppliers put near bankruptcy). I'd bet Tata are still there and embedded even further in 5 years.
If AI provides some day-to-day running cost reduction that looks good on quarterly financial statements it will be fully embraced, despite the odd "act of god".
to be clear, tata owns JLR.
My take is more optimistic. This could be an off ramp to stop putting critical business workflows in spreadsheets. If people start to learn that general purpose programming languages are actually easier than Excel (and with LLMs, there is no barrier), then maybe more robust workflows and automation will be the norm.
I think the world would be a lot better off if excel weren’t in it. For example, I work at business with 50K+ employees where project management is done in a hellish spreadsheet literally one guy in Australia understands. Data entry errors can be anywhere and are incomprehensible. 3 or 4 versions are floating around to support old projects. A CRUD app with a web front end would solve it all. Yet it persists because Excel is erroneously seen as accessible whereas Rails, Django, or literally anything else is witchcraft.
> all LLMs that I know of rely on entropy and randomness to emulate human creativity
Those are tuneable parameters. Turn down the temperature and top_p if you don't want the creativity.
> Claude is not going to improve this.
We can measure models vs humans and figure this out.
To your own point, humans already make "rampant" mistakes. With models, we can scale inference time compute to catch and eliminate mistakes, for example: run 6x independent validators using different methodologies.
One-shot financial models are a bad idea, but properly designed systems can probably match or beat humans pretty quickly.
> Turn down the temperature and top_p if you don't want the creativity.
This also reduces accuracy in real terms. The randomness is used to jump out of local minima.
> Those are tuneable parameters. Turn down the temperature and top_p if you don't want the creativity.
Ah yes, we'll tell Mary from the Payroll she could just tune them parameters if there is more than "like 2%" error in her spreadsheets
Is this just a feeling you have or is this downstream of actual use cases you've applied AI to observed and measured reliability on?
Not the parent poster, but this is pretty much the foundation of LLMs. They are by their nature probabilistic, not deterministic. This is precisely what the parent is referring to.
All processes in reality, everywhere, are probablistic. The entire reason "engineering" is not the same as theoretical mathematics is about managing these probabilities to an acceptable level for the task you're trying to perform. You are getting a "probablistic" output from a human too. Human beings are not guaranteeing theoretically optimal excel output when they send their boss Final_Final_v2.xlsx. You are using your mental model of their capabilities to inform how much you trust the result.
Building a process to get a similar confidence in LLM output is part of the game.
Yup. It becomes clearer to me when I think about the existing validators. Can these be improved, for sure.
It’s when people make the leaps to the multi-year endgame and in their effort to monetise by building overconfidence in the product where I see the inherent conflict.
It’s going to be a slog… the detailed implementations. And if anyone is a bit more realistic about managing expectations I think Anthropic is doing it a little better.
> All processes in reality, everywhere, are probablistic.
If we want to go in philosophy then sure, you're correct, but this not what we're saying.
For example, an LLM is capable (and it's highly plausible for it to do so) of creating a reference to a non-existent source. Humans generally don't do that when their goal is clear and aligned (hence deterministic).
> Building a process to get a similar confidence in LLM output is part of the game.
Which is precisely my point. LLMs are supposed to be better than humans. We're (currently) shoehorning the technology.
> Humans generally don't do that when their goal is clear and aligned (hence deterministic).
Look at the language you're using here. Humans "generally" make less of these kinds of errors. "Generally". That is literally an assessment of likelihood. It is completely possible for me to hire someone so stupid that they create a reference to a non-existent source. It's completely possible for my high IQ genius employee who is correct 99.99% of the time to have an off-day and accidentally fat finger something. It happens. Perhaps it happens at 1/100th of the rate that an LLM would do it. But that is simply an input to the model of the process or system I'm trying to build that I need to account for.
When humans make mistakes repeatedly in their job they get fired.
Not OP but using LLMs in any professional setting, like programming, editing or writing technical specifications, OP is correct.
Without extensive promoting and injectimg my own knowledge and experience, LLMs generate absolute unusable garbage (on average). Anyone who disagrees very likely is not someone who would produce good quality work by themselves (on average). That's not a clever quip; that's a very sad reality. SO MANY people cannot be bothered to learn anything if they can help it.
The triad of LLM dependencies in my view: initiation of tasks, experience based feedback, and consequence sink. They can do none of these, they all connect to the outer context which sits with the user, not the model.
You know what? This is also not unlike hiring a human, they need the hirer party tell them what to do, give feedback, and assume the outcomes.
It's all about context which is non-fungible and distributed, not related to intelligence but to the reason we need intelligence for.
> Anyone who disagrees very likely is not someone who would produce good quality work by themselves (on average).
So for those producing slop and not knowing any better (or not caring), AI just improved the speed at which they work! Sounds like a great investment for them!
For many mastering any given craft might not be the goal, but rather just pushing stuff out the door and paying bills. A case of mismatched incentives, one might say.
I would completely disagree. I use LLMs daily for coding. They are quite far from AGI and it does not appear they are replacing Senior or Staff Engineers any time soon. But they are incredible machines that are perfectly capable of performing some economically valuable tasks in a fraction of the time it would have taken a human. If you deny this your head is in the sand.
Capable, yeah, but not reliable, that's my point. They can one shot fantastic code, or they can one shot the code I then have to review and pull my hair out over for a week, because it's such crap (and the person who pushed it is my boss, for example, so I can't just tell him to try again).
That's not consistent.
Why do you frame the options as "one shot... or... one shot"?
Because lazy people will use it like that, and we are all inherently lazy
It's not much better with planning either. The amount of time I spent planning, clarifying requirements, hand-holding implementation details always offset any potential savings.
Have you never used one to hunt down an obscure bug and found the answer quicker than you likely would have yourself?
You can ask your boss to submit PRs using Codex’s “try 5 variations of the same task and select the one you like most though
Surely at that point they could write the code themselves faster than they can review 5 PRs.
Producing more slop for someone else to work through is not the solution you think it is.
You can do it cursor style
> Mistakes are rampant in spreadsheets
To me, the case for LLMs is strongest not because LLMs are so unusually accurate and awesome, but because if human performance were put on trial in aggregate, it would be found wanting.
Humans already do a mediocre job of spreadsheets, so I don't think it is a given that Claude will make more mistakes than humans do.
But isn't this only fine as long someone who knows what they are doing has oversight and can fix issues when they arise and Claude gets stuck?
Once we all forget how to write SUM(A:A), will we just invent a new kind of spreadsheet once Claude gets stuck?
Or in other words; what's the end game here? LLMs clearly cannot be left alone to do anything properly, so what's the end game of making people not learn anything anymore?
Well the end game with AI is AGI of course. But realistically the best case scenario with LLM’s is having fewer people with the required knowledge, leveraging LLM’s to massively enhance productivity.
We’re already there to some degree. It is hard to put a number on my productivity gain, but as a small business owner with a growing software company it’s clear to me already that I can reduce developer hiring going forward.
When I read the skeptics I just have to conclude that they’re either poor at context building and/or work on messy, inconsistent and poorly documented projects.
My sense is that many weaker developers who can’t learn these tools simply won’t compete in the new environment. Those who can build well designed and documented projects with deep context easy for LLM’s to digest will thrive.
I assume all of this applies to spreadsheets.
Why isn't there a single study that would back up your observations? The only study with a representative experimental design that I know about is the METR study and it showed the opposite. Every study citing significant productivity improvements that I've seen is either:
- relying on self-assessments from developers about how much time they think they saved, or
- using useless metrics like lines of code produced or PRs opened, or
- timing developers on toy programming assignments like implementing a basic HTTP server that aren't representative of the real world.
Why is it that any time I ask people to provide examples of high quality software projects that were predominantly LLM-generated (with video evidence to document the process and allow us to judge the velocity), nobody ever answers the call? Would you like to change that?
My sense is that weaker developers and especially weaker leaders are easily impressed and fascinated by substandard results :)
Okay, and now you give those mediocre humans a tool hat is both great and terrible. The problem is, unless you know your way around very well, they won't know which is which.
Since my company uses Excel a lot, and I know the basics but don't want to become an expert, I use LLMs to ask intermediate questions, too hard to answer with the few formulas I know, not too hard for a short solution path.
I have great success and definitely like what I can get with the Excel/LLM combo. But if my colleagues used it the same way, they would not get my good results, which is not their fault, they are not IT but specialists, e.g. for logistics. The best use of LLMs is if you could already do the job without them, but it saves you time to ask them and then check if the result is actually acceptable.
Sometimes I abandon the LLM session, because sometimes, and it's not always easy to predict, fixing the broken result would take more effort than just doing it the old way myself.
A big problem is that the LLMs are so darn confident and always present a result. For example, I point it to a problem, it "thinks", and then it gives me new code and very confidently summarizes what the problem was, correctly, that it now for sure fixed the problem. Only that when I actually try the result has gotten worse than before. At that point I never try to get back to a working solution by continuing to try to "talk" to the AI, I just delete that session and do another, non-AI approach.
But non-experts, and people who are very busy and just want to get some result to forward to someone waiting for it as quickly as possible will be tempted to accept the nice looking and confidently presented "solution" as-is. And you may not find a problem until half a year later somebody finds that prepayments, pro forma bills and the final invoices don't quite match in hard to follow ways.
Not that these things don't happen now already, but adding a tool with erratic results might increase problems, depending on actual implementation of the process. Which most likely won't be well thought out, many will just cram in the new tool and think it works when it doesn't implode right away, and the first results, produced when people still pay a lot of attention and are careful, all look good.
I am in awe of the accomplishments of this new tool, but it is way overhyped IMHO, still far too unpolished and random. Forcing all kinds of processes and people to use it is not a good match, I think.
This is a great point. LLMs make good developers better, but they make bad developers even worse. LLMs multiply instead of add value. So if you're a good developer, who is careful, pays attention, watches out for trouble, and is constantly reviewing and steering, the LLM is multiplying by a positive number and will make you better. However, if you're a mediocre/bad developer, who is not careful, who lacks attention to detail, and just barely gets things to compile / run, then the LLM is multiplying by a negative number and will make your output even worse.
Or you could, you know, read the article before commenting to see the limited scope of this integration?
Anyway, Google has already integrated Gemini into Sheets, and recently added direct spreadsheet editing capability so your comment was disproven before you even wrote it
> The key is really this - all LLMs that I know of rely on entropy and randomness to emulate human creativity. This works pretty well for pretty pictures and creating fan fiction or emulating someone's voice.
I think you need to turn down the temperature a little bit. This could be a beneficial change.
I don't trust LLMs to do the kind of precise deterministic work you need in a spreadsheet.
It's one thing to fudge the language in a report summary, it can be subjective, however numbers are not subjective. It's widely known LLMs are terrible at even basic maths.
Even Google's own AI summary admits it which I was surprised at, marketing won't be happy.
Yes, it is true that LLMs are often bad at math because they don't "understand" it as a logical system but rather process it as text, relying on pattern recognition from their training data.
Seems like you're very confused about what this work typically entails. The job of these employees is not mental arithmatic. It's closer to:
- Log in to the internal system that handles customer policies
- Find all policies that were bound in the last 30 days
- Log in to the internal system that manages customer payments
- Verify that for all policies bound, there exists a corresponding payment that roughly matches the premium.
- Flag any divergences above X% for accounting/finance to follow up on.
Practically this involves munging a few CSVs, maybe typing in a few things, setting up some XLOOKUPs, IF formulas, conditional formatting, etc.
Will AI replace the entire job? No...but that's not the goal. Does it have to be perfect? Also no...the existing employees performing this work are also not perfect, and in fact sometimes their accuracy is quite poor.
> “Does it have to be perfect?”
Actually, yes. This kind of management reporting is either (1) going to end up in the books and records of the company - big trouble if things have to be restated in the future or (2) support important decisions by leadership — who will be very much less than happy if analysis turns out to have been wrong.
A lot of what ties up the time of business analysts is ticking and tying everything to ensure that mistakes are not made and that analytics and interpretations are consistent from one period to the next. The math and queries are simple - the details and correctness are hard.
Is this not belligerently ignoring the fact that this work is already done imperfectly? I can’t tell you how many serious errors I’ve caught in just a short time of automating the generation of complex spreadsheets from financial data. All of them had already been checked by multiple analysts, and all of them contained serious errors (in different places!)
No belligerence intended! Yes, processes are faulty today even with maker-checker and other QA procedures. To me it seems the main value of LLMs in a spreadsheet-heavy process is acceleration - which is great! What is harder is quality assurance - like the example someone gave regarding deciding when and how to include or exclude certain tables, date ranges, calc, etc. Properly recording expert judgment and then consistently applying that judgement over time is key. I’m not sure that is the kind of thing LLMs are great at, even ignoring their stochastic nature. Let’s figure out how to get best use out of the new kit - and like everything else, focus on achieving continuously improving outcomes.
There’s actually different classes of errors though. There’s errors in the process itself versus errors that happen when performing the process.
For example, if I ask you to tabulate orders via a query but you forgot to include an entire table, this is a major error of process but the query itself actually is consistently error-free.
Reducing error and mistakes is very much modeling where error can happen. I never trust an LLM to interpret data from a spreadsheet because I cannot verify every individual result, but I am willing to ask an LLM to write a macro that tabulates the data because I can verify the algorithm and the macro result will always be consistent.
Using Claude to interpret the data directly for me is scary because those kinds of errors are neither verifiable nor consistent. At least with the “missing table” example, that error may make the analysis completely bunk but once it is corrected, it is always correct.
Speak for yourself and your own use cases. There are a huge diversity of workflows with which to apply automation in any medium to large business. They all have differing needs. Many excel workflows I'm personally familiar with already incoporate a "human review" step. Telling a business leader that they can now jump straight to that step, even if it requires 2x human review, with AI doing all of the most tediuous and low-stakes prework, is a clear win.
>Speak for yourself and your own use cases
Take your own advice.
I'm taking a much weaker position than the respondent: LLMs are useful for many classes of problem that do not require zero shot perfect accuracy. They are useful in contexts where the cost of building scaffolding around them to get their accuracy to an acceptable level is less than the cost of hiring humans to do the same work to the same degree of accuracy.
This is basic business and engineering 101.
>LLMs are useful for many classes of problem that do not require zero shot perfect accuracy. They are useful in contexts where the cost of building scaffolding around them to get their accuracy to an acceptable level is less than the cost of hiring humans to do the same work to the same degree of accuracy.
Well said. Concise and essentially inarguable, at least to the extent it means LLMs are here to stay in the business world whether anyone likes it or not (barring the unforeseen, e.g. regulation or another pressure).
There is another aspect to this kind of activity.
Sometimes there can be an advantage in leading or lagging some aspects of internal accounting data for a time period. Basically sitting on credits or debits to some accounts for a period of weeks. The tacit knowledge to know when to sit on a transaction and when to action it is generally not written down in formal terms.
I'm not sure how these shenanigans will translate into an ai driven system.
> Sometimes there can be an advantage in leading or lagging some aspects of internal accounting data for a time period.
This worked famously well for Enron.
That’s the kind of thing that can get a company into a lot of trouble with its auditors and shareholders. Not that I am offering accounting advice of course. And yeah, one can not “blame” and ai system or try to ai-wash any dodgy practices.
Checking someone elses spreadsheet is a fucking nightmare. If your company has extremely good standards it's less miserable because at least the formatting etc will be consistent...
The one thing LLMs should consistently do is ensure that formatting is correct. Which will help greatly in the checking process. But no, I generally don't trust them to do sensible things with basic formulation. Not a week ago GPT 5 got confused whether a plus or a minus was necessary in a basic question of "I'm 323 days old, when is my birthday?"
I think you have a misunderstanding of the types of things that LLMs are good at. Yes you're 100% right that they can't do math. Yet they're quite proficient at basic coding. Most Excel work is similar to basic coding so I think this is an area where they might actually be pretty well suited.
My concern would be more with how to check the work (ie, make sure that the formulas are correct and no columns are missed) because Excel hides all that. Unlike code, there's no easy way to generate the diff of a spreadsheet or rely on Git history. But that's different from the concerns that you have.
> Yes you're 100% right that they can't do math.
The model ought to be calling out to some sort of tool to do the math—effectively writing code, which it can do. I'm surprised the major LLM frontends aren't always doing this by now.
MS Office Tools menu has a "Spreadsheet Compare" application. It is quite good for diffing 2 spreadsheets. Of course it cannot catch logic errors, human or ML.
> Most Excel work is similar to basic coding
Excel is similar to coding in BASIC, a giant hairy ball of tangled wool.
I've built spreadsheet diff tools on Google sheets multiple times. As the needs grows I think we will see diffs and commits and review tools reach customers
hey Collin! I am working on an AI agent on Google Sheets, I am curious if any of your designs are out in the public. We are trying to re-think how diffs should look like and want to make something nicer than what we currently have, so curious.
Hi! Nothing public nor generic enough to be a good building block. I found myself often frustrated by the tools that came out of the box but I believe better apis could make this slightly easier to solve.
The UX of spreadsheet diffs is a hard one to solve because of how weird the calculation loops are and how complicated the relationship between fields might be.
I've never tried to solve this for a real end user before in a generic way - all my past work here was for internal ability to audit changes and rollback catastrophes. I took a lot of shortcuts by knowing which cells are input data vs various steps of calculations -- maybe part of your ux is being able to define that on a sheet by sheet basis? Then you could show how different data (same formulas) changed outputs or how different formulas (same data) did differently?
Spreadsheets are basically weird app platforms at this point so you might not be able to create a single experience that is both deep and generic. On the other hand maybe treating it as an app is the unlock? Get your AI to noodle on what the whole thing is for, then show diff between before and after stable states (after all calculation loops stabilize or are killed) side by side with actual diffs of actual formulas? I feel like Id want to see a diff as a live final spreadsheet and be able to click on changed cells and see up the chain of their calculations to the ancestors that were modified.
Fun problem that sounds extremely complicated. Good luck distilling it!
So do it in basic code where numbering your line G53 instead of G$53 doesn't crash a mass transit network because somebody's algorithm forgot to order enough fuel this month.
proficient != near-flawless.
> Most Excel work is similar to basic coding so I think this is an area where they might actually be pretty well suited.
This is a hot take. One I'm not sure many would agree with.
Maybe LLMs will enable a new type of work in spreadsheets. Just like in coding we have PR reviews, with an LLM it should be possible to do a spreadsheet review. Ask the LLM to try to understand the intent and point out places where the spreadsheet deviates from the intent. Also ask the LLM to narrate the spreadsheet so it can be understood.
That first condition "try to understand the intent" is where it could go wrong. Maybe it thinks the spreadsheet aligns with the intent, but it misunderstood the intent.
LLMs are a lossy validation, and while they work sometimes, when they fail they usually do so 'silently'.
Maybe we need some kind of method, framework to develop intent. Most of things that go wrong in knowledge working are down to lack of common understanding of intent.
> The one thing LLMs should consistently do is ensure that formatting is correct.
In JavaScript (and I assume most other programming languages) this is the job of static analysis tools (like eslint, prettier, typescript, etc.). I’m not aware of any LLM based tools which performs static analysis with as good a results as the traditional tools. Is static analysis not a thing in the spreadsheet world? Are there the tools which do static analysis on spreadsheets subpar, or offer some disadvantage not seen in other programming languages? And if so, are LLMs any better?
Just use a normal static analysis tool and shove the result to an LLM. I believe Anthropic properly figured that agents are the key, in addition to models, contrary to OpenAI that is run by a psycho that only believes in training the bigger model.
Sysadmin of a small company. I get asked pretty often to help with a pivot table, vlookup, or just general excel functions (and smartsheet, these users LOVE smartsheet)
Indeed, in a small enough org, the sysadmin/technologist becomes support of last resort for all the things.
> these users LOVE smartsheet
I hate smartsheet…
Excel or R. (Or more often, regex followed by pen and paper followed by more regex.)
Last time, I gave claude an invoice and asked it to change one item on it, it did so nicely and gave me the new invoice. Good thing I noticed it had also changed the bank account number..
The more complicated the spreadsheet and the more dependencies it has, the greater the room for error. These are probabilistic machines. You can use them, I use them all the time for different things, but you need to treat them like employees you can't even trust to copy a bank account number correctly.
We’ve tried to gently use them to automate some of our report generation and PDF->Invoice workflows and it’s a nightmare of silent changes and absence of logic.. basic things like specifically telling it “debits need to match credits” and “balance sheets need to balance” that are ignored.
Yeah, asking llm to edit one specific thing in a large or complex document/ codebase is like those repeated "give me the exact same image" gifs. It's fundamentally a statistical model so the only thing we can be _certain_ of is that _it's not_. It might get the desired change 100% correct but it's only gonna get the entire document 99 5%
Something that Claude Sonnet does when you use it to code is write scripts to test whether or not something is working. If it does that for Excel (e.g. some form of verification) it should be fine.
Besides, using AI is an exercise in a "trust but verify" approach to getting work done. If you asked a junior to do the task you'd check their output. Same goes for AI.
The use cases for spreadsheets are much more diverse than that. In my experience, spreadsheets just as often used for calculation. Many of them do require high accuracy, rely on determinism, and necessitate the understanding of maths ranging from basic arithmetic to statistics and engineering formulas. Financial models, for example, must be built up from ground truth and need to always use the right formulas with the right inputs to generate meaningful outputs.
I have personally worked with spreadsheet based financial models that use 100k+ rows x dozens of columns and involve 1000s of formulas that transform those data into the desired outputs. There was very little tolerance for mistakes.
That said, humans, working in these use cases, make mistakes >0% of the time. The question I often have with the incorporation of AI into human workflows is, will we eventually come to accept a certain level of error from them in the way we do for humans?
>Does it have to be perfect? Also no.
Yeah, but it could be perfect, why are there humans in the loop at all? That is all just math!
I don’t see the issue so much as the deterministic precision of an LLM, but the lack of observability of spreadsheets. Just looking at two different spreadsheets, it’s impossible to see what changes were made. It’s not like programming where you can run a `git diff` to see what changes an LLM agent made to a source code file. Or even a word processing document where the text changes are clear.
Spreadsheets work because the user sees the results of complex interconnected values and calculations. For the user, that complexity is hidden away and left in the background. The user just sees the results.
This would be a nightmare for most users to validate what changes an LLM made to a spreadsheet. There could be fundamental changes to a formula that could easily be hidden.
For me, that the concern with spreadsheets and LLMs - which is just as much a concern with spreadsheets themselves. Try collaborating with someone on a spreadsheet for modeling and you’ll know how frustrating it can be to try and figure out what changes were made.
I don't trust humans to do the kind of precise deterministic work you need in a spreadsheet!
Right, we shouldn’t use humans or LLMs. We should use regular deterministic computer programs.
For cases where that is not available, we should use a human and never an LLM.
I like to use Claude Code to write deterministic computer programs for me, which then perform the actual work. It saves a lot of time.
I had a big backlog of "nice to have scripts" I wanted to write for years, but couldn't find the time and energy for. A couple of months after I started using Claude Code, most of them exist.
That’s great and the only legitimate use case here. I suspect Microsoft will not try to limit customers to just writing scripts and will instead allow and perhaps even encourage them to let the AI go ham on a bunch of raw data with no intermediary code that could be reviewed.
Just a suspicion.
"regular deterministic computer programs" - otherwise known as the SUM function in Microsoft Excel
Most real-world spreadsheets I've worked with were fragile and sloppy, not precise and deterministic. Programmers always get shocked when they realize how many important things are built on extremely messy spreadsheets, and that people simply accept it. They rather just spend human hours correcting discrepancies than trying to build something maintainable.
Usually this is very hard because the tasks and the job often subtly shifts in somewhat unpredictable and unforeseen ways and there is no neat clean abstraction that you can just implement as an application. Too hererogeneous, too messy, too many exceptions. If you develop some clean elegant solution, next week there will be something that your shiny app doesn't allow and they'd have to submit a feature request or whatever.
In Excel, it's possible to just ad hoc adjust things and make it up as you go. It's not clean but very adaptable and flexible.
"I don't trust LLMs to do the kind of precise deterministic work" => I think LLM is not doing the precise arithmetic. It is the agent with lots of knowledge (skills) and tools. Precise deterministic work is done by tools (deterministic code). Skills brings domain knowledge and how to sequence a task. Agent executes it. LLM predicts the next token.
Sure, but this isn't requiring that the LLM do any math. The LLM is writing formulas and code to do the math. They are very good at that. And like any automated system you need to review the work.
Exactly, and if it can be done in a way that helps users better understand their own spreadsheets (which are often extremely complex codebases in a single file!) then this could be a huge use case for Claude.
> I don't trust LLMs to do the kind of precise deterministic work you need in a spreadsheet.
I was thinking along the same lines, but I could not articulate as well as you did.
Spreadsheet work is deterministic; LLM output is probabilistic. The two should be distinguished.
Still, its a productivity boost, which is always good.
Do you trust humans to be precise and deterministic, or even to be especially good at math?
This is talking about applying LLMs to formula creation and references, which they are actually pretty good at. Definitely not about replacing the spreadsheet's calculation engine.
I trust humans to not be able to shoot the company on the foot without even realizing it.
Why are we suddenly ok with giving every underpaid and exploited employee a foot gun and expect them to be responsible with it???
They're not great at arithmetic but at abstract mathematics and numerical coding they're pretty good actually.
It's widely known LLMs are terrible at even basic maths.
Claude for Excel isn't doing maths. It's doing Excel. If the llm is bad at maths then teaching it to use a tool that's good at maths seems sensible.
> I don't trust LLMs to do the kind of precise deterministic work you need in a spreadsheet.
Rightly so! But LLMs can still make you faster. Just don't expect too much from it.
If LLMs can replace mathematica for me when I'm doing affine yield curve calculations they can do a DCF for some banker idiots
LLMs are just a tool, though. Humans still have to verify them, like with very other tools out there
Eh, yes. In theory. In practice, and this is what I have experienced personally, bosses seem to think that you now have interns so you should be able to do 5x the output.. guess what that means. No verification or rubber stamp.
>I don't trust LLMs to do the kind of precise deterministic work
not just in a spreadsheet, any kind of deterministic work at all.
find me a reliable way around this. i don't think there is one. mcp/functions are a band aid and not consistent enough when precision is important.
after almost three years of using LLMs, i have not found a single case where i didn't have to review its output, which takes as long or longer than doing it by hand.
ML/AI is not my domain, so my knowledge is not deep nor technical. this is just my experience. do we need a new architecture to solve these problems?
ML/AI is not my domain but you don’t have to get all that technical to understand that LLMs run on probability. We need a new architecture to solve these problems.
you might trust when the precision is extremely high and others agree with that.
high precision is possible because they can realize that by multiple cross validations
ChatGPT is actively being used as a calculator.
I couldn’t agree more. I get all my perfectly deterministic work output from human beings!
If only we had created some device that could perform deterministic calculations and then wrote software that made it easy for humans to use such calculations.
ok but humans are idiots, if only we could make some sort of Alternate Idiot, a non-human but every bit as generally stupid as humans are! This A.I would be able to do every stupid thing humans did with the device that performed deterministic calculations only many times faster!
Yes and when the AI did that all the stupid humans could accept its output without question. This would save the humans a lot of work and thought and personal responsibility for any mistakes! See also Israel’s Lavender for an exciting example of this in action.
My concern is that my insurance company will reject a claim, or worse, because of something an LLM did to a spreadsheet.
Now, granted, that can also happen because Alex fat-fingered something in a cell, but that's something that's much easier to track down and reverse.
They already doing that with AI, rejecting claims at higher numbers than before .
Privatized insurance will always find a way to pay out less if they could get away with it . It is just nature of having the trifecta of profit motive , socialized risk and light regulation .
> They already doing that with AI, rejecting claims at higher numbers than before
Source?
Haven't risk based models been a thing for the last 15-20 years ?
If you think that insurance companies have "light regulation", I shudder to think of what "heavy regulation" would look like. (Source: I'm the CTO at an insurance company.)
Light did not mean to imply quantity of paperwork you have to do, rather are you allowed to do the things you want to do as a company.
More compliance or reporting requirements usually tend to favor the larger existing players who can afford to do it and that is also used to make the life difficult and reject more claims for the end user.
It is kind of thing that keeps you and me busy, major investors don't care about it all, the cost of the compliance or the lack is not more than a rounding number in the balance, the fines or penalties are puny and laughable.
The enormous profits year on year for decades now, the amount of consolidation allowed in the industry show that the industry is able to do mostly what they want pretty much, that is what I meant by light regulation.
I'm not sure we're looking at the same industry. Overall, insurance company profit margins are in the single digits, usually low single digits - and in many segments, they're frequently not profitable at all. To take one example, 2024 was the first profitable year for homeowners insurance companies since 2019, and even then, the segment's entire profit margin was 0.3% (not 3% - 0.3%).
https://riskandinsurance.com/us-pc-insurance-industry-posts-...
It's an accounting 101 thing to use all tricks in the book to reduce the reported profit, to avoid paying taxes on that profit.
The total profit of ALL US health insurance companies added together was $9bln in 2024: https://content.naic.org/sites/default/files/2024-annual-hea.... This is a profit margin of 0.8% down from 2.2% in the previous year.
Meta alone made $62bln in 2024: https://investor.atmeta.com/investor-news/press-release-deta...
So it's weird to see folks on a tech site talking about how enormous all the profits are in health insurance, and citations with numbers would be helpful to the discussion.
I worked in insurance-related tech for some time, and the providers (hospitals, large physician groups) and employers who actually pay for insurance have signficant market power in most regions, limiting what insurers can charge.
They have too much regulation, and too little auditing (at least in the managed healthcare business).
I agree, and I can see where it comes from (at least at the state level). The cycle is: bad trend happens that has deep root causes (let's say PE buying rural hospitals because of reduced Medicaid/Medicare reimbursements); legislators (rightfully) say "this shouldn't happen", but don't have the ability to address the deep root causes so they simply regulate healthcare M&As – now you have a bandaid on a problem that's going to pop up elsewhere.
I mean even in the simple stuff like denying payment for healthcare that should have been covered. CMS will come by and out a handful of cases, out of millions, every few years.
So obviously the company that prioritizes accuracy of coverage decisions by spending money on extra labor to audit itself is wasting money. Which means insureds have to waste more time getting the payment for healthcare they need.
> It is just nature of having the trifecta of profit motive , socialized risk and light regulation.
It's the nature of everything. They agree to pay you for something. It's nothing specific to "profit motive" in the sense you mean it.
I should have been clearer - profit maximization above all else as long it is mostly legal. Neither profit or profit maximization at all cost is nature of everything .
There are many other entity types from unions[1], cooperatives , public sector companies , quasi government entities, PBC, non profits that all offer insurance and can occasionally do it well.
We even have some in the US and don’t think it is communism even - like the FDIC or things like social security/ unemployment insurance.
At some level government and taxation itself is nothing but insurance ? We agree to paying taxes to mitigate against variety of risks including foreign invasion or smaller things like getting robbed on the street.
[1] Historically worker collectives or unions self-organized to socialize the risks of both major work ending injuries or death.
Ancient to modern armies operate on because of this insurance the two ingredients that made them not mercenaries - a form of long term insurance benefit (education, pension, land etc) or family members in the event of death and sovereign immunity for their actions.
Couldn't they accomplish the same thing by rejecting a certain percentage of claims totally at random?
That would be illegal though, the goal is do this legally after all.
We also have to remember all claims aren't equal. i.e. some claims end up being way costlier than others. You can achieve similar % margin outcomes by putting a ton of friction like, preconditions, multiple appeals processes and prior authorization for prior authorization, reviews by administrative doctors who have no expertise in the field being reviewed don't have to disclose their identity and so and on.
While U.S. system is most extreme or evolved, it is not unique, it is what you get when you end up privatize insurance any country with private insurance has some lighter version of this and is on the same journey .
Not that public health system or insurance a la NHS in UK or like Germany work, they are underfunded, mismanaged with long times in months to see a specialist and so on.
We have to choose our poison - unless you are rich of course, then the U.S. system is by far the best, people travel to the U.S. to get the kind of care that is not possible anywhere else.
> While U.S. system is most extreme or evolved, it is not unique, it is what you get when you end up privatize insurance any country with private insurance has some lighter version of this and is on the same journey .
I disagree with the statement that healthcare insurance is predominantly privatized in the US: Medicare and Medicaid, at least in 2023, outspent private plans for healthcare spending by about ~10% [1]; this is before accounting for government subsidies for private plans. And boy, does America have a very unique relationship with these programs.
https://www.healthsystemtracker.org/chart-collection/u-s-spe...
It is more nuanced, for example Medicare Advantage(Part C) is paid by Medicare money but it is profitable private operators who provide the plans and service it a fast growing part of Medicare .
John Oliver had an excellent segment coincidentally yesterday on this topic.
While the government pays for it, it is not managed or run by them so how to classify the program as public or private ?
Why does saying "AI did it" make it legal, if the outcome is the same?
>>They already doing that with AI, rejecting claims at higher numbers than before .
That's a feature, not a bug.
This is a great application of this quote. Insurance providers have 0 incentive to make their AI "good" at processing claims, in fact it's easy to see how "bad" AI can lead to a justification to deny more claims.
The question is how you define good. They surely want the Ai to be good in the sense that it rejects all claims that they think can get away with rejecting. But it should not reject those where rejection likely results in litigation and losing and having to pay damages.
Wait until a company has to restate earnings because of a bug in a Claudified Excel spreadsheet.
Yes it's surprising to see so much cynicism for something that has a real possibility of making so many people so much more productive. My mental model of the average excel user is of someone who doesn't care about excel, but cares about their business. If Claude can help them use excel and learn about their business faster, then this should make the world more productive and we all get richer. Claude can make mistakes, but it's not clear to me why people think that the ratio of results to mistakes will get worse here. I think there are many possible reasons why this could not work out, but many of the comments here just seem like unfounded cynicism.
Because people will be deeply affected by this, and not in the positive way. We already had this with copilot: https://i.imgur.com/nguIAsv.jpeg
Just as with copilot, this combines LLM's inability to repeatably do math correctly with peoples' overassurance in LLM's capabilities.
HN has a base of strong anti-AI bias, I assume is partially motivated by insecurity over being replaced, losing their jobs or having missed the boat on the AI.
I use AI every day. Without oversight, it does not work well.
If it doesn't work well, I will do it myself, because I care that things are done well.
None of this is me being scared of being replaced; quite the opposite. I'm one of the last generations of programmers who learned how to program and can debug and fix the mess your LLM leaves behind when you forgot to add "make sure it's a clean design and works" to the prompt.
Okay, that's maybe hyperbole, but sadly only a little bit. LLMs make me better at my job, they don't replace me.
Based on the comments here, it's surprisingly anything in society works at all. I didn't realize the bar was "everything perfect every time, perfectly flexible and adaptable". What a joy some of these folks must be to work with, answering every new technology with endless reasons why it's worthless and will never work.
I think perhaps you underestimate how antithetical the current batch of LLM AI's is to what most programmers strive for every day, and what we want from our tools. Its not about losing our job, its about "correctness". (or as said below - deterministic)
In a lot of jobs, particularly in creative industries, or marketing, media and writing, the definition of a job well done is a fairly grey area. I think AI will be mostly disruptive in these areas.
But in programming there is a hard minimum of quality. Given a set of inputs, does the program return the correct answer or not? When you ask it what 2+2, do you get 4?
When you ask AI anything, it might be right 50% of the time, or 70% of the time, but you can't blindly trust the answer. A lot of us just find that not very useful.
I am a SWE myself and use LLMs to write ~100% of my code. That does not mean I fire and forget multiplexed codex instances. Many times I step through and approve every edit. Even if it was nothing but a glorified stenographer - there are substantial time savings in being able to prototype and validate ideas quickly.
> But in programming there is a hard minimum of quality. Given a set of inputs, does the program return the correct answer or not? When you ask it what 2+2, do you get 4?
Whether something works or not matters less than whether someone will pay for it.
Modt of the time when using AI I have a lot more than 1 shot to ensure everything is correct.
> HN has a base of strong anti-AI bias
HN constantly points out the flaws, gaps, and failings of AI. But the same is true of any technology discussed on HN. You could describe HN as having an anti-technology bias, because HN complains about the failings of tech all day every day.
> HN has a base of strong anti-AI bias
Quite the opposite, actually. You can always find five stories on the front page about some AI product or feature. Meanwhile, you have people like yourself who convince themselves that any pushback is done by people who just don't see the true value of it yet and that they're about to miss out!! Some kind of attempt at spreading FOMO, I guess.
> HN has a base of strong anti-AI bias
If anything, HN, has a pro-AI bias. I don't know of any other medium where discussions about AI consistently get this much frontpage time, this amount of discussion, and this many people reporting positive experiences with it. It's definitely true that HN isn't the raging pro-AI hypetrain it was two years ago, but that shouldn't be mistaken for "strong anti-AI bias".
Outside of HN I am seeing, at best, an ambivalent reaction: plenty of people are interested, almost everyone tried it, very few people genuinely like it. They are happy to use it when it is convenient, but couldn't care less if it disappeared tomorrow.
There's also a small but vocal group which absolutely hates AI and will actively boycott any creative-related company stupid enough to admit to using it, but that crowd doesn't really seem to hang out on HN.
>but couldn't care less if it disappeared tomorrow.
Wonder how true that is. Some things incorporate in your life so subtly that you only become aware of them when totally switched off.
I really don’t think this is accurate. I think the median opinion here is to be suspicious of claims made about AI, and I don’t think that’s necessarily a bad thing. But I also regularly see posts talking about AI positively (e.g. simonw), or talking about it negatively. I think this is a good thing, it is nice to have a diversity of opinions on a technology. It's a feature, not a bug.
HN has an obsession with quality too, which has merit, but is often economically irrelevant.
When US-East-1 failed, lots of people talked about how the lesson was cloud agnosticism and multi cloud architecture. The practical economic lesson for most is that if US-East-1 fails, nobody will get mad at you. Cloud failure is viewed as an act of god.
Anti-AI bias is motivated by the waste of natural resources due to a handful of non-technical douchebag tech bros.
Everything isn't about money, I know that status and power are all you ai narcissists dream about. But you'll never be Bill Gates, nor will you be Elon Musk.
Once ai has gone the way of "Web3", "NFTs", "blockchain", "3D tvs", etc; You'll find a new grift to latch your life savings onto.
Non-reproducability is the biggest issue here. You deliver a report in 5 minutes to CFO, he comes back after lunch, gives you updated data to adjust a bit of a report and 5 minutes later gets a new report that has some non related to update number changed and asks why? what do you do?
The vast majority of people in business and science are using spreadsheets for complex algorithmic things they weren't really designed for, and we find a metric fuckton of errors in the sheets when you actually bother looking auditing them, mistakes which are not at all obvious without troubleshooting by... manually checking each and every cell & cell relation, peering through parentheses, following references. It's a nightmare to troubleshoot.
LLMs specialize in making up plausible things with a minimum of human effort, but their downside is that they're very good at making up plausible things which are covertly erroneous. It's a nightmare to troubleshoot.
There is already an abject inability to provision the labor to verify Excel reasoning when it's composed by humans.
I'm dead certain that Claude will be able to produce plausibly correct spreadsheets. How important is accuracy to you? How life-critical is the end result? What are your odds, with the current auditing workflow?
Okay! Now! Half of the users just got laid off because management thinks Claude is Good Enough. How about now?
LLMs are getting quite good at reviewing the results and implementations, though
Not really, they're only as good as their context and they do miss and forget important things. It doesn't matter how often, because they do, and they will tell you with 100% confidence and with every synonym of "sure" that they caught it all. That's the issue.
I am very confident that these tools are better than the median programmer at code review now. They are certainly much more diligent. An actually useful standard to compare them to is human review, and for technical problems, they definitely pass it. That said, they’re still not great at giving design feedback.
But GPT-5 Pro, and to a certain extent GPT-5 Codex, can spot complex bugs like race conditions, or subtly incorrect logic like memory misuse in C, remarkably well. It is a shame GPT-5 Pro is locked behind a $200/month subscription, which means most people do not understand just how good the frontier models are at this type of task now.
Yeah, this could be a pretty big deal. Not everyone is an excel expert, but nearly everyone finds themselves having to work with data in excel at some time or other.
The lady doth protest too much. People see every AI limitation crystal clear, but zero self awareness of their own fallibility.
> What is with the negativity in these comments?
A lot of us have seen the effects of AI tools in the hands of people who don't understand how or why to use the tools. I've already seen AI use/misuse get two people fired. One was a line-of-business employee who relied on output without ever checking it, got herself into a pretty deep hole in 3 weeks. Another was a C suite person who tried to run an AI tool development project and wasted double their salary in 3 months, nothing to show for it but the bill, fired.
In both cases the person did not understand the limits of the tools and kept replacing facts with their desires and their own misunderstanding of AI. The C suite person even tried to tell a vendor they were wrong about their own product because "I found out from AI".
AI right now is fireworks. It's great when you know how to use it, but if you half-ass it you'll blow your fingers off very easily.
Yeah, the danger lies not in AI itself, but in inexperienced users treating it as a magic solution.
> but these jobs are going to be the first on the chopping block as these integrations mature.
I'm not even sure that has to be true anymore. From my admittedly superficial impression of the page, this appears to be a tool for building tools. There are plenty of organizations that are resource constrained, that are doing things the way they have always done thing in Excel, simply because they cannot allocate someone to modify what is already in place to better suit their current needs. For them, this is more of a quality of life and quality of out improvement. This is not like traditional software development, where organizations are far more likely to purchase a product or service to do a job (and where the vendors of those products and services are going to do their best to eliminate developers).
It is bad in a very specific sense, but I did not see any other comments express the bad parts instead of focusing merely on the accuracy part ( which is an issue, but not the issue ):
- this opens up ridiculous flood of data that would otherwise be semi-private to one company providing this service - this works well small data sets, but will choke on ones it will need to divvy up into chunks inviting interesting ( and yet unknown ) errors
There is a real benefit to being able to 'talk to data', but anyone who has seen corporate culture up close and personal knows exactly where it will end.
edit: an i saying all this as as person, who actually likes llms.
Indeed. Take the New Zealand Department of Health as an example; it managed its entire NZD$28 billion budget (USD$16B) in a single Excel spreadsheet.
https://www.theregister.com/2025/03/10/nz_health_excel_sprea...
[edit: Added link]
Probably because many people here are software developers, and wrapping spreadsheets in deterministic logic and a consistent UI covers... most software use cases.
What does scaffolding of spreadsheets mean? I see the term scaffolding frequently in the context of AI-related articles and not familiar with this method and I’m hesitant to ask an LLM.
Scaffolding typically just refers to a larger state machine style control flow governing an agent's behavior and the suite of external tools it has access to.
Anthropic now has all your company's data, and all you saved was the cost of one human minus however much they charge for this. The good news is it can't have your data again! So starting from the 163rd-165th person you fire, you start to see a good return and all you've sacrificed is exactitude, precision, judgement, customer service and a little bit of public perception!
> but these jobs are going to be the first on the chopping block as these integrations mature.
Perhaps this is part of the negativity? This is a bad thing for the middle class.
in the short run. In the long run, productivity gains benefit* all of us (in a functional market economy).
*material benefit. In terms of spirit and purpose, the older I get the more I think maybe the Amish are on to something. Work gives our lives purpose, and the closer the work is to our core needs, the better it feels. Labor saving so that most of us are just entertaining each other on social networks may lead to a worse society (but hey, our material needs are met!)
agree with you, but it cannot be stopped. development of technology always makes wealth distribution more centralized
I kind of get what you're saying but can you explain your reasoning or provide a source?
My theory: a lot of software we build is the supposed solve for a 'crappy spreadsheet'. a) that isnt' much of a moat, b) you're watching generalization of software happen in real time.
Crappy spreadsheet is just the codification of business processes. Those are inherently messy and there's lots of assumptions, lots of edge cases. That's why spreadsheets tend towards crappy on a long enough timeline. It's a fundamentally messy problem.
Spreadsheets are an abstraction over a messy reality, lossy. They were already generalizing reality.
Now we generalize the generalization. It is this lossy reality that people are worried about with AI in HN.
The biggest problem with spreadsheets is that they tend to be accounts for the accumulation of technical debt, which is an area that AI tools are not yet very good at retiring, but very good at making additional withdrawals from.
I have to admit that my first thought was “April’s fool”. But you are right. It makes a lot of sense (if they can get it to work well). Not only is Excel the world’s biggest “programming language”. It’s probably also one of the most unintuitive ways to program.
If you exclude macros with IO it’s actually the most popular purely functional programming language (no quotes) on the planet by far.
Why unintuitive?
Can’t speak for everyone, but the reason I’m negative in the context of this idea is that it’s a stupid idea.
It's like the negativity whenever a post talks about hiring or firing. A lot of people are afraid that they are going to lose their jobs to AI.
I used to live in excel.
The issue isn’t in creating a new monstrosity in excel.
The issue is the poor SoB who has to spelunk through the damn thing to figure out what it does.
Excel is the sweet spot of just enough to be useful, capable enough to be extensible, yet gated enough to ensure everyone doesn’t auto run foreign macros (or whatever horror is more appropriate).
In the simplest terms - it’s not excel, it’s the business logic. If an excel file works, it’s because theres someone who “gets” it in the firm.
I used to live in Excel too. I've trudged through plenty of awful worksheets. The output I've seen from AI is actually more neatly organized than most of what I used to receive in outlook. Most of that wasn't hyper-sophisticated cap table analyses. It was analysis from a Jr Analyst or line employee trying to combine a few different data sources to get some signal on how XYZ function of the business was performing. AI automation is perfectly suitable for this.
How?
Neat formatting didn't save any model from having the wrong formula pasted in.
Being neat was never a substitute for being well rested, or sufficiently caffeinated.
Have you seen how AI functions in the hands of someone who isn't a domain expert? I've used it for things I had no idea about, like Astro+ web dev. User ignorance was magnified spectacularly.
This is going to have Jr Analysts dumping well formatted junk in email boxes within a month.
Whats with claiming negativity when most of the comments here are positive?
I have to remember this one. Waltz into the room and proclaim, why is everyone so negative? It's great because x, y and z. It looks pretty great.
> offense to these people but Sonnet 4.5 is already at the level where it would be able to replicate or beat the level of analysis they typically provide.
No offense, but this is pure fantasy. The level of analysis they typically provide doesn't suffer from the same high baseline level of completely made up numbers of your favorite LLM.
You would be far better off using an LLM to replace a complex spreadsheet with a Python script and SQLite.
this will push the development of open source models.
people think of privacy at first regards of data, local deployment of open source models are the first choice for them
Honestly as a dev I hate Excel its a whole mess I dont understand. I will gladly use Claude for Excel. It will understand the business needs from the data more than I a mere developer just trying to get back to regular developer work.
> What is with the negativity in these comments?
> these jobs are going to be the first on the chopping block as these integrations mature.
Those two things are maybe related? So many of my friends don't enjoy the same privileges as I do, and have a more tenuous connection to being gainfully employed.
I think excel is a dead end. LLM agents will probably greatly prefer SQL, sqlite, and Python instead of bulky made-for-regular-folks excel.
Versatility and efficiency explode while human usability tanks, but who cares at that point?
Database might be the future, but viable solution on excel are evidence to prove that it works
> Even just basic automation/scaffolding of spreadsheets would be a big productivity boost for many employees.
When most of it is wild hallucinations? Not really.
For many employees leveraging Excel for manipulating important data, it could cripple careers.
For spreadsheets that influence financial decisions or touch PPI/PII, it could lead to regulatory disasters and even bankruptcies.
Purge hallucinations from LLMs, _then_ let it touch the important shite. Doing it in the reverse order is just begging for a FAFO apocalypse.
> What is with the negativity in these comments?
Some people - normal people - understand the difference between the holistic experience of a mathematically informed opinion and an actual model.
It's just that normal people always wanted the holistic experience of an answer. Hardly anyone wants a right answer. They have an answer in their heads, and they want a defensible journey to that answer. That is the purpose of Excel in 95% of places it is used.
Lately people have been calling this "syncophancy." This was always the problem. Sycophancy is the product.
Claude Excel is leaning deeply into this garbage.
It seems like to me the answer is moreso "People on HN are so far removed from the real use cases for this kind of automation they simply have no idea what they're talking about".
This is so correct it hurts
It's actually really cool. I will say that "spreadsheets" remain a bandaid over dysfunctional UIs, processes, etc and engineering spends a lot of time enabling these bandaids vs someone just saying "I need to see number X" and not "a BI analytics data in a realtime spreadsheet!", etc.
> How teams use Claude for Excel
Who are these teams that can get value from Anthropic? One MCP and my context window is used up and Claude tells me to start a new chat.
MCPs and context window sizing, putting the engineering into prompt engineering.
I second this. Spreadsheets are the primary tool used for 15% of the U.S. economy. Productivity improvements will affect hundreds of millions of users globally. Each increment in progress is a massive time save and value add.
The criticisms broadly fall between "spreadsheets are bad" and "AI will cause more trouble than it solves".
This release is a dot in a trend towards everyone having a Goldman-Sachs level analyst at their disposal 24/7. This is a huge deal for the average person or business. Our expectation (disclaimer: I work in this space) is that spreadsheet intelligence will soon be a solved problem. The "harder" problem is the instruction set and human <> machine prompting.
For the "spreadsheets are bad" crowd -- sure, they have problems, but users have spoken and they are the preferred interface for analysis, project management and lightweight database work globally. All solutions to "the spreadsheet problem" come with their own UX and usability tradeoffs, so it'a a balance.
Congrats to the Claude team and looking forward to the next release!
> Each increment in progress is a massive time save and value add.
Based on the history of digitalization of businesses from the 1980s onwards, the spreadsheets will just balloon in number and size and there will be more rules and more procedures and more forms and reports to file until the efficiency gains are neutralized (or almost neutralized).
We'll hit a new plateau somewhere, for sure. Still, I'm glad I'm not doing my spreadsheets on paper so net win so far!
I'm a co-founder of Calcapp, an app builder for formula-driven apps using Excel-like formulas. I spent a couple of days using Claude Code to build 20 new templates for us, and I was blown away. It was able to one-shot most apps, generating competent, intricate apps from having looked at a sample JSON file I put together. I briefly told it about extensions we had made to Excel functions (including lambdas for FILTER, named sort type enums for XMATCH, etc), and it picked those up immediately.
At one point, it generated a verbose formula and mentioned, off-handedly, that it would have been prettier had Calcapp supported LET. "It does!", I replied, "and as an extension, you can use := instead of , to separate names and values!") and it promptly rewrote it using our extended syntax, producing a sleek formula.
These templates were for various verticals, like real estate, financial planning and retail, and I would have been hard-pressed to produce them without Claude's domain knowledge. And I did it in a weekend! Well, "we" did it in a weekend.
So this development doesn't really surprise me. I'm sure that Claude will be right at home in Excel, and I have already thought about how great it would be if Claude Code found a permanent home in our app designer. I'm concerned about the cost, though, so I'm holding off for now. But it does seem unfair that I get to use Claude to write apps with Calcapp, while our customers don't get that privilege.
(I wrote more about integrating Claude Code here: https://news.ycombinator.com/item?id=45662229)
Seems everyone is speculating features instead of just reading TFA which does in fact list features:
- Get answers about any cell in seconds: Navigate complex models instantly. Ask Claude about specific formulas, entire worksheets, or calculation flows across tabs. Every explanation includes cell-level citations so you can verify the logic.
- Test scenarios without breaking formulas: Update assumptions across your entire model while preserving all dependencies. Test different scenarios quickly—Claude highlights every change with explanations for full transparency.
- Debug and fix errors: Trace #REF!, #VALUE!, and circular reference errors to their source in seconds. Claude explains what went wrong and how to fix it without disrupting the rest of your model.
- Build models or fill existing templates: Create draft financial models from scratch based on your requirements. Or populate existing templates with fresh data while maintaining all formulas and structure.
If this can reliably deal with the REF, VALUE, and NA problems, it'll be worth it for that alone.
Oh and deal with dates before 1900.
Excel is a gift from God if you stay in its lane. If you ever so slightly deviate, not even the Devil can help you.
But maybe, juuuuust maybe, AI can?
"not even Devil can help you.
But maybe, juuuuust maybe, AI can?"
Bold assumption that the devil and AI aren't aligned ;)
The greatest trick the devil ever pulled was convincing the world he didn't exist
Nah, the greatest trick the devil ever pulled was convincing the world that Machine Learning is a legitimate field of study, and not just thinly veiled demon summoning.
I feel similarly about MS Word. It can actually produce decent documents if you learn how to use it, in particularly if you use styles consistently and never, ever touch the bold, italic, colours etc. (outside of defining said styles, although the defaults are probably all most people need). Unfortunately I think the appeal of Word is that you don't have to learn this and it will just do what you want. Is AI the panacea that will both do what you want and give you the right answers every time?
Also people complaining about AI inaccuracy are just technical people that like precision. The vast majority of the world is people who dont give a damn about accuracy or even correctness. They just want to appear as if not completely useless to people that could potentially affect their salary
I can pretty reliably guess that approximately 100% of all companies in the world use excel tables for financial data and for processes. Ok, this was a joke. It's actually 99.99% of all companies. One would think that financial data, inventory and stuff like that should be damn precise. No?
How precise do they really need to be? If there's 3 of a widget on the shelf in the factory, and the factory uses 1000 per day, is it crucial to know that there's 3 of them, and not 0 or 50? Either way, the factory ain't running today or tomorrow or until more of those things come in. Similarly, what's $3 missing from an internal spreadsheet when the company costs $5,000 an hour to operate (or $10 million a year). Obviously errors accumulate so the books need to be reconciled, but apl that stuff only need to be sufficiently directionally accurate with enough precision. If precision is free, then sure, but if a good enough job is cheaper? We all make that call every day.
If you have 2000 hectares of land you need to buy the exact amount of seeds to sow them. If you buy less you are losing money, if you buy more it is useless and you are losing money. If you have trucks or other machinery in the company you need to report exact amount of fuel needed/used, or either they won't run or you lose money on machinery missing fuel. If you need to tax a company, it is pretty important if there are 100 tons of steel used or 1000 tons. Or if the company has 5 factories to be taxed or 15. Etc.
You are anthropomorphizing LLM programs, you assume that if a number in a spreadsheed is big, then program can somehow understand it that it is a big number and if it will make an error it will be a small order error like a human would make. Human process: "hmm, here is a calculation where we divide our imports by number of subsidiaries, let me estimate this in my head, ok, looks like 7320." (actual correct answer was 7340, bu human made a small, typical mistake in the math) LLM program process: it literally uses heat maps and randomization to arrive at each particular character in a row. So it may be 7340, or it may be 8745632, or 1320, or whatever. There is a comment here at a top, from another user, where he queried LLM to make a change of value in the document and it did it correctly. But at the same time it replaced bank account number with a different bank account number. Because to LLM it is the same - sixteen digit in the field, or another sixteen digits in a field, it is the same for LLM. Because it is not AI and doesn't "understand" what it does.
If you have 2000 hecatares land, there is no way you're buying the exact right amount of seeds. You overbuy seeds by as little as you can, but seeds get loaded via tractor bucket, which is fairly messy. You're going to lose a decent amount of seeds. Thus, a lb or kilo of seeds or < 1% in the scheme of things isn't even going to be noticed, much less cause the demise of your farm
For fuel, similarly, you're going to lose militers to evaporation on a hot day, so similarly, being off my ml isn't material.
If you tax a company, fine, sure, the company is going to want it to be right, but 1 or two tons in a 10,000 ton order is again, < 1%. There is some threshold below which precision is extra unnecessary work, though if you have problems with thieves and corruption, you're going to want additional precision that isn't necessary elsewhere.
As to where in my comment I'm anthropomorphizing LLMs, you're going to have to point out where I did that, as the word LLM doesn't appear anywhere in my comment, so it feels like you're projecting claims my comment does not make, as it is LLM neutral and merely point of that 100% exact precision doesn't come without a cost.
"just" technical people who like precision are the reason we are here, typing this, and why lots of parts of our world is pretty cool and comfortable. I wouldn't say that's useless and "just" some people when it clearly is generating unmistakable value
The best thing that can come from this is unit tests for Excel.
LLMs work best when they can call tools (edit the sheet) and test their results in a loop.
It's like the "value seek" thing Excel has had since forever; "adjust these values until this cell is X"
Excel doesn't have any way to verify that every formula in that 60k line sheet is correct and someone hasn't accidentally replaced one with a static number for example.
They can try, but doubt anyone serious will adopt it.
Tried integrating chatgpt into my finance job to see how far I can get. Mega jikes...millions of dollars of hallucinated mistakes.
Worse you don't have the same tight feedback loop you've got in programming that'll tell you when something is wrong. Compile errors, unit tests etc. You basically need to walk through everything it did to figure out what's real and what's hallucinations. Basically fails silently. If they roll that out at scale in the financial system...interesting times ahead.
Still presumably there is something around spreadsheets it'll be able to do - the spreadsheet equivalent of boilerplate code whatever that may be
I'm bad with spread sheets so maybe this is trivial but having an llm tell me how to connect my sheet to whatever data I'm using at the moment and it coming up with a link or sql query or both has allowed me to quickly pull in data where I'd normally eyeball it and move on or worst case do it partially manually if really important.
It's like one off scripts in a sense? I'm not doing complex formulas I just need to know how I can pull data into a sheet and then I'll bucketize or graph it myself.
Again probably because I'm not the most adept user but it has definitely been a positive use case for me.
I suspect my use case is pretty boilerplatey :)
Good to know that it works well for that.
>I'm not doing complex formulas
Neither am I frankly. Finance stuff can get conceptually complicated even with simple addition & multiplication though. e.g. I deal with a lot of offshore stuff, so the average spreadsheet is a mix of currencies, jurisdictions and companies that are interlinked. I could probably talk you through it high level in an hour with a pen & paper, but the LLMs just can't see the forest for all the trees in the raw sheet.
AI slop eaters will still eat it up and ask for seconds. Pigs in oats seeing dollar signs.
Anthropic is in a weird place for me right now. They're growing fast , creating little projects that i'd love to try, but their customer service was so bad for me as a max subscriber that I set an ethical boundary for myself to avoid their services until such point that it appears that they care about their customers whatsoever.
I keep searching for a sign, but everyone I talk to has horror stories. It sucks as a technologist that just wants to play with the thing; oh well.
> I keep searching for a sign, but everyone I talk to has horror stories. It sucks as a technologist that just wants to play with the thing; oh well.
The reason that Claude Code doesn't have an IDE is because ~"we think the IDE will obsolete in a year, so it seemed like a waste of time to create one."
Noam Shazeer said on a Dwarkesh podcast that he stopped cleaning his garage, because a robot will be able to do it very soon.
If you are operating under the beliefs these folks have, then things like IDEs, cleaning up, and customer service are stupid annoyances that will become obsolete very soon.
To be clear, I have huge respect for everyone mentioned above, especially Noam.
> Noam Shazeer said on a Dwarkesh podcast that he stopped cleaning his garage, because a robot will be able to do it very soon.
We all come up with excuses for why we haven't done a chore, but some of us need to sound a bit more plausible to other members of the household than that.
It would get about the same reaction as "I'm not going to wash the dishes tonight, the rapture is tomorrow."
I want to make it very clear that this was a lighthearted response from Noam to the "AGI timeline" question.
Noam does not do a lot of interviews, and I really hope that stuff like my dumb comment does not prevent him from doing more in the future. We could all learn a lot from him. I am not sure that everyone understands everything that this man has given us.
"Noam Shazeer said on a Dwarkesh podcast that he stopped cleaning his garage, because a robot will be able to do it very soon".
How much is the robot going to cost in a year? 100k? 200k? Not mass market pricing for sure.
Meanwhile, today he could pay someone $1000 to clean his garage.
I would do it for free, just to answer the question of what does a genius of his caliber have in his garage? Probably the same stuff most people do, but it would still be interesting.
I don’t think the point was about having a clean space, it was in response to a question along the lines of: when do you think we will achieve AGI?
Trust me, I’m a genius of his caliber. Want to clean my garage? You free next week?
There is this homogenization happening in AI. No matter what their original mission was, all the AI companies are now building AI-powered gimmicks hoping to stumble upon something profitable. The investors are waiting...
Customer service at B2C companies can only go downhill or stay level. See Google, Apple, Microsoft etc. At B2B it maaaybe can improve, but only when a ten times bigger customer strongarms a company into doing it.
bad customer service comes from low priority. I think anthropic prioritize new growth point over small number of customer’s feedback, that’s why they publish new product, features so frequently, there are so much possible potential opportunities for them to focus
What happened? I'm a Max subscriber and I'd like to know what to look out for!
Best way to think of it is this: Right now you are not the customer. Investors are.
The money people pay in monthly fees to Anthropic for even the top Max sub likely doesn't come closer to covering the energy & infrastructure costs for running the system.
You can prove this to yourself by just trying to cost out what it takes to build the hardware capable of running a model of this size at this speed and running it locally. It's tens of thousands of dollars just to build the hardware, not even considering the energy bills.
So I imagine the goal right now is to pull in a mass audience and prove the model, to get people hooked, to get management and talent at software firms pushing these tools.
And I guess there's some in management and the investment community that thinks this will come with huge labour cost reductions but I think they may be dreaming.
... And then.. I guess... jack the price up? Or wait for Moore's Law?
So it's not a surprise to me they're not jumping to try and service individual subscribers who are paying probably a fraction of what it costs them to the run the service.
I dunno, I got sick of paying the price for Max and I now use the Claude Code tool but redirect it to DeepSeek's API and use their (inferior but still tolerable) model via API. It's probably 1/4 the cost for about 3/4 the product. It's actually amazing how much of the intelligence is built into the tool itself instead of just the model. It's often incredibly hard to tell the difference bertween DeepSeek output and what I got from Sonnet 4 or Sonnet 4.5
I've been playing around with local LLMs in Ollama, just for fun. I have an RTX 4080 Super, a Ryzen 5950X with 32 threads, and 64 GB of system memory. A very good computer, but decidedly consumer-level hardware.
I have primarily been using the 120b gpt-oss model. It's definitely worse than Claude and GPT-5, but not by, like, an order of magnitude or anything. It's also clearly better than ChatGPT was when it first came out. Text generates a bit slowly, but it's perfectly usable.
So it doesn't seem so unreasonable to me that costs could come down in a few years?
You are bang on.
Every AI company right now (except Google Meta and Microsoft) has their valuations based on the expectation of a future monopoly on AGI. None of their business models today or in the foreseeable horizon are even positive let alone world-dominating. The continued funding rounds are all apparently based on expectation of becoming the sole player.
The continuing advancement of open source / open weights models keeps me from being a believer.
I’ve placed my bet and feel secure where it is.
The thing really missing from multi-megabyte excel sheets of business critical carnage was a non-deterministic rewrite tool. It'll interact excitingly with the industry standard of no automated testing whatsoever.
I 100% believe generative AI can change a spreadsheet. Turn the xslx into text, mutate that, turn it back into an xslx, throw it away if it didn't parse at all. The result will look pretty similar to the original too, since spreadsheets are great at showing immediately local context and nothing else.
Also, we've done a pretty good job of training people that chatgpt works great, so there's good reason for them to expect claude for excel to work great too.
I'd really like the results of this to be considered negligence with non-survivable fines for the reckless stupidity, but more likely, it'll be seen as an act of god. Like all the other broken shit in the IT world.
This is going to be massive if it works as well as I suspect it might.
I think many software engineers overlook how many companies have huge (billion dollar) processes run through Excel.
It's much less about 'greenfield' new excel sheets and much more about fixing/improving existing ones. If it works as well as Claude Code works for code, then it will get pretty crazy adoption I suspect (unless Microsoft beats them to it).
> I think many software engineers overlook how many companies have huge (billion dollar) processes run through Excel.
So they can fire the two dudes that take care of it, lose 15 years of in house knowledge to save 200k a year and cry in a few months when their magic tool shits the bed ?
Massive win indeed
You think it's better for the company to have "two dudes" that are completely indispensable and whose work will be completely useless if they die / leave?
I think you're making an argument for LLMs, not against.
These two dudes can train the next generation, you know, like we've been doing since humans exist... instead of relying on some centralised point of failure somewhere thousands of km away which might or might not break your company whenever they decide to update something.
You're one of the people who saw nothing wrong with moving all our industries to asia right ? "It's cheaper so it's obviously better", if you don't think about any of the externalities and long term consequences sure...
Management have been executing this genius plan for decades without Ai.
If the company is half baked, those "two dudes" will become indispensable beyond belief. They are the ones that understand how Excel works far deeper, and paired with Claude for Excel they become far far more valuable.
At my org it more that these AI tools finally allow the employees to get through things at all. The deadlines are getting met for the first time, maybe ever. We can at last get to the projects that will make the company money instead of chasing ghosts from 2021. The burn down charts are warm now.
> This is going to be massive if it works as well as I suspect it might.
Until Microsoft does its anti-competitive thing and find a way to break this in the file format, because this is exactly what copilot in excel does.
That said, Copilot in Excel is pretty much hot garbage still so anything will be better than that.
What do you mean, what is copilot in excel doing exactly?
From the signup form mentioning Private Equity / Venture Capital, Hedge Fund, Investment Banking... this seems squarely aimed at financial modeling. Which is really, really cool.
I've worked alongside sell-side investment bankers in a prior startup, and so much of the work is in taking a messy set of statements from a company, understanding the underlying assumptions, and building, and rebuilding, and rebuilding, 3-statement models that not only adhere to standard conventions (perhaps best introed by https://www.wallstreetprep.com/knowledge/build-integrated-3-... ) but also are highly customized for different assumptions that can range from seasonality to sensitivity to creative deal structures.
It is quite common for people to pull many, many all-nighters to try to tweak these models in response to a senior banker or a client having an idea! And one might argue there are way too many similar-looking numbers to keep a human banker from "hallucinating," much less an LLM.
But fundamentally, a 3-statement model and all its build-sheets are a dependency graph with loosely connected human-readable labels, and that means you can write tools that let an LLM crawl that dependency graph in a reliable and semantically meaningful way. And that lets you build really cool things, really fast.
I'm of the opinion that giving small companies the ability to present their finances to investors, the same way Fortune 500 companies hire armies of bankers to do, is vital to a healthy economy, and to giving Main Street the best possible chance to succeed and grow. This is a massive step in the right direction.
Presenting your finances to investors via a tool designed for generation of plausible looking data is fraud.
Presenting false data to investors is fraud, doesn't matter how it was generated. In fact, humans are quite good at "generating plausible looking data", doesn't mean human generated spreadsheets are fraud.
On the other hand, presenting truthful data to investors is distinctly not fraud, and this again does not depend on the generation method.
> doesn't matter how it was generated
is there precedent for this supposed ruling?
US v Simon 1969, see [0] for a review.
Establishes that accountants who certify financials are liable if they are incorrect. In particular, if they have a reason to believe they might not be accurate and they certify anyway they are liable. And at this stage of development it’s pretty clear that you need to double check LLM generated numbers.
Obviously no clue if this would hold up with today’s court, but I also wasn’t making a legal statement before. I’m not a lawyer and I’m not trying to pretend to be one.
[0] https://scholarship.law.stjohns.edu/cgi/viewcontent.cgi?arti...
Fascinating thank you for the link
If humans "generate plausible looking data" despite any processes to ensure data quality they've likely engaged in willful fraud.
An LLM doing so needn't even be willful from the author's part. We're going to see issues with forecasts/slide decks full of inaccuracies that are hard to review.
I think my main point is just because an LLM can lie, doesn’t necessarily mean an LLM generated slide is fraud. It could very easily be correct and verified/certified by the accountant and not fraud. Just cuz the text was generated first by an LLM doesn’t mean fraud.
That being said, oh for sure this will lead to more incidental fraud (and deliberate fraud) and I’m sure it already has. Would be curious to see the prevalence of em-dash’s in 10k’s over the years.
You might have accidentally described what accounting is.
Completely understand the sentiment, but it doesn't apply here, because what's being generated are formulas!
Standardized 3-statement models in Excel are designed to be auditable, with or without AI, because (to only slightly simplify) every cell is either a blue input (which must come from standard exports of the company's accounting books, other auditable inventory/CRM/etc. data, or a visible hardcoded constant), or a black formula that cannot have hardcoded values, and must be simple.
If every buyer can audit, with tools like this, that the formulas match the verbal semantics of the model, there's even less incentive than there is now to fudge the formula level. (And with Wall Street conventions, there's nowhere to hide a prompt injection, because you're supposed to keep every formula to only a few characters, and use breakout "build" rows that can themselves be visually audited.)
And sure, you could conceivably use any AI tool to generate a plausible list of numbers at the input level, but that was equally easy, and equally dependent on context to be fraudulent or not, ever since that famous Excel 1990 elevator commercial: https://www.youtube.com/watch?v=kOO31qFmi9A&t=61s
At the end of the day, the difference between "they want to see this growth, let's fudge it" and "they want to see this growth, let's calculate the exact metrics we need to hit to make that happen, and be transparent about how that's feasible" has always been a matter of trust, not technology.
Tech like this means that people who want to do things the right way can do it as quickly as people who wanted to play loose with the numbers, and that's an equalizer that's on the right side of history.
I'm not excited about having LLMs generate spreadsheets or formulas. But, I think LLMs could be particularly useful in helping me find inconsistent formulas or errors that are challenging to identify. Especially in larger, complex spreadsheets touched by multiple people over the course of months.
For once in my life, I actually had a delightful interaction with an LLM last week. I was changing some text in an Excel sheet in a very progromatic way that could have easily been done with the regex functions in Excel. But I'm not really great with regex, and it was only 15 or so cells, so I was content to just do it manually. After three or four cells, Copilot figured out what I was doing and suggested the rest of the changes for me.
This is what I want AI to do, not generate wrong answers and hallucinate girlfriends.
Thanks for reminding me to check if the REGEXEXTRACT, REGEXREPLACE, and REGEXTEST functions had landed for me yet. They have! Good, because sometime in 2027 the library providing RegEx in VBA will be yanked. https://youtu.be/pGH9LdgkJio
One approach is to produce read-only data in BI tools: users are free to export anything they want and make their own spreadsheets, but those are for their own use only. Reference data is produced every day by a central, controlled process and cannot in any circumstance be modified by the end user.
I have implemented this a couple of times and not only does it work well, it tends to be fairly well accepted. People need spreadsheets to work on them, but generally they kind of hate sending those around via email. Having a reference source of data is welcomed.
So cool, I hope they pull it off. So many people use Excel. Although, I always thought the power of AI in Excel would come from the ability to use AI _as_ a formula. For example, =PROMPT("Classify user feedback as positive, neutral or negative", A1). This would enable normal people (non-programmers) to fire off thousands of prompts at once and automate workflows like programmers do (disclaimer: I am the author of Cellm that does exactly this). Combined with Excel's built-in functions for deterministic work, Claude could really kill the whole copy-pasting data in and out of chat windows for bulk-processing data.
I can't wait until someone does this, then autofills 50k rows down, then gets a $50k bill for all the tokens.
Reminds me of when our CIO insisted on moving to the cloud (back when AWS was just getting started) and then was super pissed when he got a $60k bill because no one knew to shutdown their VMs when leaving for the day.
If someone is processing 50k rows, that means they found real value and the UX is working. That's the whole point.
Also, 50k rows wouldn't cost $50k. More like $100 with Sonnet 4.5 pricing and typical numbers of input/output tokens. Imagine the time needed to go through 50k rows manually and math doesn't really work for a horror story.
On the first glance this seems to be a very bad idea. But re-readig this:
> Get answers about any cell in seconds: Navigate complex models instantly. Ask Claude about specific formulas, entire worksheets, or calculation flows across tabs. Every explanation includes cell-level citations so you can verify the logic.
this might just be an excellent tool for refactoring Excel sheets into something more robust and maintainable. And making a bunch of suits redundant.
Weird to see so much discussion when it’s still behind a waitlist. And it seems aimed at “enterprise” only
On a related note, has anyone found a good local LLM option for working with Excel files?
Here's my use case: I have a set of responses from a survey and want to perform sentiment analysis on them, classify them, etc. Ideally, I'd like to feed them one at a time to a local LLM with a prompt like: "Classify this survey response as positive, negative, or off-topic...etc".
If I dump the whole spreadsheet into ChatGPT, I found that because of the context window, it can get "lazy"; while with a local LLM, I could just literally prompt it one row at a time to accomplish my goal, even if it takes a little longer in terms of GPU and wall-clock time.
However, I can't find anything that works off the shelf like this. It seems like a prime use case for local models.
Don't know about excel, but for Google Sheets. You can ask chatgpt to write you a appsscript custom function e.g CALL_OPENAI. Then you can pass in variables into. =CALL_OPEN("Classify this survey response as positive, negative, or off-topic: "&A1)
Sheets also has an `AI` formula now that you can use to invoke Gemini models directly.
When I tried the Gemini/AI formula it didn’t work very well, gpt-5 mini or nano are cheap and generally do what you want if you are asking something straightforward about a piece of content you give them. You can also give a json schema to make the results more deterministic.
Cellm + Ollamma?
https://docs.getcellm.com/models/local-models
Maybe this is how we get code versioning for Excel.
Git LFS for workbook + the following prompt :
“Create a commit explains what has changed in the workbook since the last commit. Be brief, but explain the change in business terms as well as code change terms.”
HA!
I've worked at MULTIPLE million dollar firms whose entire business relies on 10 Excel workbooks that were created 30 years ago by a person who is either passed on or retired.
Give users who aren't intimately knowledgeable with their source material ai, and you're asking for trouble.
The undo function has a history limit.
The real issue is: at what point are we going to stop chasing efficiency and profit at the sake of humanity?
Claude and OpenAI are built on stretched truths, stolen creativity and what-if statements.
Version control and meaningful diffs for .xlsx will be in high demand in a few months
Honestly those things are well past due - if this tips the scales then I hope we can all benefit.
I’m decent at excel, but not amazing. I’ve tried again and again to use LLMs including Claude to solve specific, small, well defined problems in excel with a 0% success rate. My experience so far has been if I can’t do it LLMs can’t either.
If LLMs are a 6/10 right now at basic coding then they’re a 3/10 at excel from my experience.
IMO, a real solution here has to be hybrid, not full LLM, because these sheets can be massive and have very complicated structures. You want to be able to use the LLM to identify / map column headers, while using non-LLM tool calling to run Excel operations like SUMIFs or VLOOKUPs. One of the most important traits in these systems is consistency with slight variation in file layout, as so much Excel work involves consolidating / reconciling between reports made on a quarterly basis or produced by a variety of sources, with different reporting structures.
Disclosure: My company builds ingestion pipelines for large multi-tab Excel files, PDFs, and CSVs.
That's exactly what they're doing.
https://www.anthropic.com/news/advancing-claude-for-financia...
"This won't work because (something obvious that engineers at Anthropic clearly thought of already)"
Not really. Take for example:
item, date, price
abc, 01/01/2023, $30
cde, 02/01/2023, $40
... 100k rows ...
subtotal. $1000
def, 03/01,2023, $20
"Hey Claude, what's the total from this file? > grep for headers > "Ah, I see column 3 is the price value" > SUM(C2:C) -> $2020 > "Great! I found your total!"
If you can find me an example of tech that can solve this at scale on large, diverse Excel formats, then I'll concede, but I haven't found something actually trustworthy for important data sets
That's a basic tool call that current models already can do well. All the sql query generation LLMs can do this for example.
So more or less like what AI has been doing for the last couple of years when it comes to writing code?
Tough day to be an AI Excel add-in startup
Ask Rosie is actually shutting down right now: https://www.askrosie.ai/
I would love to learn more about their challenges as I have been working on an Excel AI add-in for quite some time and have followed Ask Rosie from almost their start.
That they now gone through the whole cycle worries me I‘m too slow as a solo building on the side in these fast paced times.
its a great time for your ai excel add-in to start getting acquired by a claude competitor though
Not OpenAI, though, because they already gave $14M to an AI Excel add-in startup (Endex)
That seems to be true for any startup that offers a wrapper to existing AIs rather than an AI on their own. The lucky ones might be bought but many if not most of them will perish trying to compete with companies that actually create AI models and companies large enough to integrate their own wrappers.
Actually just wrote about this: https://aimode.substack.com/p/openai-is-below-above-and-arou...
not sure if it binary like that but as startups we will probably collect the scraps leftover indeed instead
I wonder if this will be more/less useful than what we have with AI in software development.
There's a lot less to understand than a whole codebase.
I don't do spreadsheets very often, but I can emphasize with tracking down "Trace #REF!, #VALUE!, and circular reference errors to their source in seconds." I once hit something like that, and I found it a lot harder to trace a typical compiler error.
I guess Claude maybe useful for finding errors in large Excel Workbooks. May also help beginners to learn the more complex Excel functions (which are still pretty easy). But if you are proficient at building Excel models I don't see any benefit. Excel already has a superb very efficient UI for entering formulas, ranges, tables, data sources etc I'm sceptical that a different UI especially a text based one can improve on this.
I understand the sentiment about a skilled user not needing this, but I think having a little buddy that I can use to offload some menial tasks would be helpful for me to iterate through my models more efficiently; even if the AI is not perfect. As a highly skilled excel user, I admit the software has terrible ergonomics. It would be a productivity boon for me if an AI can help me stay focused on model design vs model implementation.
For some reason, I find that these tools are TERRIBLE at helping someone learn. I suspect because turning one on, results in turning the problem solving part of ones brain off.
Its obviously not the same experience for everyone. ( If you are one of those energized while working in a chat window, you might be in a minority - given what we see from the ongoing massacre of brains in education. )
Paraphrasing something I read here "people don't use ChatGPT to do learn more, they use it to study less".
Maybe some folk would be better off.
How is this different from the existing Claude skill, that uses a prompt and pandas to edit an Excel file?
https://github.com/anthropics/skills/blob/main/document-skil...
This isn't built for Excel users who use Github and Claude Skills, it's built for Excel users who would run away from Git commands.
The Claude skill I linked to is built into the Claude desktop client. You just attach an Excel file to your chat and ask away.
I linked to the skill prompt just to more clearly explain the approach that's currently available to all Claude users.
It requires zero familiarity with git or command line.
Pretty cool idea — AI inside spreadsheets makes sense since most of our work already lives there. But I’m a bit cautious too — spreadsheets are messy enough, and adding probabilistic AI could make mistakes harder to spot. Useful if done right, risky if not.
Gemini already has its hooks in Google Sheets, and to be honest, I've found it very helpful in constructing semi-complicated Excel formulas.
Being able to select a few rows and then use plain language to describe what I want done is a time saver, even though I could probably muddle through the formulas if I needed to.
I would recommend trying TabTabTab at https://tabtabtab.ai/
It is an entire agent loop. You can ask it to build a multi sheet analysis of your favorite stock and it will. We are seeing a lot of early adopters use it for financial modeling, research automation, and internal reporting tasks that used to take hours.
I have had the opposite experience. I've never had Gemini give me something useful in sheets, and I'm not asking for complicated things. Like "group this data by day" or "give me p50 and p90"
Last time I tried using Gemini in Google Sheets it hallucinated a bunch of fake data, then gave me a summary that included all that fake data. I'd given it a bunch of transaction data, and asked it to group the records into different categories for budgeting. When asking it to give the largest values in each category, all the values that came back were fake. I'm not sure I'd really trust it to touch a spreadsheet after that.
you should:
-stop using the free plan -don't use gemini flash for these tasks -learn how to do things over time and know that all ai models have improved significantly every few months
Or not use it.
I forgot to add, you can try TabTabTab, without installing anything as well.
To see something much more powerful on Google Sheets than Gemini for free, you can add "try@tabtabtab.ai" to your sheet, and make a comment tagging "try@tabtabtab.ai" and see it in action.
If that is too much just go to ttt.new!
Gemini integratoins to Google workspace feels like it's using Gemini 1.5 flash, it's so comically bad at understanding and generating
I'd like to see it compete in the Financial Modeling World Cup, say in Las Vegas this December. https://excel-esports.com
Just spent an hour trying to figure out how to create a waterfall chart. ChatGPT's python interpreter failed.
If this works right, this could be a game changer.
https://chatgpt.com/share/69005eec-6ee0-8009-a8d3-ebb1c30e72...
took me four prompts to do generate a waterfall chart using d3 js because it didn't want to run it. obviously with real numbers and not generated data, you'd need to check the results thoroughly.
Interesting their X post mentions "pre-built Agent Skills" but it's not on the webpage. I wonder if they will give you the ability to edit/add/delete Skills, that would be phenomenal.
Edit: found it on their other blog post https://www.anthropic.com/news/advancing-claude-for-financia...
You can add and customize skills in claude.ai and other surfaces
I really hope that all these kinds of integrations:
* Claude for Chrome * Gemini for Chrome * ChatGPT Atlas * ...
will be built on top of the ACP protocol, so that these “AI extensions” to everything can become standardized
From their FAQ “Claude doesn’t have advanced Excel capabilities including pivot tables, conditional formatting, data validation, data tables, macros, and VBA. We’re actively working on these features.”
Last week OpenAI hired ex-investment bankers to train a model to build financial models, now anthropic coming for excel.
Sounds like there's some sort of an AI race after finance people and businesses?
George Hotz said there's 5 tiers of AI systems, Tier 1 - Data centers, Tier 2 - fabs, Tier 3 - chip makers, Tier 4 - frontier labs, Tier 5 - Model wrappers. He said Tier 4 is going to eat all the value of Tier 5, and that Tier 5 is worthless. It's looking like that's going to be the case
That is a common refrain by people who have no domain expertise in anything outside of tech.
Spend a few years in an insurance company, a manufacturing plant, or a hospital, and then the assertion that the frontier labs will figure it out appears patently absurd. (After all, it takes humans years to understand just a part of these institutions, and they have good-functioning memory.)
This belief that tier 5 is useless is itself a tell of a vulnerability: the LLMs are advancing fastest in domain-expertise-free generalized technical knowledge; if you have no domain expertise outside of tech, you are most vulnerable to their march of capability, and it is those with domain expertise who will rely increasingly less on those who have nothing to offer but generalized technical knowledge.
yeah but if Anthropic/OpenAI dedicate resources to gaining domain expertise then any tier 5 is dead in the water. For example, they recently hired a bunch of finance professionals to make specialized models for financial modeling. Any startup in that space will be wiped out
I dont think the claim is exactly that tier 5 is useless more that tier 5 synergizes so well with tier 4 that all the popular tier 5 products will eventually be made by the tier 4 companies.
George Hotz says a lot of things. I think he's directionally correct but you could apply this argument to tech as a whole. Even outside of AI, there are plenty of niches where domain-specific solutions matter quite a bit but are too small for the big players to focus on.
Tier 5 requires domain expertise until we reach AGI or something very different from the latest LLMs.
I don’t think the frontier labs have the bandwidth or domain knowledge (or dare I say skills) to do tier 5 tasks well. Even their chat UIs leave a lot to be desired and that should be their core competency.
Andrew Ng argumented in 2023 (https://www.youtube.com/watch?v=5p248yoa3oE ) that the underlying tiers depend on the app tier‘s success.
That OpenAI is now apparantly striving to become the next big app layer company could hint at George Hotz being right but only if the bets work out. I‘m glad that there is competition on the frontier labs tier.
People were saying the same thing about AWS vs SaaS ("AWS wrappers") a decade ago and none of that came to pass. Same will be true here.
Claude is a model wrapper, no?
Anthropic is a frontier lab, and Claude is a frontier model
Anthropic models are Sonnet / Haiku / Opus
https://docs.claude.com/en/docs/about-claude/models/overview
Interesting. I found a reference to this in a tweet [1], and it looks to be a podcast. While I'm not extremely knowledgable. I'd put it like this: Tier 1 - fabs, Tier 2 - chip makers, Tier 3 - data centers, Tier 4 - frontier labs, Tier 5 - Model wrappers
However I would think more of elite data centers rather than commodity data centers. That's because I see Tier 4 being deeply involved in their data centers and thinking of buying the chips to feed their data centers. I wouldn't be so inclined to throw in my opinion immediately if I found an article showing this ordering of the tiers, but being a tweet of a podcast it might have just been a rough draft.
1: https://x.com/tbpn/status/1935072881425400016
Hope it’s better than what MS is currently shipping as AI. Everything I try to do something, the response is “sorry, I can’t do this”.
Copilot is getting better - I'm getting fewer of those than I used to - but it's still significantly more stupid than other agents, even when in theory it's using the same model.
As I was reading through the post, and the comments here, and pondering my own many hours with these tools, I was suddenly reminded of one of my favorite studio C sketches: An Unfortunate Fortune
https://www.youtube.com/watch?v=SF-psoWdSpo
Curious, if others see the connection. :D
I have just launched a product (easyanalytica.com) to create dashboards from spreadsheets, and Excel is on my to-do list of formats to be supported. However, I'm having second thoughts. Although, from the description, it seems like it would be more helpful on the modeling side rather than the presentation side. I guess I'll have to wait until it's publicly available
Why second thoughts?
everyone will use claude if they support it why would they use my product. so i will have to find some other angle to differentiate.
There's already a language for this, or multiple, that isn't English. Not having to use this language is NOT going to make anything better.
It will, however, make people resort more quickly to "I guess it's just not possible if Claude can't figure it out".
“Copilot in Excel is a global financial crisis waiting to happen.”
— Zack Korman, <https://x.com/ZackKorman/status/1974828240679166396>
Anyone understand how this could work? My mental model for llm is predictive text but here how can it understand cell A1 which has a string is the “header” for all values under it? How does it learn to understand table data like that?
LLMs already understand table data. "Predictive text" is somewhat true but so reductive that it leads to that kind of misconception.
HN is going to mangle this but here's a quick table:
| Type of Horse | Average Height | Typical Color | |----------------|----------------|-----------------| | Arabian | 15 hh | Bay, Gray | | Thoroughbred | 16 hh | Chestnut, Bay | | Clydesdale | 17.5 hh | Bay with White | | Shetland Pony | 10.5 hh | Black, Chestnut |
And after a prompt "pivot the table so rows are colors":
| Typical Color | Type of Horse | Average Height | |----------------|----------------------------------------|-----------------------| | Bay | Arabian, Thoroughbred, Clydesdale | 15 hh, 16 hh, 17.5 hh | | Gray | Arabian | 15 hh | | Chestnut | Thoroughbred, Shetland Pony | 16 hh, 10.5 hh | | Bay with White | Clydesdale | 17.5 hh | | Black | Shetland Pony | 10.5 hh |
> Anyone understand how this could work? My mental model for llm is predictive text but here how can it understand cell A1 which has a string is the “header” for all values under it? How does it learn to understand table data like that?
I imagine it uses the new Agent Skills features
https://www.anthropic.com/news/skills
Dumb question, but is this Claude for Excel the.. app? The webapp? Does it work on Google sheets? etc
There are quite a few spreadsheet apps out there, just curious what their implementation is or how it's implemented to work with multiple apps.
I always find Excel (and the Office ecosystem) confusing heh.
Modern Excel add-ins work in desktop Windows, macOS, and web. They're just a bit of XML that Excel looks at to call a whatever web endpoint is defined in the XML.
It’s interesting to me that this page talks a lot about “debugging models” etc. I would’ve expected (from the title) this to be going after the average excel user, similar to how chatgpt went after every day people.
I would’ve expected “make a vlookup or pivot table that tells me x” or “make this data look good for a slide deck” to be easier problems to solve.
I think actually Anthropic themselves are having trouble with imagining how this could be used. Coders think like coders - they are imagining the primary use case being managing large Excel sheets that are like big programs. In reality most Excel worksheets are more like tiny, one-off programs. More like scripts than applications. AI is very very good at scripts.
The issue is that the average Excel user doesn’t quite have the skills to validate and double-check the Excel formulas that Claude would produce, and to correct them if needed. It would be similar to a non-programmer vibe-coding an app. And that’s really not what you want to happen for professionally used Excel sheets.
IMO that is exactly what people want. At my work everyone uses LLMs constantly and the trade off of not perfect information is known. People double check it, etc, but the information search is so much faster even if it finds the right confluence but misquotes it, it still sends me the link.
For easy spreadsheet stuff (which 80% of average white collars workers are doing when using excel) I’d imagine the same approach. Try to do what I want, and even if you’re half wrong the good 50% is still worth it and a better starting point.
Vibe coding an app is like vibe coding a “model in excel”. Sure you could try, but most people just need to vibe code a pivot table
I think this is aiming to be Claude Code for people who use Excel as a programming environment.
On the one hand, most financial companies have a lot of processes in Excel that could be made better with something like Claude.
Banking secrecy laws + customer identifying data + AI tool = No bueno.
That’s it. 1T EV added.
As an inveterate Excel lover, I can just sense the blinding pain wafting off the legions of accountants, associates, seniors, and tech people who keep the machine spirits placated.
lies, damn lies, statistics, and then Excel deciding cell data types.
This could be invaluable for reverse engineering complex workbooks with multiple data sources and hundreds or thousands of formulas.
If it has a concept of data sources and can digest them, sure. Anecdotally, most issues with Excel at my job are caused by data sources being renamed, moved or reformatted, by broken logins, or by insufficient access rights.
I'm excited to see what national disasters will be caused by auto-generated Excel sheets that nobody on the planet understands. A few selections from past HN threads to prime your imagination:
Thousands of unreported COVID cases: https://news.ycombinator.com/item?id=24689247
Thousands of errors in genetics research papers: https://news.ycombinator.com/item?id=41540950
Wrong winner announced in national election: https://news.ycombinator.com/item?id=36197280
Countries across the world implement counter-productive economic austerity programs: https://en.wikipedia.org/wiki/Growth_in_a_Time_of_Debt#Metho...
Especially combined with the dynamic array formulas that have recently been added (LET, LAMBDA etc). You can have much more going on within each cell now. Think whole temporary data structures. The "evaluate formula" dialog doesn't quite cut it anymore for debugging.
from my experience in the corporate world, i'd trust an excel generated / checked by an LLM more than i would one that has been organically grown over years in a big corporation where nobody ever checks or even can check anything because its one big growing pile of technical debt people just accept as working
Fodasse a Rows é pelo menos 3x melhor
100%
Ok, they weren't confident enough to let the model actually edit the spreadsheet. Phew..
Only a matter of time before someone does it though.
How well does change tracking work in Excel... how hard would it be to review LLM changes?
AFAIK there is no 'git for Excel to diff and undo', especially not built-in (aka 'for free' both cost-wise and add-ons/macros not allowed security-wise).
My limited experience has been that it is difficult to keep LLMs from changing random things besides what they're asked to change, which could cause big problems if unattackable in Excel.
I thought there was track changes on all office products. Most Office documents are zip files of XML files and assets, so I'd imagine it would be possible to rollback changes.
When I think how easy I can misclick to stuff up a spreadsheet I can't begin to imagine all the subtle ways LLMs will screw them up.
Unlike code where it's all on display, with all these formulas are hidden in each cell, you won't see the problem unless click on the cell so you'll have a hard time finding the cause.
I wish Gemini could edit more in Google sheets and docs.
Little stuff like splitting text more intelligently or following the formatting seen elsewhere would be very satisfying.
I use excel but not for financial modelling, I’ll use it
R.I.P. global economy
If Claude is going to work on the underpinning technology of every business in the Capitalist world, We should let Claude loose on the COBOL code out there too, I can't imagine anything going wrong.
Anthropic knows almost nothing about their own products and you guys know even less.
I just want Claude inside of Metabase.
https://www.metabase.com/features/metabot-ai
Yeah now tell the Auditors that the financial spreadsheet we have here has AI touching it left and right. "I did not cook the books I promise it is the AI that made our financials seem better than they actually are trust me bro!", said Joe from Accounting.
Yet more evidence of the bubble burst being imminent. If any of these companies really had some almost-AGI system internally, they wouldn’t be spending any effort making f’ing Excel plugins. Or at the very least, they’d be writing their own Excel because AI is so amazing at coding, right?
You make a great point. Where is all of the complex applications? They haven't been able to create than own office suite or word processor or really anything aside from a halloween matching game in js. You would think we would have some complex application they can point to but nothing.
Excel is living business knowlege stuck in private SharePoint Sites, tappimg into it might kick off a nice data flywheel not to speak of the nice TAM.
The fine tuning will continue until we reach AGI.
The fine tuning will continue until we reach the torment nexus, at best
You wouldn't believe the amount of shit that runs on Excel.
Yes. I once interviewed a developer who’s previous job was maintaining the .NET application that used an Excel sheet as the brain for decisions about where to drill for oil on the sea floor. No one understood what was in the Excel sheet. It was built by a geologist who was long gone. The engineering team understood the inputs and outputs. That’s all they needed to know.
Years ago when I worked for an engineering consulting company we had to work with a similarly complex, opaque Excel spreadsheet from General Electric modeling the operation of a nuclear power plant in exacting detail.
Same deal there -- the original author was a genius and was the only person who knew how it was set up or how it worked.
I think you’re misunderstanding me. This might be something somewhat useful, I don’t know, and I’m not judging it based on that.
What I’m saying is that if you really believed we were 2, maybe 3 years tops from AGI or the singularity or whatever you would spend 0 effort serving what already seems to be a domain that is already served by 3rd parties that are already using your models! An excel wrapper for an LLM isn’t exactly cutting edge AI research.
They’re desperate to find something that someone will pay a meaningful amount of money for that even remotely justifies their valuation and continued investment.
I spotted a custom dialog in an Excel spreadsheet in a medical context the other day, I was horrified.
Sic
This. I work in Pharma. Excel and faxes.
The current valuations do not require AGI. They require products like this that will replace scores of people doing computer based grunt work. MSFT is worth $4 trillion off the back of enterprise productivity software, the AI labs just need some of that money.
A program that can do excel for you is almost AGI
If this works very well and reliable, it might not kill programming as such, but it might put a lot of small businesses who do custom software for other small businesses out of work.
The HN bubble might not realize the implications.
Been working with Claude Code lately and been pretty impressed. If this works as well could be a nice add on. Its probably a smart market to enter as Excel is essentially everywhere.
Just like Claude Code allows 1 dev to potentially do the work of 2 or 3, I could see this allowing 1 accountant or operations person to do the work of 2 or 3. Financial savings but human cost
If AI turns out to be the powerhouse it is claimed to be, AI's impact will be corporations replacing corporate dependencies upon 'Excel projects' created by self-taught assistants to department managers.
Cool but now companies POs will be like "you must add the Excel export for all the user data!" and when asked why, will basically be "so I can do this roundabout query of data for some number in a spreadsheet using AI (instead of just putting the number or chart directly in the product with a simple db call)"
This could be huge! Very exciting!
Checkmate, Altman
Can we get it in Sheets?
for sheets, Gemini exists already natively.
Alternatively, Perplexity Comet browser (or OpenAI Atlas) would presumably provide sidebar functionality to act within your spreadsheets.
[flagged]
"Eschew flamebait. Avoid generic tangents."
https://news.ycombinator.com/newsguidelines.html
Okay. But then you could say the same for a human, isn't your brain just a cloud of matter and electricity that just reacts to senses deterministically?
> isn't your brain just a cloud of matter and electricity that just reacts to senses deterministically?
LLMs are not deterministic.
I'd argue over the short term humans are more deterministic. I ask a human the same question multiple times and I get the same answer. I ask an LLM and each answer could be very different depending on its "temperature".
If you ask human the same question repeatedly, you'll get different answers. I think that at third you'll get "I already answered that" etc.
We hardly react to things deterministically.
But I agree with the sentiment. It seems it is more important than ever to agree on what it means to understand something.
I'm having a bad day today. I'm 100% certain that today I'll react completely different to any tiny issue compared to how I did yesterday.
OK then. Groks?
I mean - try clicking the CoPilot button and see what it can actually do. Last I checked, it told me it couldn't change any of the actual data itself, but it could give you suggestions. Low bar for excellence here.