If AI winning means that data center companies win out, then the wins for Azure will more than make up for the death of Office.

I am surprised that Microsoft's own copilot product is so far behind though.

There is a significant difference in experience between Copilot Basic for a M365 user whose IT admins have blocked integration capabilities with Sharepoint content vs Copilot Premium for a M365 user whose IT admins have allowed integration capabilities with Sharepoint content.

Chatgpt for Excel is still an office add-in running in the same sandbox though. strongpigeon described the exact bottleneck upthread, process boundary crossings, context.sync() roundtrips that take seconds on web. That's a platform limitation, not a model limitation. Swapping AI behind the add-in doesn't fix the fundamental constraint that third-party add-ins can't deeply integrate with Excel's runtime the way a native feature can. If copilot is bad despite having more access to excel internals(I don't like how Copilot is designed or implemented tho), an add-in with less access is likely not be better.

Do you mean that you worked on the Excel Add-Ins platform in Excel (and not on a specific Add-In)?

If you were working on the platform itself, then I would be interested in hearing your more detailed thoughts on the matters you mentioned (especially since I am developing an open source Excel Add-In Webcellar (https://github.com/Acmeon/Webcellar)).

What do you mean with a "OM" call? And why are they especially costly on Excel web (currently my add-in is only developed for desktop Excel, but I might consider adding support for Excel web in the future)?

In any case, `context.sync()` is much better than `context.executeAsync()`.

A recent funny story on this topic: https://idiallo.com/blog/what-is-copilot-exactly

HN discussion: https://news.ycombinator.com/item?id=47603231

Why are they irrelevant? You do not say anything.

I know the firm - it's RenTech.

This might be the first time I've seen a HN comment in a GPT thread that actually reflects what the average business user sees in GPT products.

They don't do the job, reliably or well. No amount of wishful thinking or extra tokens will change that.

Microsoft is better off not allowing copilot basic because of the reputational harm it will do. Not that they are thinking through copilot rationally

I worked on the Excel Add-Ins platform at Microsoft, yes. By OM call I mean "Object Model" call, basically interacting with the Excel document.

The reason those calls are expensive on Excel Web is that you're running your add-in in the browser, so every `.sync()` call has to go all the way to the server and back in order to see any changes. If you're doing those calls in a loop, you're looking at 500ms to 2-3s latency for every call (that was back then, it might be better now). On the desktop app it's not as bad since the add-in and the Excel process are on the same machine so what you're paying is mostly serialization costs.

Happy to answer more questions, though I left MSFT in 2017 so some things might have changed since.

Adding a tangential anecdote.

I asked GPT-5.4 High to draw up an architecture diagram in SVG and left it running. It took over an hour to generate something and had some spacing wrong, things overlapping, etc. I thought it was stuck, but it actually came back with the output.

Then I asked it to make it with HTML and CSS instead, and it made a better output in five seconds (no arrows/lines though).

SVG looks similar to the XML format of spreadsheets. I wonder if LLMs struggle with that?

Based on my discussion with DSEs, enterprises have not been impressed in the results of "Copilot", i.e. OAI models. MSFT has been replacing (or changing the default) to Claude across a variety of Copilot endpoints.

I care not to. I hope Anthropic and OAI keep burning money on stuff that's DOA.

I know there are employees of those firms here that would love to know. But nah lmao.

No surprise really.

Remember when Steve said 'The computers for the rest of us'?

I suppose it isn't a surprise. Are researchers/generally geeky people meant to be able to relate to the average person's day-to-day beyond their sphere? Lmao.

You can't produce stuff for people you don't understand. Understand being a very key term.

Engineer on ChatGPT for Excel here. Useful feedback. We have improved the latency inside the add-in a lot and a lot more to come. We also have the Fast, Standard and Heavy thinking modes, where you can adjust the thinking time depending on the task complexity. Curious to hear your feedback once you try this out!

nah the firm in question has much higher AUM.

There’s a magic button you have to press to make it integrate fully. Everybody is confused about why this isn’t the default behavior.

> This looks bad for Microsoft.

Maybe(?) from a product catalogue perspective... But from a strategic perspective less so because they own ~27% of OpenAI.[0]

[0] - https://openai.com/index/next-chapter-of-microsoft-openai-pa...

Microsoft has rights to all this IP. So, it might look bad for their product folks, but for the corporation this is great, to the extent it works.

Gemini is very good with SVG, but I don't really see the similarity to spreadsheets.

> though I left MSFT in 2017 so some things might have changed since.

Honestly, I struggle to think about what has actually changed between Office 2013 and Office 2024 (and their Office 365 equivalents); I know the LAMBDA function was a big deal, but they made the UI objectively worse by wasting screen-space with ever-increasingly phatter non-touch UI elements; and the Python announcement was huge... before deflating like a popped party balloon when we learned how horribly compromised it was.

...but other than that, Excel remains exactly as frustrating to use for even simple tasks - like parsing a date string - today just as it was 15 years ago[1].

[1]: https://stackoverflow.com/questions/4896116/parsing-an-iso86...

There's no real moat in feature set anymore. Within a given timeframe, any company should be able to copy some features from other companies. Thus the whole "distribution, marketing and sales are the only things that matter nowadays" joke.

Obviously doesn't apply to everything, and there are some features that are very hard to replicate. But still.

> They added a Copilot button to all their products but it doesn't do much more than open a chat side panel.

I was hyped when I heard about Copilot. "I can tell it to make pivot tables now!" When I tried to use it I was shocked how underbaked it was. Below even my worst expectations. This really was someone shoving ChatGPT into Excel with almost zero additional effort. Copilot can't DO anything useful.

I've had the same experience. Copilot for Excel can't even parse basic cell references. Meanwhile Claude handles document formatting in one pass. The catch is it works externally, not inside the app, but at least it works.

The MCP ecosystem is what makes this interesting. Claude isn't just a chat panel bolted onto existing software, it's building integrations that actually manipulate the files. Microsoft had the distribution advantage but they're losing on capability.

Claude for excel is already amazing. Fully capable of doing junior work. Formatting is great. Can refactor large multi-tab spreadsheets. It just burns tokens. If OpenAI is going to subsidize this on the monthly enterprise plans for a while then it's a game changer.

Claude for Excel (I work in finance) was one of the absolutely critical reasons we added Anthropic enterprise licenses. But they've turned out to be quite expensive ($100/day for heavy users). We'll see what OpenAI's quotas are.

Your comment is recognized as low effort, but Copilot has been OAI models behind the scenes. For enterprise customers, quickly being replaced by Sonnet as a default.

Would love for you to try both copilot and ChatGPT for Excel. Agreed on the limitations - but in our experience, ChatGPT for Excel does really well on complex sheets.

Citadel

MSFT also has a stake in Anthropic (although much less than 27%) and they host Anthropic models in Foundry now. The end game for MSFT has always been being the compute provider, so MSFT is just as happy to use any model as long as it's running in Foundry.

There is an irony here that this would be more performant with a 2002 coding model. A native plugin, COM, OLE, whatever. C++, crash prone, but fast.

As a former investment banker, I mostly agree. This is probably 10% of the work

Hi, engineer on this add-in. Fair concern but we never train on any of our business or enterprise user data, or if you have opted-out of training on your ChatGPT account.

That's the nature of these add-ins. Modern Add-ins are all little XML frames with some JS or whatever. All processing occurs server-side, hosted by the add-in publisher.

This is counter to the old (security nightmare) COM model where processing could be local.

Because the challenge is in the space between apps, not in the apps themselves.

stride.microsoft.com -> this is a virtual machine instance with developer tools that allow for same sort of work Claude cowork does. Copilot in excel has to access the excel document through excel provided APIs and can’t completely redo the document like cowork does everytime running developer scripts to generate it because the document instance is open. The model of work is entirely different.

I work with large files a lot, running claude code on it is not token intense at all. Probably because it does a lot with scripts. But its a bit more raw, but i think in the end more powerful. Have to pick a good excel library and language. I do node, maybe python can work as well

Work in a firm similar to yours and we have been going to though the motions of figuring ways for the bullpen to make use of these tools and would love to hear your thoughts if you would be willing to share!

Do you also have Google Spreadsheets on your radar?

I've had a spreadsheet integrated with ChatGPT API for a few years already. It really was not until GPT-5.4 that the models were able to actually be useful.

What is the data model that you use for the spreadsheet itself? I found I could create a chat completion persona that believed it is one of the developers of a popular open source spreadsheet, and I put this "agent" directly inside the open source spreadsheet. I did this before tool calling was available at all, so I made my own system for that, and the "tools" are the API of that open source spreadsheet. My agent(s) that operate like this can do anything the spreadsheet can do, including operate the spreadsheet engine from the inside.

What API/approach does it use to edit sheets?

I made a CLI (+skill) so agents could edit files with verbs like `insert A1:A3 '[1,2,3]'`, but did some evals and found it underperformed Anthropic's approach (just write Python).

I probably could find some really useful things for it to help me … but all software nowadays only works outside my earth region :(

This time even for pro.

I tried it the other day to work on some exported CSVs when doing my taxes. I was finally able to get it to do what I wanted, but it was definitely an exercise, feeling like I was talking to Chat GPT from a couple of years ago. (as in a really smart but easily distracted and confused child)

Does Excel for Web still spin up an actual copy of Excel.exe on a machine somewhere? I heard that is how the initial version worked.

Yeah, that makes sense. For some reason, I was under the impression that all calculations run locally in the browser, which would have been comparable to how Excel desktop works (i.e., local calculations). Is there a reason for why the Excel calculations run on the server (e.g., excessive workload of a browser implementation, proprietary code, difficult to implement in JavaScript, cross browser compatibility issues, etc.)? Furthermore, if the reason for this architecture is (or was) limitations in JavaScript or browsers, do you find it plausible that the Excel calculations will some day be implemented in Webassembly?

Regardless, I have always preferred Excel desktop over Excel web (and other web based spreadsheet alternatives). This information makes me somewhat less interested in Excel web. Nonetheless, I find Excel Add-Ins useful, primarily because they bring the capabilities of JavaScript to Excel.

it was a good name when chosen. too bad they have burned bob, clippy, cortana, sydney, and copilot already.

stride.microsoft.com -> microsoft has this stuff you just don’t know it unless you are an M365 power user

How’s that been in practice ? From what I’ve been following - Claude in finance results in models with errors that an analyst won’t make.

You get models that are formatted and structured and which balance - but there are errors introduced which an analyst / human wouldn’t make.

Stuff like hard coded values, or incorrect cell logic which guarantees the model balances.

Cheaper to get M365 Copilot licenses for the Claude models in Excel.

How well does this work compared with using GPT-5.4 in Nicopreme’s Pi for Excel?

Thank you for high effort response!

I would never use Copilot for anything useful, but I do use OpenAI products.

It doesn't matter when you use something else wholesale under the covers, if you botch the token spent...

I'm having great luck having Claude Code generate, read, and update spreadsheets by writing Python code that uses gspread.

I love Sheets, but I dont care for using Gemini to interact with Sheets. It seems like a recipe for disaster. Do I really want it to muck around with thousands of rows and no intuitive way to diff its changes? Nope, sticking with basic Sheets

Agreed; I was also shocked by how limited it was. Same with the Slides integration.

Yeah, the Sheets integration is weird. It's usually ok when it wants to place something down the first time. But then it seems incapable of making any changes to it. Or even acknowledging the data in the sheet. What's up with that?

I mean, you're wrong. As a Xoogler, everything was in Sheets. Our roadmap was in Sheets. It's more they don't care.

Parametric scenes is the whole of Houdini and any node based compositor etc. so there is some applications no?

Try https://rowzero.com ? We have written a much faster spreadsheet engine and regularly work with 10M+ row datasets

Maybe but not drastically so. My guess is that most of the slowness comes from the tool calls round tripping+processing on Anthropic/OpenAI’s servers rather than the app latency.

That’s without talking about the poor UI and security story of COM add-ins and the inability to run on Excel for iOS.

Even 10% might be generous.

I bet the bozos at OAI and Anthropic think a person who deals with stuff like cost of capital is going to go ask an LLM for it.. when in reality the individual needs to know how/why they chose what they did.

Comical stuff.

Yeah, I was expecting that you do not train on business or enterprise user data. However, I am not just worried about "training", but also about "sharing". Furthermore, I am worried about cases where an individual has chosen to integrate an add-in and then inadvertently leaks sensitive data.

However, it may be important to note that these security considerations are relevant for most Office Add-Ins (and not just the ChatGPT add-in).

My local models interact with Sheets exclusively over the API with Python scripts I been curating for years

Given how well the API works, that we are discussing Googlers, my guess is that's how they dog food their services. Programmers don't get hired by Google for mouse skills.

The GUI is for spot checking results, final presentation.

If you're sitting there point-n-clicking everything into place perhaps consider you are doing it wrong.

You should try MS Copilot which uses open source Python libraries to interact with Office file formats.

The libraries themselves are OK, but MS uses them stupidly. If you want to fill out some form in DOCX or XSLX format you will get broken formatting. And this is from Office company.

Forgive my ignorance. How do you folks manage context retention? Say if someone had a sensitive excel document they wanted inference done over, how is that data actually sent to the model and then stored or deleted?

It seems one of the biggest barriers to people's adoption is concern over data leaving their ecosystem and then not being protected or being retained in some way.

Is this is an SLA that a small or medium sized company could get?

To clarify: add-ins are essentially web pages. They can do some processing client side if they want, but yeah in the case of a ChatGPT add-in it's not like they're running the model in a web frame.

No, as the other comment mentioned. But I’ve heard of more than a few customers running their own “server excel workflow” where they have an instances of excel.exe running a VBA macro that talks to a web server (and does some processing).

Don't forget Tay!

I don’t think Excel web will ever be running the calc engine browser side, no. The only way I could see this happen would be via compiling the core to wasm, which I don’t think is worth the engineering effort.

Excel has this legacy (but extremely powerful) core with very few people left that knows all of it. It has legacy bugs preserved for compatibility reasons as whole businesses are ran on spreadsheet that break if the bug is fixed (I’m not exaggerating). The view code for xldesktop is not layered particularly well either leading to a lot of dependencies on Win32 in xlshared (at least back then).

Is it doable? I’m sure. But the benefits are probably not worth the cost.

Never did this. WAC was the original version (integrated with SharePoint Server). Everything was server-side.

I would consider myself an M365 power user and I was not aware of this. It is not well promoted--and after all the Copilot crap, I would be annoyed even if it was.

Regardless, I just tried to log in with my work MS account, and I can't do so.

The backend of Copilot is still called Sydney AFAIK

Just my experience, it’s not a solution but rather a productivity tool. I mostly use it for tasks I can do myself but it would probably take 20-30min to dial in - now Claude can do it in 2-3min. (E.g. in a data table - add a new column that checks column a if the data is a, do x, if the data is b, do y, if the data is c, do z - then combine that with the word after the hyphen in column b —- or another example —- create a new sheet that is the same format as sheet one but show calculates the difference between column a and b bot for sheets 1-12 in a summary)

I don’t get good results when I just have Claude build things on its own - but for these types of specific productivity tasks I can save a couple of hours here and there.

From my experience, LLM performance in these areas is being massively oversold. I have repeatedly tried using Claude to modify a range of models typical of investment banking / private equity / sellside research contexts, and the results have been generally disastrous. On multiple occasions, the xlsx would no longer open.

I tried looking this up but wasn't able to find info on this on Microsoft's website. Do you have a link for this?

What are the costs on that?

Does this remove (or at least increase) the upload limit?

It also works fine with Ruby and the "caxlsx" gem. Codex works fine with it as also.

Surprisingly even small models can do this quite well. I have Sonnet on a claw-like generate something based on my emails, airbnb receipts, and so on, and it was perfect and it could edit fields and whatnot, but the Gemini tool can't do anything.

Obviously. Because they didn't train the model on proprietary MS code. Which is bad but also good in some way, as it might force MS to support better their formats in the open source world.

I recently experimented with trying to generate a passable slide deck from a script and outline I had written beforehand. The ChatGPT integration built into Powerpoint was abysmally bad. Like to the point it was embarrassing as a product.

Claude one-shot something with a Python script that was pretty okay.

How? This argument would make sense if sheets wasn’t targeted at a general audience.

Can it work with comments in sheets as well? When I looked into it, that seemed like a limitation.

Exactly. My company's add-in is a React app, which can be fun to debug with the limited developer tools in Word/Excel's browser.

have you got a link to this?

Token expenditure isn't a concern for Copilot users. They don't see that form of cost model, just a flat monthly (or yearly) price for a user license.

That sounds like an extremely narrow use case, compared to what the vast majority of Sheets users will be comfortable with

If you're concerned, you don't send it outside of the M365 boundary and presumably your admin has Purview Sensitivity Labels in place covering the document to prevent such activity.

Surely picking that apart is a good idea. Even if it costs a fortune, there have to be benefits.

Thanks for the interesting info! Yeah, maybe Excel web will someday support local calculations via wasm, but for now I think I will stick with Excel desktop with add-ins.

> WAC

Now that’s an acronym that I had forgotten about.

Enterprise is Copilot Cowork one of the frontier agents. Has to be enabled by your organization I believe

> Excel has this legacy (but extremely powerful) core with very few people left that knows all of it.

Would love to hear more about this. Especially history and comparison to Lotus etc.

While working at MS I remember someone in the office team saying that the original version of Excel online spun to the actual Excel backend and had it output HTML instead of the usual win32 UI. Was I misinformed by chance?

$200-something per user per year. Will vary based on license type and seat count.

No limits.

Sorry, I got the author wrong.

It's here: https://github.com/tmustier/pi-for-excel

At the same time, it makes some sense... the programmers for a system aren't always the best users of a system. So if you're expecting them to dogfood their own system (Google Sheets), you might find that they test/interact with the system primarily through the API and not the GUI.

I have no idea if they do or not, but it's a plausible explanation...

Use case feels like the wrong term.

Do you mean restricted workflow? Googles APIs are pretty much 1:1 to the GUI

And using Python makes it trivial to copy-paste out of files and other APIs with one run of Python

Versus all the fiddling in browser tabs with a mouse, it actually affords an incredibly wide set of options to quickly collate and format data

So the first thing that's important to understand is that Excel is the product of another era. One where resources like memory were very constrained and compilers and optimizers weren't as good as they are today (so much so that the Excel team at one point wrote their own because MSVC sucked, but I digress)...

And so, a lot of the core code is used to that. Cell formatting data for example is super tightly packed in deeply nested unions to ensure that as little memory is used to store that info. If something only needs 3 bits, it'll only use 3 bits. The calc engine compiles all of your formulas to its own (IIRC variable-instruction-width) bytecode to make sure that huge spreadsheets can still fit in memory and calc fast.

And a lot of it still carries the same coding and naming practices that it started with in the 80s: Hungarian notation, terse variable and file names, etc. Now, IMO, Hungarian notation by itself is pretty harmless (and even maybe useful in absence of an IDE), but it seemed to encourage programmers to eschew any form of useful information from variable names, requiring you to have more context to understand "why" something is happening. Like, cool, I have a pszxoper now (pointer to zero terminated string of XOper), but why?

So the code is tight, has a lot of optimization baked in and assumes you know a lot about what's happening already.

But more importantly, a lot of "why" information also just lives in people's head. Yes, some teams would have documentation in an ungodly web of OneNote notebooks, or spread across SharePoint pages, which had the least useful search functionality I've ever witnessed, but finding anything you wanted was hard. But that didn't use to matter, since the core team had been there for a long time, so you could ask them question.

That being said, I joined MSFT in 2012 and started working on Excel closer to 2014. At that point, heavyweight like DuaneC (who wrote like 10% of Excel and I don't think I'm exaggerating) had already retired and while others people were very knowledgeable in some areas, nobody seemed to have a good cross view of the whole thing.

You have to understand that I was in the Office Extensibility team. We were building APIs for the whole thing. I had to touch the calc system, the cells and range, the formatting, tables, charts and images (the whole shared oart system was interesting), etc. Answering "How do I do X" was always a quest because you would usually:

- Find 3 different ways of achieving it

- One of them was definitely the wrong way and could make the app crash in some situations (or leak memory)

- All the people on the "blame" had left

- One of them was via the VBA layer which did some weird stuff (good ol' pbobj)

- Be grateful that this wasn't Word because their codebase was much worse

And so, a lot of the API implementation was trial and error and hunting down someone who understood the data structures. The fact that full sync and rebuild took about 6 hours (you ran a command called `ohome` and then you went home) meant that experimenting was sometimes slow (at least incremental builds were OK fast). The only lifeline back then was this tool called ReSearch2 that allowed you to search the codebase efficiently.

But the thing is, once you got thing to work, they worked really well. The core was solid and performant. Just slightly inscrutable at time and not the kind of code you're use to reading outside of Excel.

Excel Online was a component of WAC. It was an ASP.NET (and C++???) web application that used OAuth between SharePoint Server and Exchange Server.

So I mean yes, you viewed Excel docs through a webpage just like you do today via ODSP or OneDrive consumer. The backend is completely different in the cloud service, though.

Well other than the limits of Copilots usefullness.

Exactly, and how do you think it's rigged in the setup? You're not getting top tier OpenAI service with Copilot was my point.

Doesn't that mean you can't actually use it for those sensitive documents?

> No limits.

Yet.

Microsoft runs the model, not OAI.

Correct.

lol

{EDIT} English and or the concept of written word may be foreign to you. Thank you for your assistance.

Not sure why you'd state that. 'Correct' is a grammatically correct and complete sentence to your question.