Lately my company has been doing a lot of complex accounting and reporting in spreadsheets. Overall was surprised by how well both GPT and Claude handled some of these extremely tedious tasks. Not uncommon to have an hours-long task compressed to minutes.
My anecdotal experience is GPT 5.2 Pro is decently ahead of Claude Opus 4.5 in this category when it gets to the tricky stuff, both in presentation and accuracy. The long reasoning seems to help a lot. But, apparently the benchmarks do not agree.
The deterministic part (calculations) is done by Excel.
The non-deterministic part is turning human instructions ("calculate the NPV over 10 years for X given Y") into Excel.
This is already a non-deterministic process (humans are non-deterministic!). The question is if an AI model can be more reliable than humans, and I can't see any reason why it wouldn't be.
The correct path is pretty clear, so the logits for following that path are going to be a long way from off-path.
For something like this the real problem is training the model to use Excel (which will show up by it being confused which sheet it is on or trying to use the wrong window or things like that), not the non-determinism.
so basically what you're saying is that: it doesn't do the math, it tells the math-doing-thing what math to do. Basically, instead of humans using Excel, imagine AI using Excel?
Yet I don't understand the aha moment here? It might save analyst time but aren't there already enough automation that you don't really need to tell the AI to tell the math-doing-thing to do the math because the math-doing-thing is already optimized for most general functions? What are we gaining from adding the non-deterministic process here when the real non-deterministic process is still the human being prompting what to do?
Seems like a solution to a non-problem from my pov.
Sure in the sense that you're setting up a program that has inputs and outputs etc. But then all math is programming. All language is too, even speaking can be considered programming if you're stretching the definition enough. But I will disagree that setting up spreadsheets = basically software engineering.
The LLM just needs to make sure it uses it appropriately. Doing that bit is the non-deterministic part, but the NPV calculation itself is completely deterministic.
Just a reminder that an accountant who might say "I use Currency format" is still working in binary floating point
as the format is just used as a display mask. And using VBA macros with the Currency type will hit problems at the boundary, when values move between the worksheet and the macro. The tool is broken in a way that proper accounting software is not...
Based on the article... is this basically just making Claude better at formatting and data presentation, or does it also get better at analysis? I get the impression it's the former.
The benchmarks look good. Slide decks and spreadsheets look better. The people must use Claude Cowork and have their Claude Code moment and figure out the consequences. It will be really interesting to see articles like this (https://mitchellh.com/writing/my-ai-adoption-journey) written by people who actually care about accuracy in places like KPMG to get their perspective on things.
I remember over hearing some normal people on the bus talking about essentially orchestrating some agent scraper to pull and summarise news from 40 different sites he identified as important which put him quite ahead of his peers. These were non-technical people orchestrating an agent workflow to make them better at work.
Though there’s not much that tickles my software brain here. But the agents are coming for us all.
And then you hand it to your boss who takes a 20 second look at it and asks why you made a projection that assume massive revenue growth and 3 years of perfectly flat utilities, insurance, G&A - no inflation etc.
It does look really promising as a skeleton starting point though. Like generate it, delete numbers and populate by hand.
Not unlike the boilerplate start we saw in AI coding a couple years back
> The side-by-side outputs below show how output quality has improved from Claude Opus 4.5 to Opus 4.6.
Disclaimer: I use AI to code (and I code for finance) and I love Anthropic.
But: for f-ck's sake, I cannot click on the picture and have it show up in full. It stays at its tiny size, impossible to read the numbers. I had to right-click and "open in a new tab".
AI is, somehow, definitely still not fully there yet.
You'll use a ton of AI but it won't wipe the humans out. In the end you'll have a compositional change, likely nothing catastrophic imo. In part because there is a buck to stop and Claude ain't got no hands...
Anthropic does anything to keep the Claude hype going; from fearmongering ("AI bad, need government regulations") to wishful thinking ("90% of code will be written by AI by the end of 2025" —Dario) to using Claude in applications it has no business being in (Cowork, accessing all your files, what could go wrong?) to releasing "research" papers every now and then to show how their AI "almost got out" and they stopped it (again, to show their models are "just that good") to prescribing what the society should do to adapt to the new reality to doing worthless surveys on "how AI is reshaping economy, but mostly our AI not others".
Now this is going to be interesting to watch to see if the finance bros financing this AI wave to get rid of SW engineers will keep financing getting rid of their own.
My anecdotal experience is GPT 5.2 Pro is decently ahead of Claude Opus 4.5 in this category when it gets to the tricky stuff, both in presentation and accuracy. The long reasoning seems to help a lot. But, apparently the benchmarks do not agree.
Edit - noticed OpenAI specifically focuses on finance use cases in their gpt-5.3-codex blog as well https://openai.com/index/introducing-gpt-5-3-codex/