Will Claude Eat Excel?
RIP to your friendly investment banker
This week presented a Claude Conundrum.
Should I write about the viral Clawdbot (which has everyone buying Mac Minis)?
Or should I write about the new Claude Excel Plug-in that is supposedly making first year analysts obsolete?
Since Excel pays the bills, I’ll save Clawdbot for next week.
Join our Clawdbot demo on Friday
The AI Accelerator is your unfair advantage for upskilling in a rapidly changing AI world. Yes, Claude Code was discovered in Jan of 2026, but our members have been building their own apps and workflows on it since last June. Join our community of executives, entrepreneurs and investors as we build our AI fluency skills with courses, workshops and coaching.
Introducing the Claude Excel Plug-in
Excel and spreadsheets have been a thorn in AI’s side for a long time. And for those familiar with the “left-to-right” tokenization approach of LLMs, it makes sense that the venerable Excel would be a tough nut to crack.
Even Satya Nadella admitted that AI in Excel is hard:
Currently solutions are just a “wrapper” around an LLM. What you need is a “middle tier model” that understands the “artifacts of Excel” as opposed to just the pixels.
Well, Anthropic’s Claude wants to take yet another crack at the enterprise and recently made their Excel plugin available to all pro accounts.
According to their official documentation this plug-in can do the following:
Ask questions about your workbook and get answers with cell-level citations
Update assumptions while preserving formula dependencies
Debug errors and identify their root causes
Build new models or fill existing templates
Navigate complex multi-tab workbooks seamlessly
Claude, build me a model
As a long-time Mac user, I reluctantly downloaded Excel1 and put this to the test.
Now my financial modeling skills are quite dusty, so I thought up the hardest Excel Problem that I could think of — yet still verify myself.
That was a detailed buy versus rent calculator that incorporated a lot of different variables.
Create a buy vs. rent calculator
2 million purchase price, 20% look up the interest rate, actually 6%, 30 years. Include all of those costs. Search the web for a good estimate. Include rent increases, include HPA increases. Make the monthly rent $5,000. Factor the opportunity cost. Claude came out guns blazing and built this summary:
While it passed the sniff test, I definitely needed to see more of the supporting calculations. The summary results were also wrong (either because of calculation issues or definition issues).
I realized that Claude had made an assumption about the monthly principal vs. interest payment and prodded.
Did you create a mortgage amortization table for the monthly payment?It made me realize that I wanted to see a full breakdown:
I would like to audit this. Can you create a table with a month-by-month comparison that breaks down all of the costs and opportunity costs? Do it on a separate worksheet. This breakdown was pretty quick (< 1 minute) and directionally accurate.
I say directionally because there were a few random cells that were just blank! When prodded, it said:
Fixed! Row 3 now has all the formulas. The cells R3 and T3 were empty because when I originally created row 3 separately from row 2, and then later fixed column S, I didn't include R and T in those updates.It still had the wrong formula (not a random typo, but it fundamentally misunderstood the calculation).
It looks like buy net wealth, column S is wrong. Shouldn't it include the down payment value? A similar version of the issue persisted:
It looks like column I needs to update the cumulative cost to exclude the down payment. Eventually, it seemed to get all the calculations correct. BUT, I probably spent as much time troubleshooting the output as it would have taken me to build it from scratch.
Feedback from other AI practitioners
Next up, I collected feedback from my peers on their experiences with the Claude Excel plug-in.
Brett Caughran (Ex-PM at multiple hedge funds) didn’t even use the plug-in, but asked Claude (web app) to pull online sports gambling data.
Brett’s experience:
To my delight, Claude actually built me a 6-tab tracker in Excel that rolled up total OSB handle & OSB gross gaming revenue. It wasn't perfect (while spot check of the industry numbers looked solid, it was holding a steady 32% share at DKNG which made the DKNG estimate not reliable).
His (early) read:
Is this doing anything that I couldn't do by hand / manually? No. Did it do something in 12 minutes that would take me maybe 3 hours to build then 30 minutes to update every month? YES.
Next, venture capitalist (and one of my favorite bloggers) Jack Raines gave it a spin. He fed a sample Series A Pro Forma with the following prompt:
“Create a waterfall showing how the funds will be distributed on a $50 million exit, assuming 1x non-participating liquidation preferences from all VCs. Generate your graphic in Cell J25 and work right / down from there.”Here’s the output:
Jack admits, this isn’t a particularly rigorous task and adds:
But it still cut a ~25-30 minute task down to ~4 minutes, and actually freed up more time to spot check that the right cells were referenced, inputs weren’t hard-coded, etc. Basically, you can get to a testable final product much, much faster.
Finally, I saved the best for last.
Michael Yuan (whom I only know from following on X) is a “Goldman and TPG Alum” who is building AI for IB/PE professionals. Yuan’s post I wouldn’t use Claude in Excel (yet) describes his task:
Convert a raw transaction database -> customer cube.2 Just a simple version that an Associate could create in ~15 minutes.
Yuan goes on to go back-and-forth with Claude for 45 minutes and getting close to his desired result.
In his words:
[Claude for Excel] gets ~60-70% of the way there, in an industry where you get fired for 90%.3
Will we be living in a banker-free world?
The early read is that Claude for Excel performs well both easy and intermediate tasks.
If you’re getting paid 6 figures out of college to build Excel models, there’s probably some human judgement or nuance that could be hard for AI to replicate — and industries that require utmost precision might be a bit slow on the uptake.
But everybody else…
Not only did I download Excel, I paid for it just so I could write this post!
I had heard of the term, but had to look it up. It’s a “pivot table on steroids” looking at data across time period, customer segment, product, etc.
Yuan does not hide that he is building a competing product called Waverly.ai.
What's your take on the new Cloud Excel plugin?