← All posts

AI Strategy

Claude for Excel: find the broken formula before your CFO does.

Aditya JhaPartner — Strategy & Growth7 min readBeginner
Share
On this page
  1. 01Why tracing a formula was always the slow part
  2. 02What actually changed
  3. 03Trace the error inside the actual workbook
  4. 04Then build the report, not just fix it
  5. 05Turn it into a skill so nobody re-explains it next month
  6. 06What it replaces, and what it doesn't
  7. 07Where it's the wrong tool

The board deck is due tomorrow morning, and the numbers on the summary tab don't tie out. Somewhere across forty tabs of a model someone else built, a cell that used to feed the total got overwritten, deleted, or quietly changed to a hard-coded number three revisions ago. You're the one who has to find it — not because you built the model, but because you're the one still awake.

Why tracing a formula was always the slow part

Spreadsheets don't explain themselves. Excel's own "trace precedents" tool shows you one arrow at a time, and a real model runs that chain through six or seven tabs before it reaches a number you can check by hand. If you didn't build the workbook, you're re-learning someone else's logic under a deadline. And every AI assistant you'd normally reach for — a chat window, a copilot bolted onto a document — could only see whatever text you pasted in. None of them could see the live formula, the named range, or the tab three sheets over that the formula was quietly pointing at.

What actually changed

Claude for Excel is a sidebar add-in — an extra panel built into Excel itself, not a separate app you switch to — that reads your open workbook directly: every tab, every named range, every formula, all at once. Anthropic ran it as a research preview for Enterprise customers from October 2025, then opened it to every Pro subscriber in January 2026. It now sits on the Pro, Max, Team and Enterprise plans, inside Excel on the web, Excel on Windows with a current Microsoft 365 subscription, and Excel on Mac version 16.46 or later.

  • Works: Excel on the web, Excel on Windows (Microsoft 365), Excel on Mac 16.46+.
  • Doesn't work: Excel 2016 or Excel 2019 as a perpetual or volume-license copy — there's no subscription connection for the add-in to use.

Trace the error inside the actual workbook

Instead of clicking through precedents tab by tab, you ask it directly, and it answers with a citation — a clickable reference to the exact cell it's talking about, the same way a footnote points to a source. Anthropic's own documentation shows the kind of chain it will walk: a Q3 revenue figure traced back through an assumptions tab, a pricing tab, a forecast tab, and finally the summary cell that pulled from all three. It can trace `#REF!`, `#VALUE!`, `#DIV/0!` and circular-reference errors the same way, ending in a plain-English sentence about which upstream cell caused the break and why.

The citation chain, tab by tab: Inputs → Pricing → Forecast → Summary.
prompt
Summary!C22 should equal Forecast!B14 but it's off by about 8%.
Trace the formula chain between them and tell me exactly which
upstream cell has changed, and what it changed from.

Then build the report, not just fix it

The same access to live formulas works in the other direction. Describe the report you need in plain English, and it builds the structure with the formulas already linked — not a static table of numbers you'd have to wire up yourself. Change an assumption afterwards and the downstream cells recompute correctly, because it's editing real formulas, not overwriting them with values.

prompt
Build a monthly budget-vs-actual variance report from the Budget
and Actuals tabs. One row per line item, a variance column in both
currency and percentage, and flag anything over 10% variance in a
separate summary section. Link every cell to the source tabs.

Turn it into a skill so nobody re-explains it next month

In March 2026, Anthropic added Skills to the Excel and PowerPoint sidebars — the same idea we covered in Claude Skills, applied inside a spreadsheet instead of a coding agent. Once someone on your team has the variance-report prompt above working exactly right, they save it as a one-click action in the sidebar, and it's available to everyone else on the plan — nobody re-types it, and nobody's version drifts from anyone else's.

What it replaces, and what it doesn't

It replaces the hours someone spends as the team's de facto formula-tracer, the Slack thread trying to work out who broke the model, and the ritual of rebuilding the same report structure from scratch every month. It doesn't replace deciding which assumptions are the right ones, or what a number actually means for the business — that judgement stays yours. And it doesn't turn Excel into a database: a workbook with hundreds of thousands of rows, or one that five people need to query at once, belongs in an actual database, no matter how well an assistant reads around Excel's limits.

Free, Pro and Max train on your conversations by default. Enterprise doesn't. Team — ask.

Check before you paste in real numbers

Anthropic's default is to use Free, Pro and Max conversations to train future models unless you switch off "Help Improve Claude" in privacy settings; Enterprise accounts are exempt by default. If your company is on Team, don't assume either way — check the admin console before you open payroll, a client's model, or anything else you wouldn't want retained.

Where it's the wrong tool

Skip it for a five-line spreadsheet — fixing it yourself is faster than describing it. Skip it, too, the moment the real problem is that the workbook has outgrown Excel; moving the data is the fix, not a better way to read around the limit. And whatever it hands back, check it the way you'd check a junior analyst's first draft before it goes in front of a client or a board — it will rebuild a wrong assumption just as confidently as a right one.

The useful version of this is that the spreadsheet finally explains itself. The part that doesn't change is that someone still has to be the one who checks.
Aditya Jha, Partner — Strategy & Growth

The skill that used to matter in a finance or ops role was knowing where the numbers were hiding. That's worth less every month a tool like this improves. What's worth more is knowing which number to ask about in the first place, and catching it when the answer is wrong — and that's a judgement no add-in makes for you.

If you're not sure which of your team's recurring reports are worth wiring up this way, and which ones still need a real engineer, that's exactly the kind of call we help make in an AI Opportunity Assessment.

Next step

Want this built into your business, not just explained?

Our AI Opportunity Assessment maps where AI saves you time and money, and prices the build — $199, a written report, 7–10 days. If the answer is that AI is not worth it for you yet, we will say so in writing.