How to use Claude for Excel on a real model
Pasting a table into a chat window and asking what it means is a different thing from having something read the workbook. The Claude for Excel add-in does the second: it works inside the file you have open, answers with citations you can click, and changes values without breaking the formulas that depend on them.
Here is what that is good for, and where it stops. Checked against Anthropic's published documentation in September 2026.
Getting it installed
Install it from the Claude for Microsoft 365 listing on Microsoft AppSource, then open Excel, activate the add-in and sign in. It is generally available on Pro, Max, Team and Enterprise plans.
It runs on Excel for the web, on Windows with a Microsoft 365 subscription (build 16.0.13127.20296 or later), and on Mac from version 16.46. It does not run on the 2016 or 2019 perpetual and volume licences, or on Android. It does not run on iPad either — the add-in needs SharedRuntime support, which iPad does not provide.
Start from a copy, not the original
Before the interesting part, the boring part that saves the afternoon.
Anthropic's own best-practice guidance says to start with a trusted copy of the workbook before asking Claude to edit widely, and it is right. The add-in does warn before overwriting existing data, which catches the obvious cases — but a warning you approve is still an edit, and a model has a lot of cells whose importance is not obvious from the cell itself.
Duplicate the file. Then experiment freely, which is the point.
Cell-level citations are the real feature
Ask a question about the workbook and the answer comes back with cell-level citations you can click to jump to the cell it came from. That sounds like a convenience and is actually the whole thing: it means you can audit the answer instead of trusting it.
The requests this unlocks are the ones that used to mean an hour with the trace-precedents arrows:
Where does the revenue figure on the summary tab actually come from? Walk me through it.
What assumptions is this model making about churn, and which cells set them?
Explain this formula in plain language, then tell me what breaks if the input is zero.
It works across multi-tab workbooks, which is where tracing by hand gets genuinely painful.
Changing values without breaking the model
When Claude updates cell values it keeps formula relationships intact, so downstream cells recompute the way they should. That makes flexing assumptions a conversation rather than a careful manual edit:
Set the growth rate to 8% and tell me what happens to the year-three margin.
Run this at 5%, 8% and 12% growth and give me the three outcomes side by side.
The discipline that makes this trustworthy: ask it to tell you which cells it changed, every time. A model that recalculates correctly and a model that was changed where you expected are two different claims.
Building something from nothing
It will populate an existing template from a description, or generate a model from one — Anthropic's own examples are an LBO template and a three-statement model built from a trial balance.
The two-step is worth using here in the same way it is in a chat window. Ask for the structure before the numbers:
I need a three-statement model from this trial balance. Lay out the tab structure and the key schedules first. Do not build anything until I approve it.
Arguing with a described structure is minutes. Arguing with a built model is an afternoon.
Debugging what is already broken
It locates the root cause of calculation errors and suggests fixes — the #REF! that appeared after somebody deleted a row, the #DIV/0 that only shows up in one scenario. Ask for the cause rather than the fix, and you learn something the next time it happens:
This cell shows #REF!. Find the root cause, explain what broke it, then fix it.
The Excel-native operations people forget to ask for
It does the ordinary spreadsheet work too: sorting, filtering, editing pivot tables, applying conditional formatting, creating data validation dropdowns. These are the tasks most likely to be done by hand out of habit, and they are the ones with the least risk in handing over.
Two things it will not do
The documentation names exactly two unsupported capabilities, and both are worth knowing before you plan around them: data tables, and macros and VBA operations.
If your model leans on either, that part stays manual. It is a short list, but it is a real one.
Persistent instructions worth setting
Settings has an Instructions field that applies to every conversation in Excel, separate from the ones in Word and PowerPoint. The obvious candidates are formatting conventions you would otherwise repeat — thousand separators, bold column headers, currency and locale, how you like schedules laid out.
And the one worth setting on day one: tell it never to invent a figure, and to leave a visible placeholder when a number is not in the source. In a spreadsheet a plausible invented number is far more dangerous than a wrong sentence, because nothing about a cell says where it came from.
Working across Excel, PowerPoint and Word
The Excel, PowerPoint, Word and Outlook add-ins can share one conversation, so the figures in an open model can populate a slide or a memo without a copy-paste round trip. The setting is "Let Claude work across files", in each add-in's settings. It is on by default for Pro and Max, off by default for Team and Enterprise, and the toggle is per device.
Two limits before you rely on it: Claude can only read and write files that are currently open, and it cannot create, open, close or switch files itself.
What to check before you send it
Anthropic's guidance is explicit that the add-in is not recommended for final client deliverables without human review, or for audit-critical calculations without verification. Treat what follows as the minimum.
Every cell it changed. Ask for the list, then spot-check the ones that feed a headline number.
The formulas, not just the outputs. A number that looks right and a formula that is right are not the same finding.
The edges. Zero, negative, empty and the largest plausible input. Plausible-looking output that fails at the edges is the most common failure there is.
Files from outside. Anthropic warns that spreadsheets from external sources — downloaded templates, vendor files, data imports — can carry hidden instructions designed to manipulate the add-in into extracting data or modifying records. When Claude asks you to confirm an operation, read the confirmation rather than clicking through it.
Where to go from here
Most of what separates a useful answer from a disappointing one is not specific to Excel — it is the habits that make any request land: saying who the output is for, giving it the raw material instead of letting it invent, asking for the format you actually want, and iterating instead of restarting. Claude Chat — The Beginner's Handbook covers those in 42 pages, along with the document skills that produce Word, Excel, PowerPoint and PDF files from a plain conversation.