Claude for Excel: install it, use it on a real model, and check its work

Open a workbook you didn't build, open the Claude add-in, and type "Walk me through how the revenue number in cell C42 is calculated." The answer comes back in a sidebar with cell-level citations: click one and Excel jumps to the input it names. That's the core of Claude for Excel, and the citations are why we'd use it on a model we have to defend.

What Anthropic documents, as of 22 October 2026:

Sources: Anthropic's Claude for Excel docs, the Claude for Microsoft 365 page and claude.com/pricing.

Anthropic doesn't publish a file-size or row limit for the add-in. It does publish what it shouldn't be used for, which we quote below.

For yourself

1. Open the Claude for Microsoft 365 listing on Microsoft AppSource and select "Get it now".
2. Open Excel, activate the add-in (Home → Add-ins on Windows, Tools → Add-ins on Mac), and sign in with your Claude account.
3. Open Settings in the add-in sidebar and fill in the Instructions field: number formats, sign conventions, currency, your fiscal year. These apply to every Excel conversation, and only to Excel. The add-in doesn't read your Claude projects.

For your organisation

An admin deploys it from the Microsoft 365 admin centre (Anthropic's steps):

1. Settings → Org Settings → User owned apps and services: turn on "Let users access the Office Store".
2. Settings → Integrated apps → Add-ins, and search for "Claude for Microsoft 365".
3. Assign it to the organisation, or to specific users or groups.

Two snags Anthropic documents. If your admin roles are activated through Microsoft Entra Privileged Identity Management, the Integrated apps page doesn't recognise them and deployment fails; deploy from an account with the role assigned permanently. And where the Office Store is switched off, deploy the custom manifest XML instead. Companies that route AI through Amazon Bedrock, Google Cloud Vertex AI, Azure AI Foundry or an LLM gateway can deploy it without individual Claude accounts.

It won't run on Excel 2016 or 2019 perpetual licences, on iPad, or on Android.

Already paying for Copilot?

Fair question. Microsoft's Copilot in Excel needs a Copilot-eligible Microsoft 365 subscription, and with a commercial Copilot subscription or Microsoft 365 Premium it can switch between OpenAI and Claude models, once an admin enables Anthropic as a sub-processor (Microsoft Support FAQ). If you already have it, try it on the walkthrough below before buying anything. We'd still pick Anthropic's add-in for model work, for the cell citations and the documented guardrails; Copilot's advantage is that it's already in your tenant and your IT team knows how to govern it.

Whichever add-in you choose, check the numbers before they leave the workbook. In Claude chats, Count does that claim by claim, tagging each one as a citation, an assumption, model knowledge or a hallucination.

The most useful paragraph in Anthropic's documentation is its own list of where to stop. Under "Current limitations", Claude for Excel is not recommended for:

• Final client deliverables without human review.
• Audit-critical calculations without verification.
• Models containing highly sensitive or regulated data without proper controls.

It also lists two unsupported capabilities: "Data tables" (Excel's what-if tables) and "Macros and VBA operations". If your model's sensitivity analysis is a data table, Claude can't build or edit it.

Then the warning that matters most for anyone who receives workbooks from outside: "Only use Claude for Excel with trusted spreadsheets. Files from external sources can contain hidden instructions that manipulate the add-in into extracting data, modifying records, or performing destructive actions." Anthropic says its testing found exactly those scenarios. A vendor's pricing template or a downloaded model is an external file.

For compliance teams, the data handling is unusual and worth knowing before rollout:

• Inputs and outputs are deleted on Anthropic's side within 30 days, with exceptions set out in its retention policy.
• Chat history is stored in your browser, isn't synced across devices, and can be cleared from Settings.
• The add-in "does not inherit custom data retention settings" and its activity "is not included in Enterprise audit logs". Enterprise organisations using the Compliance API do get Excel sessions there, in public beta.

An illustrative example: a 40-person SaaS company's FY26 budget workbook with four tabs. TB_Oct holds the October trial balance, Budget the monthly plan, Actuals a SUMIFS layer mapping accounts to budget lines, and Summary the variance table finance sends to the CEO. Summary!D12 says operating expenses are £84k over budget for October. You've been asked why, by tomorrow.

1. Work on a copy. Anthropic's best practices say to start from a trusted copy before asking Claude to edit widely.
2. Trace the number. Ask how Summary!D12 is calculated, listing every input cell. You're looking for the path from TB_Oct through Actuals to Summary, with citations at each step. Click two of them.
3. Find the break. Ask which accounts in TB_Oct don't map to any budget line in Actuals. Unmapped accounts are the commonest cause of a variance that "isn't real".
4. Split the variance. Ask for the £84k by budget line, then by account within the largest line, using only Actuals and Budget.
5. Ask for the explanation last, and only from the workbook: anything it can't see in a cell gets labelled as unexplained.
6. Check before it leaves, using the steps at the end of this page.

The order matters. Explanation is where an assistant is most fluent and least grounded, so make it the final step, after the arithmetic is pinned to cells you've clicked.

Seven jobs, each with a prompt to paste and the check that makes it safe. All illustrative; fill the brackets.

1. Audit a formula chain

Walk me through how [Summary!F18] is calculated. List every input
cell, any number typed directly into a formula, and any link to
another workbook.

Check: click three citations and read the formulas yourself. Hard-coded numbers inside formulas are the classic trap.

2. Find hard-codes and inconsistencies across a row

On [Forecast], find formulas in rows [10-40] that differ from the
others in the same row, and any cell where a formula has been
overwritten with a typed value.

Check: spot-check two flagged cells with Trace Precedents on the Formulas tab. Ask what it didn't flag in row 10; an empty list deserves suspicion.

3. Reconcile two tabs

Compare [Actuals] with [TB_Oct]. For each budget line, show the
total in both, the difference, and the accounts causing it. List
any account in TB_Oct that maps to no budget line.

Check: the differences should sum to the gap between the two tab totals. If they don't, something is double-counted.

4. Build a model from a trial balance

Build a three-statement model on new tabs from [TB_Oct]. Put every
assumption on one Assumptions tab, labelled with its source or
ASSUMPTION. No numbers typed into formulas.

Check: the balance sheet balances, cash ties between statements, and every ASSUMPTION is one you'd defend. This prompt adapts one from Anthropic's own docs.

5. Explain a variance

Explain the variance in [Summary!D12] using only data in this
workbook. Split it by budget line and account. Label anything you
can't trace to a cell as UNEXPLAINED.

Check: the parts must add up to the total. Then read every sentence that gives a reason and find the cell behind it.

6. Flex an assumption

On this copy, change [churn] from [2%] to [3%] and update dependent
calculations. List every Summary cell that changed, before and after.

Check: the direction. Higher churn should lower revenue; anything that moved the other way needs explaining.

7. Debug an error

Find the root cause of the #REF! errors on [Summary] and propose a
fix. Don't change anything until I confirm.

Check: after the fix, compare the totals that weren't broken before with their old values. A fix that moves them changed more than the error. More prompts for this kind of work are in prompts for financial analysis.

A cell citation tells you where a number came from. It doesn't tell you the formula in that cell is right, or that it was the right cell to use. And spreadsheet work is still hard for AI: the best published score on SpreadsheetBench V2's 321 end-to-end business workflows is 34.89% (checked October 2026). Microsoft says the same of its own tool: "Review, edit, and verify anything Copilot creates before you rely on it."

A two-minute check before a number leaves the workbook

1. Pick the three cells the answer depends on and click through to them.
2. Check the basis: period, units, currency, and whether a SUMIFS range stops short of the last row.
3. Recompute one figure on a scratch tab, by hand.
4. Compare one total with the system it came from: the ledger, the billing export, last month's pack.
5. Ask Claude: "Which statements in your answer aren't backed by a cell in this workbook?"

Step 5 is the one that catches an invented reason for a variance, the most persuasive and least supported part of most answers. Count asks that question of every claim: connect it to Claude, ask it to audit the answer, and each statement comes back labelled as a citation, an assumption, model knowledge or a hallucination, with the reasoning chain behind it. For the wider finance setup, see Claude for finance; for data teams, Claude for data.

Is Claude for Excel free?

How do I add Claude to Excel?

Does Claude for Excel work on Mac?

Can Claude for Excel write macros or VBA?

Is Claude for Excel better than Copilot in Excel?