Claude for Excel is the right place for three finance jobs inside the open workbook: building a three-statement model from actuals, tying the cash flow to the balance sheet, and gathering every control onto one tab, each with a check you run before sign-off. Public tests cover the first two. Two more, journal-entry testing and turning a folder of receipts into a schedule, run better as a task in a Claude conversation. As of October 2026 the add-in is generally available on Pro and above (Pro is $20 a month, or $17 a month billed annually) and not on the Free plan. Public tests show that the speed is real, and so are the faults: balance checks typed as the word "OK", a blank workpaper tab, a rounding threshold Claude chose for itself. Each workflow below comes with the check that catches its fault before sign-off.

What Claude for Excel is, and who can use it

Many people call it "Claude in Excel". The product is Claude for Excel, one of the add-ins installed from the "Claude for Microsoft 365" listing on Microsoft AppSource, alongside the PowerPoint and Word add-ins. IT can deploy it from the Microsoft 365 Admin Center instead. Once it is installed, open it from Home > Add-ins on Windows or Tools > Add-ins on Mac, and sign in with a Claude account.

As of 31 October 2026, Anthropic's pricing page includes Claude for Microsoft 365 on Pro, Max, Ultra, Team, and Enterprise, and not on Free. It runs on Excel on the web, Windows (Microsoft 365), and Mac, but not on Excel 2016 or 2019 perpetual licences, iPad, or Android, and it does not support macros, VBA, or data tables.

Anthropic's documentation says it answers questions about a workbook "with cell-level citations", adjusts assumptions "while keeping formula relationships intact", identifies and resolves errors and their root causes, and generates new models or populates existing templates. Elsewhere on the same page, Anthropic says the add-in runs on a curated subset of Claude's models.

On 16 September 2026 Anthropic folded Cowork into ordinary chat: describe a job in any conversation, and Claude either answers it or takes it on as a task that hands back finished files. The merged experience is rolling out first to Pro, Max, and Ultra. Since 6 October 2026, new tasks on those plans run in Anthropic's cloud, and when a task needs a file from your computer, Claude fetches a copy of that file. That matters for workflows 3 and 4, which handle ledgers and client receipts.

The 5 workflows at a glance

WorkflowWhere it runsWhat the public test showedThe check before you sign
1. Three-statement model from actualsClaude for ExcelThe forecast matched a course model; the first balance checks were typed "OK"Click into every check cell: each must be a formula
2. Cash flow tied to the balance sheetClaude for ExcelClaude classed a $2 difference as rounding and set its own thresholdYou set the threshold, and you open every plug Claude lists
3. Journal-entry testingA task in Claude28,160 ledger lines tested in about 10 minutes; the testing-results tab came back blankOpen every tab; tie counts to counts and amounts to amounts
4. Receipts to a reconciled scheduleA task in ClaudeFiles renamed and linked; three vendors with no receipt foundCount files in against rows out; open at least 10 receipts
5. The control tabClaude for ExcelNo public test of this exact promptBreak one subtotal on a copy and watch every flag turn red

The tests used earlier versions of the add-in and its models, and workflows 3 and 4 were recorded in Cowork. Read them as evidence of a pattern, not a benchmark.

Workflow 1: a three-statement model from your actuals

In the Corporate Finance Institute's test, Tim Vipond deleted the forecast from a course model, gave the blank to the add-in with a two-sentence prompt, and got back statements whose net income, balance sheet, ending cash, and retained earnings matched the course model. AI for month-end close covers that test in detail.

For a reviewer, Josh Aharonoff's test (Your CFO Guy) is more useful, because it shows where the build goes wrong. On a monthly model with a drivers tab, he found three faults:

  • Typed checks. The first balance-sheet checks were cells containing the word "OK", not formulas. Claude replaced them with formulas later in the build.
  • Typed subtotals. Some historical subtotals stayed as typed values. When he changed a historical security-deposit balance to 10, nothing flagged it.
  • Hidden checks. The checks sat on one tab, in green, where a reviewer could miss them.

On the credit side, Claude caught its own wrong column reference mid-build, a raise in revenue growth from 5% to 10% updated the income statement, and a $10,000 purchase of computers appeared as an outflow on the cash flow statement. The Runbook's flex-test prompt checks the rest by asking Claude to list every line on all three statements that moved.

The Runbook's prompt adapts Aharonoff's and closes the gaps his test found:

Build a three-statement financial model from the actuals in this workbook,
[Jan 2024] to [Dec 2025]. Project monthly through [Dec 2027].

1. Add a Drivers tab showing the historical and projected value for every
   account, and the basis of each projection (growth %, % of revenue, fixed).
2. Link the income statement, balance sheet, and cash flow so the balance
   sheet balances and a change to any driver flows through all three.
3. Build every subtotal as a SUM formula and every check as a live formula.
   No typed values in any calculation or check row.
4. Roll net income into retained earnings correctly at each year end.
5. Colour hard-coded inputs blue and formulas black.
6. Add error checks wherever a figure should tie.

Do not edit the source tabs. When you finish, list every assumption you made,
every cell you were unsure of, and every check you added, with its cell.

The check. Click into every check cell: a cell that reads OK and contains the word OK is not a check. Type a wrong number into one historical line; if no check fires, a subtotal is typed. On Windows, F5, then Special, then Constants with only Numbers ticked selects every typed number on a sheet; inputs belong on the Drivers tab, and anything else needs a reason. Then agree or change each assumption on Claude's list: a driver you did not choose is still yours once you sign.

Workflow 2: tie the cash flow to the balance sheet, and name every plug

A model that balances is not yet a model that reconciles. If Claude derives cash as the balancing figure, the balance sheet balances by construction, which proves nothing. This workflow replaces the plug with an indirect-method cash flow, links balance-sheet cash to it, and has Claude list every remaining plug with its cell.

In Aharonoff's test, Claude built and linked the cash flow, found a circular reference on its own, and found a $2 difference between the old plug and the new cash flow. It classed the $2 as rounding and added a threshold of its own. It may have been right; it was still Claude's call.

The Runbook's prompt therefore fixes the threshold and the sign conventions before the build, and takes net income and depreciation from the income statement, never from movements in retained earnings or accumulated depreciation.

The check. The tolerance in the check cell is the figure you chose, in your reporting currency. Open every plug on Claude's list and ask what it would be if it were calculated. Then ask again, in a new message, whether any other cell balances by subtraction. Pick one asset and one liability that moved and confirm that the sign of each cash effect is right. It is the same habit as tying every figure in commentary to a cell, which How to check AI-drafted variance commentary sets out step by step.

Workflow 3: journal-entry testing as a task in Claude

Journal-entry testing covers the whole population, where code beats a person scrolling, so it runs as a task in Claude, with 4 files in one working folder: the ledger extract, the system log, the test procedures, and the workpaper template.

The evidence is a public demonstration by the accountant who publishes as Pythonic Accountant. Given a ledger of 13,295 entries on 28,160 lines, worth $928.3 million, Claude recommended the tests the data could support, and wrote and ran Python against every line. In about 10 minutes, set-up included, it reported 4 unbalanced entries, 12 sequence gaps, 109 round-dollar amounts, and 99 entries flagged by the business-hours test.

Then he opened the workpapers. The index had no preparer or reviewer dates, the lead sheet was missing the log counts, two sample-selection columns were empty, and the testing-results tab was blank. Asked, Claude filled the gaps, but the second pass wrote over the sample-selection headings and compared the ledger's line counts with the log's dollar amounts. His verdict: good, not perfect, and not yet something he would use on anything important.

The Claude Finance Runbook splits this job into two prompts: the first reads the procedures, describes the population, and stops for your approval; the second runs the approved tests and fills every tab. It also sets the conversation's permission to Manual, so Claude asks before each action. That is a choice you make: Auto is the default on Pro, Max, Ultra, and Team.

The check. Read every tab against the procedures, not Claude's summary. Tie the entry count, line count, and total debits and credits to the system log, counts against counts and amounts against amounts. Confirm one flagged entry from each test in the ledger yourself. Compare the headings with the original template, file Claude's script with the workpaper, and leave preparer, reviewer, and dates to people. The UK Financial Reporting Council's AI guidance of 30 March 2026 is blunt on the last point: "the human auditor is always accountable". For US audits, the PCAOB's amendments on technology-assisted analysis apply to fiscal years beginning on or after 15 December 2025.

Workflow 4: receipts to a reconciled schedule

As a task in Claude, with a working folder of receipts connected from the desktop app, Claude reads each file, renames it to a pattern you set (vendor, date as YYYY-MM-DD, category, amount), builds an Excel schedule with a link to every file, and lists what is missing against an expense list or card statement.

The CPA Hector Garcia ran a version of this on a folder of generically named receipts. Claude renamed every file as payee, date, category, and amount, and the two he opened by hand matched. Given a client's vendor list, it named the three vendors with no receipt in the folder. It checked restaurant receipts item by item for alcohol and flagged a supermarket receipt it had already categorised as liquor; Garcia said it would also list any receipt with no itemisation. It then built a schedule with totals by category and a working link to each file. He noted that its categories could be wrong.

The check. Count first: the number of files in the folder equals the number of rows in the schedule. Open at least 10 receipts, and every one over your threshold, and compare vendor, date, and amount. Confirm every category comes from your chart of accounts, clear each blank or unreadable receipt by hand, and click three links. The reconciliation is only as complete as the folder and the list you gave it, so confirm both cover the whole period.

Client receipts are client data, and a task fetches a copy of each file it opens, so use a folder holding only this engagement's receipts.

Workflow 5: the control tab

Workflows 1 and 2 leave checks wherever Claude put them, and a green check on one tab gets missed while a typed subtotal fails without a sound. The fifth moves every control onto one tab, with a PASS or FAIL flag on every sheet, copying how Aharonoff builds checks into his own models.

Add a Checks tab that consolidates every control in this model:

- Assets = liabilities + equity, for every period.
- Each subtotal = the sum of its parts.
- Each inter-statement tie: net income, cash, and retained earnings.
- Any other check already in the model, moved here and linked.

Give each check a live formula, a tolerance of [amount], and a PASS or FAIL
result. Add one master flag that fails if any check fails. Put a cell linked
to the master flag at the top of every sheet, formatted to turn red on FAIL.

Then list every hard-coded number outside the Drivers tab and the historical
actuals, with its sheet, its cell reference, and what it appears to be.

No public test that we found has run this exact prompt, so test the result harder, not less.

The check. Break the model on purpose, on a copy. Type a wrong number over one historical subtotal: the flag on every sheet should turn red. Change one month's balance-sheet cash by more than your tolerance: the cash tie and the master flag should both fail. Undo both and confirm every flag returns to PASS. A check you have never seen fail is untested.

What Anthropic says the add-in is not for

Anthropic's own Excel documentation lists three uses it does not recommend: final client deliverables without human review, audit-critical calculations without verification, and models containing highly sensitive or regulated data without proper controls. The 5 checks above deal with the first two. The third is a firm decision, and these terms, as of October 2026, belong in it:

  • Retention. Add-in inputs and outputs are deleted on Anthropic's side within 30 days, except in the cases Anthropic's data-retention article lists, and the add-in does not inherit an organisation's custom retention settings.
  • Audit trail. Add-in activity is not in Enterprise audit logs. Enterprise organisations with the Compliance API enabled do get add-in sessions included.
  • Chat history. It is stored locally in your browser and not synced across devices. Keep the prompts in a notes tab in the workbook so they travel with the file.
  • Outside files. Use only trusted spreadsheets: Anthropic warns that templates, vendor files, and data imports can carry hidden instructions. Claude asks before a risky operation, so read each confirmation.
  • Training. On personal plans Anthropic labels model training "Opt-out". Switch off "Help Improve our AI models" in Settings > Privacy before a ledger or trial balance goes in.

Your firm's policy on where client data may go comes before all of it. The four-gate governance model is one way to write that policy down.

Questions finance teams ask

Is Claude for Excel available on the free plan?

No. As of October 2026, Claude for Microsoft 365 is included on Pro, Max, Ultra, Team, and Enterprise, and Pro costs $20 a month billed monthly, or $17 a month billed annually.

If Claude's forecast matches a known model, is it right?

It means the formula logic held on that file. Aharonoff's test shows the same tool typing "OK" where a check belongs. The controls, the threshold, and the assumptions still need a named reviewer.

Where to go next

The Claude Finance Runbook is free with your email address: 25 pages covering all 5 workflows, with 8 copy-ready prompts and 28 sign-off checks. For Claude in depth across professional finance work, from contracts to reconciliations and analysis, The Claude Playbook is $47, paid once.

Sources

Vendor pages checked on 31 October 2026. The 4 public tests were confirmed live on 22 October 2026.