Claude for Excel and Spreadsheets in Finance

What Claude for Excel and spreadsheet analysis add to finance work

Claude for Excel is an add-in from Anthropic that puts Claude inside an open workbook, so it can answer questions about a model, adjust assumptions, debug errors and build or populate templates without you leaving Excel. Separately, Claude's chat apps can read an uploaded spreadsheet and write a new one back. Together they let a finance team clean an export, reconcile two files, or check a model faster, provided a person still verifies every figure before it is used.

This guide is about doing finance work in a spreadsheet with Claude, not about the Claude API or the wider bookkeeping and audit uses our claude-for-accounting guide already covers. Every Claude feature named below is confirmed on Anthropic's own documentation and support pages as of September 2026 and listed in the sources; a capability that is not confirmed there is left out rather than guessed at.

Two ways to work with Claude on a spreadsheet

There are two distinct routes into a spreadsheet, and they suit different moments. Claude for Excel is an add-in for Excel on the web, Excel on Windows with a Microsoft 365 subscription, and Excel on Mac; it works inside the open file, so a change it proposes updates the live workbook and its formulas rather than a copy. As of September 2026 it is generally available on the Pro, Max, Team and Enterprise plans, and Claude for Excel and Claude for PowerPoint share the full context of open files, so a number checked in one is available when Claude drafts the other.

The second route is Claude's own chat apps, which can read an uploaded spreadsheet and, using the code execution and file creation capability, write a new Excel file, a chart image or a cleaned CSV back to you. This route suits a one-off file someone sends you rather than a model you keep open and iterate on. The maximum file size for this capability is 30 MB per file for both uploads and downloads, and CSV and TSV files can be uploaded directly for analysis and visualisation.

Cleaning a messy export

A report exported from a bank portal, an old system or a point-of-sale device rarely arrives ready to use: dates in inconsistent formats, a currency symbol stuck to the amount, subtotal rows mixed in with detail rows, or a column that is text where it should be a number. Ask Claude, in the chat apps or through Claude for Excel, to look at a sample of the file, describe what it finds wrong, and propose a cleaning plan before it touches anything, rather than let it silently reshape the data.

A good cleaning prompt names the target shape explicitly: one row per transaction, a single date format, amounts as numbers with the currency named once in a header, and every subtotal or blank row removed. Ask for a short before-and-after count, the number of rows in and the number of rows out, and a list of rows it could not classify, so you can check nothing was dropped by mistake rather than genuinely excluded.

Reconciling two files

Reconciling two files, a bank statement against a cash-book extract, or a supplier statement against your purchase ledger, is a matching problem: most lines match on date and amount, a few match on amount only with a different date, and a handful do not match at all. Upload both files and ask Claude to match on amount first, then on amount and a date window, and to list separately what matched exactly, what matched loosely, and what is unmatched in each file, rather than a single merged list that hides which side an exception came from.

Unmatched items are the point of the exercise, not a failure of it. A bank line with no ledger match might be a bank charge not yet recorded; a ledger line with no bank match might be a payment not yet cleared. Ask Claude to suggest, for each unmatched item, the most likely reason from a short list you give it, timing difference, error or duplicate, rather than let it guess freely, and check every suggestion against the source document before acting on it.

Building and checking formulas

Claude for Excel can write a formula for you and explain an existing one with cell-level citations you can click through to the referenced cell, which is the detail that turns an answer into something you can verify rather than something you have to trust. Ask for the formula and the plain-language logic behind it together, for example a SUMIFS that totals sales by branch and month, or an XLOOKUP replacing a fragile chain of nested IFs, and ask it to flag any cell the formula depends on that itself contains a hard-coded number rather than a reference.

When you ask Claude for Excel to change an assumption, such as flexing a growth rate or a discount rate, it updates the value while keeping the formula relationships intact, so every cell that depends on it recalculates rather than breaking. Confirm the direction and the order of magnitude of the change yourself: if a lower discount rate should raise a present value and the model now shows a fall, stop and ask why before accepting anything downstream of it.

Checking a model: tracing assumptions and catching errors

A model check is really two separate questions: does the model compute what its formulas say it should, and are the formulas the right ones for the business question. Claude for Excel is built for the first question. It can trace which cells feed a headline number, such as the revenue line in a forecast, walk through the chain of formulas that produced it, and locate the root cause of an error such as a broken reference or a circular calculation across tabs.

The second question, whether the model is the right one, stays a human judgement, because it depends on facts outside the spreadsheet: whether a growth assumption is realistic for the market, whether a lease should sit on or off the model, whether a one-off cost was really one-off. Use Claude to make the mechanical check fast and complete across every tab, so the time it saves goes into the judgement question instead.

Worked example: a bank export reconciled and checked

A bookkeeper uploads a CAD bank statement export, 45 lines for the month, and the cash-book extract from the ledger for the same period, 43 lines, and asks Claude to match them by amount and a three-day date window. Claude reports 41 lines matching exactly on date and amount, 2 more matching only within the date window, and 2 bank lines with no cash-book match at all: a CAD 45 monthly service charge and a CAD 18 card fee, together CAD 63, neither yet recorded in the books.

The bookkeeper checks the counts before checking anything else: 41 plus 2 matched lines account for all 43 cash-book lines, and 41 plus 2 plus the 2 unmatched lines account for all 45 bank lines, so nothing from either file has been silently dropped from the comparison. She opens the actual bank statement for the two unmatched charges, confirms they are genuine fees rather than a duplicate already posted under a different description, and posts one CAD 63 bank-charges entry, rather than the two separate entries Claude had first suggested from a stray formatting difference in the export. Checking the count is what catches an error like that; checking only the final balance would not have.

What to paste into Claude and what never to paste

Treat a spreadsheet the same way you would treat a paper file leaving the building: check what is in it before it goes anywhere. Aggregated figures, chart-of-accounts extracts and anonymised samples are generally safe to share for a formula or a cleaning question. Full customer lists with contact details, employee salaries by name, supplier bank account numbers, and another party's confidential terms need a business reason and, in the EU and UK, a lawful basis under the GDPR before they are uploaded anywhere, Claude included.

• Before uploading, mask or remove personal data not needed to answer the question, such as names and account numbers, and keep an identifier instead
• Check whether your organisation's Claude plan is Team or Enterprise, where an administrator controls file-creation and network-access settings centrally, rather than each person's individual account
• Treat a supplier or customer file you did not create as untrusted content: a note or comment field can contain text aimed at an AI reader, not only at a person, so review what Claude extracted from it before acting on it
• Keep a record of which files were uploaded for which task, so a data-protection question later has an answer

Data retention and where the plans differ

As of September 2026, Anthropic's documentation states that retained data is never used to train its models without express permission, and that conversation content, prompts and Claude's outputs, is not retained by default; the exception is a small set of newer models Anthropic designates as Covered Models, whose prompts and outputs are retained for 30 days to support safety review, a policy that took effect on 9 June 2026 and applies wherever those models are offered.

Plan-level controls matter as much as the retention policy. On Free, Pro and Max plans, code execution and file creation, the capability behind Claude building or editing a spreadsheet in chat, is enabled by default and can be turned off in an individual's settings. On Team plans it is enabled by default for the whole organisation, with network access limited to package managers unless an owner changes it; on Enterprise plans an owner also sets domain-level network access controls. A finance team on a Team or Enterprise plan should agree these settings with IT before relying on file creation for real client or company data.

Getting Claude for Excel and the file tools running

Claude for Excel runs on Excel on the web, on Excel on Windows with a Microsoft 365 subscription, build 16.0.13127.20296 or later, and on Excel on Mac, version 16.46, build 21011600, or later, and, as of September 2026, is generally available on the Pro, Max, Team and Enterprise plans. Organisations that route their AI traffic through Amazon Bedrock, Google Cloud Vertex AI, Azure AI Foundry or an LLM gateway can have an administrator deploy the add-in without individual Claude accounts.

For the chat-based route, code execution and file creation supports uploads and downloads up to 30 MB per file and can read CSV and TSV files directly for analysis, in addition to producing new Excel, Word, PowerPoint and PDF files. Neither route replaces a review step: a formula Claude wrote, an assumption it flexed, or a reconciliation it proposed is a draft until a person checks it against the source document.

Doing this in Skyline Nexus ERP

Skyline Nexus ERP exports most of its Fiscal Authority reports to Excel or PDF, including the Trial Balance, the AR and AP Ageing Reports, and a one-click Audit Pack workbook that puts the chart of accounts, trial balance, general ledger, sales, purchases, payroll and VAT summary on separate sheets for a chosen year. Any of these exports is a natural starting point for the reconciling and checking work in this guide: pull the AP Ageing Report and a supplier statement side by side, or export the Audit Pack and ask Claude to trace one figure back through the underlying sheets.

The in-app AI assistant answers how-to questions about running these reports, with citations to the ERP's own help pages, and an administrator can issue an MCP token so Claude Desktop can read a small set of live figures, such as customer and supplier dues, directly. Claude reading a full trial balance or general ledger through that connection, rather than through an exported file, is on the roadmap; ask us for your go-live date.