Keep Your Books Readable With AI

Most small-business books are not wrong. They are unreadable. Four hundred lines a month with descriptions like SQ *THE COFFEE 4471, half of them landing in whatever category the bank feed guessed, and a profit figure nobody can explain without opening six screens.

AI is genuinely good at the readable part. Reading, summarising, a decent first pass at sorting. Your records are a separate thing and they stay exactly where they are. The US Internal Revenue Service puts the standard plainly on its recordkeeping page for small businesses: "You may choose any recordkeeping system suited to your business that clearly shows your income and expenses." And on what backs it up: "Purchases, sales, payroll, and other transactions you have in your business generate supporting documents. These documents contain information you need to record in your books." The same page states that records of employment taxes must be kept for at least four years, and that you keep records as long as needed to prove the income or deductions on a return.

Nothing below changes any of that. The documents stay. The model reads them.

Export the month, then ask three questions of it

Whatever you keep the books in, it exports a CSV. Pull one month of transactions with at least date, description, amount and category, and drop it into Claude.

Two settings first. Anthropic's page on creating and editing files with Claude states the ceiling as 30MB per file for both uploads and downloads, which no monthly transaction CSV will come close to. And if you want a workbook back rather than a wall of text, the capability has to be switched on: Free, Pro and Max users go to Settings > Capabilities and turn on Code execution and file creation; Team and Enterprise admins use Organization settings > Capabilities. The page says the feature is available across Free, Pro, Max, Team and Enterprise, and that Claude "can create Excel spreadsheets (.xlsx), PowerPoint presentations (.pptx), Word documents (.docx), and PDF files."

Then ask three questions, in this order, every month:

1. Total by category for this month. Next to each, the same category last month and the change as a percentage. Return it as an .xlsx with live formulas, not typed totals.

2. List every supplier or payee that appears this month and did not appear in either of the two previous months.

3. List every transaction that is uncategorised, categorised as Other, appears twice with the same amount and date, or is more than three times the median amount for its category.

Question one is the summary you wanted. Question two catches the subscription somebody signed up for. Question three is the one that finds actual errors, and it is the one to run first if you only run one.

The instruction about live formulas matters more than it looks. Ask for a total and you get a number the model wrote. Ask for a workbook with formulas and the arithmetic is done by Excel, where you can click a cell and see what went into it.

Categorise with a formula you can audit

For the sorting job itself, a spreadsheet formula beats a chat window, because every result lands in a cell next to the row it belongs to. Google documents the syntax as AI("prompt",[optional range]) on its page for using the AI function in Google Sheets, with published examples including =AI("Classify the sentiment of the realtor analysis.", D2).

Paste your uncategorised rows into Sheets and put this alongside the first one:

=AI("Classify this bank transaction description into exactly one of: materials, subcontractor, vehicle, software, insurance, professional fees, other. Answer with one word only.", B2)

Select the cells and click Generate and insert. Refresh and insert re-runs them if the underlying text changes. The constraints on that page, all relevant to a four-hundred-row month:

• It needs an eligible Google Workspace or Google AI plan.
• A maximum of 350 selected cells can generate at once, so a long month is two passes.
• Responses are text only.
• There is no undo. You regenerate instead.
• The function cannot access your entire spreadsheet or your Google Drive files, so it only sees the range you hand it.
• Generated content is attributed to you in version history.

That last point is the reason to prefer this over a chat window. Six months later, the sheet shows which cells were generated and when. A chat transcript does not sit inside your books.

Set the classification against a fixed list, always. An open-ended "categorise this" produces category names that drift month to month, and a category that exists in March and not in April makes your comparison useless.

The add-in that cites cells, and what it refuses to do

If the books already live in Excel, there is a route that skips the export entirely. Anthropic describes Claude for Excel on its Excel add-in help page as "an add-in that integrates Claude into your Excel workflow", built "for professionals who work extensively with spreadsheets, particularly in financial analysis and modeling".

Availability is stated as Pro, Max, Team and Enterprise plans, so a free account cannot use it. To install, the page says to go to the Claude for Microsoft 365 listing on Microsoft Marketplace, click Get it now, then open Excel, activate the add-in and sign in. For a team, an administrator goes to the Microsoft 365 Admin Center, Settings > Org Settings > User owned apps and services, confirms Let users access the Office Store is enabled, then deploys it from Settings > Integrated apps > Add-ins.

The capability that fits bookkeeping is being able to "ask questions about your workbook and get answers with cell-level citations", plus navigating multi-tab workbooks and debugging errors. A cited cell is checkable in four seconds.

The refusals are published and they are unusually direct for vendor documentation. Unsupported file types are listed as data tables, macros and VBA. And the page states it is "Not recommended for: Final client deliverables without human review" or for "Audit-critical calculations without verification". Anthropic also warns to "Only use Claude for Excel with trusted spreadsheets and not spreadsheets from external untrusted sources", because a spreadsheet can carry instructions aimed at the model. A workbook a supplier emailed you qualifies.

Read that second refusal slowly if you were planning to use this on year-end figures. The vendor has told you, in its own help centre, not to.

Two numbers no vendor publishes

How often the categorisation is right. Neither Google's AI function page nor Anthropic's Excel page publishes an accuracy rate for classification on financial data. There is no benchmark to quote, so there is no way to say "expect 94% and check the rest". The only defensible process is to check all of it the first month, count how many you changed, and use your own number to decide how much you check in month two.

How much you can actually run. Google's page tells you that long-term generation limits exist and that hitting one means waiting 24 hours, without publishing the threshold. Plan a busy month around an unpublished quota and you will meet it on the day the accounts are due.

What to do with all this. Run question three on last month tonight, before you set anything else up. It takes ten minutes and it is the fastest way to find out whether your books are merely unreadable or actually wrong. And whatever the model tells you about last month, the receipts and invoices behind those lines are still what a tax authority asks for. Keep them.