How to let an AI agent work in your finance spreadsheets

Let the agent read your finance spreadsheets, work in its own copy, and hand back proposed changes as a list a person can accept cell by cell. Writing to the master file is held for its owner. Treat any figure that came from a spreadsheet as declared, not sealed, until it is checked against the ledger.

Which spreadsheets, and who owns them?

Finance runs on spreadsheets that no system replaces: reconciliation workbooks, accrual calculations, budget templates, forecast models, board pack backing schedules, tax computations. Each has an owner who built it or inherited it. Many have no version control beyond the file name, and some feed postings or payments directly.

The owner is accountable for what the spreadsheet says. The controller is accountable for what gets posted from it. An agent working in these files can save hours: tracing formulas, checking ties, rolling a model forward, drafting a new schedule. It can also break a model nobody fully understands. The answer is not to keep the agent out but to change where it writes.

Recurring refreshes of fixed reporting workbooks are a narrower case with tighter locks, covered in how to let an AI agent update spreadsheets used for reporting.

Which capabilities do you grant, and at what depth?

1. Agent, depth 1: reads workbooks in the finance file store. Within the folders the owner grants, never the whole drive.
2. Agent, depth 1: reviews. Traces formulas, finds hard-coded numbers in formula areas, broken links, hidden rows, totals that do not tie, and inconsistent ranges.
3. Agent, depth 2: works in a copy. Rolls a schedule forward, adds a month, builds a new tab, drafts a reconciliation. The copy is clearly named and stored apart from the master.
4. Agent, depth 2: proposes changes to the master as a diff. Cell address, current value or formula, proposed value or formula, and the reason. The owner accepts or rejects each line.
5. Agent, held: any direct write to a master file. Only the owner can apply the accepted diff, or let the agent apply it once accepted.
6. Agent, refused: enabling or running macros, changing sharing, sending files outside the finance team, deleting files.

Why are spreadsheet figures only "declared"?

A spreadsheet cell holds whatever someone typed or calculated. Nothing in the file proves it matches the ledger, the bank or the contract. When the agent reads an accrual schedule, it is reading the preparer's claim. That is fine for drafting and review. It is not fine as the basis for posting a journal or preparing a payment.

So set a rule in the agent's bounds: a spreadsheet value may support a draft, but any action at depth 3 or above must rest on the system of record. If an accrual schedule says 12,000 and the agent parks a journal for 12,000, the reviewer sees that the amount is declared and looks at the support. Where a figure appears in both the spreadsheet and the ledger, the agent checks they agree and reports any difference.

How do held writes, revocation and the record work?

A held write is a diff waiting for the owner. It lists every cell, its reason and the effect on key outputs. The owner can accept some lines and reject others.

Revocation is by folder and by capability. The owner can remove a workbook from scope, and the controller can withdraw the "work in copy" capability across finance if, for example, copies are piling up and confusing people. A failed certification run drops the agent to reading only.

Each session is filed: files opened, copies made, diffs proposed, lines accepted and by whom. That record answers the audit question that spreadsheets usually cannot: who changed this cell and why.

Which controls matter, and what evidence does each leave?

Copy naming and location. Copies sit in a separate folder with an agent prefix. Evidence: file paths on the session record.

Diff acceptance. Evidence: accepted and rejected lines with the owner's identity.

Formula integrity scan. Run on every master before and after changes. Evidence: scan results.

Declared source flag. Evidence: source class on any draft that used a spreadsheet figure.

What should you measure?

• Errors found in master files, by type. These often justify the work on their own.
• Diff lines accepted unchanged, as a sign of draft quality.
• Orphan copies older than a set number of days.
• Spreadsheet figures that disagreed with the ledger.

Where does it usually go wrong?

Copies become masters. A manager opens the agent's copy, likes it and starts using it. Now two versions exist. Mark copies clearly in the file and on every tab.

Macros in old models. A legacy model only calculates with macros enabled. The agent cannot run it, so it guesses. It should say that it cannot calculate and stop.

Too much read access. The agent is given the whole finance drive, including HR files and board papers. Grant folders by job.

Costs misjudged. Spreadsheet work is open-ended and can run up usage. Set a budget per session; how much an AI finance agent costs covers the drivers.

Where Pryme fits

In Pryme, reading, working in copies and proposing diffs are separate capabilities on the finance agent's depth matrix, each certified on your own files. Spreadsheet values reach the agent as declared under the Pryme Context Protocol, so a draft shows when it rests on a spreadsheet rather than the ledger. The wider limits on which finance tasks an agent can do on its own are part of our guide to AI agents for finance.