AI Spreadsheet Changes Need a Review Gate Before Board Reporting

Tools & Technical Tutorials

27 October 2026 | By Ashley Marshall

Your next spreadsheet error may arrive wrapped in a confident explanation. The question is whether anyone can prove what changed before that number reaches the board.

The spreadsheet has become an AI workflow

Microsoft now describes Copilot in Excel as a tool for suggesting formulas, cleaning data, analysing trends, creating charts and summarising workbooks. That is useful, but it changes the status of a spreadsheet. A workbook that used to be edited by one analyst can now be altered through a conversational instruction, copied into a report and circulated before anyone has inspected the calculation behind it. The business risk is not that AI is uniquely bad at arithmetic. It is that a plausible-looking summary can hide a changed range, a missing row or an old assumption. Microsoft's August 2026 Excel update also introduced a change history skill that can summarise who made edits and how people and AI modified a workbook. That is a practical signal: provenance is now part of the product experience, not an optional audit exercise. For UK firms, the first question is which workbooks influence invoices, cash decisions, customer outcomes or board papers. Inventory lists and informal planning sheets can tolerate a lighter process. Payroll reconciliations, revenue forecasts and regulatory returns cannot. Create a register of decision-critical workbooks, with an owner and a named reviewer for each. Record where the raw data came from, which sheet is the approved output and when it was last checked. This need not become a vast governance programme. A single page per critical workbook can prevent a team from treating a chat response as an authorised financial result. The operating principle is simple: let AI draft an analysis, but make a person accountable for the number that leaves the workbook. Microsoft's Excel guide itself advises checking formula references, data-cleaning changes and summaries against source data. That advice becomes more important, not less, when the interface makes edits feel effortless.

Start with a frozen baseline and a change log

Before an assistant touches a critical workbook, save a dated baseline with known totals. Those totals should include the figures a reviewer can independently reproduce: row count, total revenue, total cost, key balances and the number of excluded records. Then record the request made to the assistant, the proposed change and the cells or tables affected. This gives the reviewer a before-and-after comparison rather than a memory test. Microsoft's August 2026 release says its change history skill can identify who made edits and can help restore formulas to a previous working state. That can support investigation, but do not mistake a generated summary for your sole audit record. Keep the workbook's native version history and the source extract as well. A change log should answer four plain questions: what changed, who requested it, who checked it and what evidence proves the output stayed right. For example, if Copilot removes duplicates from a customer ledger, ask it to list the removed rows, then compare that list with an agreed definition of duplicate. Two customers sharing a name are not necessarily the same customer. If it changes a formula, compare the formula's range with the original specification and test the result on deliberately awkward rows. Set a rule that AI can propose edits in a working copy, while the approved reporting workbook is updated only after review. In practice, this prevents a draft analysis from silently becoming the board number. Teams often object that version history already exists. It does, but a reviewer still needs a clear decision trail showing why an edit was accepted. The combination of baseline, change record and approval gives that trail without requiring specialist software.

Test the edge cases, not just the headline total

A workbook can pass a quick glance and still fail exactly where the business is exposed. An AI-generated formula might omit the last row of a growing table, treat a blank as zero, combine net and gross values or count a cancelled invoice. Test these cases deliberately. Build a small set of reference rows with expected outputs: a normal transaction, a blank, a negative adjustment, a duplicate, a late date and a boundary value. Keep the expected answers outside the AI-edited workbook. Run the same rows before and after the change. For a sales forecast, compare both the total and the distribution across months, regions or product groups. A stable grand total can conceal a category shift that changes a staffing or stock decision. Microsoft says its Excel assistant can suggest formulas, explain error codes and generate summaries, but its guidance also tells users to read suggestions and confirm cell references. That is the right standard. An assistant may help create the tests, but it should not mark its own output correct. Give the human reviewer a one-page checklist and enough time to inspect exceptions. A lightweight rule might require independent recalculation of two material figures and a review of every changed formula that feeds a board chart. If that sounds slower than clicking Accept, compare it with the time needed to retract a wrong client invoice or explain a misstated forecast. The point is not to ban AI from spreadsheets. It is to spend review effort where an error would change a decision. Low-risk exploratory analysis can stay fast; decision-critical calculations deserve a release test.

Keep sensitive data within an approved route

The spreadsheet is also a data container. It may hold names, salaries, account details, support notes or health information, even when the task sounds as harmless as summarising a table. Before staff paste workbook contents into a general chat tool, identify whether the approved Excel environment already offers the capability and what data the assistant will process. The ICO's AI and data protection guidance says there is no one-size-fits-all security measure: controls depend on the nature of the data and the risks to individuals. It also asks technical teams to document movements and storage of personal data and to delete intermediate files when they are no longer required. Its guidance is under review following the Data (Use and Access) Act, so teams should check the current ICO position when setting policy. In practice, start with a data map. Which columns are required for the analysis? Could direct identifiers be removed or replaced with pseudonymous IDs? Who can access the workbook and its version history? Does the assistant send data to a processor covered by the organisation's existing contract? Make these questions part of the workbook approval, not a separate compliance form completed after rollout. If a task involves only public or synthetic data, the review can be short. If it involves employee pay or customer records, require the data owner to approve the route and record any sharing with third parties. The common misconception is that a familiar Excel interface makes the data transfer risk disappear. It does not. The relevant boundary is where the data goes and who can see it, not whether the prompt was typed inside a workbook.

Plan for features to change or disappear

The AI feature itself can be transient. Microsoft's own post about the in-cell =COPILOT function says it became unavailable from 14 September 2026 and directs users to the Copilot side pane for many of the same tasks. That is a concrete reminder not to build an enduring finance process around an experimental formula or interface without a fallback. The same post described 100 calls every 10 minutes and up to 300 per hour for its original preview, figures that illustrate how usage limits can shape a workbook design, but should not be treated as current entitlements. If a workflow depends on an AI function, list that dependency alongside ordinary external links and macros. Test what happens if the function is removed, a licence changes or a workbook opens under another account. Can the reporting team reproduce the essential figure with standard formulas or a documented manual step? For a critical workbook, retain the source data, the accepted values and the method used to calculate them. Do not allow an unverified generated text result to become the only explanation for a board metric. A simple continuity drill is valuable: ask another analyst to reproduce last month's output without using the original author's chat history. If they cannot, the process is not yet reliable. The counterargument is that AI tools evolve so quickly that documenting each workflow is wasted effort. The opposite is more persuasive for important decisions: rapid product change is precisely why the business needs a stable specification and a repeatable test. Document the outcome required, not every button in the user interface.

Make approval proportionate to the decision

A sensible policy distinguishes exploration from release. Analysts should be free to ask an assistant for a draft chart, a suggested formula or a plain-English explanation in an approved environment. Approval is required when a result will be used to pay someone, change a customer record, file a return or advise the board. Define the boundary in terms of consequence, not in terms of whether an AI button was pressed. For a weekly operational report, the owner might check row counts, two sample calculations and any anomalous movements. For a statutory or investor-facing figure, require a second reviewer and a retained evidence pack. Put the checks into the existing month-end or management reporting routine so they are actually done. Measure three things for the first month: the number of AI-assisted changes, the proportion reviewed before circulation and the number of material corrections caught. This gives leaders a clearer signal than asking how many employees used Copilot. Include the time spent reviewing and fixing errors in any productivity claim. A faster first draft is not a saving if the finance team spends longer rebuilding confidence in it. Microsoft's release notes show change history and chat history becoming more visible in Excel, but those features support a control process; they do not replace ownership. In practice, appoint one accountable owner for each critical workbook, give reviewers a short exception-based checklist and keep a route to roll back a bad edit. The business gets the benefit of faster analysis while preserving the thing that matters most: a defensible decision based on numbers someone has checked.

Frequently Asked Questions

Should we ban Copilot from financial spreadsheets?

No. Use it to draft and explain, but require review before material figures are released.

Which workbooks need the strongest controls?

Those affecting payments, customer outcomes, regulatory filings, forecasts or board decisions.

Is Excel version history enough for an audit?

It helps show changes, but you also need a record of why a change was approved and how it was tested.

What should a reviewer check first?

Compare source row counts and key totals, then inspect changed formulas and known edge cases.

Can AI validate its own spreadsheet output?

It can suggest tests, but an independent expected result and a human reviewer should decide whether the output is correct.

What if a workbook contains personal data?

Use an approved processing route, minimise unnecessary fields and document where the data moves.

Does a Microsoft 365 licence settle the data protection question?

No. Check the actual processing route, supplier terms, access controls and the nature of the data.

What if an AI spreadsheet feature is withdrawn?

Keep source data, a documented calculation and a manual or conventional-formula fallback.