On 27 October 2026, Microsoft announced a data-cleaning skill for Copilot in Excel. It checks for issues including stray spaces, inconsistent dates, duplicate values and numbers stored as text, then proposes changes for you to review. Microsoft says it is rolling out to general availability across Windows, Mac and the web.
For an operations team preparing a sales report or combining supplier exports, that is worth a look. I would start with a modest question: can it help someone find problems they would otherwise have to hunt for by hand?
The important decision is what counts as a problem. A repeated reference might be a mistake, or it might identify separate lines on the same order. Tidying the appearance of a spreadsheet and preserving its meaning are different jobs. I would keep the person who understands the records in charge of accepting changes.
Start with a copy and a question
Choose a spreadsheet whose purpose you can explain in a sentence. A monthly order report is a clearer starting point than a workbook that feeds several departments and contains years of unexplained formulas.
Use a separate copy for the trial and keep the original export unchanged. Write down where it came from, the reporting period and what the finished report is meant to answer. If you cannot establish those basics, ask the person responsible for the source system before cleaning anything.
Check with whoever manages Microsoft 365 that the feature is available under your business’s licence and settings. A rollout announcement is a reason to check, rather than a reason to assume everyone already has it. Follow your existing rules about which business data belongs in approved AI tools.
Then make the request specific. For example, ask Copilot to identify inconsistent dates and explain its proposed corrections before you approve them. Keep the first trial narrow enough that you can inspect the result yourself.
Decide the rules before accepting changes
I would ask the person responsible for the report to write down what each important column means. You do not need a technical specification. You need enough detail to distinguish an error from a valid record.
Consider these questions before approving a suggested fix:
- What makes a row unique?An order number alone may be insufficient if each product has its own line. Agree which fields identify a separate record before removing duplicates.
- Which values must stay as text?Treat account codes, product references and identifiers as labels unless the source system says otherwise. Check whether leading zeroes matter before changing their format.
- How should dates be read?Establish the source system's date convention. Where a date could mean different things, compare it with the original record rather than guessing from nearby rows.
- What does a blank mean?It could mean unknown, not applicable or not yet entered. Do not turn it into zero merely to make a calculation easier.
These are review questions, not a claim that Copilot will make those mistakes. My point is that a proposed correction needs a business rule behind it. If nobody can explain the rule, leave that value unchanged and resolve the question with its owner.
Check the report still answers the same question
Before accepting changes, note the checks you will use afterwards. For an order report, I would compare the number of rows, the total order value and a selection of records against the untouched export. Choose checks that fit the workbook; a contact list needs different checks from a financial report.
Totals alone are insufficient. Matching totals would not tell you whether a customer reference had lost its leading zeroes or a date had been interpreted incorrectly. Inspect the columns you changed, including unusual entries and records that were flagged as duplicates.
Microsoft describes the feature as proposing corrections that users can review and choose to apply. Use that review stage deliberately. Accept a change because you can explain why it is right, and retain a short record of the rules you used. Ask a colleague to check anything that will affect an invoice, a customer commitment or a management decision.
Use the trial to improve the next export
If the same issue appears in every export, I would look upstream. Ask whether the source system can produce the required format, or whether the reporting process needs a documented conversion. Cleaning the same column repeatedly should prompt a conversation about where the inconsistency starts.
For a trial during November, pick one report, name its owner and agree the checks before you begin. Afterwards, ask whether the suggestions were useful, how much review they needed and whether the report was easier to prepare correctly. Include checking time in that judgement.
You do not need to expand the trial just because the feature exists. Keep using it where a person can verify the changes and explain the result. Where the meaning is unclear, the next useful step is a conversation with the person who owns the data.