Make the task reversible

Never start by asking AI to “fix this spreadsheet” in place. Duplicate the file and record its row count, column names and a few known totals. Define the problem precisely: inconsistent dates, extra spaces, duplicate customer records or values in the wrong column. Decide which fields are identifiers; a leading zero in a product code may be meaningful even if software wants to treat it as a number. If the sheet contains personal or financial data, use an approved environment or anonymize a representative sample.

Ask the assistant for a proposed transformation plan before it changes anything. A useful request is: “List each cleaning rule, a before-and-after example, the number of rows it might affect and cases requiring manual review. Do not delete rows.” This separates an understandable rule from an opaque automation. Apply one transformation at a time, preferably with formulas or a script that can be rerun.

Check the exceptions

Start with whitespace and capitalization, then date formats, then duplicate detection. Treat duplicates as candidates, not automatic deletions. Two customers can share a name; two invoices can share an amount. Use stable IDs and compare multiple fields before merging. For dates, test ambiguous values such as 03/04/2026 under the correct locale. Keep an exception sheet with every record the system could not classify confidently.

After each step, compare row counts and control totals with the original. Sample changed rows from the beginning, middle and end of the file. If a tool outputs a new spreadsheet, verify that formulas, formatting and sheet references survived. An assistant can help write the cleaning logic, but a deterministic check should decide whether data was preserved.

Produce a handoff, not just a file

Save three artifacts: the untouched original, the cleaned copy and a short change log describing the rules and unresolved exceptions. Ask a second person to review a sample if the data feeds payroll, taxes or customer communications. The cleanest sheet is not necessarily the best sheet; the best one is accurate, traceable and reversible. AI accelerates the repetitive work when you keep the human definition of “correct” in control.