Practical knowledge. Smarter decisions.AR  ·  EN  ·  FR

Using AI with Excel without exposing client data: redaction, testing and reconciliation

A practical workflow for identifying sensitive workbook content, building synthetic samples and verifying every transformation.

AI assistants can explain Excel formulas and suggest Power Query transformations, but their usefulness is not permission to upload a client's workbook. This guide was reviewed on 28 October 2026 against Microsoft guidance on hidden workbook information, the UK Information Commissioner's Office guidance on pseudonymisation, and NIST's generative-AI risk framework. The central method is to classify the information, work with synthetic examples wherever possible, and independently reconcile the result. Replacing a person's name with a code does not necessarily anonymise the underlying record, and a polished AI explanation is not evidence that a calculation is correct.

Two colleagues discussing information displayed on a laptop at an office desk
Real editorial photograph — Vitaly Gariev / Unsplash. Illustrative photograph, not live market data or a product endorsement.

Classify the workbook before opening an AI tool

Start with a field inventory rather than a prompt. Record who owns the workbook, why it exists, which sheets and connections it contains, and which columns hold names, contact details, account numbers, salaries, addresses or invoice references. Distinguish personal data from commercially confidential figures and from summaries genuinely authorised for distribution. A small company's sales figures may be sensitive even if no individual appears. A contract or employer policy that prohibits external disclosure still applies when an AI service offers an easy upload button.

Write down the precise task: detect duplicate invoices, explain a formula, reconcile two totals or design a repeatable transformation. Many such questions can be answered using column names and ten invented rows instead of the real file. If production processing is unavoidable, use an organisation-approved environment with documented access rights, retention controls and an appropriate data-processing arrangement. Do not infer confidentiality guarantees from a marketing label or assume every consumer and enterprise product uses identical data handling.

Pseudonymisation does not erase re-identification risk

Replacing a customer's name with C001 while keeping a separate lookup table is pseudonymisation, not necessarily irreversible anonymisation. The ICO explains that additional information and the possibility of linking records matter when assessing whether people can still be identified. A combination of city, age and transaction date may reveal an individual even when direct identifiers disappear. Review indirect identifiers, rare values and unique combinations, and protect any lookup key separately with restricted access.

For a safe demonstration, create a new workbook containing invented customers, fictional amounts and dates unrelated to actual transactions. Changing five names in a copy of a live customer database is not enough: comments, references, hidden sheets and identifiers may survive. When the question only requires monthly trends, consider aggregates rather than person-level rows, while checking whether very small groups could still reveal someone. The appropriate protection depends on the context and likely attackers, not on a universal masking recipe.

Inspect information that a clean-looking sheet hides

Microsoft recommends running Document Inspector on a copy before sharing a workbook. Excel files can contain hidden sheets and rows, comments, author metadata, external links, embedded objects and cached information associated with analytical features. In supported desktop versions, inspect File, Info, Check for Issues, then Inspect Document; examine each reported category before removing anything. The warning matters: deleting hidden material can alter calculations or references, and removal may not be reversible. Keep the original intact.

Inspect queries, connections, named ranges and PivotTables as well as the visible worksheet. A blank-looking tab does not prove that confidential content is absent from a cache or embedded file. Save a dedicated sharing copy under a new name, reopen it, and test its essential calculations. If an external connection still points to an internal server or you cannot predict the consequences of removing hidden content, stop and request technical review rather than producing a misleading or damaged financial record.

Ask AI for a method, then test the method yourself

An assistant can propose a SUMIFS formula or a deduplication routine. Give it a synthetic test set that deliberately includes a duplicate, a missing date, a negative amount and two similar-looking identifiers. Specify data types, matching rules and expected outputs. Test any generated Power Query code on a tiny example whose answer you already know. Then challenge it with blanks, currencies, changing date formats and repeated amounts. Plausible code is not the same as validated code.

Microsoft's Power Query profiling tools report column quality and value distributions, but the documentation notes that profiling may default to the first 1,000 rows. Switch to the entire data set when the task requires complete coverage, or errors near the end may be overlooked. Log row counts before and after each transformation, distinct key counts and amount totals by currency. Do not remove legitimate transactions merely because they share an amount or supplier name.

A reconciled example and a release checklist

Consider a fictional workbook with 1,000 invoice rows totalling 250,000 currency units. A controlled test identifies ten genuinely duplicated records totalling 3,000. Removing those records should leave 990 rows and a total of 247,000. These figures are an illustration, not measured client results. The duplicate key must be explicit, such as invoice number plus supplier plus date, and exceptions should be documented. If the AI output disagrees with the independently calculated total, investigate the affected records instead of forcing an unexplained balancing entry.

Before releasing the result, verify data-owner authorisation, the absence of sensitive identifiers in the sample, hidden-content inspection, a recorded transformation sequence, reconciliation of counts and totals, an independent human review, and controlled storage and deletion. Preserve a test log containing prompts, expected outputs and discovered errors without copying secrets into that log. NIST's generative-AI profile emphasises structured risk management and ongoing evaluation. Sometimes the correct outcome is not to upload the workbook at all.

Official resources

Try a practical free calculator or builder ↗
Important notice

No online income or trading system guarantees profits. Results depend on skills, costs and circumstances.

Disclaimer ↗

Search the library