Guide

How to Use Copilot in Excel Safely With Sensitive Data

Published 2026-09-01 · 5 min read

You open a workbook with last quarter's churn numbers, real account names, real revenue figures, and ask Copilot to build a summary pivot. It answers in a couple of seconds, reading straight from the cells you already have open. Nothing gets uploaded to a chat window first. That's the part people miss when they try to apply the same anonymize-before-you-paste habit they've built for ChatGPT.

Copilot in Excel and Claude for Excel don't work like a chat window. There's no separate step where you paste text and could swap in a masked version instead. The add-in reads the live sheet through its own connection, and that connection is a fixed part of how the feature works.

Why you can't mask the data in flight

The obvious fix people ask for is intercepting the traffic between Excel and the vendor's API and masking it on the way out. We looked at building that into ShiftMask and decided against it. It's not that it's impossible in a lab — it's that it would break every time Microsoft or Anthropic touched their API, and it puts you in the position of quietly proxying network traffic you don't control, on a connection whose terms you agreed to unmodified. That's not a trade we're willing to make, and we wouldn't trust a vendor who made it either.

The workflow that actually works

1. Make an anonymized copy of the workbook with ShiftMask — same tabs, same column headers, same relationships between sheets, none of the real values.
2. Open the masked copy and let Copilot or Claude for Excel do its normal work: pivot tables, formulas, cleanup, chart suggestions.
3. Take the formula or the calculated column it wrote — not the specific numbers it typed into a summary cell.
4. Paste that formula into the real workbook. Because column names, data types, and table relationships are identical between the masked copy and the original, it runs without edits.

This works because a formula references structure, not content. =SUMIFS(Sales[Amount],Sales[Region],"West") doesn't care whether the amounts are real or masked — it cares that a column called Amount and a column called Region exist. Ask Copilot to compute something, and the masked copy and the real file give you the identical formula.

What breaks the pattern

Ask Copilot to interpret rather than compute — "which customer is our biggest churn risk, and why" — and the answer names specific rows that only exist in the masked copy. You can't check "Customer_00231" against the real business without resolving it back to a real name first. That's a different problem from formula reuse, and it's covered in a separate guide on preparing data for chat-based analysis.

When it's fine to skip masking entirely

If a sheet has no customer names, no employee records, and nothing you'd hesitate to put in a Slack message to the whole company, masking it first is wasted effort. Internal budget scenarios, template files, and synthetic planning numbers don't need a step that exists specifically to protect data you don't have in the sheet.

FAQ

Does Copilot or Claude for Excel use my data to train their models?

Training-data policies differ by plan and change over time — a commercial Microsoft 365 tenant has different data commitments than a personal subscription, and Anthropic's terms differ between individual accounts and Claude for Excel used under a business agreement. Check the current policy for your specific plan rather than assuming either way; that's exactly the kind of detail worth confirming before a sensitive workbook is ever open in the add-in.

Does masking the data change how well Copilot understands my formulas?

No. Formula suggestions depend on structure — headers, data types, key relationships — not on the literal content of cells. If ShiftMask preserves the shape of the workbook, formula suggestions come out the same either way. The one exception: open-ended prompts like "write a summary paragraph for the exec team" will read differently against placeholder values than against real ones. Generate that kind of narrative summary against the real file directly, using the same formula or prompt — not against the masked copy.

Can I use ShiftMask output with the chat-based Claude or ChatGPT instead of the Excel add-in?

Yes — that's actually the more common use case, and it's covered in a separate guide on preparing data for chat-based AI analysis.