Audit what Copilot infers from your work data

Ask Copilot, ChatGPT and Claude to list what they know or have inferred about you, then log each item on one sheet as correct, wrong or shouldn't know, with its source. Delete what you can, then let the sheet decide: any shouldn't-know item you can't remove one by one means turning Copilot's work insights off.

Steps follow Microsoft's Message Center post MC1492154 and each assistant's memory help page, checked 12 October 2026. The Harbourline example is made up.

For Claude on its own, see Review and correct what Claude remembers about your work.

What's changing in Microsoft 365 Copilot in October 2026?

The Message Center post says Microsoft 365 Copilot will use inferences from your Microsoft 365 work data as context for more personalised answers. Microsoft calls these work insights. The post, published 8 October 2026, says the worldwide rollout begins in early October 2026 and should finish by late October 2026, so your tenant may not have it yet. The linked roadmap item, 551195, is tagged for the Worldwide (Standard Multi-Tenant) cloud and gives November 2026 as its public availability date, which is a month later than the post.

Microsoft's support page says work insights only apply when you're signed in with a work or school account, and depend on which Copilot plan your organisation has. Admins can switch the feature off for everyone with the enhancedPersonalizationSetting control, so if you can't find the toggle described below, ask your Microsoft 365 admin whether it's been turned off or hasn't reached you.

Where can you see what each assistant keeps?

Each assistant gives you a different view, and only some of them list every item. You'll work from whatever the settings show, plus each assistant's answer to the prompt further down.

Run every prompt signed in to the same account you use for work. A personal Copilot or ChatGPT account will tell you about the wrong memory.

• Copilot: in Copilot Chat, select ... Settings and more > Chat settings > Personalization. With a work account the toggle may read Chat history & work insights. Microsoft's pages describe the toggle and deletion but don't describe a list of individual work inferences to browse, so asking Copilot may be the only way to see them.
• ChatGPT: Settings > Personalization > Memory. OpenAI says controls vary by plan, region and workspace, and that the memory summary is a high-level view that may not include every detail or source.
• Claude: Settings > Memory lists everything under Topics, where you can edit or delete each one, according to Anthropic. On Team and Enterprise plans an owner has to make memory available first, and it's off for each member until they turn it on.

Fill in the sheet

Use a spreadsheet, a table in a doc or plain text. One row is one item from one assistant, using the five columns in the checklist on this page. Put all three assistants on the same sheet, so the decision at the end uses one pass.

Add a last line under each assistant: list complete or list may be incomplete. Mark it incomplete when the assistant said so, when there's no settings view to compare against, or when the settings view showed items the answer missed.

Then act on each wrong or shouldn't-know row where the tool lets you. In ChatGPT, OpenAI says you can correct or delete items in the memory summary or saved memories, and that deleting a chat doesn't always delete a saved memory taken from it. In Claude, delete or edit the topic in Settings > Memory. For Copilot, the only control Microsoft documents is the toggle, so any Copilot row you want gone goes in the action column as can't remove one by one.

• Assistant: Copilot, ChatGPT or Claude, plus the account it was signed in to.
• What it says it knows: the item in its own words.
• Source it gave: the source from the answer, or source unknown.
• Verdict: correct, wrong or shouldn't know. Shouldn't know means the item may be true but you don't want an assistant using it, such as health, pay, HR matters or something from a thread you weren't meant to see.
• Action taken: deleted, corrected, left, or can't remove one by one.

Example: one pass across Copilot, ChatGPT and Claude

A Harbourline delivery lead runs the prompt on 12 October 2026 and logs five rows.

Copilot (work account) | Leads the Harbourline cutover | Teams meetings | Correct | Left

Copilot (work account) | Reports to Nina Patel | Outlook email | Wrong | Can't remove one by one

Copilot (work account) | Has a hospital appointment on 20 October | Outlook calendar | Shouldn't know | Can't remove one by one

ChatGPT (work account) | Writes status updates in British English | Saved memory | Correct | Left

Claude (Team plan) | Is planning a move to a new role | Past chat | Shouldn't know | Deleted in Settings > Memory

Copilot's list is marked may be incomplete, because there was nothing to compare it with. The Claude row was removed one by one, so it doesn't count against Claude. The Copilot appointment row can't be removed on its own, so the rule below turns Copilot's work insights off. ChatGPT has no wrong or shouldn't-know rows, so its memory stays on.

Keep personalisation on or turn it off?

Decide per assistant, using only the rows on the sheet.

Otherwise keep it on, and run the same prompt again after a month, or sooner if Copilot's rollout reaches your tenant after this audit.

A list marked may be incomplete doesn't count as a clean result. It means the decision rests on the rows you could see, so rerun the prompt sooner.

If you turn Copilot's toggle off, leave it off for at least 30 days. Microsoft says the inferences are deleted 30 days after you turn it off, and come back if you turn it on again within that window. To check the change took, open a new Copilot chat after 30 days, run the prompt again and look for the rows you marked shouldn't know.

• Turn it off if any shouldn't-know row is left that you couldn't delete one by one. For Copilot that means turning off Chat history & work insights.
• Turn it off if wrong rows still outnumber correct ones after you've corrected what the tool allows.

Sources

• MC1492154, Enhanced M365 Copilot Memory: Copilot is now personalized with work data, Microsoft 365 Message Center, published 8 October 2026, checked 12 October 2026.
• Microsoft 365 roadmap item 551195, checked 12 October 2026.
• How Microsoft Copilot Chat history works, Microsoft Support, checked 12 October 2026.
• Memory in ChatGPT, OpenAI Help Center, checked 12 October 2026.
• Use Claude's chat search and memory to build on previous context, Claude Help Center, checked 12 October 2026.
• Related: Review and correct what Claude remembers about your work.