AI in Microsoft 365 and Google Workspace

In-suite copilots help where the file already lives

Copilots inside Microsoft 365 and Google Workspace help where the file already lives. Permissions and DLP still decide whether that is safe. Microsoft 365 Copilot can draft in Word, propose work in Excel, summarise Outlook and Teams, and reach files Graph allows the user to see. Gemini for Google Workspace can sit in Docs, Gmail, Sheets, and Meet with a similar promise. The attraction is obvious: less paste, less export, help next to the artefact. The risk is equally obvious. A copilot with broad search will treat a mis-shared drive as knowledge. Convenience does not rewrite access control. It rides it.

Names multiply and confuse buying. Copilot can mean GitHub, Windows, Bing, or Microsoft 365. Gemini can mean a consumer app or a Workspace add-on. Write the product you mean, the SKU, and the admin who configures it. A policy that says Copilot is allowed without that detail is how a consumer chat and a tenant copilot get the same blessing. They should not. The in-suite product is still a model. It will invent a clause, misread a table, and smooth dissent in a summary. The suite does not make fluency into evidence. It makes the file location saner if, and only if, sharing was already sane.

This essay is how to use those copilots as part of a stack, not as a magic layer over a messy tenant. You will match the job to the app, inspect permissions and DLP, keep a human on numbers and send, and refuse to export a confidential file into a consumer chat to see if it sounds nicer. You will also train people that tenant search is not corporate memory. It is their access plus everyone else's mistakes. Clean sharing is part of AI adoption. If that sentence surprises the room, start there before you buy more licences.

Stay in the app when the artefact and the class allow it

If the work object is a Word paper, an Excel model, a deck, or a mail thread in an approved tenant, prefer the in-app copilot over exporting. Export is how classification dies. If the work is a think pass on notes you are allowed to paste, an approved separate chat may be fine. If the work is code, use the coding copilot your engineering standard named. Write those defaults. People otherwise open whichever window is already in the browser. Browser habit is not a data-class decision. It is how a customer workbook leaves Microsoft 365 for a personal ChatGPT tab because someone wanted a rewrite.

Permissions and DLP are the real governors. Over-shared sites will feed the copilot. Labels that do not block will not save you. Test with a labelled file and a user who should not see it. If the copilot still summarises it, you have a sharing or product-configuration problem, not a training problem. Google and Microsoft both offer admin controls; neither is on by wish. Assign an owner for those controls the way you assign an owner for identity. A copilot licence without that owner is a new search box over the same mess.

Configure, default, then practise in the live file

Before training, confirm licences, identity, retention, DLP, and who can disable web grounding if that is a requirement. Write a one-page default: which app for which artefact, which labels are in scope, who reviews numbers. Train in the live file, not in a slide. Show a bad summary that dropped a footnote, and a good pass that asked for section pointers. Show the sharing pane. People should leave able to refuse a copilot action that would widen access. If they cannot find the sharing pane, they are not ready to summarise a site.

Measure time-to-trusted-output inside the suite, including the check against the file. Watch for duplicate export. Sample whether labelled files stayed in tenant. When Excel is involved, recalculate. When mail is involved, check that the draft did not accept a meeting or a discount. The copilot is a colleague with access, not a signer. Keep ChatGPT or Claude, if approved, for jobs the suite does not hold. Do not keep them as a second copy of the same confidential paper. One copy, one boundary, one review. If you need a second opinion, use an approved workspace on allowed notes, not a copy of the labelled file.

Role: You are an IT and operations lead setting defaults for in-suite AI.
Task: Draft a one-page rule for Microsoft 365 Copilot and Gemini in Google Workspace.
Context: I will paste which suite we use, data labels, and the artefacts we produce.
Constraints:
- Prefer staying in the file when the class allows it.
- Forbid export of confidential files to consumer chats.
- If sharing or DLP facts are missing, write missing and do not allow broad summarise.
- Name who owns admin controls and who signs output.
Output: Defaults by artefact, a sharing check, and a review rule for numbers and mail.
Quality checks: Where could this page still allow a copilot to draft from a mis-shared drive?

Put the one-page rule in the same place people open Word or Docs, not only in an intranet article. When someone exports a tenant file to compare tone, treat it as a classification miss and fix the default, not as a taste debate. Recheck sharing on the sites the copilot can search. The quality of in-suite AI is mostly the quality of your permissions. That work is unglamorous and it is the work that makes the licences usable.

Tenant-wide summarise, silent export, and the wrong Copilot

Tenant-wide summarise is the fantasy that the copilot knows the company. It knows what the account can read. If that is too much, the draft will be a leak in polite language. Silent export is the second risk: a download, a paste, a personal window, a nicer paragraph, and a second copy outside DLP. The wrong Copilot is the third: GitHub habits in an office file, or a consumer Gemini tab that people think is Workspace because the logo is familiar. Write the product names. Test the labels. Watch the download.

Over-trusting in-app formulas is a quieter miss. A suggested Excel formula still needs a person who understands the sheet. A Gmail draft still needs a person who knows whether the apology was authorised. Suite AI reduces friction. It does not reduce accountability. If your DLP cannot see copilot prompts, ask whether you are ready to put the labelled class in that product. Readiness is a control question, not a feeling that Microsoft or Google are big enough to be safe. Size of vendor is not a substitute for a setting you can show.

Related reading on StudyGrid

Read next: ChatGPT vs Microsoft Copilot at Work A Tool Stack That Fits What Not to Paste into ChatGPT. Those essays sit beside this one. Use them when you need the neighbouring skill, not as a substitute for the check you still have to make.

What to do this week

Name the SKU you actually have in Microsoft 365 or Google Workspace. Write which artefacts must stay in-app. Test one labelled file with a user who should not see it. Fix sharing if the copilot still reaches it. Ban export of that class to consumer chats. Practise one live rewrite in the file with the review rule next to it. Keep the licence for that job only if the check still happens.