Using AI inside Excel, Google Sheets, Outlook, Word and PowerPoint.
In Word, Excel and PowerPoint the check runs at file open, not continuously, so a label applied mid-session doesn’t block anything until you reopen the file. Plus what the coverage does and doesn’t include.
Twenty source items total, and images, recordings and OneNote still aren’t readable. Plus the permission trimming that makes one agent answer the same question differently depending only on who’s asking.
A named action item is the highest-cost place for a wrong attribution to survive unchecked. Two documented causes, one hard 20-person ceiling, and the checklist to run before a recap goes to people who’ll act on it.
There are three separate clocks, not one: a 120 day default for recordings and transcripts, a shorter 60 day window for audio recaps, and a recycle bin period on top of that. Changing the policy today does nothing for a recording that’s already 119 days old.
Copilot won’t work in a meeting hosted by another organisation, no matter your own licence. The three states of one setting decide everything else, and most attendees never check which one is active.
The template is the foundation Copilot learns from, not the prompt. Two documented limits that break a real workflow, and the checklist that works around both before a deck goes in front of anyone.
A better prompt does not fix content Copilot under-weighted because it sat in the middle of a long file. Three documented causes for one symptom, checked in the order that actually rules them out.
A scanned page with no text layer is read as nothing, not as an error. Here are the four documented reasons Acrobat’s AI Assistant refuses a file, and what actually fixes each one.
The setting has to be on before the edit, not after. Compare is the only tool that catches an AI edit made before Track Changes was switched on, because it reads two saved files rather than an edit history that may never have started.
A comment can be deleted if Copilot rewrites the paragraph it’s anchored to. This is documented Microsoft behaviour, not an edge case, and it’s one of six stated exclusions that decide whether you can rely on the feature at all.