Each Microsoft 365 app has one habit worth learning and one place it needs a human check, and a simple rule tells you when to use Copilot, when to use Claude, and when neither settles the question.
You have Copilot across the Microsoft 365 apps but no sense of which habit to build in which app, and you are being asked whether to use it or a separate assistant such as Claude. A feature tour does not answer either question. A habit per app, and a rule for choosing tools, does.
The notes below come from a few months of daily use, described as of September 2026. Tool behaviour changes quickly, so test each habit against your current version.
After a week of travel an inbox reached three hundred unread messages. An AI pass of about twenty minutes (the source does not say which tool) sorted subject lines and senders into "needs a reply today", "can wait" and "doesn't need one", drafted the easy replies, and set the two or three hard messages aside, marked. The point is that an inbox needs sorting first so the order becomes obvious. Anything with real stakes, such as an unhappy client or a commitment of money or time, is read in full.
Inside Outlook, the habits that paid back fastest:
Tone was the weak point. A reply that read as professional to Copilot sometimes read as cold to the recipient, so every external email got a human tone check.
Ask for decisions and who owns what next, not a transcript summary. Ask whether your name came up and why. Follow up with a specific question ("what did they decide about the budget") instead of scrubbing the recording.
Check anything that sounded like a joke. Sarcasm and side comments that carried meaning in the room were sometimes flattened into neutral statements.
In Word, give Copilot something to react to, ask for structure before prose, and use it to rewrite. In PowerPoint, the same logic applies:
Data slides were the weak point. Copilot formats a chart competently but cannot tell which number matters to this audience. That judgment stays human.
Multi-sheet models with cross-references and named ranges confused it more than a flat table. Any formula touching more than one sheet got a manual check, every time.
Mamdouh reports that a half-day pack became about ninety minutes with the same order each time:
A relationship detail, such as an offhand remark a major donor made, was added by hand.
| Situation | Choice |
|---|---|
| Work living inside a Word file, Excel model or email thread | Copilot, because you never leave the document |
| Open-ended questions, or a document needing real restructuring | Claude, whose longer responses held up better |
| Either could do it | Whichever already holds the context |
| A hard judgment call with no clean factual answer | Neither: asking a second AI did not help |
Switching tools mid-task to chase a marginally better answer cost more than it saved. Both tools sound confident either way, and confidence is not evidence.
Teams get value from Microsoft 365 Copilot when they learn to name the audience, goal and format, treat the first output as a draft, know its failure modes, and keep a hard line on sensitive data.
Team AI Training · 5 minFramework · 29 September 2026Three questions about the task in front of you pick between Copilot, Claude, ChatGPT and Gemini faster than any feature chart, and a team should settle its workflow before debating which tool to standardise on.
Team AI Training · 5 minExplainer · 11 September 2026A plain list of what a language model cannot do, from not knowing your organisation to not being accountable, so staff can judge each task rather than rely on belief.
Team AI Training · 4 minIf this is the question on your desk, a thirty-minute call tells you whether the service fits, or that you do not need us yet.