How to use Microsoft Copilot well

What Microsoft Copilot is shaped for

Work inside Microsoft 365 — documents, spreadsheets, mail, and meetings you already have.

• Reads your organisation's own content under your existing permissions
• Works directly inside Word, Excel, Outlook, and Teams
• Turns a meeting transcript into concrete actions and owners
• Strong at your own numbers when it can see the actual spreadsheet

Where it gets its information

Its distinguishing input is your organisation's content, under your existing permissions. It can only see what your account can already see — which is both the safety property and the main source of confusion.

How to prompt it well

• Use it inside the app, not beside it. In Word or Excel it has the document; in the standalone chat it does not. Most disappointing answers trace to this.
• Reference documents and people by exact name, since it resolves against your tenant's files and directory.
• In spreadsheets, ask for the method, not just the number — "show the formula you'd use" is checkable in a way a bare figure is not.
• Ask meeting questions concretely: "what was I asked to do and by when?" is more useful than "summarise the meeting."

The fundamentals — goal, context, constraints, format — work the same everywhere. Build one in the prompt lab →

Giving it context

• Permissions are the boundary — if you cannot open a file, neither can it.
• Say which document you mean when several have similar names.
• A long thread mixing several documents produces blended answers; keep tasks separate.

Getting the output format you want

• Ask for output in the format of the app you're already in.
• For mail, ask for the draft only, without commentary around it.
• Specify tone explicitly for anything customer-facing; its default register is corporate-neutral.

Files, documents, and research

Best on your own organisation's material: summarising a long internal document, pulling actions out of a meeting, working across a spreadsheet. Verify numbers it produces from spreadsheets by checking one case by hand — a plausible wrong total is the characteristic failure.

Example prompts

Draft a reply to this email. Accept the first request, decline the second with a reason, propose Thursday for the third. Keep it under 120 words.

Across our team's documents from this quarter, where has this topic been discussed, and what was decided?

In this spreadsheet, write the formula that flags rows where the date is after the deadline column, and explain it.

Using our internal onboarding docs, explain how this process works here, and tell me what the docs don't cover.

Summarise this 40-page document into a one-page brief for someone who has to make a decision. Include what's unresolved.

Based on the notes in this document, what options haven't we considered?

From today's meetings, list everything assigned to me with deadlines.

Common mistakes

• Using the standalone chat and wondering why it lacks context the in-app version would have had.
• Expecting it to see a file you have no permission for.
• Trusting a spreadsheet total without checking one row by hand.
• Vague meeting prompts producing a vague recap instead of concrete actions.

What it will not do well

• Cannot access anything outside your existing permissions
• Not built to match a specialised writing tool for long creative work
• Gives thin answers about documents it has not been pointed at directly

Privacy and what you send it

In a managed organisation, your admin's configuration governs, including whether content stays within your organisation's boundary — often the strongest enterprise data story among these tools, but entirely dependent on how your organisation configured it. The consumer version is a different product with different handling; know which you are using.

How to check its work

• Open the source document it cites and confirm the passage yourself.
• Recompute one spreadsheet result by hand before relying on the rest.
• For meeting recaps, confirm attributed actions with the person named.