Microsoft 365 Copilot · Practical walkthrough

Work out why Copilot gives two people on the same team different answers

You and a colleague ask Copilot the same question about the same policy and get different answers. A different sentence is usually less interesting than a different source, so start with what each answer cites.

Task by Task13 min read + practice

What you'll make

a side-by-side record of two Copilot answers, the sources each one cited, and a short list of likely causes to take to your administrator.

What you'll need

your work account, a colleague willing to run one prompt at the same time, one non-sensitive document you can both open, and about 20 minutes. You don’t need admin rights.

Example result.

Priya (Premium label): "Wednesday 12:00" – cites Equipment booking v4, §2.
Tom (Basic label): "usually Thursday" – no citation.
Likely cause: Tom's experience doesn't search work files automatically.
Next: Tom attaches v4 and repeats the prompt.

1. Check which Copilot each of you is using

“Copilot” is three licensed experiences, and the difference explains a lot on its own. Microsoft’s documentation names Copilot Chat (Basic), Microsoft 365 Copilot (Basic) and Microsoft 365 Copilot (Premium). Only Premium grounds answers automatically in your work data through Microsoft Graph and Work IQ. On the two Basic experiences, Copilot uses work content only if you supply it: upload the file, work with content open in Outlook or Teams, or use a pay-as-you-go agent.

Look at the label each of you sees in Word, Excel, PowerPoint, OneNote or the Copilot app, and write it down. Then pick a question with an answer you can find yourself, such as “Which day does the equipment-booking handbook say requests close?”, and open the source in both accounts. If one of you can’t open it, stop and ask the owner; don’t copy restricted text across to make the test work.

Why this works: two people on different experiences are asking the same words of two differently equipped systems.

What you should see: two labels, one question, and a source you’ve both opened, with its title and version.

2. Ask the same bounded question at the same time

Start fresh conversations in the same app and run this within a few minutes of each other. Where you can attach the file, do it rather than naming it.

Use only [exact document title and attached file or link] to answer:
[one question].

Give the answer, the source title and version if visible, and the
heading or short passage supporting it. If you cannot read the source,
say so. Do not substitute a general answer or a public-web answer.

List anything the source leaves unclear. Do not guess missing details.

[exact document title and attached file or link] is the source both of you opened; [one question] is the question from step 1.

Why this works: the “say so” line matters most. An account that can’t reach the document will otherwise produce a fluent answer from the web, and nobody will notice. Timing matters too: Basic users get standard access, which can vary through the day, so a gap recorded two hours apart is weaker evidence than one recorded two minutes apart.

What you should see: two answers, each with a citation or a plain statement that the source couldn’t be read.

3. Compare the evidence, not the wording

Copilot picks a model per prompt by default (Auto, with Quick response and Think deeper in the selector), so wording will differ. Phrasing is noise; citations are signal. Open each cited passage and sort the pair into one of three outcomes:

Same source, same meaning: different wording alone isn’t a problem.
Different or missing sources: investigate the causes in step 4.
Same passage, different facts: an answer-quality problem. Read the passage yourself; a licence change won’t fix it.

Why this works: it separates “Copilot found different things” from “Copilot read the same thing differently”, which have different fixes.

What you should see: one of the three outcomes, with each cited passage checked.

4. Walk the five-point checklist

When the sources differ, work through these in order. Don’t change settings yourself.

Licence. If one label says Basic, that account doesn’t search work files automatically. This explains most gaps.
Permissions. Copilot only shows content each person can access. If you’re in a SharePoint group your colleague isn’t, you’ll get an answer they can’t.
Grounding switched off. On Premium, Work IQ can be turned off, and answers then lose work context. Admins also control web search grounding for all licences.
Discovery restrictions. An admin can restrict a SharePoint site from organisation-wide search and Copilot without changing permissions. You can then open a document that Copilot won’t find.
Freshness. New documents on SharePoint sites shared with two or more people are indexed daily; edits to indexed documents are indexed straight away. A policy published this morning may not be findable yet.

One more: delegated and shared mailboxes aren’t in the user index, so asking about “our emails” in a shared mailbox behaves differently from what anyone expects.

Why this works: each item is documented behaviour with a recognisable symptom, so you can say which one fits before asking IT.

What you should see: one or two likely causes, written as possibilities, not conclusions.

Check the result

Both labels, both prompts and both times are recorded.
Every cited passage was opened and read.
An answer with no citation is treated as unsupported, not as an alternative reading.
Causes are written as possibilities for an admin to confirm.

The usual failure: the uncited answer gets treated as a second opinion. “Usually due on Thursday” sounds reasonable, so people argue about which day is right instead of noticing it cites nothing. Spot it by checking which answer you can trace to a passage; correct the other before anyone repeats it.

The source passage says: [exact relevant passage].
Your answer said: [incorrect statement].
Correct the answer using that passage only. Keep any conditions or
uncertainty in it. Do not change the source document.

[exact relevant passage] is the text you found; [incorrect statement] is the line to fix. A corrected answer you can trace to the source is useful, but it doesn’t explain why the first one failed.

Worked example: two equipment-booking answers

Fictional example. Priya and Tom ask about “Equipment booking, version 4”. Both can open it. Section 2 says:

Submit requests by Wednesday at 12:00 for collection on Friday.
Requests after Wednesday at 12:00 need separate confirmation.

Priya’s answer says Wednesday at 12:00 and cites section 2. Tom’s says requests are “usually due on Thursday” and cites nothing. Tom’s label shows Microsoft 365 Copilot (Basic), and he only named the document rather than attaching it, so Copilot never had it. He attaches v4, repeats the prompt and gets Wednesday at 12:00 with a citation.

What mattered:

The uncited answer was the clue. “Usually Thursday” wasn’t a misreading of the policy; it was a general answer to a general question.
The licence was a plausible cause, not a proof. Attaching the file fixed Tom’s route; the admin can confirm the rest.
The condition still matters. “Late requests are refused” would also be wrong: separate confirmation isn’t refusal.

Task kit for this job: Download (ZIP) · Reusable instructions

Go deeper

When one account gives two answers

Sometimes the discrepancy is between Tuesday and Thursday, not between two people. The most common cause is several copies of the same policy in different libraries, such as v4 in the team site and v3 in an old project site. Re-run the step 2 prompt and add:

Also give the last modified date and location of the document you
used, and say whether you found more than one file with that name. If
there are several versions, list them and say which one you used.

The fix is in SharePoint, not in the prompt: ask the owner to retire the old copy.

Write a support message an admin can act on

Send the prompt, both labels, the times, the document link and what each answer cited, without the document’s contents. Ask one question: is this a licence difference, a permissions difference or a discovery restriction? Microsoft’s AI Administrator role can manage Copilot settings without full Global Administrator rights, so that person may be the quickest route if your organisation has assigned it.

What this doesn’t tell you

The comparison shows that two accounts retrieved differently and narrows the likely cause. It doesn’t prove which cause applies; only someone who can see licences, permissions and SharePoint settings can confirm that. And a correct citation doesn’t make a summary faithful, so read the passage. Use a published policy for the test, not a sensitive one.