Lesson 01
Ground every prompt in something specific
Copilot's advantage over a general chatbot is that it can see your work: your files, your mail, your meetings. A prompt that does not point at any of that is throwing away the only thing that makes it worth the licence.
Try this
Using the Q3 board pack and the two emails from the
finance manager last week, draft the operations section
of this month's report. Match the structure of the Q3
document. Where the numbers have moved, say by how much
and note where I need to explain why.
What good looks like: a draft with references back to the source files, so you can click through and check anything that looks off.
The gotcha: "write me a report about our operations" gets you generic filler, and people conclude Copilot is weak. It is not weak, it is unbriefed. Name the file, the meeting or the thread every time.
Lesson 02
Regenerate instead of rewriting your prompt
Copilot Chat has a Regenerate action with Try Again and Switch Model. When an answer is nearly right, the instinct is to retype the whole request. Usually you just need another pass, or a different model behind the same prompt.
Switch Model matters more than it sounds. Different models are genuinely better at different jobs, and the same prompt can produce a noticeably better answer simply by being handled by a model suited to it.
The gotcha: regenerating an answer to a bad prompt gives you a different bad answer. If two passes both miss, the problem is upstream in what you asked, not downstream in how it replied.
Lesson 03
You can now pick Claude inside Copilot Chat
As of July 2026, Anthropic's Claude is available as a model option in Copilot Chat, for complex analysis, document understanding and structured content generation.
This quietly ends the "which one should we buy" argument for a lot of businesses. If your work lives in Microsoft 365 and you occasionally need harder reasoning over a long document, you can now reach for that without leaving the place your work already is.
Try this
Switch to Claude for this one. Read the attached 60-page
tender document and give me: what they are actually asking
for, the mandatory requirements we would have to meet, and
anything in the evaluation criteria that favours an
incumbent. Cite section numbers.
The gotcha: this does not make the two products interchangeable. Model choice inside Copilot changes the reasoning behind an answer, not what Copilot is allowed to see or do. Read
Claude or Copilot for where the line actually falls.
Lesson 04
Copilot Notebooks: a place to put the context
Notebooks give you a shared space to gather the material for a piece of work, then work from it repeatedly rather than re-attaching the same files every time. Since July you can add Outlook emails as references, which matters because in most businesses the actual decisions are sitting in a mail thread rather than in a document.
Try this
Add the supplier contract, the three emails where we
agreed the variation, and the meeting notes from 14 July.
Then: summarise what we actually committed to, and flag
anything in the emails that contradicts the contract.
What good looks like: the contradiction found, because the email thread and the contract are finally being read together instead of separately.
The gotcha: a Notebook with everything in it is as useless as one with nothing in it. Gather the material for one piece of work, not the whole department.
Lesson 05
Copilot Cowork: ask for the finished thing
Cowork became generally available worldwide in June 2026. It is agentic: you define the task, it plans, executes and returns a completed deliverable rather than a draft or a recommendation.
The change in how you phrase things is the whole skill. Stop asking for the next step and start describing the finished state, including what "done" means.
Try this
Task: produce the monthly client report for [account].
Done means: a document using our standard template, with
ticket volumes and response times for the month, the three
issues that took longest, and a short plain-English summary
a non-technical client can read. Pull the numbers from the
usual sources. Tell me what you could not find.
The gotcha: "tell me what you could not find" is the line that saves you. An agentic system that cannot find something will often produce a complete-looking deliverable with a quiet hole in it. Ask it to declare the gaps.
Lesson 06
Vision: show it, do not describe it
Vision arrived in July 2026. You can show Copilot what is on your screen, or point your phone camera at something, and ask about it in a voice conversation. It uses that visual context to explain what you are looking at and suggest next steps.
The obvious business use is the one nobody predicted: error messages, unfamiliar software, and equipment. Describing a dialog box accurately is harder than showing it.
Try this
(sharing screen, by voice)
"I have never used this screen before. Walk me through what
these fields mean and which ones I actually need to fill in
for a standard supplier record."
The gotcha: be deliberate about what is on screen before you share it. Vision sees what you show it, including the client data in the window behind the one you meant to share.
Lesson 07
Copilot in SharePoint is already on
From 16 June 2026, Copilot in SharePoint began rolling out on by default for everyone with a Microsoft 365 Copilot licence, with new capabilities for creating content, automating work and getting answers out of what is stored there.
Worth saying plainly: this is not something you switch on when you are ready. If you are licensed, it is on. Which makes lesson 9 urgent rather than optional.
Try this
From the documents in this site, what is our current
process for onboarding a new client? If there is more
than one version, tell me which documents disagree and
how they differ.
The gotcha: this is the prompt that discovers your document estate is a mess. That is not Copilot failing. That is Copilot reporting accurately, and it is usually the most useful thing it does in the first month.
Lesson 08
Meetings, where most teams get their first real win
Meeting recap is the feature that converts sceptics, because the value is obvious and immediate: what was decided, who owns what, and what got missed. It is also the safest place to start, since a summary of a meeting you attended is easy to sanity-check.
Try this
From this meeting: list only the decisions and the actions,
each with an owner and a date. Then separately list the
things that were raised and never resolved. Do not summarise
the discussion.
What good looks like: a short, boring, useful list. The instruction not to summarise the discussion is what makes it usable.
The gotcha: the default recap is a narrative of the meeting, which almost nobody reads. Ask for decisions and actions specifically and it becomes the thing you paste into the project channel.
Lesson 09
The lesson that decides whether the other eight work
Copilot reads what your permissions let it read. That single sentence explains almost every Copilot rollout that has gone badly.
If your SharePoint and Teams sharing has drifted over the years, and in most businesses we take over it has, then Copilot does not create an oversharing problem. It surfaces the one you already had, instantly, to everyone with a licence. Someone asks an innocent question and gets an answer sourced from a document they were never meant to see.
We insist on fixing sharing before a rollout, and it is the least popular advice we give, because it is unglamorous and it delays the exciting part. It is also the difference between a rollout that lands and one that gets paused after a fortnight.
The short version: review who has access to what, tighten the defaults, deal with the sites that have grown broad access by accident, and only then hand out licences broadly.
The gotcha: people treat this as an IT chore to do afterwards. It is a prerequisite. Since Copilot in SharePoint is now on by default for licensed users, "afterwards" may already have arrived.