Picture two people in the same team asking Microsoft 365 Copilot to do the same job. Same licence, same tenant, same Monday morning. One ends up with something she can send to a customer. The other ends up with something he quietly deletes.
The tool is identical. Everything around it is not.
Where the difference shows up
The task is simple: a customer has sent a long Outlook thread about a delayed project, and the team needs a recovery update for a Teams meeting that afternoon.
The first person starts by thinking. She knows what the customer is worried about, and which commitments are real and which are wishful. She asks Copilot to summarise the thread, then to separate the decisions from the complaints. She opens the project plan in Planner, checks the dates against what Copilot said, and spots a milestone stated with far too much confidence. She corrects it, then asks for a draft in Word, matching the tone of an earlier update stored in SharePoint.
The second person types “write an update about the delay” and pastes the answer into a document. It reads fluently, is polite, and is wrong in two places. He doesn’t notice, because he never decided what a good answer looked like.
Same wand. The difference was framing the problem, checking the result, and knowing where the real data lives.
It goes beyond prompts
I’m wary of the idea that this is all about prompt wording. Prompts matter, but they’re the visible tip. Underneath sits judgement: knowing which Excel workbook is the trusted one and which is last year’s copy with a different name. Knowing that a SharePoint site full of duplicates gives Copilot a confused view of your business. Knowing when a sensible-sounding answer should make you suspicious rather than relaxed.
None of that arrives with a licence. It comes from experience, a healthy scepticism, and someone being willing to say “that doesn’t look right” out loud.
It changes how I think about buying decisions, too. A business that rolls Copilot out to everyone and calls it done has bought wands. A business that invests in tidy data, clear ownership and shared habits for checking AI output has started building wizards. I’d take the second approach with half the licences over the first with all of them. A skilled person with a plain tool beats a novice with a fancy one most days.
What I’d do on Monday
Pick one real task, such as a weekly report or a customer reply, and have your most careful person and your most enthusiastic person both do it with Copilot. Then compare more than the output. How did each of them get there? What did they ask first? What did they check? What did they ignore?
That comparison tells you more than any vendor demo ever will. It shows you what to teach, what to document and where your data needs attention.
If you buy services from an MSP or adviser, apply the same test to them. Ask less about which products they resell and more about which problems they’ve solved that look like yours. Ask who will actually work with you, and what they did the last time it went sideways.
Tools will keep changing every few months. Framing a problem, questioning an answer and understanding your own information will still be worth having when the next one arrives. That’s what I’d invest in first, and I suspect it quietly decides whether any of this pays off.
