Copilot in 90 minutes: a first encounter for staff

For whom: staff with no experience of generative AI, in a group that also holds people who already use it daily

Length: 90 minutes (a 60-minute version is at the end) · Group size: 8 to 20. Above 20 you can no longer get round everyone during the hands-on part, and that is what this session runs on.

Prior knowledge: none required, but expect wide differences within the group.

What participants can do afterwards

• Say in their own words what Copilot for Web is, and what the VU has and has not licensed.
• Build a prompt out of task, persona, steps, context, purpose and format, and point to what each part changes in the output.
• Name one task from their own work that suits this and one that does not.
• Say which data does not belong in here, and why.
• Check an answer before using it.

What to arrange beforehand

• Everyone on a laptop with a VU account. Put it in the invitation, or someone will arrive without one.
• Yourself logged in at m365.cloud.microsoft, in a clean window with none of your own conversation history on screen.
• Browser zoom at 150%. What you can still read, the third row cannot.
• A real, dull task of your own for the demo. Not an invented example: a room can tell, and it is the drudgery that convinces.
• Four or five scenarios on paper, for those who cannot invent a task on the spot.
• A flip chart or a stretch of whiteboard, empty, for the parking space. See the section on objections.
• Walk through the environment the day before. Copilot changes without notice, and a demo that stalls on a moved button costs you ten minutes.

Timeline

1. 0 to 10 · Welcome and taking stock. Dutch or English: ask straight away. Then gauge experience, and do it visibly: have people stand up, or raise hands at four levels from "never" to "daily". This is not an icebreaker. You are drawing your map for the rest of the hour, and showing the group how wide the range is, so that nobody later feels like the only one lagging.
2. 10 to 20 · What it is, and what the VU has. Briefly. The VU has Copilot for Web, not Copilot for Microsoft 365. That distinction is the most common misunderstanding and it pays to explain it once, properly: it is not inside Word and Outlook, it is a separate place you go to. State the rule about data here too (see below for why here and not at the end).
3. 20 to 35 · Demo. One real task, thinking aloud. See the section below.
4. 35 to 45 · Warming up. Two tasks that cannot fail: put the letters of a name back in order, and pull email addresses out of a slab of text into a table. The point is not the result but the first time typing, and noticing that the answer arrives at once.
5. 45 to 55 · The prompt model. Built in front of them, not shown as a finished diagram.
6. 55 to 75 · Hands on. Three tracks side by side. See the section on mixed levels.
7. 75 to 85 · Considerations, and clearing the parking space. This is the part people think back on. Protect it: if the hands-on runs over, cut there, not here.
8. 85 to 90 · One thing. Everyone names in a sentence what they will try on Monday. "Nothing" is allowed, with a reason. That is a valid answer and it keeps the last round honest.

The demo

Pick a task you genuinely have: a policy note on changing intake, a set of minutes, an email you have been putting off for two days. Work from role, task and the tone you want.

Three things separate a demo that impresses from a demo somebody learns from:

• Let your first attempt fail. Type the lazy prompt everyone would type, show the mediocre answer, and then improve it. Showing the polished result straight away teaches the room only that it is magic.
• Think aloud. Why you are adding that sentence, what you expect to change, and whether it did.
• Let it be wrong once. Ask about something you know and the model does not, and let the room catch it. It is the cheapest lesson in checking there is, and it lands better than a slide about it.

Then take a few minutes to show what else is there: uploading files (three at a time, up to a megabyte, landing in your own OneDrive), producing Word or Excel files, forcing tables with "use Markdown and tables", building and sharing an agent, and copying output into Loop. Showing is enough; nobody will remember this today.

The prompt model

The slide says TAPESTCODOFO, which is not a word but the first syllables of six parts: Task, Persona, Steps, Context, Do as in purpose, Format. Say so out loud, or people copy it down as though it means something.

Build it with one running example, minutes of a meeting for instance. Start with "write minutes" and add one thing at a time. Actually run it after each addition, and show what changed:

• Task alone produces something generic.
• Persona ("as an experienced minute-taker") changes the tone.
• Steps ("first an overview, then a summary, then the decisions") makes the structure predictable.
• Context (what it was about, who was there) is where it really turns. This is the step participants skip.
• Purpose (who is this for, who was absent) changes what gets left out.
• Format (at most 300 words, three paragraphs, plain language) makes it usable.

The mnemonic is not the point. The point is that they see each addition do something. Show the finished prompt last, because starting with the whole thing makes it look like an unreachable amount of work.

Hands on, with wildly mixed levels

This is the hardest part of the session, and one task for everybody will not solve it. Run three tracks side by side and say out loud that they exist. That is the whole trick: someone who knows there is a beginners' track does not have to pretend.

1. Warming up. The two tasks from the timeline. Everyone starts here, the advanced included, and it takes five minutes.
2. Scenarios on paper. For those who want a realistic task without having to invent one. This is the busiest track; bring enough.
3. Own work. For those who already know what they want to try. The most to gain here, and the most disappointment, because real work is rarely as tidy as a scenario.

A few things that make the difference in practice:

• Let the beginner type. Pair an experienced participant with an inexperienced one and the experienced one takes over while the other watches. Say beforehand who has the keyboard.
• Give the advanced ones work. Anyone finished after ten minutes gets the harder moves: ask the model to improve your own prompt, ask it to put a question to you before it starts, have it produce a Word or Excel file, or upload an old document and ask for that same tone. And ask them to walk around; they are your second pair of hands.
• Do not stay at the front. Walk. Someone who has stalled is usually stuck on logging in rather than on prompting, and you only see that up close.
• Close with two screens. Ten minutes before the end, ask for someone whose attempt worked and someone whose came to nothing. The second is more instructive, and it gives the sceptic something true to hold on to.

Objections to generative AI

Almost every group holds someone who wants none of this, and the objections they raise are not silly: the power of a handful of companies, the environmental cost, unease about who we are tying ourselves to, and the suspicion that we are making ourselves stupider. Treat those as resistance to be overcome and you lose the room, deservedly.

Three things beforehand that make the rest easier:

• Say in the first ten minutes that this is not a sales pitch. The VU took out a licence; you are not defending that decision, you are helping people judge what to do with it, and that can end in "do not use it". A good deal of the friction disappears before it forms.
• Make a parking space. Literally: an empty sheet where you write objections as they arrive, with the promise that you will deal with them at 75 minutes. It works because it is honest, not because it is a trick. So keep the promise, even if you have to cut elsewhere.
• Give whoever will not practise a role. Observer. Ask them to note what goes wrong for their neighbour, and to report it at the end. That is a genuine contribution and often the sharpest of the morning.

Per objection, what is true and where it can go:

• Dependence on one company. True, and the choice was made institutionally. You cannot undo it in this room and should not try. What does sit here: do not build a way of working you can no longer do without. Ask what they will do the day the licence ends or the thing changes. The latter happens, without notice.
• Environmental cost. True and measurable, which makes it usable rather than vague: the cost sits per query, so "could this be done without" is at once the ethical and the frugal question. That is why it heads the list of considerations. For anyone who wants to go further, the exercise What uses more: the footprint of a prompt has participants estimate first and look it up after.
• Unease about the companies and their politics. The heaviest objection, and the one a trainer is most tempted to talk around. Do not. It is a sound question about who an institution ties itself to, and it will not be settled in a workshop. Say that. Say also that someone may decide not to use it, and that the session has still given them what they need to judge how colleagues use it. In a policy department that is not the consolation prize but the more useful skill.
• Dulling the mind. The objection least easily waved away at a university, and there is research on offloading thinking. The turn that works: this is a design question about your own work. Which part of this task was the thinking, and did you just give it away? That question outlives the workshop; the tool may not. Further reading: Cognitive offloading.

Two warnings. Do not let it become a debate between the enthusiast and the sceptic, because those two start performing and the rest watch. And ask the sceptic to open the closing discussion: they have thought about it most, and it turns an objection into a contribution.

Frequently asked questions

• Is Copilot safe? As safe as it can be within the VU environment, but do not put personal or sensitive information into it. It also searches the internet through Bing, so do not treat it as a closed system.
• Is this the same as Copilot in Word and Outlook? No. The VU has Copilot for Web, not Copilot for Microsoft 365. This is the question that comes back most often.
• Can it summarise a meeting? Yes, by several routes. What works: record the Teams meeting, take the transcript, and have Copilot summarise that. Mind the data such a transcript contains.
• Where do my files go? Uploaded files end up in Microsoft Copilot Chat files in your own OneDrive.
• May I put student data in here? No. Answer this one briefly and without hesitation, even when someone has a reason why it would be convenient.

What goes wrong

• People end up in a different Copilot. The button in the edge of Edge, or a private account. Go to m365.cloud.microsoft together at the start and have people glance at their neighbour's screen.
• Someone uploads a file with personal data during the exercise. Which is why the rule about data sits before the hands-on here rather than with the considerations at the end. On the original slides it comes last; that is one slide too late.
• The demo stalls. Good. Say what you expected and what happened. It is the most honest minute of the session.
• The discussion disappears because the hands-on ran over. The most common mistake, and the most expensive: the closing conversation is what people take away.

Variants

In 60 minutes. Drop the warm-up and go from the demo straight to the prompt model; cut the hands-on to fifteen minutes with the paper scenarios only. Keep the considerations at full length. What you cut is practice time, not thinking time.

Online. It works, but walking around disappears and with it your view of who is stuck. Replace it with breakout rooms of three including one experienced participant, and ask people to share their screen the moment they stall rather than typing in the chat.

With ChatGPT instead of Copilot. The shape holds; only the first eight minutes and the frequently asked questions change. The prompt model and the section on objections are not tool-specific.

Material and origin

• Based on the workshop ChatGPT / Copilot for SOZ by Silvester Draaijer, VU Centre for Teaching & Learning, 21 slides.
• The slides carry no speaker notes. This plan is a reconstruction of what happens between them; depart from it where your group asks you to.
• For anyone who wants to talk about teaching rather than office work: the AI literacy exercises are aimed at students, and the didactic tips on generative AI are shorter.