Split the hour into five blocks: a five-minute opening, 10 minutes on the rules and the approved tool, 15 minutes of live demos using your team's real work, 20 minutes of paired practice, and 10 minutes collecting ideas and agreeing a follow-up. Budget about three hours of preparation in the week before.
One hour won't make anyone skilled, so don't design it as if it could. Its job is narrower: every person leaves having used the approved tool safely on their own work at least once, knowing what they must never paste into it, and having named one task they'd like to try it on. Below you'll find the run sheet, the preparation list, three demo prompts, a practice card, a follow-up email and an illustrative cost for a 14-person engineering consultancy.
The run sheet, minute by minute
| Time | Block | What happens | Materials |
|---|---|---|---|
| 0:00 to 0:05 | Opening | The owner says why the business is doing this, which tool is approved, and what AI is and isn't for here. | None |
| 0:05 to 0:15 | The rules | Walk through the one-page data card: what can go into the tool, what can't, who to ask. | Printed data card |
| 0:15 to 0:30 | Three live demos | The presenter runs three prepared tasks from your own work, including one where the AI gets something wrong. | Demo prompts, cleaned sample documents |
| 0:30 to 0:50 | Paired practice | Pairs pick a real weekly task, run the practice prompt, improve the result and score it. | Practice card, laptops already logged in |
| 0:50 to 1:00 | Harvest and close | Each pair names one task worth trying every week; you say who follows up and by when. | Sticky notes or a shared document |
The order matters. Rules come before demos so nobody pastes a client's drawings into the tool during practice. Demos come before practice so people have seen what a good result looks like and how to judge it. Practice gets the biggest block because using the tool once, on something you actually do, is what changes behaviour. Watching a presenter doesn't.
Keep the owner's opening to five minutes and make it honest. If you have no plans to cut roles, say so. If you're not sure yet, don't promise; say that the aim of the session is to find the dull parts of everyone's week.
The difference between a weak and a useful opening is mostly specifics. A weak one: "AI is changing everything, and we need to keep up or we'll be left behind." It tells people nothing and makes the anxious ones more anxious. A useful one, from an owner who has no plans to cut roles: "We're using Copilot Chat, which comes with our Microsoft plan, so nobody needs a new account. Today is about finding the parts of your week you'd happily hand to it, like writing up site notes. Nobody's job is changing because of this session, and nothing you try today is being scored."
Three hours of preparation, in order
- Confirm the tool and every login (45 minutes). The session runs on the tool you've approved for business use, on business accounts. If you're on a Microsoft 365 business plan, Copilot Chat is included at no extra cost; Google Workspace business plans now include Gemini. If you're unsure what the free tier covers, read what Copilot Chat gives you compared with paid Microsoft 365 Copilot. Ask every attendee to log in two days before and reply "done". Failed logins eat more workshop time than anything else.
- Collect real tasks (20 minutes). Send two questions: "What's one task you do every week that involves writing, reading or copying information?" and "Roughly how long does it take?" The replies feed your demos and the practice block.
- Build three demos (60 minutes). See the next section.
- Clean the sample material (30 minutes). Take real documents and swap client names, fee figures and personal details for placeholders. It shows the team what "removing identifiers" looks like in practice.
- Print the data card and practice card (15 minutes).
- Rehearse each demo twice (30 minutes). The same prompt can give a different answer on each run, so you need to know the range of what might appear on screen. Why AI gives different answers each time covers the causes and what to do about them.
Step four catches people out because names and numbers are the easy part. Take a paragraph from one of the consultancy's site letters: it named the client and the site address, gave the agreed fee as $38,000, and mentioned that the client's neighbour was disputing the party wall. The cleaned version read "[CLIENT] at [SITE], fee [FEE], and a neighbour dispute is ongoing." The neighbour detail is the one presenters forget, because it isn't a name or a figure, yet it identifies the job to anyone in the local trade. Showing that before and after on screen during the rules segment teaches more than the rule itself.
Picking demos from your own inbox, not from a video
Generic demos (a poem, a holiday itinerary) get a laugh and change nothing. Each demo should be a task most of the room does, with a result anyone can judge in under a minute. One of the three should go wrong in a way people can see, because learning to catch the error is the most useful skill in the room.
For an illustrative 14-person engineering consultancy, a good set looks like this. Demo one turns a client's 30-page specification into clarification questions. Demo two drafts a fee proposal cover email from five bullet points. Demo three asks the AI to total a 14-line fee schedule with expenses, then checks the answer against a spreadsheet in front of everyone.
DEMO 1: Clarification questions from a specification
You are helping a structural engineering team review a client specification.
Read the attached document. List up to 12 questions we should ask the client
before we can price the work. For each question, quote the section number it
relates to. Group them under: scope, programme, site information, deliverables.
If something could be read two ways, give both readings.
Do not invent section numbers. If you cannot find one, write "no reference".DEMO 2: Fee proposal cover email
Draft a cover email sending our fee proposal to a client. Facts:
- Client: [CLIENT]; project: [two-storey extension, steel frame]
- Fee: in the attached schedule. Do not restate any figures in the email.
- Programme: survey within 10 working days of instruction
- Tone: plain and confident, no sales language, under 150 words
Give me two versions: one formal, one friendlier.Rehearsing demo one shows why the last line of that prompt is there. An illustrative extract from a rehearsal run:
SCOPE
1. Section 2.4 lists "all temporary works". Does this include the
propping to the existing party wall, or only new openings?
2. Section 3.1 refers to "the attached survey". Can you send it?
It isn't in the pack we received.
PROGRAMME
3. Section 5.2 requires "design sign-off before the planning
decision". Is that a fixed date, or tied to the decision?
SITE INFORMATION
4. Section 6.7 mentions ground conditions. Has a site investigation
been carried out?Questions 1 to 3 are good, and question 2 is the kind people miss. Question 4 cites a section 6.7 that doesn't exist; the specification stops at 6.3. It's a sensible question with an invented reference, which is the most convincing kind of error. Leave it in the live demo and ask the room to check the references against the document. Someone always finds it, and that's the lesson.
For demo three, don't script the failure. Ask for the total, then add the column in a spreadsheet while the room watches. If the AI gets it right, say so, and point out that you only know because you checked. If it gets it wrong, you have your lesson. Either way the rule lands: totals, dates, clause numbers and names get checked against the source every time.
The same structure works in any business; only the tasks change. For an illustrative three-chair hair salon, the demos might be: turning a week of booking notes into a list of clients due a colour appointment who haven't rebooked; drafting a reply to a client unhappy with a cut, from the stylist's notes; and asking the AI how much colour product the salon used last month from a pasted stock sheet, then checking the total against the sheet. The third is the failure demo again, because stock quantities are exactly where AI arithmetic slips.
The rules segment: ten minutes and one card
Don't read the whole policy aloud. Hand out a card with three colours and talk through one example for each.
- Green, fine in the approved tool: internal notes, public information, your own drafts, templates, meeting agendas.
- Amber, approved tool only, identifiers removed first: client documents, specifications, fee proposals, project correspondence.
- Red, never: passwords and access codes, personal data about staff, anything a client contract says must stay in named systems, and any business material in a personal or free account.
Pick your three examples from the replies people sent, so they recognise their own work. For the consultancy, green was "tidy up the agenda for Friday's project meeting"; amber was "summarise this client specification", done only after swapping the client's name and site for placeholders; and red was "fix the formatting in the staff holiday and sickness spreadsheet". The red one always gets a reaction, because it seems harmless: it's only formatting. That reaction is the moment to explain that the tool sees every cell, not just the ones you want tidied.
Add one sentence people remember: "You're responsible for anything you send, whether you or the AI wrote it." If you haven't written the full policy yet, how to write an AI usage policy covers it, and rolling out an AI policy so staff follow it covers the part a workshop can't.
If you sell to customers in the EU, Article 4 of the EU AI Act asks businesses that use AI systems to take measures supporting their staff's AI literacy. A dated attendance list, the data card and the prompts are sensible evidence of a first measure, though one hour isn't a literacy programme. What staff need to know under AI literacy rules sets out the rest.
The practice block: pairs, one real task, a scoring card
Work in pairs: one person types, the other reads and judges. Nervous people talk more with one colleague than in a group, and the reader catches errors the typist skims past. Each pair picks a task from the replies you collected. The card below keeps them moving without you standing over them.
PRACTICE CARD (20 minutes)
1. Pick one task one of you does every week: ______________________
2. Run this prompt, filling the brackets:
"I work at an engineering consultancy as a [ROLE]. I need to [TASK].
Here is the material: [PASTE THE CLEANED SAMPLE]. The result should be
[FORMAT, LENGTH, TONE]. Before you start, ask me up to three questions
if anything is unclear."
3. Answer its questions. Read the whole result.
4. Ask for one improvement: "Make it shorter", "Use our headings",
or "List every assumption you made".
5. Score it (circle one per line):
Accurate? yes / mostly / no
Usable after edits? yes / heavy edits / no
Faster than by hand? yes / about the same / no
6. Would you use this next week? If not, which task instead?Here's a card as one pair actually filled it in, an office administrator and a junior engineer (illustrative):
1. Task: turn site-visit notes into a short summary for the client
2. Prompt run with: ROLE = junior engineer; FORMAT = 5 bullet
points, plain English, no recommendations
3. Its questions: "Should I include the photos' references?"
"Is the client technical?" We said: yes; no.
4. Improvement asked for: "List every assumption you made"
-> it had assumed the crack was "historic". The notes don't say.
5. Accurate? mostly Usable after edits? yes
Faster than by hand? yes (about 8 min vs 20)
6. Next week? Yes, but the engineer checks every summary first.Line 4 is the one worth reading out in the harvest. Asking the AI to list its assumptions found a word that would have changed what the client thought the engineer had concluded.
Walk the room and watch for three things. Pairs pasting red-category material: stop them kindly and show how to strip it. Pairs accepting the first answer without reading it: ask them what it got wrong. Pairs stuck on a login: hand them a spare laptop that's already signed in rather than troubleshooting.
Handling the sceptic, the enthusiast and the quiet one
The sceptic usually has a point. Agree where they're right, then make them the checker in their pair. Their instinct to distrust the output is exactly the habit you want everyone to copy.
The enthusiast already uses five tools. Thank them, then ask privately which of those run on personal accounts, because that's where client material leaks. They may be the person to own AI day to day later, once they're working inside the rules.
The quiet one is often worried about their job and won't say so. Pair them with someone patient, not the enthusiast, and make sure the task they practise on is one they dislike doing. Saving someone from a chore they hate is the fastest way to change how they feel about the tool.
When a demo misfires in front of everyone
- A login fails: switch to the presenter's screen. Don't troubleshoot live for more than a minute.
- A demo you expected to work gives a wrong answer: carry on and show how you'd catch it. It's the most useful minute of the hour.
- You're running over: cut demo two. Never shorten the practice.
- Someone asks something you can't answer, such as whether the tool trains on your data: write it down and answer in the follow-up email within two days. Don't guess in the room.
The seven days after the session
Before the sort, sharpen the sticky notes. Harvests are full of vague entries such as "emails" or "reports", which can't be scored or tried. Ask the pair who wrote each one two questions, ideally as you collect the notes at the close and at the latest the next morning: which emails, and how many a week? At the consultancy, "emails" became "replies to contractors asking when revised drawings will be issued, about ten a week", which is small, checkable and easy to template. "Reports" turned out to mean monthly progress reports for two long-running clients, a different and much bigger job.
Most of the value is lost if nothing happens in the following week. On day one, send the follow-up email. By day three, sort the tasks the pairs named: score each one on hours spent per week and how easy the output is to check, using a one-page use case template, and pick one or two to try properly for a month. On day five, check who has used the tool since; if your plan's admin view shows usage, look there, otherwise ask at the weekly meeting. On day seven, give one person five minutes to show something that worked.
For the consultancy, the day-three sort might look like this once the sticky notes are typed up:
| Task named by a pair | Hours a week (team) | Easy to check? | Pick? |
|---|---|---|---|
| Site-visit notes to client summary | 6 | Yes, against the notes | Yes |
| Clarification questions from specifications | 4 | Yes, against the document | Yes |
| Checking calculations | 5 | No: needs an engineer to redo them | No |
| Minutes of internal meetings | 2 | Yes | Later |
| Rewording CVs for bids | 1 | Yes | Too small |
"Checking calculations" scored high on hours and was still dropped, because a check that needs the work redone saves nothing. The two picks are both frequent and checkable against a source document, which is the pattern to look for.
Subject: AI session: the card, the prompts and what happens next
Thanks for Thursday. Three things:
1. The data card is attached. If you're unsure whether something is
amber or red, treat it as red and ask me.
2. The prompts we used are in [SHARED FOLDER]. Copy them, change them,
and add any that work well for you.
3. You named [NUMBER] tasks. [NAME] and I will pick two to try properly
over the next four weeks. We'll tell you which by Friday.
Questions I couldn't answer on the day:
- [QUESTION]: [ANSWER]What the hour costs a 14-person consultancy (illustrative)
Say the consultancy's average loaded staff cost is $55 an hour. Fourteen people for one hour is $770 of time. Preparation is about three hours for the presenter plus an hour of the owner's time, another $220. If the firm is already on a Microsoft 365 business plan, Copilot Chat costs nothing extra, so the session costs roughly $990 in all.
That's about 18 hours of staff time. If the two tasks chosen afterwards save the team three hours a week between them, the session pays for itself in six weeks. If they save nothing, you've still bought a shared set of data rules and a list of real tasks, which is where any sensible choice about what to automate first starts.
The measure of success isn't whether people enjoyed the hour. It's how many of the 14 used the approved tool on real work in the fortnight after, and whether any client material went somewhere it shouldn't. Count both.
Questions owners ask before running the session
Should the workshop be in person or on a video call?
In person works better for the practice block, because pairs can share one screen and you can walk round and see where people get stuck. On a video call, use breakout rooms of two, ask one person per pair to share their screen, and add five minutes to the practice block, since instructions take longer to land remotely. Record the demos, not the practice.
Who should run it if I'm not confident with AI myself?
Whoever on the team already uses the approved tool most, with you opening and closing the session so it's clear the business backs it. The presenter doesn't need to be an expert; they need to have rehearsed the three demos on real work. Being honest when the AI gets something wrong teaches the room more than a polished performance.
What if some staff don't have a licence for the approved tool?
Sort access before the day. If a paid seat isn't justified for everyone yet, check whether your office suite already includes a business chat assistant, such as Copilot Chat on Microsoft 365 business plans or Gemini on Google Workspace plans. Don't let anyone practise on a personal free account with business material, even for an exercise.
How often should we repeat the workshop?
Run a second hour four to six weeks later, built around the tasks the first session produced and what people learnt trying them. After that, a 15-minute slot in a monthly team meeting, where one person shows one thing that worked, keeps momentum better than another full workshop.
Further reads
- How to Train Staff to Use AI in a Small Business — The longer programme that should follow this first hour.
- How to Get Your Staff to Actually Use AI Tools — What to do when people nod in the session and never open the tool again.
- How to Build a Shared Prompt Library for Your Team — Turn the workshop prompts into a library the team keeps adding to.
- How to Choose and Support an AI Champion in a Small Team — Your keenest attendee may be the right person to own this.
- How to Talk to Staff Who Fear AI Will Take Their Job — For the quiet attendee worried about their job.
- How Much Time and Money Does AI Staff Training Take? — Budget the hours beyond the first workshop.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: Microsoft 365 business plan pages (Copilot Chat inclusion), Google Workspace pricing page (Gemini in business plans), EU AI Act Article 4 as amended by the Digital Omnibus on AI.