Run it in three layers over about a month: one 60-90 minute foundation session on what AI does well and badly and on your rules, then short role-based sessions where each person practises on two or three of their own tasks, then a follow-up two weeks later. Most small teams need four to six hours per person.
Generic "introduction to ChatGPT" courses tend to teach features. What actually changes how people work is practice on their own tasks, plus the habit of checking what comes back. Below are learning outcomes to aim for, a session-by-session schedule, a foundation agenda, an error-spotting exercise you can copy, and a way to tell whether the training landed.
What everyone should be able to do afterwards
Write the outcomes before you plan any session. These seven suit almost any small business:
- Explain in a sentence how the tool produces answers: it predicts likely text from patterns, so it can sound confident and still be wrong.
- Give it context and a clear request, rather than a one-line question.
- Check every fact, number, name, date and price before anything leaves the building.
- Know which information must never go into it.
- Recognise tasks where AI shouldn't be used at all, or only with extra checks.
- Improve a weak result by refining the request instead of giving up.
- Know who to ask for help and how to report a mistake that got through.
Outcome three matters most. Staff who understand why AI invents plausible details are far better checkers; AI hallucinations explained for business owners is good background reading for whoever runs the sessions.
The training plan, session by session
| Layer | Length | Who | Content | People leave with |
|---|---|---|---|---|
| Optional pre-work | 1 to 4 hours | Keen volunteers | A free vendor course | Background and vocabulary |
| 1. Foundation | 75 minutes | Everyone together | What AI is and isn't, your rules, how to ask well, first attempts | One task done with AI, and the rules in their head |
| 2. Role practice | Two sessions of 45 minutes | Small groups by role | Two or three real tasks each, plus the error-spotting exercise | Working prompts for their own tasks |
| 3. Practice on the job | Two weeks, about an hour in total | Individually, with a buddy | Use AI on the practised tasks in normal work | A habit, and questions |
| 4. Follow-up | 30 minutes | Everyone | What worked, what went wrong, prompts to share | Improved prompts; problems fixed |
That's about four and a quarter hours per person before pre-work. For the pre-work, two free options: Anthropic's AI Fluency: Framework and Foundations course on Claude Academy (listed at about four hours over 14 lessons) and OpenAI Academy's free self-paced courses. Keep them optional; they're general, and your own sessions are where the business-specific learning happens. For a detailed budget, including facilitator time, see how much time and money AI staff training takes.
Running the 75-minute foundation session
If you've never run a session like this, the format in how to run a 60-minute AI workshop is a good template. For a first training session, this agenda works:
- 10 minutes: what it is and isn't. Live demo. Ask the tool something about your own business it can't know, such as your opening hours or your refund terms, and watch it answer confidently anyway. Nothing teaches checking faster.
- 15 minutes: your rules. Which tool and account to use, and a simple traffic light for data: green (fine to use), amber (only with names removed), red (never). The method in how to classify business data before using AI tools gives you the categories.
- 20 minutes: how to ask well. Show a vague request and its bland result, then a four-part request (below) and its much better result.
- 20 minutes: everyone tries. Each person does one real task from their week. Walk round and help.
- 10 minutes: the checking habit and where to get help.
The traffic light only sticks if every colour has examples from your own work. Here's a filled-in slide for an illustrative six-person property-maintenance firm:
| Colour | What it means | Examples from this business |
|---|---|---|
| Green | Fine to use in the business account | Price list, job descriptions without addresses, website copy, general questions about materials |
| Amber | Only with names, addresses and phone numbers taken out | Customer complaint emails, a tenant's description of a fault, landlord job summaries |
| Red | Never, in any account | Key-safe and alarm codes, bank details, photos showing a house number or car registration, anything about a tenant's health |
Key-safe codes are the kind of item a generic policy never mentions and a maintenance firm handles every day. Asking the team "what's the most sensitive thing you type in a week?" during the session usually produces two or three red items the owner hadn't thought of.
The four-part request
CONTEXT: "We're a language school. Our students are adults,
mostly working professionals, at intermediate level."
TASK: "Write a 250-word reading text about remote working,
with 6 comprehension questions and an answer key."
FORMAT: "Title, then the text, then numbered questions,
then answers on a separate section."
CONSTRAINTS: "Use everyday vocabulary; no idioms. Don't include
statistics or real company names."Show the vague version first. "Write a reading text about remote working" typically comes back, as an illustration, at about 600 words of advanced vocabulary, with a sentence like "According to a recent study, 70% of employees prefer working from home." That statistic has no source and may not exist, which makes it the perfect teaching moment. The four-part request above gets a much closer draft, but still show what the teacher fixed: one idiom ("a double-edged sword") slipped through despite the constraint, and question 4 had two defensible answers, so the answer key needed a rewrite. Even a good request gives you a draft rather than a finished piece.
The constraints line is the one people forget. Telling the AI what to avoid (invented statistics, real names, promises) prevents most of the errors staff would otherwise need to catch.
Role sessions built on real tasks
Ask each person to bring two real tasks from the coming week. Practising on invented examples teaches the tool; practising on real work teaches the job. For the language school in the worked example, the groups looked like this:
- Teachers: the same reading text at two levels for a mixed class; warm-up activities for a lesson topic; written feedback comments on a student's essay, drafted from the teacher's own notes and then checked against the essay.
- Administrators: replies to enrolment enquiries using the current course list; timetable change notices; a notice translated into the main languages students speak, flagged for checking.
- Director of studies: course descriptions for the website, and a first pass at summarising end-of-course feedback forms.
Whoever facilitates doesn't need to be an expert, just a step ahead: often the person who already uses AI well in that role. When the second session finishes, each person should have a saved prompt for each of their tasks that has worked at least twice.
Teaching the second attempt on a live task
Outcome six, improving a weak result instead of giving up, is easiest to teach on one administrator task run in front of the group. The enquiry: an adult learner asks whether there's a Saturday class at intermediate level and what it costs. The first attempt, from a one-line request, came back like this (illustrative):
Thank you so much for reaching out! We would be absolutely delighted to help you on your language-learning journey. Our courses are designed to suit learners of every level, and our friendly, experienced teachers will support you every step of the way. Please don't hesitate to get in touch if you have any questions.
Sixty words, and neither of her two questions answered. Rather than starting a new chat, the administrator typed a follow-up into the same one:
Too generic. Answer her two questions only: yes, there's a
Saturday Intermediate class, 10am to 12pm, starting 3 October,
and the fee is on the attached course list. Under 80 words.
Friendly but plain, no exclamation marks. End by offering a
free placement test.The second draft was nearly usable, with one catch: it quoted the weekday fee, because the course list showed both prices on one row. Point that out to the group. The cure for a bland reply is a more specific follow-up; the cure for a wrong figure is the human check, and no amount of prompt-polishing replaces it.
A spot-the-errors exercise you can copy
This is the most useful 15 minutes of the whole programme. Hand out a realistic AI draft with errors planted in it, and ask people to find them in five minutes. Adapt the business details to your own:
AI-drafted reply to an enquiry (find the problems)
"Dear [customer name],
Thank you for your interest in our Intermediate Evening Course.
The next course starts on Monday 5 October and runs for 10 weeks,
Tuesdays and Thursdays from 6.30 to 8pm. The fee is $540, which
includes all materials and an internationally recognised
certificate on completion. If you're not satisfied after the
first two lessons, we'll refund the full fee, no questions asked.
We look forward to welcoming you!"The planted problems: the course "starts on Monday" but runs on Tuesdays and Thursdays; the fee should be checked against the current price list (in this exercise it's wrong); the school doesn't award an "internationally recognised certificate"; and the no-questions refund policy doesn't exist. The AI invented the last two because they're the kind of thing such emails often say.
Afterwards, ask which errors a customer would hold you to. The answer, usually the refund and the certificate, is why a human check is non-negotiable for anything that makes a promise.
Who should deliver it: you, a colleague or an outsider
In a small business the trainer is usually one of three people, and each has a trade-off:
- The owner or manager. Your involvement signals that this matters and that the rules are real. The risk is that staff won't admit confusion in front of the boss. Run the foundation yourself and hand role sessions to someone else.
- A confident colleague. Peers explain things in the team's own language and are easy to ask later. Give them preparation time and the outcomes list, and check their examples don't include data that breaks your own rules. That check is easy to skip. In an illustrative ten-person estate agency, the negotiator running the role session demonstrated with a real seller's email, complete with the property address and the seller's reason for moving, pasted into her personal free account and projected on the meeting-room screen. It was exactly what the foundation session had called red. Prepare two or three cleaned-up practice examples in advance, with names and addresses swapped out, so nobody reaches for the inbox on the day.
- An outside trainer. Useful when nobody inside feels ready, or for a whole-team rollout of a new tool. Insist that sessions use your tasks and your rules, and that you keep the materials afterwards.
Whoever runs it, keep groups to six or fewer for practice sessions. Beyond that, the quieter people stop typing and start watching.
Worked example: training a 12-person language school
Suppose, as an illustration, a language school has an owner, a director of studies, eight teachers (several part-time) and two administrators. The school runs on Google Workspace Business Standard, where Gemini is built into Gmail, Docs and the other apps, so the training needs no extra licence.
- Timing: the foundation session runs in the quiet week between course terms; role sessions run in 45-minute slots before evening classes, paid at normal rates.
- Staff time: 12 people × about 4.5 hours = 54 hours. At an illustrative average loaded cost of $25 an hour, that's about $1,350 of time, plus around six hours of the director of studies' time to prepare.
- At the follow-up: 10 of 12 were using AI weekly. Teachers reported lesson-prep time for a typical lesson falling by about a third.
- Two problems surfaced. A teacher used AI-written dialogues for a listening lesson, and students noticed phrasing no native speaker would use. New rule: model texts get a read-through by a teacher who's a fluent speaker before use. And an administrator's translated notice contained an error in one language, which a student pointed out politely. New rule: translations of anything important are checked by a speaker of that language before going out.
Both rules came from real use in the practice fortnight, which is exactly what the follow-up session is for.
How to tell whether the training landed
Attendance proves nothing. Two weeks after the role sessions, check four things:
- Rules: ask three people, casually, what they'd never put into the tool. They should answer without looking it up.
- Output: spot-check five recent AI-assisted pieces of work per person. Look for unchecked facts and invented promises.
- Refinement: has each person improved at least one prompt on their own? That's the sign they understand it rather than following a script.
- Reporting: has anyone reported a mistake? Zero reports usually means nobody is looking, not that nothing went wrong.
A spot-check earns its time when it finds a pattern rather than a culprit. At the language school, one administrator's five emails included a reply quoting last term's fee. The cause was a saved prompt with the old price typed into it, and two colleagues had copied the same prompt. The fix was to rewrite every saved prompt to say "use the fee from the attached current course list" instead of containing a number, so a price change now means updating one document rather than hunting through a dozen prompts.
Keep a simple training record: who attended what, when, and which rules were covered. A few rows from the language school's record might read (illustrative):
| Date | Session | Attended | Covered | Materials kept |
|---|---|---|---|---|
| 6 Jul | Foundation, 75 min | 11 of 12 (one teacher on leave) | Data traffic light, checking rule, four-part request | Slides, rule sheet v1 |
| 13 Jul | Foundation catch-up, 40 min | 1 teacher | Same, shortened | As above |
| 14 and 16 Jul | Role practice, teachers | 8 teachers | Levelled texts, feedback comments, error-spotting exercise | Saved prompts for 3 tasks |
| 30 Jul | Follow-up, 30 min | 12 of 12 | Two new rules (dialogues, translations) | Rule sheet v2 |
If you sell to customers in the EU, Article 4 of the EU AI Act asks businesses using AI to help their staff build AI literacy. Since the July 2026 Digital Omnibus amendments the duty is about taking reasonable steps, not reaching a guaranteed standard. A record of short, role-based training is sensible evidence. AI literacy requirements for staff covers what that duty asks for.
Keeping skills current without another course
AI tools change their features every few months, so one round of training fades. Three light habits keep skills current:
- A 15-minute slot at a monthly team meeting: one new trick, one mistake caught, one prompt updated.
- A short version of the foundation session for new starters in their first week. Onboarding new hires onto your AI tools and rules sets out what to include.
- A quick refresh when you change tools or plans, focused only on what's different.
Save the prompts that work where everyone can find them, and name an owner for them. Training teaches people how to ask; a shared set of tested prompts means they don't have to start from scratch every time.
Questions about training a small team
Should I pay for an external AI training course?
Usually not first. Free vendor material covers the general ideas well, and the part that changes behaviour, practice on your own tasks with your own rules, is something an outside course can't do without learning your business. Consider paying for help when you're rolling out a tool to the whole team at once, or when nobody inside has enough confidence to run the sessions.
Do part-time staff need the same training as full-timers?
They need the same foundation, especially the data rules and the checking habit, because a mistake from a part-timer reaches customers just the same. Role practice can be shorter: one session on their single most common task. Pay them for the time, and schedule it within their usual hours where you can.
How do I train someone who's nervous about technology?
Start with a task they know inside out, so they can judge the output with confidence, and sit beside them for the first attempt. Avoid jargon and tool tours. Let them type. Nervous learners often become the best checkers, because they don't assume the output is right.
Further reads
- How to Get Your Staff to Actually Use AI Tools — If people are trained but still not using the tool.
- How to Build a Shared Prompt Library for Your Team — Keep the prompts from training sessions in one place.
- How to Choose an AI Training Provider for Your Team — If you decide to bring in outside help.
- How to Give AI Your Business Context So Answers Fit — The skill that most improves results after training.
- How to Write an AI Usage Policy for Your Small Business — The rules the foundation session should teach.
- Is Machine Translation Good Enough for Client-Facing Documents? — When AI translation is safe to send, and when it isn't.
- How to Handle Staff Who Over-Rely on AI — Signs of AI over-reliance, a conversation script, a three-rule standard to put in writing, and when a pattern needs a formal process.
- AI Change Management for Small Teams: A Practical Plan — An eight-week plan sized for five to twenty people, with a meeting script, a resistance table and a worked nursery example.
- How to Talk to Staff Who Fear AI Will Take Their Job — Prepare your honest answer, then use a six-part conversation outline, better phrasing and a physiotherapy clinic example to talk it through.
- How to Choose and Support an AI Champion in a Small Team — A weighted scoring sheet for candidates, a one-page remit template, a monthly check-in agenda and a music school example with time costs.
- How to Survey Your Staff Before an AI Rollout (With Questions) — Fifteen ready-to-use questions, an invitation email and anonymity settings for surveying a small team before you choose any AI tools.
- How to Test Job Candidates' AI Skills in an Interview — An AI skills test for small-firm interviews: one realistic task with a planted error, a scoring rubric, and follow-up questions that reveal judgement.
- Will AI Replace My Employees? An Honest Answer for Small Firms — An honest answer for owners: what the evidence shows, a task-by-task method to score each role, a bookkeeping firm example and what to tell staff.
- How to Train Front-of-House Staff to Work Alongside AI — Three short sessions, six role-play cards and a one-page counter card for teaching front-of-house teams to work with AI bookings, chat and phone assistants.
- Hire an AI Specialist or Upskill Your Agency Team? — When an agency should train the team it has, when a specialist hire pays for itself, and the hybrid most agencies under twenty people end up with.
- Why AI Adoption Stalls in Accounting Firms, and How to Restart It — The five reasons AI stalls in accounting practices, a twenty-minute diagnosis, and a 30-day restart timed around the deadline calendar.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: Claude Academy course page for AI Fluency: Framework and Foundations; OpenAI Academy course information; EU AI Act Article 4 as amended by the Digital Omnibus on AI; Google Workspace plan information (checked September 2026).