The short version: Staff learn AI by doing three or four real tasks from their own job with a person watching over their shoulder, not by sitting through a slideshow. Give them one tool, one week of practice on real work, and a manager who uses it too, and adoption sticks. Skip any of those three things and you'll be paying for a course nobody uses by March.
What goes wrong first
I ran a "ChatGPT afternoon" for my own team back when everyone was still calling it a chatbot and not a colleague. Twelve people, a slide deck, a demo, and a promise that this would save them hours a week. Three weeks later I checked usage. Two people had opened it more than twice. One had used it to write a birthday card message. That was it.
The mistake wasn't the tool. It was that I'd taught them what AI could do in general instead of what it could do for the specific report, the specific client email, the specific spreadsheet they were dreading on a Tuesday afternoon. General demos create curiosity for about four days. Then the old workflow wins because it's familiar and the new one takes effort to remember.
This is the bit most people writing about AI training won't say plainly: staff don't resist AI because they're scared of it or too old for it or stubborn. They resist it because using it well takes more short-term effort than not using it, and nobody has shown them the specific payoff on their specific desk. Fix that one thing and most of the "change management" problem disappears.
Start with the job, not the tool
Before you book a single training session, sit down with three people in three different roles and ask them to talk you through their Tuesday. Not their job title. Their actual Tuesday. The customer service manager who drafts forty similar emails a week. The bookkeeper who summarises supplier invoices by hand. The marketing coordinator who writes the same three social captions with slightly different words each time.
Write down five to ten tasks like that across the business. These become your training curriculum. You are not teaching "AI". You are teaching "how Sandra drafts a response to a late-delivery complaint in ninety seconds instead of twelve minutes."
This is the same principle behind good ERP system training: nobody remembers a walkthrough of every feature, but everyone remembers the one screen they use every single day. AI training works the same way. Narrow it down or lose them.
The four-week plan I use now
After that first flop, I rebuilt the approach and I've used a version of this with client teams ranging from four people to around sixty. It works whether you're training a small law firm's admin staff or a marketing department.
Week one: the audit
Pick five real tasks (as above). Time how long each one currently takes. Write the number down. You need this number later to prove anything happened.
Week two: the pilot pair
Don't train everyone at once. Pick two people, ideally one enthusiast and one sceptic, and spend ninety minutes with them doing the actual task inside ChatGPT, Claude, or Gemini (whichever your business has already licensed, and yes, pick one and stop debating it). Write the prompt together. Redo it three times until the output is usable, not just impressive. Save the final prompt.
Week three: the playbook
Turn each saved prompt into a one-page instruction sheet: what to paste in, what to check before sending, what to never let the AI decide on its own (pricing, legal wording, anything customer-facing without a human read-through). Five tasks, five one-pagers. This becomes your internal manual, and it's worth more than any external course because it's built from your own business, not a generic template.
Week four: the wider rollout
Now bring in the rest of the team, in groups of four or five, and have them work through the playbook on their own real tasks with the pilot pair sitting alongside them answering questions. This is the session that sticks, because it's peer to peer and it's their actual work on the screen.
Four weeks, five tasks, two pilot staff, one playbook. That's the whole structure. It costs almost nothing beyond time and whatever AI subscription you're already paying for, and it beats a £2,000 external workshop nine times out of ten, because external trainers don't know that Sandra's Tuesday involves forty complaint emails.
The uncomfortable part: you have to use it too
Here's the thing nobody likes hearing. If the manager running this rollout isn't using AI themselves for their own work, staff clock it within a fortnight and the whole thing quietly dies. I've watched this happen at two client businesses. The staff aren't stupid. If the boss says "use AI to draft your reports" but still asks his PA to type his own emails word for word with no AI involved, everyone reads that correctly: this is optional, and optional things get skipped when things get busy.
The businesses where this worked long-term were the ones where the owner or a senior manager could point to their own screen and show a prompt they used that morning. That's not a soft leadership tip. It's the single biggest predictor of whether adoption survives past month two.
If you're a business owner reading this thinking "I don't have time to learn this myself, that's what the training is for", I'd gently push back. My piece on AI training for business owners goes into why most courses fail at exactly this point: owners buy training for the team and skip it for themselves, and the gap shows up fast.
What to teach (beyond "how to prompt")
Prompting is the easy 20 percent. The harder, more useful 80 percent is teaching people what to check before they hit send. Build this into every session:
- How to verify a fact or figure the AI gives you before it goes in a client document (AI gets numbers wrong confidently, more often than most people expect)
- How to feed it your own examples so the tone sounds like your business, not like a generic press release
- Where the line is on data: what customer information should never be pasted into a public tool without checking your data policy first
- How to iterate: the first draft is almost never the one you send, and teaching people to push back with "make this shorter" or "this sounds too formal, redo it" is often the whole skill
Skip teaching people "prompt engineering frameworks" with acronyms. In practice almost nobody uses them six weeks later. What they do remember is "type it like you're explaining it to a new colleague and then ask it to tighten it up."
Want AI doing the heavy lifting in your marketing?
I build the systems that handle the boring 80 percent, so you get your week back. Done properly, with the human kept in.
Different roles need different training, not one workshop for all
A finance team needs training on data handling and verification far more than they need creative prompting. A marketing team needs training on brand voice and where AI replaces hours of work versus where it just produces average content that needs heavy editing. I've written before about where the line sits between using AI and hiring marketing staff, and the honest answer is that AI trained well can absorb the repetitive 60 percent of a marketing job, but someone still needs the judgement for the other 40 percent, and that person needs proper training on strategy, not just tool clicks.
Customer-facing staff need the heaviest emphasis on the "check it before you send it" rule, because a wrong AI-generated answer to a customer does more damage than a slow human one. Back-office staff can move faster and looser because the stakes of a rough draft are lower.
Measuring whether it worked
Go back to the timings you took in week one. Four to six weeks after rollout, time the same five tasks again. In the businesses I've worked with, a task that took twelve minutes typically drops to four or five once someone is comfortable with the tool and the playbook, not because the AI is magic but because half the twelve minutes was staring at a blank page deciding how to start.
Track two other numbers: how many people are still using it weekly at the six week mark (not day one, week six is where the truth shows up), and how many playbook prompts have been updated or added by staff themselves. That second number matters more than people think, because staff adding their own prompts to the playbook is the clearest sign the training has landed rather than just been attended.
Training in general only earns its cost if it changes behaviour weeks later, not just satisfaction scores on the day. That's true of AI training and it's true of any staff training, which is a point I've made before about why training programs are necessary in the first place: the goal was never the session, it was the Tuesday afterwards.
Common mistakes I keep seeing
- Training everyone in one big session instead of a small pilot first, so mistakes get repeated at scale instead of fixed early
- Choosing three different AI tools because different departments had preferences, which fragments the playbook and doubles the support burden
- No written policy on what data can and can't go into the tool, which either scares staff into not using it at all or, worse, means nobody's thought about it and something sensitive ends up pasted somewhere it shouldn't
- Treating the training as a one-off event rather than a living playbook that gets updated as staff find better prompts
- Skipping the "why" and going straight to the "how", so staff comply without understanding what problem this solves for them personally
If your business is growing fast and training keeps getting pushed to "next quarter" because everyone's underwater, that's worth naming honestly rather than pretending it isn't happening. It's the same pattern I see in the funding and growth conversations I write about in scaling with strain: the things that would make growth easier get postponed precisely because you're too busy growing, and AI training is usually one of them.
When to bring in outside help
If you've got under ten staff and one clear use case, you don't need external help, the four-week plan above will get you there. If you've got multiple departments, a compliance or data sensitivity issue, or you simply know your own team won't sit still long enough to build the playbook themselves, that's when working with an AI implementation coach earns its fee, because someone with no politics in the room and a repeatable process behind them will move faster than an internal owner trying to do it alongside their day job.
Frequently asked questions
How long does it take to train staff to use AI well?
Plan for four to six weeks from first audit to confident daily use, based on real timings before and after rather than a gut feeling. A single afternoon workshop rarely changes behaviour past the first fortnight, which is why the pilot-then-rollout structure works better than one big session.
Which AI tool should we train staff on?
Pick one main tool for most staff (ChatGPT, Claude, or Gemini are the realistic choices for a small or mid-size business in 2026) and stick with it rather than letting departments each choose their own. One tool means one playbook, one support process, and one thing to get good at instead of three things everyone half-knows.
What should we do if staff are resistant to using AI?
Resistance almost always means "I haven't seen this save me time on my actual work yet," not "I refuse to try new things." Sit with the resistant person for one real task from their own day and build the prompt together rather than showing them a generic demo. Resistance usually drops once there's a specific, personal payoff on the table.
Do managers need AI training too, or just front-line staff?
Managers need it more, not less. If a manager isn't visibly using AI in their own work, staff read that as permission to quietly skip it whenever things get busy, and adoption falls apart within a couple of months. Train the managers first and have them run the pilot pair sessions themselves.
Want this done for you? See how to build an AI workflow.
Related reading: Should You Let an AI Notetaker Into Your Client Meetings? and I Tested an AI Phone Receptionist for Three Months on a Real Small Business.
