You bought AI assistant licences for the whole office. Three months later, a handful of people use it every week, mostly to polish emails. The rest tried it once, got a confident wrong answer and went back to the old way. Then the renewal invoice arrives and nobody can say what changed.
This is the normal result, and it isn’t your people’s fault. A licence gives access. It doesn’t tell an accountant which part of month-end close the assistant can help with. It doesn’t tell a support agent whether they may paste a customer’s email into it. This post covers what actually changes how people work, and a 6-week plan to get there.
Why licences alone don’t change how people work
When adoption stalls, it’s usually for one of five reasons:
- People can’t see where it fits in their job. Generic training shows how to write a poem about the company. Nobody needs that on a Tuesday morning.
- The first try went badly. A vague question, a made-up answer, and the verdict is in. Most people don’t give a tool a second chance.
- The rules are unclear. If people don’t know what data they may put in, the careful ones stay away completely.
- There’s no time to learn. Any new tool slows you down for a week or two. Nobody plans for that week.
- Fear. Some people wonder whether getting good at this makes their own job unnecessary. If nobody talks about it, they draw their own conclusions.
None of these is solved by a better tool. They’re solved by how you introduce it.
Two ways to introduce an AI assistant
Licences only
- Licences sent by email
- One generic webinar
- Everyone figures it out alone
- Unclear data rules, so careful people stay away
- Nobody knows if it helped
Licences plus a rollout
- Two or three use cases per team
- A champion in each team
- A shared prompt library
- One page of data rules
- Usage and time saved checked in week 6
What “AI literacy” means in practice
The EU AI Act has a short article on this. Since February 2025, companies that use AI systems are expected to take measures so their staff understand AI well enough for the way they use it. It doesn’t prescribe a course, an exam or a certificate. What matters is that you can show what you did: who was trained, on what, and which rules they follow.
You don’t need to turn everyone into an expert. Three levels cover most companies:
Three levels of AI training
| Everyone | Regular users | Process owners | |
|---|---|---|---|
| What they learn | What AI is good and bad at, the data rules, when to check | Prompting for their own tasks, checking outputs, the prompt library | Where AI fits in their process, measuring results, handling errors |
| Format | One session for the whole company | Hands-on sessions with their own work | A workshop on their process |
| Time needed | Under an hour | A few hours over six weeks | A day or two |
The first level matters more than it looks. Everyone should know three things: AI writes plausible text, not checked facts; the person who sends the output is responsible for it; and some data never goes into an external tool. If people only remember that, you’ve avoided most of the expensive mistakes.
Start from tasks, not from the tool
“Here is the assistant, use it” rarely works. “Here are three tasks from your week where it saves time, and here is how” does. Ask each team a simple question: what do you do every week that involves reading, writing or searching? Then pick two or three tasks per team.
Some examples from ordinary companies:
- Accounting firm: summarising a long client email thread before a call, drafting reminder letters for missing documents, turning a list of booked entries into a plain-language note for the client. Not: asking the assistant for current tax rates. It may be out of date, so use the official source.
- Distributor’s sales team: product descriptions from technical data sheets, replies to standard price enquiries, meeting notes turned into a CRM update.
- Support team: suggested replies drafted from help-centre articles, a short summary of a long ticket before handing it to a colleague.
- Recruitment agency: job ad drafts and summaries of interview notes. Not: ranking candidates, which carries legal and fairness risks.
For each task, write down what a good result looks like and what must be checked before it’s used. If you need a method for picking tasks, see how to find the first AI use cases worth building.
Name a champion in each team
Pick the person others already go to with questions, not necessarily the most technical one. Give them two hours a week for the six weeks, early access to the tools and a direct line to whoever owns the rollout.
Their job is practical. They try the team’s use cases first, sit next to colleagues while they try them, collect prompts that work and report what doesn’t. A colleague who says “this is how I do the supplier reminder now” convinces more people than any training session.
Build a shared prompt library
A prompt that works is worth more than a training slide. Keep one shared page per team, in a document, your wiki or the saved-prompts feature of your tool. Each entry has the same five parts:
- The task, in the team’s own words
- The prompt, ready to copy
- An example input and the output it gave
- What to check before using the result
- Who added it, so others can ask
Start with ten entries and delete the ones nobody uses. For how to write the prompts themselves, see writing prompts that work.
Show and tell every two weeks
Thirty minutes, three people, one real example each. Include what went wrong. A colleague showing how the assistant invented a delivery date teaches more about checking outputs than a warning in a policy. It also gives quieter people permission to admit they’re stuck.
Talk about jobs honestly
Don’t promise that nobody’s job will change if you don’t know that. Say what you do know: which tasks you’re trying to speed up, what you plan to do with the time saved (serve more clients without hiring, answer customers faster, close the books earlier), and that people will hear about any change to their role before it happens. People handle honest uncertainty better than reassurance they don’t believe.
Make the data rules clear before day one
Before anyone starts, they need one page that says which tools are approved and what data can go into them. Without it, careful people won’t use AI at all and careless people will use it with anything. We wrote a separate guide with a template: writing an AI usage policy.
The 6-week plan
A 6-week rollout
- 1
Week 0
PrepareApproved tools set up, one page of data rules, two or three use cases per team, champions named.
- 2
Week 1
Kick-off for everyoneUnder an hour: what AI is good and bad at, the data rules, a live demo on a real task.
- 3
Weeks 2–3
Hands-on in each teamChampions run short sessions on the team’s own use cases. First ten prompts go into the library.
- 4
Week 4
Show and tellTeams show what works. Drop use cases nobody uses.
- 5
Week 5
Fix the frictionAccess problems, missing data, tasks that need a proper workflow rather than a chat window.
- 6
Week 6
Measure and decideUsage, time saved on sampled tasks, quality. Decide on licences and the next use cases.
Week 5 is the one most companies skip. By then you’ll know which tasks are awkward in a chat window: the same request typed 200 times a week, or work that needs data from two systems. Those are signs that the task belongs in an automated workflow, not in more training. Our post on workflows, AI workflows and agents explains the difference.
What to cover in the kick-off
Keep the week-1 session short and concrete. A plan that fits in under an hour:
- Ten minutes: what it is. It predicts plausible text from patterns. It drafts, summarises, extracts and translates well. It doesn’t know your company unless you give it the material, and it can state wrong things confidently.
- Ten minutes: the rules. Approved tools, green, amber and red data, who checks what. Show the one-page policy on screen.
- Twenty minutes: live demo. Take a real task from the audience, such as a messy customer email, and do it live. Show a bad prompt, then a good one. Show one mistake and how you’d catch it.
- Ten minutes: what happens next. Who the champions are, where the prompt library lives, the date of the first show and tell.
Record it for people who join later.
Measuring adoption
Logins tell you little. Someone who opens the tool to rephrase one email a week counts as active. Measure these instead:
- Weekly active users per team. Most business plans show this in the admin console.
- Use cases in regular use. Ask in a two-minute survey which tasks people now do with AI at least weekly.
- Time saved on specific tasks. Time a sample of tasks before and after, rather than asking people to guess.
- Quality. Errors found in checks, corrections from colleagues, complaints from clients.
Weekly AI use by team after six weeks
Read a chart like this as a set of questions, not a ranking. Support may be high because suggested replies fit their work well. A low number in the warehouse may simply mean a chat assistant is the wrong tool for that team. Their gains are more likely to come from automating a process, such as reading delivery notes, than from more training.
Pitfalls to avoid
- Training once. People forget a one-off session within weeks. Short, regular sessions with real work stick.
- Forcing use. “Everyone must use AI every day” produces AI output nobody needed. Ask for results, not usage.
- Measuring logins. See above.
- Treating every task as a chat task. High-volume, repetitive work belongs in a process with AI built in, not in hundreds of copy-and-paste sessions.
- No owner after week 6. The prompt library goes stale and new joiners never hear about it. Name someone to keep it alive.
Before you start
Ready for a rollout
- Approved tools set up with business accounts
- One page of data rules What can go into which tool
- Two or three use cases per team Taken from real weekly work
- A champion in each team, with time set aside
- A shared prompt library, even if it starts empty
- A baseline for a few tasks How long they take today
- A date for the week-6 review
Adoption is the cheapest part of an AI project to get right, and the most often skipped. If you’re planning the bigger picture, our guide on bringing generative AI into your company puts training alongside choosing use cases, tools and measures. If some of your teams need AI built into a process rather than a chat window, tell us about the process.
