Operations
AI training for employees: a first-month plan for a small team
AI training for employees that actually sticks. A four-week plan for 5 to 30 people, one task per role, two volunteers first, and a check that holds.
Here's the version that doesn't work. It costs nothing to try, which is why everybody tries it.
You tell the team AI is fine to use. Maybe you pay for a few licenses. Two weeks later your estimator is quietly getting good at it. Two people tried it once. Everyone else is working the way they worked in March.
Nobody did anything wrong. The instruction was just impossible to follow.
"Try AI" isn't something a person can act on. "Draft the next five quotes with it and tell me whether it saved you anything" is.
Everything below is four weeks of turning the first kind of sentence into the second kind.
What training means here
Not a course. The courses near the top of this search are built for companies with a learning and development department. And they teach the tool.
Your people don't need the tool explained. They need to use it on Thursday's actual work. They need somebody to ask when the output looks wrong. And they need a reason to bother.
That's the whole plan, and the only thing it costs you is attention at the right four moments.
The four weeks
Week one: one task per role, two people
Pick one task per role. Specific enough to finish in one sitting.
Hand it over as a sentence somebody can act on, not as a topic. Not "have a go at it on follow-ups". Say: draft the next five follow-ups with it and tell me on Friday whether any of them came out worse than yours.
- Service manager. The follow-up message after a completed job.
- Office admin. Turning a voicemail and three texts into one clear job note.
- Estimator. The written explanation of why a quote came in where it did.
- Front desk at a practice. The recall message to patients who are overdue, with no patient detail pasted in.
- Account manager at an agency. The first draft of the monthly client update.
Then pick two people, and let them pick themselves. Ask who wants to try, and take the two hands that go up.
Two reasons for volunteers. They'll come back with the real problems instead of a polite nod. And when it works, word spreads sideways between your people, which is the only way it ever spreads.
Week two: the prompt sheet
By the end of week two you want one page that nobody is proud of.
It's got the prompts that worked, written out in full, with a line about when to use each one. Not clever prompts. The four or five boring ones your people will reuse forty times a month.
The four we end up writing most often, for whatever that's worth: turn these notes into a summary a customer can read, explain this quote to a homeowner, reply to this review in our voice, and rewrite this so a new person could follow it. Yours will be different. They won't be very different.
Keep it where the work happens and let anyone edit it. A prompt sheet that lives in one person's bookmarks is that person's skill, not your company's.
Week two is also when you teach the two rules that need judgment. What never goes into a tool, and what gets read before a customer sees it. Both fit on one page in the AI policy template. The harder call, which data is safe to paste, has its own list in what your team can put into ChatGPT.
Week three: somebody has to check it
This is where the month either sticks or quietly becomes a problem.
Decide who reads AI-drafted work before a customer, a supplier or an inspector sees it. Write the name down. Then check the things that cost you when they're wrong. Prices, dates, measurements, anything you're promising.
Pick the name carefully. It's usually not the owner, and it's usually whoever already catches things. One name, not a rota, because a rota means nobody.
Your team will find this annoying for about a week. Then they get fast at it, because they learn where the tool is reliable and where it invents. That's the actual skill, and no course can hand it to you.
Week four: the day-30 look
Fifteen minutes, in this order.
- Which task got genuinely faster, and who says so?The name matters. A task that got faster according to nobody in particular did not get faster.
- What did the check catch?Read one example out loud. One real caught mistake does more for week five than any amount of explaining why the check exists.
- What's the next task, and who's doing it?Decide it in the room. A next task chosen later is a next task chosen never.
Then write the next month on the same page. A team only gets a second month if they saw the first one end on purpose.
If nobody can name a task that got faster, don't run the second month on the same tasks. Run the same month on different ones. The plan is fine. The task was wrong.
Figure 1 / Matrix
Four weeks, each with one thing made and one thing checked, and week one is two people who put their own hands up.
The first month, week by week
| The question | Week one | Week two | Week three | Week four |
|---|---|---|---|---|
| Who is involved | Two people, and let them pick themselves. Take the two hands that go up. | Anyone can edit it, where the work happens. | One named reader of AI-drafted work. | The team, for fifteen minutes. |
| What gets made | One task per role, specific enough to finish in one sitting. | One page of the prompts that worked, with a line about when to use each. | A name, written down. | The next month, written on the same page. |
| What you check | That they come back with the real problems and not a polite nod. | That it lives where the work happens, not in one person's bookmarks. | Prices, dates, measurements, anything you're promising. | Which task got genuinely faster, and who says so. |
When training is the wrong fix
Training is one of three routes. It's the cheapest to start and the easiest to waste, so here's when to spend the month on something else.
| If this is true today | What a training month actually produces | Do this first instead |
|---|---|---|
| The work isn't written down | People asking a tool about a process it can't see | Get it out of the one head it lives in |
| Nobody checks anything today | More unread work, faster | Name the person who reads it before a customer does |
| Your people are on personal accounts | Skill on tools you can't see into or switch off | One page of rules, before week one |
| The person who needs it is you | A group session politely teaching around the problem | Do it on your own quotes, privately |
| The process itself is the problem | Twelve people running twelve steps faster | Look at the steps before you make them permanent |
Four of those five have their own page. Getting it out of one person's head is here, and the prompt and checklist for doing it take an afternoon. The accounts problem is solved by one page of rules. The last row is a different article altogether: what to automate first scores the process before anybody gets trained on it.
The thing nobody plans for
What to do about the person who won't.
Don't turn it into a loyalty test, and don't make it a meeting. Ask which part of their week they'd pay to never do again, and start there. There's a specific task behind the reluctance. Find the task.
If the real worry is "does this replace me", say the true thing out loud, whatever your true thing is. That conversation is yours and not a trainer's, and ducking it costs you the month.
If the true thing is that nobody's job changes here, say that sentence. If the true thing is that you don't know yet, say that one instead. What people can't work with is silence, and they will fill it with the worse answer.
Figure 2 / Flow
Behind a refusal there is a specific task, and if the worry is replacement that is the owner's sentence and not a trainer's.
The person who won't
Needs a person
Ask what they'd pay to never do again
Which part of their week. Not a loyalty test, and not a meeting.
Needs a person
Find the specific task behind the reluctance
Yours to say
Say the true thing out loud
If the worry is replacement. That conversation is yours, not a trainer's.
- Needs a person
- Affirmed
Is skill actually the problem?
Here's the question to answer before you spend four weeks of your team's time.
If AI isn't working in your business today, which of these is it?
- Your people can't use the tools well enough.
- The work isn't written down anywhere.
- Nobody reads the output before it goes out.
- It's all happening on accounts you can't see into.
Four different problems. From your chair they look identical, and only one of them is training.
The free AI readiness check is twenty questions, one tap each, about four minutes. Your finding and all five areas land on screen when you finish. It also names what it ruled out. That's the part that stops you spending a month on the wrong thing.
If it comes back that skill is the problem, this plan is the plan. If it comes back as something else, what to automate first or when to hire help beats a training month.
Figure 3 / Decision
Four problems look identical from the owner's chair, and the tree rules out three of them before it reaches training.
Which of the four problems is it?
Needs a person
Is the work written down anywhere?
- No
Outcome
Write it down first
- Yes
Needs a person
Does anybody read the output before a customer sees it?
- No
Outcome
Fix the check first
- Yes
Needs a person
Are your people on accounts you can see into?
- No
Outcome
Hand out accounts first
- Yes
Affirmed
Skill is the problem. This plan is the plan.
- No
- No
- No
- Needs a person
- Affirmed
The train route, plainly
Four weeks, your own team, on your own work. That's the train route, and it's the one we'd point a small team at first, because the skill stays in your building when we leave.
It's also the only one of the three where the thing you're buying walks around your building afterwards. That's worth something on the day a tool changes or a supplier puts its price up.
The other two are building the thing so the skill isn't needed, or running it for you. The honest way to choose is to find out which problem you've actually got.