How to train your agency team to use AI

People learn AI by doing, not by reading

When a founder decides the team needs to "get good at AI," the instinct is to reach for the familiar tools of training: write a guide, buy a course, schedule a webinar, share a long doc of best practices. It feels responsible, and it almost entirely fails. Passive learning — reading a PDF, watching someone else demo — produces recognition ("I've seen that") but not capability ("I can do that"). And capability is the only thing that drives adoption. A team that has watched a great AI training video and a team that can actually use AI in their daily work are two completely different teams, and the gap between them is practice on real tasks.

AI is a skill, and skills are built the way all skills are built: by doing the thing, getting a result, seeing what worked and what didn't, and adjusting. You didn't learn to ride a bike from a manual, and your team won't learn to use AI from one either. So the entire approach to training should be reorganised around doing rather than watching. Here's how to run training that actually changes how your team works.

A hands-on training approach

1. Teach concepts briefly, then get practical

There's a small amount of conceptual grounding worth doing first, because it reduces fear and sets realistic expectations. Spend about fifteen minutes — not more — on what AI is genuinely good at (great first drafts, summarising, reformatting, brainstorming) and what it's bad at (being confidently wrong on facts, names, numbers, anything requiring real-world verification). This framing does important work: it tells people the tool is a capable assistant that needs supervision, not an oracle, which both calms the "it'll replace me" fear and inoculates against the "it told me X so I shipped X" failure. Then stop talking and move to doing, because the concepts only really land once people apply them.

2. Train on one real task, not "the tool"

Don't teach the software in the abstract — teach a specific workflow your team actually does, using real materials. A real transcript, a real client brief, a real messy set of notes. Generic training examples don't transfer to the job, because the gap between "summarise this sample article" and "turn this client's rambling feedback into a clear action list" is exactly where people get stuck. When you train on the actual task with the actual kind of input your team faces, the training transfers directly to the work, and people leave able to do the real thing, not a toy version of it. Pick the same low-stakes, repetitive workflow you'd choose for a first rollout: high frequency, real time saved, low risk.

3. Do a live walkthrough, then have everyone try

Run the training live, not pre-recorded. Demonstrate the task end to end, narrating your thinking as you go — why you phrased the prompt that way, what you're looking for in the output, and crucially, where the AI gets it wrong and how you catch and fix it. Watching you spot a hallucination and correct it teaches the single most important habit more vividly than any rule: treat output as a draft to verify, never an answer to trust.

Then comes the part most training skips and the part that actually creates capability: have every person do the same task on their own example before the session ends, with you there to unstick anyone who gets stuck. Doing-in-the-room is where the learning happens. Everyone should leave the session having personally produced one good result. "Try it on your own time later" is where training evaporates; "do it now while I'm here" is where it takes root.

4. Build a shared prompt library

A huge amount of AI skill is really just knowing the prompts and patterns that work for your specific tasks — and if that knowledge lives only in one person's head (usually your champion or you), it doesn't scale and it's lost the day that person is busy or leaves. The fix is simple and high-leverage: capture the winning prompts in a shared, living document organised by task. "Here's the prompt that turns a transcript into show notes in our format." "Here's the one that drafts a client update from a status doc." This turns one person's hard-won skill into a team capability, makes results consistent across people, and dramatically shortens the learning curve for anyone new. The prompt library is one of the highest-return artifacts an AI-using agency can maintain.

5. Follow up at two weeks

Training isn't an event; it's the start of a process. About two weeks after the session, skills have started to fade and real questions have piled up — the edge cases people hit once they're using AI on live work that didn't come up in training. A short follow-up at that point — what's working, what's confusing, what new prompts have you found — is where proficiency actually locks in. It catches the people who quietly stalled, surfaces new winning prompts for the shared library, and signals that this is a real capability the agency is investing in, not a one-off box-tick. Skipping the follow-up is the most common reason training produces a brief flurry of usage and then nothing.

Make AI skills feel personal, not mandated

There's a motivational layer worth getting right. People invest far more effort in learning AI when they see proficiency as a career asset — something that makes them personally faster, more capable, and more valuable — rather than as a company requirement they're being made to satisfy. The framing you choose matters. "This makes you faster and more valuable, and these are skills that compound for your whole career" produces genuine investment; "we need everyone to hit an AI usage target" produces resentment and the bare minimum. Frame training as an opportunity you're giving your team, not an obligation you're imposing, and you'll get real effort instead of compliance. (More on the motivation side in getting your team to use AI.)

Common training mistakes

A few patterns reliably waste your training effort. Front-loading theory — long conceptual sessions with no doing, so nothing transfers. Training on toy examples instead of real client work, so the skills don't carry over to the actual job. One-and-done training with no two-week follow-up, so the early questions never get answered and usage fades. Letting prompts stay tribal in one person's head instead of capturing them. And framing it as a mandate, which kills the intrinsic motivation that drives real skill-building. Avoid these five and your training will do something most AI training never does: change how the work actually gets done.

Give the team one place for tools and knowledge

Training sticks far better when the things it produces — the prompt library, the AI policy, the approved workflows — live in one findable place rather than scattered across Slack threads, shared drives, and people's memories. When someone two weeks past training hits a question, the difference between "I know exactly where the prompt for this lives" and "it's in some doc somewhere" is the difference between continued use and quiet reversion.

That's one of the things Forge builds: internal team portals shaped to your agency, a single branded homebase where SOPs, the prompt library, the AI policy, and the tools themselves live together. When knowledge and tools share a home that the team actually uses, training compounds instead of decaying. See how it works →