Who in your company needs AI training first

Most companies buy one session for everyone. It is cheaper, it is easier to schedule, and it serves none of the four groups well.

Here is what each group actually needs.

Engineers#

What they need. Evaluation sets, cost control, and the failure modes that bite in production.

Not prompting. Engineers work out prompting on their own in an afternoon. What they have usually never done is build a set of real inputs with known-good outputs, cap a spend before the bill arrives, or design what the system does when it is wrong.

What goes wrong without it. They ship something that works in ten manual tests. The eleventh real user finds the edge. Nobody can tell whether last month's model change made it better or worse.

Format. Bootcamp, on your own codebase, with your own data.

Managers and team leads#

What they need. To judge whether a request is reasonable.

A manager is the person who will be asked for four hours of an engineer's week to build something. They need enough to ask two questions: is this a model problem or a rules problem, and how will we know afterwards whether it worked.

What goes wrong without it. They approve the thing that demos best, which is a different property from the thing that changes a number.

Format. Half a day, worked through real examples from your own roadmap.

Frontline teams#

What they need. A short list of tasks worth handing to a model, and a longer list that is not. Using their own documents and their own tools.

This is the group most companies train last. It is usually where the measurable change is, because they are the people doing the repetitive work every day.

What goes wrong without it. Two things, and they look opposite. Some people use nothing, because nobody told them it applied to their job. Others use it for everything, including the parts where being confidently wrong is expensive.

Format. Half a day, on their own tasks, followed by something in week three.

Executives#

What they need. To read an AI business case and find the invented number.

Not tooling. An executive does not need to prompt. They need to ask where a projection came from, and to recognise the answer "we benchmarked it" when it means a vendor's published figure rather than a measurement on your work.

What goes wrong without it. A budget gets approved on a slide, and eighteen months later nobody can say what it bought.

Format. Two hours, using your own business cases, including one that went badly.

Why not one session for everyone?#

Because the useful content for each group is boring or baffling to the other three.

Engineers switch off during "what is a large language model". Frontline teams switch off during evaluation harnesses. Executives switch off during both and leave early, which the room notices.

A single session solves a scheduling problem and creates an attention problem.

Which group first?#

Start with frontline teams, unless you are about to build something.

Two reasons. The change is measurable, because the tasks are repetitive and countable. And it is the cheapest group to get a real result from, which matters because the first result is what funds the next session.

Train engineers first only if a build is already scheduled. Then it is not training, it is part of the build.

Executives are worth two hours early, because they approve everything else.

What to do this week#

Write down which of the four groups you were about to buy a session for. Then write down which group does the most repetitive work in your company.

If those are different, look again. The second one is usually where a number moves.

What the market charges is here, how to check it worked is here, and the format question is here. If you want the split decided against your actual teams, that is a 30-minute call.