The conversation about AI training usually stops at prompting courses. Those are useful for about a month and then irrelevant, because models keep improving and less and less technique is needed to ask them for things. What does not expire is a different question: what does a person do well once the mechanical part of the job is no longer theirs.
We have been noting down what separates teams that get value out of AI from teams that merely have it installed. It comes to five capabilities, and none of them is technical.
1. Knowing what to ask
A machine answers a well-framed question very well, and answers a badly framed one with total confidence. The difference comes from the person who knows what they are actually looking for. It is an old skill — the skill of a good project manager or a good doctor — now applied hundreds of times a day.
2. Judgement to review
Accepting or rejecting what a system produces requires being able to tell what is right from what merely sounds right. That rests on craft: if nobody on the team has ever done that work by hand, nobody will be able to review it with judgement. It is the strongest argument we know for not hollowing out entry-level roles completely.
3. Deciding with incomplete information
When analysis gets cheap, it stops being the bottleneck. The bottleneck moves to the decision: choosing between two reasonable options knowing that data is missing and is not going to arrive. Many organisations discover their problem was never a shortage of reports, but a reluctance to decide with the ones they had.
4. Explaining a decision
If a client, an auditor or a colleague asks why something was done this way, “the system said so” is not an acceptable answer in any serious sector. Being able to reconstruct and defend a decision in plain language becomes part of the job, not an extra.
5. Difficult conversations
Every transformation reaches a moment where you have to talk about job fears, tasks that disappear and people who have spent fifteen years doing something that no longer needs doing that way. Middle managers who can hold that conversation carry the change. Those who avoid it sink it, however impeccable the technical project.
Nobody resists giving up a boring task. They resist not knowing what comes next.
How you train this
Not with a generic course. What works is working on the real tasks of each role, with the company's own documents and cases, in short sessions spread over several weeks, with follow-up afterwards. People do not drop off during the training; they drop off three weeks later, when the normal workload comes back.
And one detail changes the outcome a lot: saving and sharing the examples that work, in-house. One person who solves something well with AI and writes it down saves twenty other people the same discovery.