AI upskilling is a facilitation problem

Mostly a tour. Here is the interface, here are some prompts, here is a use case from a different industry. It is cheap to produce, it scales to everybody, and it reliably produces the feeling that something was done.

The measured returns for real upskilling are large enough to be worth the distinction. Organizations already operationalizing AI expect further upskilling to lift productivity by more than 30%, against 10 to 30% among those still experimenting, across nearly 1,000 organizations surveyed by i4cp. Read it as what leaders closest to the work expect, not as a measured return, and note that the expectation attaches to capability building rather than to a tool walkthrough.

Judgment about scope. Not how to write a prompt, which people work out in a week, but when the output is acceptable, when it needs a human check, and when the task should never have gone near the model.

That is a decision-making skill exercised under time pressure, usually alone, usually when the person is already behind. It is the same category of skill as deciding whether a contract needs legal review, and nobody believes you can teach that with a demonstration.

Because judgment is only observable when somebody exercises it and is corrected. A demonstration shows a correct outcome. It never shows the participant their own incorrect one, which is the only moment that changes anything.

Run it the other way. Give the room a real task from their own work, have them use the tool, then put three outputs on screen and ask which one would have been safe to send. The disagreement in the room is the curriculum. That conversation cannot be recorded in advance, because nobody knows yet what this group will get wrong.

Ninety minutes, their own artifacts, three rounds. First round, a task the model handles well, to establish that the facilitator is not there to warn them off it. Second round, a task where the output looks right and is not, which is where most real exposure lives. Third round, the participants draft the rule they will apply on Monday.

That third round is the deliverable. A team that has written its own line is far more likely to hold it than a team issued a policy, and the written version gives the learning and development team something to circulate to people who were not in the room.

Directly, and the two are often built by different people who never meet. The learning and development team runs the enablement and somebody in HR or legal writes the acceptable use policy, usually months apart, usually without reading each other's drafts.

Sequence them together. Urban Human Resources takes the position on what AI should never decide in an HR context, which is the boundary the training has to teach rather than contradict. If the session tells people to move fast and the policy tells them to escalate, the session wins, because the session had a person in it.

No, and running one generic session for the whole company is the most common design error here. The risk profile is entirely different for a recruiter screening applications, a manager drafting a performance note, and an analyst summarizing a report.

Segment by what the person could plausibly get wrong rather than by seniority. TalentLMS finds 79% of HR managers now applying a skills-based approach to training and development, and this is the topic where that framing does the most work.

Not by completion, and not by a confidence survey taken immediately afterward, which measures enthusiasm. Watch the escalation rate: how often people bring a borderline case to a human before sending it.

That number rising is the program working. It looks like friction on a dashboard, which is why it needs explaining before the program starts rather than after. How to train your team on AI covers the wider rollout.

Building a learning and development plan this quarter?

Book twenty minutes and bring the capability you are trying to move.