Role-based AI training beats all-hands demos: a practical playbook for enterprise cohorts

You can book the all-hands in ten minutes. One hour, one demo, every function on the same call, and a recording for the people who missed it. The chat fills with thanks. The recording gets twelve views. On Monday the finance manager and the customer lead stare at the same generic prompt, and neither can see their job in it.

A demo shows people that a tool exists. I have yet to see one change how a role works. The cohort is large enough that one session feels efficient, and varied enough that one session teaches nobody's actual task.

Why it fades

People remember what they tried. In an all-hands they try nothing, or they try a party trick. You chose the examples to impress a mixed room, so the examples belong to no one. You wave at risk on a final slide, because a recruiter and a claims handler carry different risks, and the hour has no room to say so.

Senior people attend so others see them attend. Junior people will avoid a basic question in front of the whole company. The recording then becomes the programme. A recording cannot sit with someone while they try the task and tell them whether the output is fit to send.

Roles, then a cohort

Start with a handful of roles where the work is frequent, the data rules are knowable, and a manager cares about the output. A mid-market firm can skip a pathway for every job title in the HR system. Build one for the jobs that touch customers, money, code, documents or decisions.

Write one page per role before you book a room. What they produce in a week. What a good one looks like. What they must keep out of a tool, under AI governance. Which steps are draft, which are decide, and which are send. If you cannot fill the page, hold the session. The all-hands will still be easy to schedule. Each person will still leave without a task they recognise.

A useful cohort is the people who do the role, plus the manager who will see the work, small enough that everyone builds something. Eight to twenty. Go larger and you are back to a demo, with breakout rooms nobody visits. Bring last week's real task, and swap sensitive data for a stand-in if you must. Each person produces an output they could defend. The manager says whether they would send it. The group names the check that was missing.

One workflow with an owner

The AI training workshop is built this way. We stay vendor neutral, we work on the team's tasks, and we leave them with a short list of uses for that role. We skip the tour of the product.

The fortnight

Choose the role with a sponsor who will look at the outputs two weeks later. No sponsor, leave it. Write the one page. What the role produces, the quality bar, the data that stays in bounds under the OECD AI Principles and the step a human still owns. Collect five real tasks, stripped of anything the policy forbids, so the facilitator stops inventing the exercises.

Teach the smallest set of moves that role needs. Ignore the features that belong to a different job. Make every person produce one output in the room, and make the manager mark it before they leave. Agree one workflow they will run for ten working days, with an owner and a place the outputs live. Meet again for an hour. Keep it, change the check, or stop. Write the decision where the next cohort can see it.

That last meeting is how you stop collecting pilots. If a workflow is still called a pilot after ten days, either nobody runs it or nobody owns it. Both get expensive if you keep them on a slide.

What remains of the all-hands

Keep a short one for the story. Say which roles are in motion, what you allow, what you forbid, and where people ask for a cohort. Twenty minutes, led by someone who has done the work, with no live prompt theatre. Point people to the role sessions and to what a corporate AI training day should produce. The all-hands can carry the story. The practice has to live in the cohort.

Sequence the cohorts so they teach you. The second role should be easier to brief than the first, because you learned what the one pager missed. Share the data rules. Share the workflows only when the work actually matches. A lawyer and a marketer both use words. They need different exercises.

Measure by role. Did the named workflow run for ten days? Did a manager reject at least one output? Do you have a sentence that says what happens next? A company wide confidence score will dodge those questions. If you cannot answer them, you ran a demo. The cohort is where the practice sits, and the recording can wait.