From compliance AI training to judgment at work: teaching teams to challenge model output

The last slide asks people to tick a box. They have read the acceptable use note. They will keep customer data out of a public tool. They tick the box before the kettle boils. The learning record shows a pass.

That record has a job. It tells an auditor the organisation said the rule out loud. It says nothing about whether anyone on the floor can look at a confident paragraph and decide the paragraph is wrong. Acknowledgement and judgment ask for different skills. Only the second one changes the email that leaves the building.

What the module is for

Compliance training draws a line. Here is the data you must leave where it is. Here is the tool we approved. Here is who you call when you are unsure. People need that line, especially on a regulated team, and they need it in plain language. If only the legal team can parse the policy, people will nod and then ignore it.

The module falls over when you ask it to teach judgment. The NIST AI Risk Management Framework is the reference I use when that boundary has to be written down. Judgment means a person stops, checks the claim against something outside the model, and will send the draft back. A tick box can confirm that someone saw the line. It cannot rehearse the habit, because the habit only shows up when the draft is theirs.

You can see it in the artefacts. Compliance leaves a completion report. Judgment leaves a marked-up draft, a question in the margin and a decision to use the output or drop it. If the programme only produces the completion report, you have trained acknowledgement.

Tuesday afternoon

An analyst asks a model for the risks in a contract summary. The model returns three, neatly numbered, in the house style. Two sit in the document. The third sounds plausible and the document never says it. The useful move is opening the source before those risks go into the email, because they sound like the sort of thing the document would say.

A customer lead asks for a reply to a complaint. The draft is warm and specific. It also offers a remedy the company never gives. Catch the offer before it leaves the building. A manager asks for a briefing on a competitor. The briefing cites a figure nobody in the room has seen. Leave the figure out of the pack until someone can point at it.

You can still like the tool while you do this. Treat it as a fast colleague who is sometimes sure and sometimes making it up. The colleague stays fine when you check. A lot of people still think a check means they are bad at the tool, so say the opposite out loud, in the session and again in the template.

A move they can repeat

Telling people to be more critical gives you a poster. Give them a move, and make them use it on their own work. Before a second reading, they write what would have to be true for this output to be safe to send. They check every name, number, date and promise against a source the model did not write. They mark the sentence they would hate to defend, and they rewrite that one themselves. If they cannot find the source in two minutes, they leave the claim out.

Keep one example a week of a catch, and one of a miss, so the team can see the pattern. A catch that stays in one person's head never becomes a team skill. Punish a miss and people hide the tool, or they hide the error. You want both on the table, where the manager already looks.

Use the real draft

A quiz will never give you judgment. The quiz has a known wrong answer and a hint. Take the last ten outputs a team actually used. Strip the names if you must. Sit for an hour and mark each one as used on arrival, edited, or rejected. Then ask why. Untouched outputs usually cluster on low stakes work, and the edit is often a tidy of the sentences rather than a check of the claims. That hour beats another module on the limits of models, because the limits are sitting in their documents.

Then change the template they already use. If a draft goes to a customer, the sender names the source they checked. If a number goes to a leader, the number has an owner. If the model proposed an action, a person records the yes. Make the unchecked path slightly harder than the checked one, or the unchecked path will win on a busy afternoon.

The critical thinking workshop is where we put a team through that rep, on a problem they brought. A slide about bias in the abstract will never substitute for the hour.

The next cycle

Keep the compliance module. Shorten it if people are clicking through. Then spend the hours you had reserved for a second generic course on judgment reps inside two or three roles. Pick roles where a wrong output is visible: advice, numbers, customer commitments, anything that leaves the building. Pair each cohort with the manager who will see the work. A learning team cannot grade judgment from a completion file. The manager can, with the moves and a fortnight of examples.

Count drafts people checked against a source. Count claims they pulled. Count outputs that reached a customer with no human edit, and decide whether that was a rule or an accident. A confidence score on its own will hide this. People can feel sure and still forward the tidy wrong answer.

Fear sends people back to the old process, or into tools nobody can see. Practise the reply in a psychological safety workshop before you add another module. Keep the boundary in the compliance module, and rehearse it in the AI governance workshop. Spend the next cycle on the habit inside it. You want a person who opens the source when the draft sounds finished, and who says so in the document the manager will read.