Coach AI-enabled work through a real task, not a prompt demonstration

Diagnose the work and its constraints before teaching a technique

Consider a synthetic internal communications team that drafts a weekly update from approved project notes. The team has an authorized AI workspace, but a new employee keeps producing polished updates that merge tentative milestones with approved dates. The coaching need is not more fluent wording. It is a way to identify approval status before asking for a draft, preserve the link between each statement and its source, and hand the final update to the accountable editor.

At intake, ask the worker for the intended audience, deliverable, deadline, source pack, permitted tool and data rules. Find one recurring failure that can be safely practiced with synthetic or approved non-sensitive material. Agree on an observable outcome: for this task, a short update whose dates and claims can be traced to approved notes, with uncertain items explicitly marked for review. If the tool or source permission is unclear, stop that particular exercise and seek the actual owner's decision; coaching does not grant access rights.

NIST's AI RMF Core includes workforce training for assigned responsibilities and defined roles for human oversight. The UK Government's AI Playbook discusses practical capability and responsible use in its public-sector setting. These sources support the importance of context, training and oversight; neither prescribes the coaching sequence here for every workplace.

Sources: [1] [2]

Demonstrate, transfer, then watch the worker make a judgment

The coach first demonstrates on a small synthetic packet: one confirmed date, one tentative date and one contradictory note. They show how to ask an assistant for a draft that labels each proposed claim with its source, then check those labels against the original packet. They deliberately reject an attractive sentence that turns the tentative milestone into a promise. The worker then repeats the exercise on a different approved packet while the coach observes only the relevant work steps. Coaching should not turn into the coach secretly correcting the final answer or treating prompt memorization as competence.

The transfer artifact for the synthetic team is a one-page “claim handoff.” It contains the draft sentence, source note and version, approval state, check performed, open question and named editor. Example: “Regional launch is 15 October” maps to a planning note labelled tentative; it is changed to “The regional launch date remains under review,” pending owner confirmation. A second sentence about a completed internal test maps to an approved status note and can remain with that scope. The worker records the assistant's suggested wording separately from the evidence that authorizes publication. The coach looks for whether the worker caught both the unsupported certainty and the supported statement, not only whether the final prose sounds smooth.

Arrange follow-up around observed friction

At the end, the worker identifies what they can repeat alone, one verification step they are still likely to miss, and the actual help route if the next task differs. A follow-up can use a new synthetic packet with a changed audience or a missing approval marker; it should reveal whether the worker transfers the source-to-claim check rather than reusing a fixed prompt. The coach records a narrow observation such as “checked both dates against the supplied approval labels in this exercise,” not a global AI-skill certification.

NIST's Govern playbook asks organizations to consider whether people have the training, resources and domain knowledge needed for their responsibilities. The UK Open Innovation Team's AI Adoption Unit describes hands-on coaching linked to real tasks in its civil-service programme. The worked handoff above is a proposed practice for this fictional case, not a reported result of either programme or a claim of measured productivity gains.

The failure to avoid is a one-off demonstration that produces a good-looking artifact but leaves the worker unable to identify a false date or request help in the next situation. This page concerns supportive practice and escalation, not a training-completion metric, platform policy or proof of workforce transformation. Two Prune can create a comparable synthetic assessment task and report task-bounded evidence for human interpretation, but it does not certify coaching effectiveness or decide anyone's employment outcome.

Sources: [3] [4]