Role-based AI training for knowledge work

Start from work, not a job title

Select a recurring task with a named output and user—for example, a customer-success manager preparing a renewal-risk brief. Observe how the task is currently performed, which sources are authoritative, where judgment occurs, what the worker may approve, and which decisions belong to someone else. Separate common AI foundations from the task-specific capability.

Interview the role holder, manager, workflow owner, data or security representative, and downstream reviewer. Their perspectives reveal different failure costs. A manager may care about consistency, while the downstream reviewer sees unsupported claims and the worker sees missing access or unrealistic time limits.

The UK government's 2026 employer guide to AI upskilling recommends practical, contextualized learning linked to real tasks and decisions. It also describes shared foundations with role-based modules and use-case libraries. The guide supplies design evidence, not proof that a particular training programme changes workplace outcomes.

Sources: [1]

Assemble the role-task-control bundle

Write the target task, acceptable output, user, source set, time constraint, AI actions, prohibited actions, review owner, stop condition, escalation, evidence to retain, and observable success anchors. Add at least one ordinary case, one ambiguity, and one material failure condition. Keep the scenario synthetic or use explicitly approved non-sensitive materials.

Teach the workflow in sequence: frame the assignment, select evidence, direct the assistant, inspect the output, run task-appropriate checks, resolve or escalate uncertainty, and deliver with limitations. Feedback should point to the observed step and evidence. Generic advice such as “be more critical” gives the learner little basis for improvement.

The NIST AI Risk Management Framework Core includes outcomes concerning roles and responsibilities, relevant training, human oversight, testing, feedback, and risk response. These outcomes help connect learning to governance, but they do not prescribe a universal curriculum or certify learner competence.

Sources: [2]

Worked artifact: a synthetic renewal-risk brief

In a synthetic exercise, a customer-success manager receives account notes, usage data, an unresolved support ticket, and an assistant-generated summary. The role-task-control bundle says the manager may draft a risk brief but may not promise a commercial concession or classify the customer as certain to churn. Material account claims require a source link, and contract interpretation must be escalated.

The first attempt repeats the assistant's claim that product use declined by 40%. The learner checks the date range and discovers that the current month is incomplete. They recalculate matched periods, label the trend uncertain, attach the support-ticket evidence, and route the contract question to the account owner. Feedback addresses denominator choice, claim strength, and authority rather than prompt style.

A transfer exercise uses a different account and replaces the incomplete month with conflicting meeting notes. The learner must apply the same principles through a different checking method. Passing one rehearsed case is treated as bounded evidence, not a universal capability claim. The example is synthetic and is not evidence of Two Prune customer outcomes.

Verify learning and transfer carefully

Compare performance against prewritten anchors and retain the deliverable plus available workflow evidence. Look for correct task framing, evidence use, appropriate AI direction, verification, escalation, and communication. Recheck on a varied task after time has passed; do not infer workplace transfer from course completion, satisfaction, or self-confidence alone.

Review the bundle whenever the tool, data, policy, workflow, authority, or failure consequence changes. Two Prune can create task-bounded assessment evidence in a controlled simulation for human interpretation. It does not certify a worker, establish future performance, or issue an employment decision.