Train-the-trainer programmes for AI-enabled work

Define what trainers must facilitate

Choose a bounded workflow that learners genuinely perform. Describe its output, user, authoritative sources, AI contribution, retained human judgment, prohibited actions, common failures, and accountable owner. Separate shared foundations from local role details so trainers know what must remain consistent and what can be adapted.

Specify trainer capabilities as observable actions: setting up the scenario, explaining boundaries, recognizing evidence use, distinguishing a correction from superficial editing, giving feedback tied to behavior, handling uncertainty, escalating policy questions, and offering accessible participation routes. Do not treat presentation confidence as evidence of these capabilities.

The UK government's 2026 employer guide to AI upskilling emphasizes practical, reachable, integrated, modular, expandable, and sustainable learning. It supports role-linked practice, local adaptation, feedback, accessibility, and refresh cycles, while local outcomes still require direct evidence.

Sources: [1]

Assemble the facilitator calibration pack

Include the learner brief, approved tools, synthetic or authorized materials, timing, expected artifact, decision rights, seeded ambiguity, verification route, stop conditions, observation form, scoring anchors, feedback examples, accessibility alternatives, escalation contacts, version history, and change log. Mark which statements are instructions and which are examples.

Run calibration with multiple trainers before delivery. Each trainer observes the same sample, records evidence independently, chooses feedback, and identifies any escalation. Compare where they noticed different behaviors or interpreted anchors differently. Revise ambiguous materials, then repeat with a new sample rather than coaching everyone to remember one answer.

Include delivery boundaries in the pack. Trainers should know which questions they may answer, which require a workflow owner, and which must go to privacy, security, legal, accessibility, or employee-representative reviewers. Give them a safe way to pause an exercise when materials, access, or instructions create an unfair or risky condition.

The NIST AI Risk Management Framework Core includes outcomes concerning roles, relevant training, human oversight, documented testing, feedback, and risk response. These outcomes help connect trainer practice to organizational governance but do not certify a facilitator or prescribe one curriculum.

Sources: [2]

Worked artifact: a synthetic evidence-checking session

Imagine three facilitators preparing to teach analysts how to verify AI-assisted briefings. The calibration sample contains a correct claim, an outdated source, a denominator error, and an unresolved policy question. One facilitator rewards the final polished memo, another counts every browser visit, and the third records whether the learner chose a suitable check and changed the recommendation.

The pack's anchors are revised to focus on observable decisions: identifying the material claim, selecting the authoritative source, rebuilding the denominator, communicating uncertainty, and escalating the policy issue. A second sample changes the domain and hides the error in a chart label. The facilitators independently reach closer evidence-based judgments and give feedback tied to the observed step.

One facilitator still treats an unobserved check as a failure. The programme adds a missing-evidence rule and a follow-up question rather than awarding or removing credit without evidence. This example is synthetic and does not establish trainer effectiveness or Two Prune customer outcomes.

Verify delivery and refresh the system

Observe live sessions, sample trainer records, review learner questions, inspect escalations, and check whether feedback references evidence rather than personality. Compare calibration decisions over time and after material changes. Refresh the pack when the tool, source set, role authority, policy, learner population, or failure consequence changes.

Track delivery separately from transfer. Attendance, trainer coverage, and learner satisfaction can describe reach and experience; they do not by themselves show that people apply the behavior later. Use a varied follow-up task, record missing participation, and avoid turning a small local exercise into a broad effectiveness claim.

Two Prune can create task-bounded evidence about participant behavior in controlled simulations for human interpretation. It does not certify trainers, prove learning transfer, or make employment decisions.