Practical guide

How to Create an AI Capability Programme

An AI capability programme should prepare people to perform named work with appropriate judgment, verification, and escalation. Start with the work and its risks, not a catalogue of tools or courses. The useful output is a controlled programme specification that connects tasks, learning experiences, workplace support, evidence, owners, and review decisions.

By Two Prune

An AI capability programme should prepare people to perform named work with appropriate judgment, verification, and escalation. Start with the work and its risks, not a catalogue of tools or courses. The useful output is a controlled programme specification that connects tasks, learning experiences, workplace support, evidence, owners, and review decisions.

Start with a task-and-decision inventory

List the decisions and deliverables that people are expected to handle with AI assistance. For each one, record the user group, system or model, permitted information, required sources, reviewer, consequence of error, and conditions that require escalation. Separate general AI orientation from the capabilities needed for a particular workflow.

NIST's AI RMF Core says that roles, responsibilities, and lines of communication should be documented and that personnel should receive training that enables them to perform their duties. Treat those outcomes as programme inputs rather than proof that a training event was effective. NIST AI RMF Core

Use a needs assessment to distinguish a skill gap from an access, policy, process, workload, or management problem. The U.S. Office of Personnel Management notes that training is not always the best or only response to a performance gap. OPM planning and evaluating guidance

Sources: [1] [2]

Write a capability contract for each workflow

Turn each priority workflow into a one-page contract. Name the task, inputs, allowed AI assistance, human decisions, required checks, expected artifact, principal failure, escalation route, and acceptable evidence. Define observable behavior rather than labels such as beginner or advanced.

For example, “uses AI well” is not testable. “Can compare a generated customer response with the approved policy, identify an unsupported exception, correct it, and retain a source-linked review record” is. The contract should also state what the observation cannot establish outside the bounded task.

The European Commission's current AI-literacy Q&A recommends adapting literacy activity to the organization's role, the systems and risks involved, people's prior knowledge, and the context of use. It also says that no single format fits every situation. That supports differentiated programme design; it does not make a particular curriculum sufficient or certify compliance. European Commission AI literacy questions and answers

Sources: [3]

Build a learning-and-support sequence

Design a sequence around the capability contracts:

The OPM Training Evaluation Field Guide recommends aligning training with performance, preparing learners for critical behaviors, building workplace drivers, and designing evaluation before delivery. Use those principles as design prompts; following them does not establish transfer. OPM Training Evaluation Field Guide

  • Orientation establishes permitted use, data handling, model limitations, accountability, and escalation.
  • Demonstration shows a complete task, including source checks and rejected AI output.
  • Deliberate practice varies evidence quality, ambiguity, and failure conditions.
  • A realistic simulation asks the participant to produce a work artifact under stated constraints.
  • Workplace support supplies checklists, source registers, review routes, and refreshers.
  • Follow-up evidence asks whether the intended behavior appears in later work, without assuming that training caused every change.

Sources: [4]

Assign programme decisions and evidence

Create a programme register with one accountable owner for scope, policy, content, system access, practice cases, assessment, operational support, privacy or security review, and revision. Contributors may be numerous, but the decision authority and escalation route should be explicit.

Keep evidence layers separate. Attendance shows participation. Knowledge checks show performance on those items. Tool telemetry shows recorded activity under a specific configuration. A simulation provides task-bounded evidence. Manager observation adds contextual evidence but may vary by opportunity and standard. Operating quality or incident measures answer still different questions. Do not collapse these into one adoption or capability score.

Set missingness rules before launch. A participant who never encountered a contradiction did not demonstrate how they handle one. A missing event may be an instrumentation failure rather than poor behavior. Report unavailable evidence as unavailable and state what follow-up would reduce uncertainty.

Worked programme: customer-guidance review

Suppose a service team wants 24 advisers to use an internal assistant when drafting customer guidance. The initial request is for a two-hour prompt course. The task inventory finds three material decisions: selecting the governing policy, deciding whether an exception applies, and approving the final guidance. Access logs also show that six advisers cannot reach the current policy library, so training alone cannot solve the problem.

The programme owner creates three capability contracts. The highest-risk contract requires an adviser to compare a draft with two policy sources, flag a fabricated exception, correct the response, and route a genuinely ambiguous case to the policy owner. Learners receive a short orientation, a worked demonstration, two practice cases, and a timed synthetic case with conflicting documents. A source-and-approval checklist remains available during later work.

The release dashboard records access fixed for six advisers, 24 orientations, 22 completed practice sets, 20 completed simulations, and 16 later reviewed work samples. It does not combine these counts into a single capability percentage. Two simulation records lack a complete event stream, so their workflow evidence is marked limited. A four-week review compares common failure modes and decides whether to repair the source-selection exercise, the policy library, or the escalation route.

Apply an acceptance and revision gate

Before launch, verify that every priority workflow has a current capability contract, accountable owners, source-controlled materials, realistic practice, a verification method, missingness rules, participant instructions, and an authorized escalation path. Record unresolved risks and the person accepting them. Schedule review when the system, policy, task, evidence source, or observed failure pattern changes, not only on a fixed annual date.

AI can help compare contracts, find inconsistencies, generate practice variations, and summarize anonymized error patterns. A human owner must approve the task, evidence, claims, and changes against current organizational requirements.

Two Prune can provide a controlled work simulation and structured, confidence-aware evidence about observed performance in that task. It does not run an organization's entire capability programme, certify AI literacy or legal compliance, prove workplace transfer, establish safer outcomes, or make employment decisions.

Sources

  1. 1.NIST AI RMF Core · National Institute of Standards and Technology
  2. 2.OPM planning and evaluating guidance · U.S. Office of Personnel Management
  3. 3.European Commission AI literacy questions and answers · digital-strategy.ec.europa.eu
  4. 4.OPM Training Evaluation Field Guide · U.S. Office of Personnel Management

Explore this topic