Practical guide
AI academy curriculum: what to include
An AI academy curriculum should prepare people to make better decisions in real work with AI, not merely complete tool demonstrations. A useful curriculum connects role-specific tasks, evidence handling, AI collaboration, verification, judgment, delivery, governance, and repeated practice. Its measurement plan should distinguish attendance and confidence from demonstrated performance. This guide.
An AI academy curriculum should prepare people to make better decisions in real work with AI, not merely complete tool demonstrations. A useful curriculum connects role-specific tasks, evidence handling, AI collaboration, verification, judgment, delivery, governance, and repeated practice. Its measurement plan should distinguish attendance and confidence from demonstrated performance. This guide uses WORKS: workflow outcomes, operating rules, representative practice, knowledge and judgment, and sustained evidence.
Source basis and limits
The UK Government's July 2026 employer guide says AI upskilling should be practical, reachable, integrated, modular, expandable, and sustainable. It recommends linking training to real tasks and decisions, combining technical, non-technical, and responsible skills, and revisiting learning over time. The companion AI Skills Tools Package separates foundational, operational, and strategic needs and is explicitly a starting point to adapt. NIST's AI RMF Core calls for defined human-AI roles, context-specific risk management, and documented testing and evaluation. These sources inform curriculum controls; they do not determine local learning needs or establish that a particular course transfers to workplace performance. Employer guide AI Skills Tools Package NIST AI RMF Core
Sources: [1] [2] [3]
1. W — Workflow outcomes
Start with work, not a catalogue of features. For each audience, document:
| Field | Curriculum decision | | --- | --- | | Representative task | A recurring, consequential task with available evidence | | Required deliverable | What useful output must be produced | | Decisions and constraints | What the learner must judge, preserve, or escalate | | Principal failure | The most important realistic error | | Evidence available | Documents, data, messages, policies, and external sources | | Transfer boundary | Contexts where the practice may not apply |
Choose a small set of task families that reveal different capabilities: framing an ambiguous request, finding and weighing evidence, collaborating with AI, checking claims and calculations, resolving tradeoffs, and communicating a decision. A generic prompting lesson can support these tasks, but it should not replace them.
2. O — Operating rules
Every module should make local rules usable inside the task:
Turn rules into decisions. Instead of asking learners to recite a policy, give them a source packet containing a restricted data field, an attractive unsupported claim, and a deadline. Ask them to select a safe route, explain the tradeoff, and preserve an audit trail.
- approved tools, data classes, and access boundaries;
- when disclosure or attribution is required;
- decision rights and human review responsibilities;
- risk tiers, escalation triggers, and incident routes;
- accessibility, privacy, security, intellectual-property, and records requirements;
- what evidence must be retained.
3. R — Representative practice
Use a progression in which the support changes while the core work remains visible:
Include both appropriate use and deliberate non-use. Learners should sometimes decide that AI adds little value or that the available tool cannot handle the information safely.
- Demonstrate: an instructor models task framing, source use, prompting, verification, and revision aloud.
- Guide: learners complete a short case with checkpoints and targeted feedback.
- Vary: change a constraint, source quality, audience, or tool behavior so memorized steps are insufficient.
- Perform: learners complete a realistic simulation with bounded assistance and a defined deliverable.
- Reflect: learners identify what they trusted, checked, changed, rejected, and would escalate.
4. K — Knowledge and judgment spine
Build common foundations, then adapt examples and depth by role.
| Module | Learner must be able to do | Evidence artifact | | --- | --- | --- | | Task and context framing | Define audience, outcome, constraints, and success | Task brief with assumptions | | Evidence navigation | Find, compare, and cite relevant sources | Claim-to-source ledger | | AI orchestration | Decompose work, prompt, critique, and iterate | Interaction and revision record | | Verification and risk | Check claims, calculations, provenance, and uncertainty | Verification log and escalation | | Judgment and synthesis | Compare options and explain tradeoffs | Decision record | | Communication and delivery | Produce a useful, audience-fit deliverable | Final artifact and rationale | | Governance in practice | Apply local rules and decision rights | Control and exception map |
Do not isolate responsible use in a final compliance module. Put a relevant risk, source-quality problem, or authority question inside each applied exercise.
5. S — Sustained evidence and improvement
Predeclare what the academy will observe and what each measure cannot establish.
| Evidence layer | Example | Interpretation limit | | --- | --- | --- | | Reach | enrolment, access, completion | does not show learning | | Experience | relevance, accessibility, confidence | self-report is not demonstrated capability | | Learning | case performance and error patterns | may be specific to the case | | Transfer | later representative task sample | affected by tools, managers, workload, and opportunity | | Operations | incident or quality indicators | cannot isolate training without comparison and context |
Use the same scoring criteria before and after a module where feasible, retain missingness and reviewer disagreement, and review curricula when tools, policies, work, or failure patterns change. A later task sample is more informative than a satisfaction survey alone, but it still needs a clear comparison and local context.
Worked curriculum: fictional client-update pathway
A professional-services team wants an academy module for AI-assisted client updates. The original syllabus contains a 90-minute tool tour, prompt library, quiz, and completion badge.
WORKS replaces this with a three-part pathway. In the demonstration, an instructor turns a messy source pack into a bounded update, showing how to distinguish confirmed progress, forecast, risk, and decision requests. In guided practice, learners receive contradictory milestone dates and an AI-generated citation that does not support its claim. They must reconcile sources, label uncertainty, and ask an authorized owner to resolve the conflict.
In the performance simulation, learners prepare a two-page update for a client sponsor under a time limit. The source packet includes a correct aggregate that hides a high-consequence exception, a restricted note that must not be pasted into the tool, and an optimistic forecast without an estimate basis. The evidence artifact includes the final update, source ledger, AI interaction record, corrections, and escalation note.
Reviewers score the six capabilities using anchored criteria and record missing evidence. The programme reports completion separately from observed performance. After six weeks, a voluntary work-sample review checks whether source traceability and exception visibility persist in comparable updates. Results are used to revise the module; they are not treated as proof of organization-wide productivity or individual employment suitability.
Curriculum QA checklist
[ ] Each module begins with a representative task, deliverable, principal failure, and evidence artifact.
- [ ] Technical, non-technical, and responsible skills are integrated in practice.
- [ ] Local tool and data rules appear as task decisions.
- [ ] Practice includes variation, feedback, verification, non-use, and escalation.
- [ ] Common foundations and role-specific adaptations are explicit.
- [ ] Measures separate reach, experience, learning, transfer, and operations.
- [ ] Missingness, reviewer disagreement, confounds, and interpretation limits are retained.
- [ ] A named owner reviews content when work, tools, policies, or failure patterns change.
Two Prune boundary
Two Prune can run realistic assessment simulations and report how a participant framed work, navigated evidence, used AI, verified outputs, exercised judgment, and delivered a business artifact. That assessment evidence can inform curriculum diagnosis and follow-up questions. Two Prune does not deliver an academy, accredit training, establish workplace transfer, make employment decisions, or replace learning, policy, or subject-matter specialists.
Sources
GOV.UK: Employer guide — What works for AI upskilling in the UK
- GOV.UK: AI Skills Tools Package
- NIST: AI Risk Management Framework Core
Sources: [1] [2] [3]
Sources
- 1.Employer guide · UK Government
- 2.AI Skills Tools Package · UK Government
- 3.NIST AI RMF Core · National Institute of Standards and Technology