Delivering AI Literacy: Keeping It Current and Showing Your Work

A well-segmented curriculum delivered badly produces the same outcome as no curriculum: completion records and unchanged behaviour.

The structural problem is that AI literacy competes for attention with every other mandatory training, and staff have learned that mandatory training is something to click through. Anything delivered in that channel inherits that response, regardless of quality.

The way out is to stop treating it as a training campaign and attach it to moments when the person already has a reason to pay attention. The strongest of these is tool rollout: someone about to be given a new AI tool wants to know how it works, what it is good for, and what will get them in trouble. Training delivered at that moment is not an interruption, it is the thing they were waiting for.

The principle generalizes from the previous cursus in this catalogue: capture and delivery both work best as side effects of something already happening, rather than as new events competing for time.

Each tier has a format that fits its content, and using the wrong one wastes both money and goodwill.

Tier 0, the universal baseline, suits asynchronous e-learning. The content is stable, the audience is everyone, and scale matters more than depth. Keep it under forty-five minutes and refresh it when the policy changes rather than on a calendar.

Tier 1, regular users, suits short tool-specific modules bundled with access. Ten minutes attached to the moment of first login beats an hour scheduled three months later. Pair it with a one-page reference that lives where the work happens.

Tier 2, oversight and decisions, needs discussion. Automation bias and the limits of oversight are not learnable by watching a video, because the point is to recognize a tendency in yourself. Workshops using real cases from the organization, including decisions that went wrong, are what move this tier.

Tier 3, builders and high-risk operators, needs system-specific instruction with assessment, because Article 26 sets a competence standard rather than an exposure standard.

The budget consequence is worth stating plainly: the cheapest format serves the largest group, and the expensive formats serve small groups where the consequences are concentrated.

The AI Office's guidance is explicit that Article 4 does not entail an obligation to measure employees' AI knowledge, and no certificates are required. Under the softened obligation of effort, that is even clearer: the test is what measures you took, not what any individual can demonstrate.

So measurement is a choice, and worth making for reasons unrelated to compliance.

Completion rates measure delivery, not learning, and should never be reported as though they measure the program.

Scenario judgement is more informative: present a realistic situation and ask what the person would do. Would you paste this into that tool? This output contradicts the source system, what now? These surface the gap between knowing the policy and applying it.

Behavioural indicators are the strongest signal and the hardest to collect: incidents reported, approval requests submitted rather than tools adopted quietly, questions escalated. A rise in reported near-misses usually indicates the program is working, not failing, which is a result worth explaining to leadership before it appears in a dashboard.

Tier 3 is the exception where assessment is genuinely warranted, because the legal standard there is competence.

Since the obligation is one of effort, the evidence is a record of the effort. The Commission's guidance accepts internal records of training initiatives; there is no prescribed format and no filing requirement.

A defensible file contains six things.

The AI system inventory, dated, showing what was in use when the program was designed.

The segmentation logic: which populations were identified, on what basis, and why each was placed where it was. This is the document that demonstrates you considered the factors the article names.

The materials themselves, versioned, so it is possible to establish what was taught at a given time.

Delivery records: who received what, and when.

The review cadence and its triggers, showing the program is maintained rather than issued once.

Decisions and their reasons, particularly the exclusions. Recording why a population was assessed as out of scope is more valuable than recording the populations you included, because that is the judgement someone would question.

Write it for a reader two years from now who was not there. That is the same standard as a decision record, and for the same reason.

The AI Office maintains a living repository of AI literacy practices, collecting examples submitted by organizations of different sizes and sectors, covering technical and non-technical roles and in some cases extending to vendors, partners and clients.

It is genuinely useful as a source of formats and framings, particularly for smaller organizations without a learning function.

One caution is carried in the guidance itself and deserves repeating: replicating a practice from the repository does not automatically grant a presumption of compliance. This matters because it is the opposite of how harmonised standards work elsewhere in EU product law, where conformity with a published standard does confer a presumption. There is no such mechanism here.

The reason is the same one that runs through this whole cursus. Appropriateness is contextual. A practice that suits a hospital's clinical staff is not automatically appropriate for a logistics firm, and adopting it without doing the inventory and segmentation yourself means you have copied someone else's answer to a question about their organization.

Use it for ideas. Do not use it as a template that substitutes for the reasoning.

Annual refresh is the default and it is poorly matched to the problem, because the thing that dates the training is not the passage of time but a change in the systems.

Bind the refresh to events that actually invalidate content.

A new AI tool is approved, or an AI feature is switched on inside existing software. This is the most frequently missed trigger, because nobody experiences it as a deployment.

A system moves into a new use, particularly one affecting individuals, which may change its risk classification and therefore the tier of the people operating it.

A model or vendor changes materially, so documented limitations and failure modes shift.

An incident occurs, internally or in a comparable organization. Incidents are the highest-value training material available and their half-life is short.

The regulatory position changes, as it did in 2026.

A supporting mechanism makes this workable: put AI literacy review into the procurement and change-management checklist, so approving a new tool automatically raises the question of who now needs training on it. Without that hook, the trigger exists on paper and fires only when someone remembers.

Ownership is where these programs quietly fail, because the work sits across three functions that each have a partial view.

Learning and development owns delivery and knows how to build training that people complete, but usually does not know which systems are deployed or what the Regulation asks.

Legal or compliance owns the interpretation and the evidence file, but tends to produce material that teaches Article numbers rather than practice.

The technical function owns the inventory and the system specifics, and is generally the least equipped to teach non-technical staff.

None of the three can deliver a working program alone, and the common failure is assigning it wholly to one. The arrangement that works is a named owner accountable for the program as a whole, drawing on all three, with the inventory maintained by the technical function as a standing artefact rather than a one-off exercise.

For a small organization this is one person wearing three hats, which is fine. What is not fine is the program having no owner, which is the state most organizations are actually in.

For an organization starting from nothing, an order of work that produces something defensible without a large budget.

Weeks 1 to 3: inventory. Survey tools in use, audit AI features inside existing software, and run an amnesty on shadow use. Expect the result to be larger than anticipated.

Weeks 4 to 5: classify and segment. For each system, record what it decides and who it affects, flag anything potentially high-risk for legal review, and map populations to tiers. Write down the reasoning as you go, because this becomes the evidence.

Weeks 6 to 8: build Tier 0 and the acceptable-use policy together. The training explains the policy, so writing them separately produces two documents that disagree.

Weeks 9 to 10: deliver Tier 0, and build Tier 1 modules for the two or three most-used tools.

Weeks 11 to 12: run the first Tier 2 workshop with the oversight population, and put the refresh triggers into the procurement checklist.

Tier 3 is scheduled separately against a specific high-risk deployment, because it is system-specific and cannot be built generically.

The sequence deliberately front-loads the inventory. Everything downstream is guesswork without it.

Worth stating plainly, because these programs are routinely asked to carry weight they cannot bear.

It cannot make an unsuitable system suitable. If a model performs badly on part of the population, training the operators to be careful is a mitigation, not a fix, and treating it as one shifts responsibility onto the people with the least power to change the system.

It cannot substitute for oversight authority. Teaching someone to spot a bad output is pointless if overriding it is career-limiting or if the throughput target makes genuine review impossible. Oversight requires permission and time, and neither is a training outcome.

It cannot make an opaque system explainable. If a deployer cannot obtain adequate information from the provider, that is a procurement and contractual failure, and no amount of internal training closes it.

It cannot create accountability. Someone must own the decision to deploy. Distributing understanding across many people is not the same as anyone being answerable.

The honest summary of the cursus: the Regulation now asks you to support the development of AI literacy rather than guarantee it, the fine that people quote does not attach to this provision, and the reasons to build a good program are mostly reasons that would exist if the Act had never been written.