How to document AI-literacy measures for an Article 4 review

Article 4 of the EU AI Act names no certificate, course, or number of hours. It asks providers and deployers to take measures that support AI-literacy development among staff and anyone operating AI on their behalf, with knowledge, experience, training, context, and affected people taken into account. It expressly avoids a guaranteed individual level. Records made at the time can show which measures you took without turning a completion score into a legal conclusion.

Who actually asks, and what they accept

Three audiences end up reading this file, and they have converging standards. A market surveillance authority investigating an incident asks for the training record early, because it colours everything else. An enterprise customer's procurement team asks during vendor review, because your literacy gap becomes their supply-chain risk. An insurer or counsel asks after something goes wrong, because the record is the difference between a defended process and negligence. None of them accepts a spreadsheet that anyone could have edited yesterday.

The five layers of an Article 4 file

1. 1Coverage: who, against which roleA roster mapping every person in scope to a role band and its required depth, with completion state and dates. Contractors operating AI on your behalf belong on this roster too. The gaps are as much a part of the record as the completions: a file that shows you knew who was untrained reads better than one that pretends everyone was.
2. 2A content register: what was actually taughtVersioned content, so the record says which material a person completed, not merely that something was completed. When the curriculum changes, the register shows who trained on the old version and when they were brought forward. Without versions, every claim collapses to an unverifiable somebody did something once.
3. 3Attestations: the person acknowledges the rulesPer-person, dated acknowledgements of the policies that govern their AI use. Small on their own, but they anchor the rest: nobody can later claim the rules were never put in front of them.
4. 4Behaviour: drills, not just lessonsRecords of simulated incidents. The learner delegated a task to an agent, a poisoned instruction appeared mid-run, and the record shows whether they stopped it, how fast, and what they did next. This shows how a practical measure performed, rather than merely showing exposure to material.
5. 5Integrity: a trail that cannot be rewrittenEvery event above, hash-chained in sequence so that removing or editing one entry breaks the chain visibly. This is what separates a record from a claim. It also protects you: an intact chain shows the programme existed before the incident, not after it.

Why attendance keeps failing the test

The instinct is to buy a video course, collect completion certificates, and file them. That produces a stack of documents proving that people pressed play. It says nothing about whether the finance clerk would recognise a manipulated invoice instruction, or whether the operator supervising an agent knows which of its actions cannot be undone. Recognition on a quiz is not behaviour under pressure, and reviewers know the difference because the incident that brought them to your door usually proves it.

Context cuts the same way. Article 4 says measures must take knowledge, experience, education, training, use context, and affected people into account. A roster helps show how the programme applied those factors, with deeper practice where systems have greater reach or consequence.

Assembling it without a project team

The file above sounds like bureaucracy when described, but none of it needs to be produced by hand. If the training system records coverage, versions content, collects attestations, runs drills, and chains its own audit log, then the evidence pack is an export, not a project. The work happens once, in choosing a system that records the right things as a side effect of training, instead of a documentation sprint every time someone asks.