All resourcesPractical guide

AI literacy: documenting measures after the 2026 change

After Article 4 of the AI Act was amended in July 2026, providers and deployers of AI systems still have to take measures that support the development of AI literacy among staff and other people operating or using AI systems on their behalf. The rule says that technical knowledge, experience, education, training and the context of use must be considered. It also explicitly says that providers and deployers do not have to guarantee a specific level of AI literacy for each individual. The practical consequence is a need for a proportionate programme and an accurate record of the measures taken. An organised record helps leaders focus follow-up on how their teams use AI at work; one universal certificate does not provide that context.

Author: Piotr Adamski · Updated:

Use the amended Article 4 as your starting point

Article 4 of the AI Act changed on 27 July 2026 through Regulation (EU) 2026/1744. The amended provision concerns measures supporting the development of AI literacy among staff and other people using AI systems on an organisation's behalf. It takes account of their knowledge, experience, education and training, the context of use and the people affected. It does not require an organisation to guarantee a specified individual competence level or obtain a prescribed AI literacy certificate [1].

This matters when designing evidence. A record should explain what the organisation actually did to support relevant learning. It should not pretend that one short course grants a universal qualification to use every AI system. The European Commission's AI literacy questions and answers also describe flexibility in the approach and the possibility of keeping internal records of initiatives [2].

Connect the programme to real AI use

Article 4 has not simply disappeared. Equally, it should not be described using an outdated obligation as though the July amendment had never happened. Use the current legal text as the primary source and treat the Commission's explanations as guidance. This article proposes an operational method; it does not determine every national enforcement or sector obligation.

The useful question is what people need to understand for the AI-assisted activities they actually perform. A customer service worker, software developer and manager reviewing an AI-supported recommendation face different outputs, decisions and consequences. Their development activities should reflect those differences, giving leaders concrete situations to discuss with each team rather than another generic attendance list to maintain.

Map AI use before assigning a generic course

Create a practical inventory of known AI-assisted work. Record the system or tool, purpose, user group, type of input, output and the decision that follows. Include approved tools and known informal use so the organisation can address actual behaviour rather than only its official procurement list. Identify whose interests can be affected. Internal summarisation, public content generation and a recommendation influencing a person's employment or access to a service have different implications. The context helps determine what users need to recognise, which checks are appropriate and when escalation is necessary.

An illustrative manufacturing office example involves AI drafting a procedure summary. The user needs to recognise that a fluent summary can omit a stop condition, alter a numerical limit or invent a reference. The content owner then needs a review method that compares the draft against the approved source. A general definition of artificial intelligence does not cover that practical need. Assign an owner to each use case or group of similar cases. The owner should work with the relevant process, security, data protection and training roles where those functions exist. A training coordinator can organise learning, but should not decide whether a particular AI-supported decision is appropriate without the responsible process owners.

Define role-specific learning objectives

Turn the inventory into observable objectives. A user might need to identify prohibited inputs under the organisation's policy, verify a generated factual statement, distinguish a draft from an approved document and report an output they cannot validate. A reviewer may need to check evidence and recognise when the system's output should not influence a decision. Keep the objectives connected to local tools and work. Users need to know what data may enter the approved system, where output is stored, who can access it and which verification steps apply. Do not claim a service offers confidentiality or a particular data location unless the relevant configuration and agreement support it.

Address limits without teaching helplessness. AI outputs can be useful while requiring verification suited to the consequence of error. A brainstorming suggestion does not need the same review as a safety instruction. Explain that relationship clearly so users understand when to apply an approved check and when to escalate. Include the people who use AI on the organisation's behalf outside a traditional employee roster. The amended provision refers to staff and other relevant persons [1]. Identify what the organisation can reasonably communicate or organise through contractor and supplier relationships. Record responsibility instead of assuming that another company has covered the specific context automatically.

Choose measures that people can use at work

Development measures can include briefings, workshops, supervised practice, role-specific instructions, examples, access to guidance and targeted courses. Choose the combination that addresses the identified need. The Commission's Q&A does not prescribe one universal certificate or identical programme for all organisations [2]. Use realistic exercises. For a team drafting customer replies, provide a sample output containing an unsupported claim and ask what needs verification before sending. For a team summarising controlled documents, include a missing exception and ask how the reviewer finds it. Label these as exercises rather than customer cases or measured evidence of product effectiveness.

Provide a route for uncertainty. People need to know who can approve a new use case, where to report a problematic output and what to do if a required verification cannot be completed. Training that says "always check" without identifying usable sources and escalation roles leaves the hardest decision unresolved. Discuss that point with the team using a real task, so the leader can coach the judgement the guidance requires. Allocate time and access. A policy link and a course invitation are measures, but their usefulness depends on whether people can understand and use them. Consider language, role, previous knowledge and available devices. An appropriate programme need not be long; it needs to address the relevant behaviour and be accessible to its intended population.

Retain evidence of measures without overstating it

Keep a concise record of the use context, objectives, intended population, material, dates and responsible owner, so a leader can prepare follow-up without repeatedly assembling evidence of the same initiative. Where a course or knowledge check is used, retain the relevant completion and assessment records. Where the measure is a workshop or updated guidance, record what was covered and how it was made available. The evidence should distinguish the organisation's action from an individual's assessed result. A briefing record can show that a measure occurred. A scenario response can show what a participant recognised within that exercise. Neither demonstrates mastery of all AI applications or guarantees compliance with every AI Act requirement.

Avoid certificates whose wording exceeds the activity. A document stating completion of a named course is different from a certificate claiming legal compliance or unrestricted AI competence. If the organisation issues a completion record, identify its scope and date accurately. Appearance does not strengthen the underlying evidence. Employee identities and results can be personal data. EDPB guidance on purpose, minimisation, accuracy and storage limitation applies to this record design [3]. Collect the information needed to explain the measure, manage follow-up and meet applicable obligations. Do not attach unrelated employee information or retain every exploratory prompt indefinitely merely because storage is available.

Check whether people can apply the guidance

Evaluation can improve the measures even though the amended Article 4 does not demand a guaranteed specified competence level for every individual. Use a focused check that reflects the local learning objective. A scenario question, discussion or supervised review can reveal unclear instructions and gaps in understanding. For an illustrative document-summary exercise, ask a participant to compare an AI draft with the source and identify an altered acceptance limit. Record the exercise's scope and result if useful. Do not extrapolate that result into a claim that the participant can safely use any AI system for any purpose.

Look for patterns. If people repeatedly submit restricted information, investigate whether the approved workflow is unclear or impractical. If reviewers cannot verify outputs, check whether they have access to authoritative sources. Another generic course may leave the underlying process problem unchanged. Evaluate the programme, not just its attendance rate. Relevant signals can include whether users know the approved tool, can identify the verification route and recognise the escalation condition. Use your own observations. Avoid invented percentages about reduced risk, productivity or legal compliance. A completion chart is a record of activity, not a causal measurement of business outcomes.

Review after changes in tools, tasks or policy

An AI literacy programme should follow meaningful changes in use. A new tool, different data category, automated workflow or changed decision can create a new development need. The owner should review the context and decide whether guidance, a briefing or further practice is appropriate. Record the trigger and decision. If a team moves from drafting internal notes to publishing generated content, the verification and approval process changes. If a system begins influencing decisions about people, the organisation needs to assess the relevant obligations separately rather than extending an old general course by assumption.

Maintain material versions. Historical records should show the guidance that applied to an earlier initiative. A later policy update must not rewrite what participants previously received. Identify affected groups and communicate the specific operational difference, including changed inputs, review requirements or escalation contacts. Review legal sources when the programme is updated. Article 4 is one part of the AI Act, and obligations connected with particular systems or roles can remain relevant. This article's general approach does not replace assessment of high-risk system oversight, sector rules, data protection or the law applicable to a particular deployment. Scope is part of the evidence, not a footnote to ignore.

Keep a proportionate and honest programme

Start with a few real use cases that matter to the organisation. Produce clear objectives, appropriate measures and a retrievable record. Check whether users can apply the guidance, then improve it where the process reveals gaps. This is more useful than buying a universal certificate and leaving local decisions unspecified.

CrewShift can support assignments, learning checks, completion history and exports for the training elements of the programme. Create an account and confirm the current workflow with a representative use case. Check whether the leader can retrieve the learning record and identify which AI-related decision needs further discussion with the team. It can help document an initiative; it does not determine the organisation's AI Act role, assess every system's risk classification or guarantee compliance.

State what your measures actually covered

Prepare an evidence summary that states the current use context, measures taken, population, dates and follow-up. Keep supporting records available with appropriate access. Use wording such as "completed the scenario workshop for this use case" where that is what occurred. Do not upgrade it to a broader claim because a marketing phrase sounds stronger.

The July 2026 amendment changes how Article 4 should be described, while leaving a practical reason to support relevant learning. People need usable guidance for the systems they operate and the decisions they make. The record should show those measures accurately, with their scope and limitations intact, and support the leader's next conversation about a difficult output or an uncertain decision.

Questions worth asking

Does the AI Act require an AI certificate for every employee?

No. Article 4 requires measures supporting AI literacy, but after the 2026 amendment it expressly does not require a provider or deployer to guarantee a specific individual level. The measures should reflect the role and context of AI use.

Is a test required after every AI training activity?

There is no single statutory test requirement for everyone. A test can be useful when it matches the purpose and risk. In other cases an acknowledgement, scenario exercise or practical review may be more appropriate.

When did the amended Article 4 take effect?

The amendment in Regulation (EU) 2026/1744 took effect on 27 July 2026. Use its current wording when describing the organisation’s measures.

Does a course guarantee AI Act compliance?

No. It can support relevant development and its record. System-specific, sector and other legal obligations need separate assessment.

Sources and scope

  1. Regulation (EU) 2026/1744 — amendment of Article 4 of the AI Act
  2. European Commission — AI literacy questions and answers
  3. European Data Protection Board — data protection basics
  4. European Commission, repository of AI literacy practices

This informational material does not replace licensed standards or individual legal advice.