Article 4 of the EU AI Act requires every provider and deployer of an AI system to take measures to support AI literacy among the people who operate and use those systems on its behalf, and it has applied since 2 February 2025. It is the broadest obligation in the Act: it is not limited to high-risk systems, so an organisation that only uses a chatbot or a coding assistant is in scope. The Digital Omnibus on AI, Regulation (EU) 2026/1744, rewrote Article 4 with effect from 27 July 2026, removing the requirement to ensure a "sufficient level" of literacy but keeping the duty to act. This guide explains who is covered, what the amended text requires, how national authorities enforce it, what it means for teams that run AI agents, and how to build a programme you can evidence. It is for security, HR and learning leads, AI governance owners and executives. For the wider Act, see the EU AI Act Compliance Guide.
Key takeaways
- It applies to almost everyone using AI at work. Any provider or deployer of an AI system, of any risk level, must take measures for its staff and for other persons using AI on its behalf, including contractors and service providers.
- The Omnibus softened the standard, not the duty. Since 27 July 2026 the text says organisations "shall take measures to support the development of AI literacy" and expressly does not require them to guarantee a specific level for any individual.
- There is no certificate and no mandated format. The Commission says no certification, specific training format, AI officer or measurement of staff knowledge is required, but an internal record of measures is the practical way to show compliance.
- Enforcement is national and started on 2 August 2026. Market surveillance authorities supervise Article 4 under penalty rules set by each Member State; Article 4 has no dedicated fine tier in Article 99.
- High-risk deployers need more than literacy. Article 26(2) separately requires human oversight to be assigned to people with the necessary competence, training and authority, and the Omnibus did not remove that.
Scope: who and what is covered
Organisations
- Providers (Article 3(3)): bodies that develop an AI system or GPAI model, or have one developed, and place it on the market or put it into service under their own name, paid or free.
- Deployers (Article 3(4)): bodies using an AI system under their authority, other than for personal non-professional activity. A company licensing a SaaS assistant for staff is a deployer.
- Any risk level: Article 4 sits in Chapter I (general provisions), so it covers minimal-risk tools as well as high-risk systems. The Commission's Q&A frames it as applying wherever someone in the organisation is directly dealing with an AI system.
People
| Group | In scope? | Basis and notes |
|---|---|---|
| Employees who use, configure or oversee AI systems | Yes | "Staff" in Article 4. |
| Contractors, consultants, outsourced service providers acting for you | Yes | "Other persons dealing with the operation and use of AI systems on their behalf"; the Commission says these are people broadly under the organisational remit. |
| Clients | Sometimes | The Commission says clients may be relevant depending on the risk, citing recital 20. |
| Affected persons (people the AI is used on) | Not as trainees | The amended Article 4 requires you to consider the persons or groups on whom systems are used when designing measures. |
| Staff with no contact with AI systems | Not required | Many organisations include them in general awareness anyway, because shadow AI use is common. |
AI literacy is defined in Article 3(56) as the skills, knowledge and understanding that allow providers, deployers and affected persons to make an informed deployment of AI systems and to gain awareness of the opportunities and risks of AI and the possible harm it can cause.
Key obligations
The amended text of Article 4
As shown in the Commission's AI Act Service Desk consolidated view, Article 4(1) now reads: "Providers and deployers of AI systems shall take measures to support the development of AI literacy of their staff and other persons dealing with the operation and use of AI systems on their behalf, taking into account their technical knowledge, experience, education and training and the context the AI systems are to be used in, and considering the persons or groups of persons on whom the AI systems are to be used. This obligation does not require providers or deployers to guarantee any specific level of AI literacy of any individual."
Article 4(2) obliges the Commission and Member States to support providers and deployers, in particular SMEs, and requires the Commission to publish practical compliance examples on the single information platform referred to in Article 62(3)(b). Article 4(3) asks the AI Board to adopt recommendations, taking into account European competence frameworks, including common objectives.
Before and after the Omnibus
| Aspect | Original text (2024/1689) | After Regulation (EU) 2026/1744 |
|---|---|---|
| Core duty | Take measures to ensure, to their best extent, a sufficient level of AI literacy | Take measures to support the development of AI literacy |
| Outcome standard | "Sufficient level" (undefined) | No specific level guaranteed for any individual |
| Factors to consider | Knowledge, experience, education, training, context, affected persons | Unchanged in substance |
| Role of the Commission and Member States | Not in Article 4 | Support and facilitate, publish practical examples, Board recommendations |
| High-risk human oversight training (Art. 26(2)) | Required | Still required |
What the Commission says good looks like
The Commission's AI literacy Q&A suggests that measures, at a minimum, cover four things.
- General understanding: what AI is, how it works, which AI the organisation uses, and its opportunities and dangers.
- The organisation's role: whether it is a provider or deployer of each system.
- Risks and mitigations: the risks of the systems provided or deployed and what staff need to know to deal with them.
- Tailoring: measures adapted to staff knowledge, the context of use and the people affected. Legal and ethical aspects, including the AI Act, are encouraged.
The Q&A also says that relying on instructions for use, or asking staff to read them, is often ineffective, and that a risk-based approach applies, with more measures for higher-risk uses.
Role-based design
| Audience | Depth | Example content |
|---|---|---|
| All AI users | Foundational | Approved tools, data that must never be entered, verifying outputs, hallucination, reporting route |
| Developers and data scientists | Technical | Prompt injection, insecure output handling, model and data supply chain, evaluation, logging |
| Agent and automation owners | Technical and operational | Tool permissions, credentials agents use, approval gates, kill switches, monitoring |
| Human overseers of high-risk systems | Competence for Art. 26(2) | System limits, automation bias, when and how to override, escalation |
| Procurement, legal, HR | Applied | Provider versus deployer roles, supplier documentation, prohibited practices |
| Board and executives | Strategic | Risk appetite, accountability, regulatory exposure, metrics |
Timeline and current status
| Date | Event |
|---|---|
| 2 February 2025 | Chapters I and II apply, including Article 4 (Art. 113) |
| 4 February 2025 | AI Office living repository of AI literacy practices published (last updated 10 August 2026) |
| 2 August 2025 | Date by which Member States were due to adopt penalty rules, according to the Commission Q&A |
| 8 July 2026 | Regulation (EU) 2026/1744 (Digital Omnibus on AI) adopted, per EUR-Lex metadata |
| 27 July 2026 | Omnibus enters into force; amended Article 4 applies |
| 2 August 2026 | Supervision and enforcement of Article 4 by national authorities begins |
| 2 December 2027 | Annex III high-risk obligations apply, including Art. 26 deployer duties for those systems (moved by the Omnibus) |
The duty has existed since February 2025, so organisations cannot treat August 2026 as a start date: an authority looking at an incident can ask what measures were in place before it. National enforcement approaches will differ, and Commission guidance and Board recommendations under Article 4(3) may still change expectations, so review this programme at least annually.
Enforcement and penalties
- Who enforces: national market surveillance authorities, not the AI Office. The AI Office and the AI Board support consistent implementation.
- Penalties: set by national law under Article 99(1), which requires penalties to be effective, proportionate and dissuasive and to take account of SMEs and small mid-caps. Article 4 is not listed in the fixed tiers of Article 99(3) to (5), so there is no Act-level maximum for it.
- How authorities will weigh it: the Commission says enforcement must be proportionate, considering the nature and gravity of the infringement and whether it was intentional or negligent. Evidence that an incident was caused by inadequate training or guidance makes enforcement more likely.
- Indirect exposure: the AI Act creates no criminal offences or compensation right itself, but poor literacy can be relevant to claims under national law, and to related breaches such as high-risk deployer duties under Article 26, which do sit in the Article 99(4) tier of EUR 15 million or 3% of worldwide turnover.
What it means for AI agents and non-human identities
Agents change the literacy question from "can staff use a chatbot safely" to "do the people who configure and approve autonomous systems understand what they have delegated". Article 4 covers the humans, not the agent, but the risk sits in the gap between them.
- Agent builders need identity literacy: the most common agent failures are excessive permissions and over-privileged credentials, described in OWASP LLM06:2025 Excessive Agency as excessive functionality, permissions and autonomy. Train owners to scope API keys, OAuth grants and service accounts to the task. See the Agentic AI Compliance Guide.
- Approvers need to recognise risky delegation: people who click "allow" on an agent's OAuth consent, or who approve its actions, are dealing with the operation of an AI system and should understand what they authorise.
- Everyday users need data boundaries: the Samsung ChatGPT source code leak is a clear example of staff pasting confidential code into a public AI tool, the kind of gap Article 4 literacy training is meant to close. See the Shadow AI Discovery Guide.
- Contractors running your agents are in scope: outsourced teams operating automations on your behalf fall under "other persons", so put literacy requirements in contracts.
Compliance steps
- Inventory AI systems and roles: list every AI system you provide or deploy, including embedded features in SaaS and agents, and record whether you are provider or deployer for each.
- Map the people: identify staff and other persons who operate or use each system, by role, and note high-risk uses that trigger Article 26(2).
- Set an AI acceptable use baseline: approved tools, prohibited data, verification expectations and a reporting route. See the AI Acceptable Use Policy Template.
- Design role-based measures: use the four Commission minimum elements, add technical modules for builders and agent owners, and competence-based training for human overseers. Mix formats: short modules, workshops, guidance in the tool, and supervised practice.
- Pilot and adjust: run with one business unit, gather feedback and incident data, then roll out by population. Keep earlier versions of materials so you can show what was in place at a given date.
- Record what you did: keep an internal register of measures, audiences, dates, completion and content versions. No format is mandated, but this is your evidence for an authority.
- Extend to third parties: add clauses requiring contractors and service providers who use AI on your behalf to meet equivalent literacy measures.
- Review: at least annually, and when you adopt a new class of AI system, after incidents, or when Board recommendations or national guidance change.
Small versus large organisations
| Element | Small organisation or SME | Large enterprise |
|---|---|---|
| Ownership | One named owner, often the operations or IT lead | Shared between AI governance, security and HR or learning |
| Delivery | Short briefing plus written guidance; Commission examples on the single information platform | Learning management system, role-based curricula, in-tool prompts |
| Evidence | Simple spreadsheet of who received what and when | LMS records linked to the AI system inventory and HR data |
| Penalty context | Article 99(1) requires authorities to consider SME interests | Higher scrutiny where AI is core to operations |
Detection and measurement
Article 4 does not require you to measure staff knowledge, but metrics show the programme is working.
- Coverage: percentage of in-scope staff and contractors who have received role-appropriate measures, by AI system.
- Currency: percentage of materials reviewed in the last 12 months.
- Shadow AI signal: unapproved AI tool usage detected through web proxy, CASB or OAuth grant reviews, trending down.
- Data loss events: DLP alerts for confidential data sent to AI services.
- Agent hygiene: share of agent credentials with a named, trained owner.
- Oversight readiness: percentage of high-risk system overseers with completed competence training.
Framework mapping
| Article 4 element | NIST CSF 2.0 | NIST SP 800-53 Rev. 5 | NIST AI RMF |
|---|---|---|---|
| General literacy for all users | PR.AT-01 | AT-2 | GOVERN 2.2 |
| Role-based depth for builders, agent owners, overseers | PR.AT-02 | AT-3 | MAP 3.4 |
| Roles and accountability for the programme | GV.RR-04 | PM-13 | GOVERN 2.1 |
| Third parties acting on your behalf | GV.SC-05 | SR-3 | GOVERN 2.2 (partners) |
ISO/IEC 42001 clauses on competence and awareness serve the same purpose for organisations running an AI management system. See the AI Governance Guide.
Common failure patterns
- Treating it as high-risk only: Article 4 applies to every AI system you provide or deploy, including office assistants.
- Handing out the vendor manual: the Commission says relying on instructions for use is often ineffective, and for high-risk systems it is not enough.
- One generic module for everyone: the text requires measures that take account of knowledge, context and affected persons; agent owners and overseers need more.
- Forgetting contractors: outsourced teams operating AI on your behalf are in scope.
- No record: without dated evidence you cannot show what measures existed when an incident happened.
- Reading the Omnibus as repeal: the duty to take measures remains and has applied since February 2025.
Practitioner checklist
- Inventory every AI system you provide or deploy and your role for each.
- Identify staff and other persons who operate or use each system.
- Publish an AI acceptable use policy with prohibited data and approved tools.
- Cover the four Commission minimum elements in foundational material.
- Add technical modules for developers and agent owners, including credential scoping.
- Give high-risk overseers competence training that meets Article 26(2).
- Keep a dated register of measures, audiences and content versions.
- Put AI literacy requirements in contracts with service providers.
- Track coverage, shadow AI and DLP metrics and report them to the board. See the Board Cyber Risk Reporting Guide.
- Review the programme at least annually and after incidents.
Standards and references
- AI Act Service Desk: Article 4, AI literacy
- European Commission: AI Literacy, Questions and Answers
- EUR-Lex: Regulation (EU) 2026/1744 (Digital Omnibus on AI)
- European Commission: Living repository to foster learning and exchange on AI literacy
- AI Act Service Desk: Article 99, Penalties
- AI Act Service Desk: Article 26, Obligations of deployers of high-risk AI systems
- NIST: AI Risk Management Framework Core
- OWASP: Top 10 for LLM Applications 2025
Related NHI Mgmt Group resources: EU AI Act Compliance Guide · EU AI Act High-Risk AI Systems Guide · EU AI Act Prohibited Practices Guide · AI Acceptable Use Policy Template. This guide summarises the AI literacy obligation in Article 4 of the EU AI Act and is not legal advice.