EU AI Act Article 4: A Practical Guide to the AI Literacy Obligation

Article 4 of the EU AI Act has been in force since 2 February 2025. National enforcement by market surveillance authorities begins on 2 August 2026. That’s about 10 weeks out.

Most European organisations I speak to haven’t started their AI literacy programme. Plenty still don’t know the obligation applies to them.

Here’s the version I keep having to explain on 30-minute calls. What the regulation actually says, who’s caught by it, what it requires in practice, what doesn’t count, what the penalties look like, and what to do before August.

What Article 4 actually says

The text:

“Providers and deployers of AI systems shall take measures to ensure, to their best extent, a sufficient level 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.”

One sentence. Four things worth pulling out.

1. The obligation falls on providers and deployers. More on what that means below.
2. It requires a “sufficient level of AI literacy”, and the regulation deliberately doesn’t define that prescriptively.
3. The scope covers staff and other persons dealing with AI systems on the organisation’s behalf. Contractors and freelancers count.
4. The standard is contextual. Training has to track staff’s actual knowledge, the context the AI is used in, and the people affected.

You satisfy this by being able to show that the people who touch your AI systems can use them responsibly and competently in the context of their actual role. The off-the-shelf “watch a 45-minute video, take a quiz, get a certificate” approach is exactly the one the Commission has gone out of its way (across at least three pieces of published guidance) to say doesn’t get you there.

Who counts as a deployer

Article 3(4):

“a natural or legal person, public authority, agency or other body using an AI system under its authority except where the AI system is used in the course of a personal non-professional activity.”

Practically: any organisation using AI systems in a professional context. No size threshold. No sector threshold. No risk-level threshold.

The one carve-out is personal, non-professional use. Everyone else is in.

The Commission’s AI Literacy FAQ is explicit that Article 4 covers every risk category, including minimal-risk uses like general-purpose chatbots.

Does this apply to you? Common scenarios

These are the questions I keep getting asked, with the answers.

A marketing agency using ChatGPT internally to draft copy. Yes. Legal person, professional context, under its authority. Deployer.

A law firm using generative AI for document review. Yes. Same logic.

A small accounting practice running Copilot in Excel. Yes. Small doesn’t matter. SMEs stay in scope of Article 4. The regulation lets them calibrate how much training is enough (“to their best extent”), without any blanket exemption.

A recruitment agency using AI to screen CVs. Yes. And the screening tool is likely a high-risk AI system under Annex III, point 4, which brings further obligations on top of Article 4.

An SME running Microsoft 365 Copilot across its business. Yes.

A charity drafting grant applications in ChatGPT. Yes. Charities are legal persons. Grant applications are professional activity.

An e-commerce shop with an AI chatbot handling customer queries. Yes. The chatbot is an AI system. The business is deploying it. Deployer.

A non-EU company selling into the EU, using AI-generated marketing copy aimed at European customers. Yes. Article 2(1)(c) reaches “providers and deployers of AI systems that have their place of establishment or are located in a third country, where the output produced by the AI system is used in the Union.” Marketing copy read by European consumers counts as output used in the Union.

A private individual using ChatGPT to plan a holiday. No. That’s the carve-out. Personal, non-professional activity sits outside the regulation.

The pattern: if your organisation uses any AI system in any professional context, Article 4 applies. Size, risk level, and informality of use don’t change the answer.

What “sufficient AI literacy” means in practice

The regulation doesn’t answer this prescriptively. The Commission has said repeatedly there’s no one-size-fits-all approach, and the AI Office doesn’t intend to mandate specific trainings.

The test is whether your training is genuinely appropriate to your staff, your tools, and your context. The Commission’s FAQ puts flesh on that in several ways.

Training must be differentiated by role

The Article 4 text explicitly requires considering “the persons or groups of persons on whom the AI systems are to be used” and “the context the AI systems are to be used in.” In practice: a marketing team’s training looks different from a finance team’s, which looks different from a customer-service team’s.

A single generic programme delivered across the whole organisation rarely satisfies the obligation, because the context the FAQ keeps pointing at really does vary from role to role.

Training must address the actual systems in use

The Commission’s FAQ talks about “what employees need to know when dealing with such AI system” and gives examples tied to specific tools, including ChatGPT-specific risks like hallucination. If your team uses Copilot in Excel, train them on Copilot in Excel. If they also use a custom GPT for invoice analysis, train them on that too.

Training must enable risk-literate use

Article 3(56) defines AI literacy as the “skills, knowledge and understanding that allow providers, deployers and affected persons… to make an informed deployment of AI systems, as well as to gain awareness about the opportunities and risks of AI and possible harm it can cause.”

What this should look like on the ground is staff who can describe, in concrete terms, the failure modes of the tools they use day to day. That includes hallucination in the case of LLMs, but also bias, data leakage, drift, and over-reliance.

Methods can vary, some don’t qualify

The Commission’s FAQ lists acceptable approaches: traditional courses, peer knowledge-sharing, internal knowledge databases, experimentation spaces. No mandated format. What’s ruled out is “simply asking staff to read an AI system’s instructions for use,” which the Commission has explicitly called insufficient.

Records matter, even if certificates don’t

The Commission says organisations “can keep an internal record of trainings and/or other guiding initiatives.” There’s no certificate requirement, but if you get challenged you need to be able to show what was delivered, to whom, and when. A spreadsheet that nobody loves the look of will, in practice, get you most of the way there, and I’ve yet to see a regulator ask for anything more elegant.

What doesn’t satisfy Article 4

The Commission has been unusually direct about a few things that don’t work.

A 45-minute generic video about what AI is, followed by a multiple-choice quiz and a certificate. Most LMS providers are currently selling exactly this, often labelled “Article 4 compliant.” The Commission’s own words: a “standardised, ‘one size fits all’ approach is not appropriate, as it may be excessive for some organisations and insufficient for others.”

Asking staff to read the user manual. Commission FAQ: “Simply relying on the AI systems’ instructions for use or asking the staff to read them might be ineffective and insufficient. Relying on the instructions of use is therefore not sufficient, further measures are necessary.”

Treating compliance as a one-off event. AI literacy is meant to be ongoing. Your tools change, the people using them change, and the obligation runs alongside both.

What a compliant programme looks like

If I were building an Article 4 programme from scratch, here’s where I’d start.

First, a role map. Which teams use AI, which systems, what for, and at what level of risk? Most organisations I work with discover they have no real picture of this until they actually sit down and write it out, and the exercise tends to surface tools that no one in compliance knew were being used.

Second, differentiated training paths. Each role group needs training built for its tools, its workflows, and its risk context. A clienteling team’s training looks nothing like a finance team’s, and a generic deck delivered to both is what the Commission’s FAQ politely calls insufficient.

Third, practice time with the real tools. People should be building, prompting, and evaluating during the session rather than watching slides about it from across the room. The Commission lists “experimentation spaces” as an acceptable method, and I’d go a step further and treat them as the spine of the whole programme, because the only literacy that survives the next quarter is the one people built with their own hands.

Fourth, risk framing tied to the specific tools in use. A hallucinated factoid in a marketing draft is awkward; the same hallucination in a regulatory submission is a different kind of problem. Your staff need to be able to articulate the failure modes of the systems they actually touch.

Fifth, documentation. An internal record of who received what training, when, with what content. The Commission doesn’t require certificates, and you don’t need an LMS to do this. You do need to be able to produce the evidence if someone asks.

Sector-specific context

Article 4 applies uniformly. The risk context varies dramatically by sector.

Financial services

European banks are among the heaviest AI deployers in the region. Microsoft 365 Copilot is being rolled out to tens of thousands of banking employees. AIB rolled out Microsoft 365 Copilot to over 10,000 employees in July 2025. BBVA announced in December 2025 it was extending ChatGPT Enterprise to all 120,000 employees globally, up from an earlier 11,000-licence deployment. LGT, BNP Paribas, and Crédit Agricole are all running generative AI widely. The EBA’s 2024 Risk Assessment Questionnaire data shows a consistent upward trend in AI adoption across EU banks, with around a third already running generative AI in at least one production use case.

Why it matters for Article 4:

• The sector already sits under intense regulatory scrutiny (PRA in the UK, MiFID II and DORA in the EU, plus each firm’s national supervisor), and AI Act non-compliance stacks on top of all of that rather than replacing any of it.
• The tools being deployed run against highly regulated activities: an analyst drafting client communications in Copilot, an ops team running trade reconciliation through an LLM, a compliance team putting KYC screening through a model. None of that is what the Act would call a low-risk application.
• Most generic AI literacy e-learning sold into banks was built for generic knowledge workers, which means it doesn’t say anything useful about MiFID II, PRA rules, DORA, or the specific ways financial services staff actually use these tools in their day.

Insurance

Insurance sits right behind banking on EU AI adoption. EIOPA’s 2024 Digitalisation report puts 50% of non-life insurers and 24% of life insurers already running AI across pricing, underwriting, fraud detection, and claims.

Credit scoring and life and health insurance pricing are explicitly Annex III high-risk. Under the Digital Omnibus political agreement of 7 May 2026, the full high-risk obligations for stand-alone Annex III systems (governance, documentation, human oversight, EU database registration) were pushed to 2 December 2027. The Article 4 literacy obligation for these use cases still applies from August.

Insurers also have governance scaffolding to work with. Solvency II Article 41, IDD Article 25, and DORA Articles 5 and 6 each already require effective governance and risk management. Mapping Article 4 literacy onto what’s already there is usually cheaper than building a parallel programme.

The literacy bar sits higher than in most sectors because underwriters, claims handlers, and pricing actuaries are making AI-assisted decisions on high-risk systems. Sufficient literacy here means being able to describe where the model is likely to be wrong, with examples drawn from the actual product.

Most generic e-learning sold to insurers doesn’t reference Solvency II, IDD, DORA, or the actual products in use, which is why I’d be careful about treating an off-the-shelf course as the full answer.

Small and mid-sized businesses

Most SMEs assume Article 4 doesn’t apply to them because the Act has “SME proportionality.” That phrase calibrates the size of your programme, but it doesn’t excuse you from running one in the first place, and I find this is the single most common misreading among smaller firms I talk to.

For a 20-person marketing agency running Copilot and ChatGPT, a defensible Article 4 programme looks like:

• A list of AI tools actually in use, including the ones staff adopted without telling anyone (shadow AI is almost always the biggest category).
• A one-page policy on acceptable use, data handling, and where human review is required.
• A short working session per role group, walking through the tools they use on real briefs.
• A written record of who attended, when, with what content.

That takes about two days to build and a half-day to run. What gets SMEs in trouble is turning up in August with nothing documented at all.

Marketing and creative teams

Marketing is usually the heaviest AI user in any organisation. Content generation, campaign variants, image work, social copy, the audience segmentation slides nobody quite wants to take ownership of: most creative and marketing staff just think of themselves as using ChatGPT to do their jobs faster, and Article 4 catches them whether they think of it that way or not.

The risks a marketing team needs to understand sit somewhere different from the rest of the business. Hallucinated product claims that land in client-facing copy come back to the agency, regardless of which model wrote them. Generic outputs dilute brand distinctiveness when staff don’t know how to steer the tool (and when to override it). IP and licensing questions around generated imagery and text get sharper once that output is in a commercial campaign. And anything pasted into a public ChatGPT session has, in the strict sense, stopped being confidential.

Generic “what is AI” e-learning doesn’t really touch any of this. A marketing team’s Article 4 programme has to sit on their actual workflows with their actual tools, which means starting from the briefs and the campaigns rather than from a slide deck.

HR and recruitment

HR is where Article 4 bites twice.

First, HR often owns the company-wide training function. When compliance, legal, or the CTO decides Article 4 needs a programme, HR builds and runs it. The HR leader reading this is often the literal buyer of Article 4 training for the rest of the organisation.

Second, HR itself becomes a high-risk deployer the moment it touches AI for recruitment. Annex III, point 4 lists AI systems used for:

• Placing targeted job advertisements.
• Analysing and filtering job applications.
• Evaluating candidates.

CV-screening tools, interview-analysis tools, video-based personality scorers: all high-risk. That stacks a full set of Article 26 deployer obligations (human oversight, impact assessment, monitoring) on top of Article 4 literacy.

An HR team running AI-assisted recruitment, then, needs differentiated training on the specific tools they use, plus two records kept in parallel: an Article 4 training record and an Article 26 deployment oversight record. In my experience HR functions tend to under-invest in documentation discipline until an incident makes the case for them. By then it’s usually too late to backfill.

Luxury and heritage brands

The dominant AI framing in market right now is mass-automation: scale, speed, efficiency, replacement. That framing is more or less the opposite of what a heritage house is selling, since craft, scarcity, judgement and taste are the assets that let it charge what it charges in the first place. A generic “what is AI” training nudges staff into the wrong relationship with the tool before they’ve even opened it.

There’s a different framing, which I’m working out in more detail in a short book I’m writing called Artisanal Intelligence: AI deployed with restraint, judgement and taste to amplify craft rather than erase it. That looks, in practice, like creative assistants that support artisans without making decisions for them, archive digitisation that makes a heritage searchable without flattening it, and demand forecasting that’s used to preserve scarcity rather than chase volume out the door.

That framing matters for Article 4 compliance in luxury because generic literacy training pushes staff to maximise AI output, when what luxury teams actually need is the opposite: judgement about when not to use the tool at all. The Commission’s contextual standard (“the context the AI systems are to be used in”) gives you a lot of room to train differently, and luxury’s context is unambiguously different from a generic enterprise’s.

In luxury, the hardest compliance risks are usually brand risks first and regulatory risks second: a VIC relationship soured by tone-deaf AI outreach, a heritage archive quietly polluted with hallucinated content, a creative team’s voice diluted by uncritical model use over a couple of seasons. The fines, if they come, tend to follow the brand damage rather than precede it.

If you run AI literacy in a luxury house, the question worth asking is whether your programme teaches staff to treat the tool with the same restraint the rest of the brand operates with.

Other regulated sectors

A few more worth noting briefly.

In healthcare, clinical decision support and medical imaging systems sit in Annex III as high-risk. AI embedded in MDR/IVDR-regulated medical devices was originally captured by an August 2027 deadline; the Digital Omnibus deal pushes that to 2 August 2028, and a December 2025 Commission proposal would, if adopted, remove medical devices from the AI Act’s high-risk scope altogether and route the regime through MDR/IVDR instead. Article 4 itself applies to hospitals and providers regardless.

In the public sector, every European civil service is a deployer. The Commission’s Apply AI Strategy (October 2025) and the GenAI4EU initiative both signal fast public-sector deployment, and the literacy obligation applies to public employees the same way it does to private ones.

In education, universities and schools are in scope. The European Commission and OECD released a draft AI literacy framework for primary and secondary education in May 2025, with the final version expected in the first half of 2026. Teacher readiness is the gap most ministries name when asked.

Penalties, enforcement, and timeline

The AI Act’s penalty provisions, in Article 99, run in three tiers:

• Up to €35 million or 7% of global turnover for violations of Article 5 (prohibited practices).
• Up to €15 million or 3% of global turnover for non-compliance with operator and notified body obligations in Articles 16, 22, 23, 24, 26, 31, 33(1)(3)(4), 34, and 50.
• Up to €7.5 million or 1% of global turnover for supplying misleading information to authorities.

Article 4 isn’t explicitly named in any of those tiers. The Commission’s FAQ is clear that “national market surveillance authorities could impose penalties and other enforcement measures to sanction infringements of Article 4”, and that enforcement rests on national laws that Member States were required to adopt by 2 August 2025.

The point worth holding onto is that the Commission has been explicit: sanctions for Article 4 failures are more likely where an incident can be traced back to missing or inadequate training. Article 4 non-compliance becomes a compounding factor when something else goes wrong. If an AI-related incident lands at your organisation and a regulator finds that staff hadn’t been properly trained for it, your exposure is meaningfully higher than it would otherwise have been.

The timeline worth keeping straight:

• 2 February 2025: Article 4 became applicable. The obligation has been live for 15 months.
• 2 August 2025: Deadline for Member States to adopt national penalty regimes and designate competent authorities. Many are late.
• 2 August 2026: National enforcement of Article 4 begins. GPAI obligations and the AI Act governance framework become fully applicable. This is the deadline most organisations have in mind.
• 2 December 2026: Synthetic content marking and disclosure obligations (Article 50) begin to apply, under the Digital Omnibus political agreement of 7 May 2026.
• 2 December 2027: High-risk obligations for stand-alone Annex III systems (recruitment, credit scoring, insurance pricing, biometrics, education, employment, migration) take effect. This was originally 2 August 2026; the Digital Omnibus deal pushed it back.
• 2 August 2028: High-risk obligations for AI embedded in regulated products (medical devices, machinery, toys) take effect.

Today, the Article 4 deadline is about 10 weeks out.

One live development worth tracking

The European Commission, Parliament, and Council reached a provisional political agreement on the Digital Omnibus package on 7 May 2026. Formal adoption is expected in the following weeks.

The Commission’s original proposal would have shifted the Article 4 obligation from companies onto the Commission and Member States, softening it from a duty of result into a duty of means. Based on the publicly available summaries of the political deal, that softening did not survive the trilogue: Article 4 appears to remain in force as originally drafted, with no substantive amendment to the obligation on providers and deployers. The final consolidated text will confirm this.

What the Omnibus deal did change matters too: most high-risk AI system obligations were delayed by 16+ months (Annex III stand-alone to 2 December 2027, embedded products to 2 August 2028), a new Article 5 prohibition was added for systems generating non-consensual intimate content or child sexual abuse material, and SME relief was broadened to companies up to 750 employees and €150 million revenue.

The takeaway for Article 4: nothing in the Omnibus deal weakens the obligation that bites on 2 August 2026. Every organisation I’ve described above is still in scope.

What to do now

If you’re in compliance, L&D, legal, HR, or operational risk at a European organisation and you haven’t started, the practical sequence runs roughly as follows.

Start with an honest audit of which AI systems are actually in use across your organisation, including the ones staff have adopted without telling anyone. The real footprint is almost always broader than the policy suggests, and shadow usage is where most of the surprises live.

Map that usage to role groups. Who uses what, and what for. This becomes the basis for a differentiated training plan that the Commission’s FAQ effectively asks you to produce.

Assess your existing training against the Commission’s contextual standard. A generic e-learning module you’ve already rolled out may satisfy baseline awareness for some roles and fall short for others; knowing exactly where it falls short is more useful than scrapping it.

Then document what’s already been delivered, even retrospectively. Informal training, brown-bag sessions, Copilot enablement walk-throughs: write them down. The Commission doesn’t require certificates, but it does expect evidence, and most organisations have done more than they think they have once they actually count.

Finally, set a realistic rollout plan for the 10 weeks before 2 August 2026. You won’t perfectly train every role group in that window if you haven’t started yet. What you can show is good-faith progress and a credible plan, which is how most enforcement will land in the first year.

When August arrives, scrutiny will focus on organisations with nothing to show at all.

A note on why I wrote this

I run a consulting practice focused on Article 4 literacy programmes for European organisations. If this post’s useful and you want to compare where your organisation sits against the regulation, happy to chat. Book a free call via the link in the navigation, or message me directly.

No sales pitch. If your setup’s already in good shape, I’ll tell you that. If it isn’t, I’ll tell you that too.