AI tools enter a company faster than the paperwork that governs them. A team subscribes to something, it works, it spreads, and the question of who is accountable for it arrives well after the invoice does.
Eventually that question lands on a desk. Usually one desk in particular: do the people using these things understand what they do?
Not the features. Whether they know what the system does with what they feed it. Whether they can tell when the output is wrong. Whether anyone has told them the company now has legal obligations because of it.
That question has a legal answer. It has applied since February 2025, and it changed in July 2026 in a way much of the available guidance has not caught up with.
What the Obligation Actually Is
It lives in Article 4 of the EU AI Act. As it stands today, after the amendment:
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.
Two paragraphs follow it. The Commission and the Member States have to support providers and deployers, small and medium businesses first, and the Commission has to publish practical compliance examples on its single information platform. The AI Board has to adopt recommendations drawing on European competence frameworks.
If you want to know what “European competence frameworks” actually points at, Recital 8 of the amending regulation names them: the Digital Competence Framework for Citizens (DigComp) and the AI Literacy Framework for Primary and Secondary Education, with the Apply AI Alliance as the forum where the wider discussion happens. Those are open in a browser tab, which is more than can be said for most of what the Act references.
Three terms in that text decide whether it lands on you.
Provider means you developed an AI system, or had one developed, and either placed it on the market or put it into service under your own name (Article 3(3)). The second half matters: a bank that builds a tool for its own staff never sells it to anyone and is still a provider. Deployer means you use an AI system under your own authority, outside a purely personal activity (Article 3(4)). If your company bought a tool and switched it on, you are a deployer. Most organizations are. The two roles carry different obligations throughout the Act, and people get this wrong constantly, which is why I wrote a separate piece on it: Provider vs. Deployer Under the EU AI Act.
AI literacy has its own definition, in Article 3(56). Stripped down: the skills, knowledge and understanding that let people deploy AI systems with awareness of the opportunities and risks, and of the harm the systems can cause. That definition covers three groups. Providers, deployers, and affected persons, meaning the people on the receiving end of what the system decides.
Article 4 itself only binds the first two, and only regarding their own people. Hold that thought, because the third group comes back later.
Three Things That Make This One Unusual
It has applied since 2 February 2025. Not August 2026 with the transparency rules. Not December 2027 with the high-risk regime. February 2025, alongside the prohibited practices. It is one of the first obligations in the Act to bite, and it arrived while many organizations were still deciding who owned AI compliance internally.
Risk level is irrelevant. Article 4 sits in Chapter I, which applies across the whole Act. It does not care whether your system is high-risk, limited-risk or trivial. Your translation tool counts. Your chatbot counts. The coding assistant counts. There is no threshold to fall under.
It reaches past your payroll. The text says staff “and other persons dealing with the operation and use of AI systems on their behalf.” Contractors. Consultants. An outsourced support team operating your tooling. If they touch your AI systems on your behalf, they are in scope.
What sits outside the Act entirely, and therefore outside Article 4:
Systems used exclusively for military, defense or national security purposes (Article 2(3))
Third-country authorities using AI under law enforcement cooperation agreements with the EU (Article 2(4))
Systems developed solely for scientific research and development (Article 2(6))
Research, testing and development before a system goes on the market, real-world testing excluded (Article 2(8))
AI released under free and open-source licenses, unless it is high-risk or caught by the prohibitions or the transparency rules (Article 2(12))
One more boundary, because it comes up. Article 4 attaches to AI systems. If you provide a general-purpose AI model, the model alone does not pull you into Article 4.
What Changed in July 2026
On 27 July 2026, Article 4 was replaced in full by Regulation (EU) 2026/1744, the Digital Omnibus on AI. Not amended around the edges. Replaced.
The version the 2025 alerts, vendor decks and training proposals were built on read: take measures “to ensure, to their best extent, a sufficient level of AI literacy.”
Both of those phrases are gone. “To their best extent” is gone. “A sufficient level of” is gone.
The duty changed shape with them. The old text was an obligation of result: get your people to a sufficient level. The current text is an obligation of means: take measures that support the development of AI literacy. The provision then says so in its own words, stating that it does not require you to guarantee any specific level for any individual.
Recital 8 of the Omnibus explains the reasoning, which is that the stricter version was producing compliance burden out of proportion to what it delivered, particularly for smaller organizations.
Which matters commercially. If someone is selling you a program on the basis that you must prove a sufficient level, or certify staff against a benchmark, they are quoting a repealed provision. The obligation is real. That version of it is not.
What survived: the February 2025 application date, the Article 3(56) definition, Recital 20, the reach to contractors, the indifference to risk level, and the tailoring clause. That clause, the one about technical knowledge, experience, education, training and context, now does all the proportionality work “to their best extent” used to do.
If You Already Have Records From 2025
Article 4 applied from February 2025. The text changed in July 2026. There is no transitional provision covering conduct in between, which leaves a question the amendment did not answer: if an authority examines what you were doing in early 2026, which version of the standard governs?
The argument for the softer text is lex mitior, the retroactivity of the more lenient law. Article 49(1) of the Charter states it in criminal terms, and the Court of Justice has applied it beyond formally criminal penalties to administrative penalties that are criminal in character.
The argument has a real problem, and it is worth knowing before you rely on it. The Court applies lex mitior only where the amendment calls into question the unlawfulness of the earlier conduct. A relaxation going forward does not by itself reach backward. In Paoletti (C-218/15) the Court found no genuinely more lenient law where the change did not disturb the wrongfulness of what had been done, and the defendants lost on exactly that point.
Recital 8 does not help either. It says the stricter obligation created disproportionate burden, and in the same breath insists that AI literacy matters whether or not a regulator is asking. That is the legislature calling the old mechanism clumsy, not calling untrained staff acceptable.
So the honest position: the argument is available, I have not found it tested, and it faces two hurdles rather than one. Whether your Member State’s Article 4 penalty is criminal in character at all, and then whether the amendment disturbs the unlawfulness of what you did.
There is a fallback that needs none of this. Under Article 99(7), what you were doing in 2025 goes to gravity, to the organizational measures you had in place, and to whether any failure was negligent rather than deliberate. That works whichever text applies, and it rewards exactly the records this article is telling you to keep.
What the Commission Says You Have to Do
The European Commission maintains a Q&A on AI Literacy, updated on 27 July 2026 to reflect the amendment. It is the most concrete official guidance available, and it got a fraction of the attention the prohibited practices guidelines did.
Its tone is deliberately loose. No mandatory certifications. No required training hours. No pass or fail tests. In its words, there is “no one size fit all when it comes to AI literacy and no strict requirements or mandatory trainings are imposed.”
The floor it describes:
Build a general understanding of AI across the organization. What AI is, roughly how it works, which systems you use, what can go wrong with them.
Account for your role. Provider and deployer need different things. A company building AI systems needs a different literacy profile from one running off-the-shelf tools.
Identify and communicate the risks specific to your systems. People need to know what failure looks like in the tools they personally use, and what to do when they see it.
Tailor it to the people and the context. A data scientist and a claims handler do not need the same session. The provision says so directly.
And the part that tends to relieve people: the Q&A confirms Article 4 “does not entail an obligation to measure the knowledge of AI of employees.” No testing requirement. No certification requirement. You do not have to quiz your head of sales on what a neural network is.
What you do need is a record. The obligation is now explicitly one of means, and means are proved by evidence that you took them.
Three Levels of Training
Germany’s Bundesnetzagentur announced its role as market surveillance authority, single point of contact and complaints body on 29 July 2026. Its AI literacy guidance sets out three stages: a basic understanding of data and AI across the organization, then advanced AI literacy, then role-specific training with individual focal points such as technical, legal or ethical aspects. Its page has since been updated to cite Article 4 as amended, so read the current version before relying on it.
The Commission’s Q&A says separately that differentiated levels of training “could be appropriate.”
What follows is neither of those. It is mine, built on both.
Level One: General Awareness
Everyone in scope. What AI is at a conceptual level, which systems the organization uses, that legal obligations attach to them, what to do and who to tell when something looks wrong.
This is not a 40-hour course. An annual e-learning module, an internal AI policy people actually read, a session that gets recorded. The format is open. The evidence that it happened is not.
One point of precision worth keeping straight. Article 4 reaches staff and others dealing with the operation and use of AI systems on your behalf. A warehouse worker who never touches one is arguably outside it. Training the whole organization anyway is good governance and it is what I would do. It is not what the provision requires. Keep the two apart in your policy, because one is a legal obligation and the other is a choice you can revisit.
Level Two: Role-Specific Competence
Anyone who operates a system, or decides anything on the strength of its output.
What the system does. What data it uses. Where it fails. When to distrust it. The claims handler needs to know that “flagged” is a probability score rather than a verdict, and needs the false positive rate that goes with it. The loan officer reviewing an AI credit recommendation needs the model’s inputs and limits, not to retrain it, but because they are about to decide something about a person’s mortgage.
This is where the real work sits, and where a program that stopped at level one will show it.
Level Three: Oversight Competence
For high-risk AI systems, the Act requires named humans in an oversight role, and it sets a standard for them.
The provision is Article 26(2): “Deployers shall assign human oversight to natural persons who have the necessary competence, training and authority, as well as the necessary support.”
What those people must be able to do is set out in Article 14(4): understand the system’s capacities and limitations, stay alert to automation bias (the human tendency to trust a machine’s output more than it deserves), interpret the output correctly, decide to disregard or override it, and intervene or stop it.
Two different duties, and the split matters. Article 14 binds the provider, who has to build the system so oversight is possible at all. Article 26(2) binds the deployer, and it is about who you put in the chair.
Timing: both apply from 2 December 2027 for stand-alone high-risk systems, and 2 August 2028 for high-risk AI embedded in regulated products. So the oversight duty is not live yet. The literacy obligation that feeds it has been live since February 2025, which is the useful asymmetry. You have roughly one budget cycle to build the competence that the 2027 deadline will assume you already have.
Worth planning around: the Article 6(4) duty to document why an Annex III system is not high-risk was deferred to the same December 2027 date. Classification work and oversight-competence work now land in one window, and they depend on each other. You cannot staff an oversight seat for a system you have not classified.
What It Costs When It Goes Wrong
Article 4 appears nowhere in Article 99, the penalties provision.
The tiers are: Article 99(3), breaches of the prohibitions in Article 5, €35 million or 7% of worldwide turnover. Article 99(4), €15 million or 3%, covering Articles 16, 22, 23, 24, 25(2) and (4), 26, the notified body duties and Article 50. Article 99(5), €7.5 million or 1%, for supplying incorrect information.
Article 4 is in none of them. If you have seen a headline fine attached to AI literacy, it was imported from a tier that does not cover it.
One nuance on those ceilings, since it cuts the other way for smaller businesses: where the Act sets a fine as an amount or a percentage of turnover, the higher figure applies to most operators, but for SMEs it is the lower of the two.
Exposure for Article 4 runs through Article 99(1) instead, which requires Member States to lay down penalties for “any infringement of this Regulation by operators,” effective, proportionate and dissuasive, under national law that was due by 2 August 2025. The Commission’s Q&A confirms national market surveillance authorities can sanction Article 4 infringements. No EU ceiling, 27 national regimes, and rather less predictability than a single number would have offered.
The mechanism that actually matters sits further down the same article. Article 99(7)(g) tells authorities to weigh “the degree of responsibility of the operator taking into account the technical and organisational measures implemented by it,” alongside 99(7)(a) on gravity and 99(7)(i) on whether the infringement was intentional or negligent.
Your literacy program is not primarily something you get fined for missing. It is something that gets weighed when you are being fined for something else.
The Q&A says the same thing more plainly: enforcement “might, however, be more likely if there is proof of an incident due to lack of appropriate training and guidance of employees or other persons.”
Two further points from that page.
The first concerns whether anyone can sue you directly. The Q&A is careful: the AI Act creates no criminal offences and no right to compensation of its own. It does not stop there, though. It raises private enforcement as a live route, describing the scenario where someone harmed by an organization’s failure under Article 4 sues for damages, and notes that a provider can be sued according to national law. So the Act closes no doors it did not open. Whether an employee or an affected person has a claim is a question of national tort and employment law, and since the AI Liability Directive was withdrawn in 2025, national law is now the entire field.
The second is smaller. The page gives two different dates for when supervision started, saying 3 August 2026 in one answer and 2 August 2026 in another. Both are on the page as of writing.
Three Situations
The Insurer
Fraud detection flagging suspicious claims, plus a model assisting underwriters with risk assessment.
Level one is an annual briefing. Level two is the real work: the claims team understanding that a flag is not a verdict, the underwriters understanding when to override. Level three is the senior underwriter in the oversight seat, from December 2027.
One qualifier before you build any of that. Annex III point 5(c) covers risk assessment and pricing in life and health insurance. General claims fraud detection is not automatically high-risk, and a good deal of the insurance AI population sits outside the category. Classify before you assume, because the whole of level three depends on the answer.
The 30-Person E-Commerce Company
Shopify’s product recommendations, ChatGPT for customer service drafts, an AI accounting tool. They built none of it. Pure deployer, and Article 4 applies anyway.
At that size: a one-page policy listing the tools and the rules, a team briefing, and making sure the person drafting customer replies knows what a hallucination is before one reaches a customer. No literacy officer, no quarterly assessment cycle. Something written down and demonstrable.
The bar is lower. It is not zero.
New since July, and useful for exactly this company: the Commission and the Member States now have their own obligation to help businesses this size, and the Commission has to publish practical examples. Check what has appeared before you build from scratch.
The Startup Building an AI Product
Five technical people building an AI legal research tool. They can explain attention mechanisms over coffee. Their literacy gap runs the other direction.
As a provider, they need to know what obligations attach to their product, what documentation the Act requires, and that their law firm customers will have obligations the product has to support. That knowledge does not need a formal program. It does need to exist, be demonstrable, and belong to a named person.
The Third Group
Article 3(56) covers providers, deployers, and affected persons. Article 4 only imposes the duty on the first two. Recital 20 goes further, saying affected persons need the knowledge to understand how AI-assisted decisions affect them.
That is not an obligation to train your customers. It connects to the transparency rules instead, and those are worth attributing correctly. Articles 50(1) and 50(2) sit on providers: telling people they are interacting with an AI system, and marking synthetic output in a machine-readable format. The deployer duties are 50(3), for emotion recognition and biometric categorization, and 50(4), for deepfakes and certain public-interest text.
The practical version holds regardless: your people need to understand the system well enough to explain its output to the person on the other end. When the policyholder asks why the claim was denied, “the model said so” is not an answer, and it is not a transparency defense either.
Why This Is the One to Get Right First
Every other obligation in the Act assumes it.
Risk classification needs someone who understands the system well enough to judge which Annex III category it falls into, if any. Conformity assessment needs people who can prepare and review technical documentation. Human oversight requires competence by name. Transparency requires knowing what the system does before you can tell anyone else. Post-market monitoring requires recognizing a problem. Incident reporting requires knowing what counts as a serious incident.
Each of those is a duty performed by a person who understands what they are looking at. Article 4 is that assumption, written down.
Which is also why it will never headline a conference panel. It does not generate €50,000 consulting proposals or dramatic LinkedIn posts. It sits underneath everything, doing structural work, and the organizations that skip it find every subsequent obligation harder, every assessment thinner, and every oversight function weaker than it looks on paper.
What to Do
Inventory first. Which AI systems, used by whom, in what context. You cannot build a program around tools you have not mapped. The same exercise feeds classification later, so none of it is wasted, though note that Article 6 was itself amended and the written non-high-risk assessment in Article 6(4) is now deferred to December 2027 with the rest of the high-risk regime.
Sort your people into levels. General awareness for everyone in scope, role-specific competence for the people working with AI daily, oversight competence for anyone who will sit in an Article 26(2) seat. The Commission’s Q&A supports differentiating.
Build the floor. What AI is, what you use, what the risks are, who to tell. E-learning, a policy document, a briefing. Format is yours. Documentation is not optional.
Then go deeper where it counts. Role-specific training for operators and decision-makers. For future high-risk systems, start on oversight competence now rather than in late 2027.
Write it down. The amendment softened the standard and did nothing to the evidence problem. An obligation of means is proved by records of the means: what training, for whom, when, on which systems, and how the program changes as your AI estate does. “We had some conversations” is not a record.
Check what your providers give you. Your ability to train anyone on a system is capped by what its documentation says about limitations and intended use. If that documentation is thin, your literacy program has a hole in it that is not your fault and is still your problem. Better to raise it with the vendor before a regulator raises it with you.
Watch the single information platform. The Commission owes practical examples under the new Article 4(2), with no deadline attached. Whether they arrive tells you something about how seriously the support half of this obligation was meant.
Where That Leaves You
The obligation got easier to satisfy and harder to point at. You can no longer be told you failed to reach a sufficient level, because there is no longer a sufficient level to reach.
What is left is whether you took measures, and the only evidence of that is what you wrote down while you were taking them.
The legislature appears to have seen the reaction coming. Having softened the obligation, Recital 8 adds that AI literacy “should be a strategic priority, regardless of regulatory obligations and potential sanctions.”
That is an unusual sentence to find in an amendment written to reduce compliance burden, and it is the most useful one in it.
The regulation gives you flexibility on the how. Not on the whether.
Like this article? The AI Governance Roadmap is something you can use in practice. It’s a company’s roadmap with steps and working documents: the classification, inventory spreadsheet, role assignment, vendor questions and AI policy template. Everything a company needs to govern AI, in one place.
I really enjoyed reading this - thank you for making the AI Act a little less legalese and a little more English ;)
I was reading the other day about AI literacy vs AI enablement and how they’re really not the same thing. Giving people tools to use is not the same as actually knowing what those tools could do. I’m also guilty of automation bias, especially when it’s something a little less cognitive - and “sure, good enough” won’t hold up in an audit!
I didn't consider the impact of the DPO requirement. Thank you!
It definitely set a clear standard of accountability to get compliance started. The tiering system you mention doesn't have the same dedicated role with an expert required.