- The EU AI Act's AI literacy duty in Article 4 applied from 2 February 2025, and national authorities only began supervising and enforcing it on 2 August 2026.
- Digital Omnibus amendments took effect in mid July 2026 and shifted the emphasis away from prescriptive obligations on businesses towards promotion by the Commission and member states.
- Training duties for human oversight of high risk systems survived that change, so the shift is a narrowing rather than a repeal.
- The rule never required you to measure what your staff know, and it never required a specific level of knowledge from any individual.
- For a shop of two or three people the useful document is one page about what may be pasted into which tool and who checks what, not a compliance manual.
- Outside Europe there is no equivalent legal duty, and the NIST framework that fills the gap is explicitly voluntary.
If you employ anyone and you use AI, you have probably seen a headline telling you that European law now requires AI training. It did, then the rule moved, and the enforcement date arrived after the amendment. The result is a subject where most of what is written is either out of date or describes an obligation designed for companies a thousand times your size.
The short answer for a small shop is that you need a page, not a programme. The longer answer is worth having because the dates are specific and the parts that survived the amendment are the parts that were always the point.
Does European law actually require an AI policy?
It requires measures that support AI literacy, which is not the same thing as a policy document and never was. Nothing in the text names a document, a course or a certificate.
The obligation sits in Article 4 of the AI Act. As the text of Article 4 on AI literacy sets out, providers and deployers must support the development of AI literacy among staff and other people operating AI systems on their behalf, taking account of those people's technical knowledge, experience, education and training, and of the context in which the systems are used. The regulation defines AI literacy as the skills, knowledge and understanding that allow someone to make an informed deployment of AI systems and to be aware of the opportunities and risks.
Read the qualifier in that sentence, because it is what makes the duty workable for a business of three. The measures have to account for the experience of the people involved and the context. A shop where one person uses a chatbot to draft product copy is not in the same position as a bank running credit decisions, and the rule was written to say so.
What changed in July 2026?
The emphasis moved from the business to the state. The Digital Omnibus amendments took effect in mid July 2026 and shifted the focus of Article 4 towards the Commission and member states promoting AI literacy, rather than towards prescriptive requirements on providers and deployers.
The European Commission's own questions and answers on AI literacy records both the change and the timeline around it. The Article 4 obligation entered into effect on 2 February 2025. Supervision and enforcement by national market surveillance authorities began on 2 August 2026. The Digital Omnibus amendments landed in mid July 2026, a fortnight before that enforcement date.
One thing did not move, and it is the part that matters most in practice. Training obligations connected to human oversight of high risk systems were maintained. So the amendment narrowed a broad duty rather than removing the idea that people supervising consequential automated decisions should understand what they are supervising.
This area is genuinely in motion, and anything you read about it, including this article, carries a date. If your business depends on the answer, the Commission's own questions and answers page is the thing to check rather than a summary of it. What follows is about what is useful to do, which has changed far less than what is formally required.
What does AI literacy mean in a shop of three people?
That the people using the tools know what the tools get wrong. Not architecture, not training data, not model names. The failure modes, and what to do when one shows up.
The Commission is unusually direct about the limits. Article 4 does not entail an obligation to measure the knowledge of AI of employees, and providers and deployers are not required to guarantee any specific level of AI literacy for any individual. So the compliance products that offer to test and certify your staff are selling something the rule did not ask for.
| Question a shop asks | What the rule says | What is actually worth doing |
|---|---|---|
| Do I need a written policy? | No document is named in the text | Write one page anyway, because it is how three people stay consistent |
| Do I need to test my staff? | No obligation to measure employee knowledge | Have one conversation about what the tool gets wrong |
| Does it apply to contractors? | Covers others operating systems on your behalf, depending on context | Include your freelancers in the same page |
| Does it apply outside the EU? | This duty is European | The practical content is identical wherever you are |
The second row is the one that saves money. A conversation about failure modes, held once and repeated when something changes, is both closer to what the regulation describes and considerably more useful than a slide deck nobody remembers.
What should actually be in the page?
Four things, and none of them are about artificial intelligence in the abstract. They are about your data, your customers and who is accountable when the output is wrong.
What may be pasted where. This is the single most valuable line in any small business AI policy. Customer names, addresses, order histories, supplier pricing, anything under a confidentiality term: say explicitly which tools these may go into and which they may not. Most small businesses have never made this decision, which means it is currently being made independently by whoever is busiest, and we worked through the reasoning in a piece on what customer data you may paste into an AI tool.
What must never be published unchecked. Prices, stock availability, delivery promises, returns terms, safety or ingredient information, anything a customer could treat as a commitment. The rule is not that AI cannot draft these. It is that a person signs them off.
Who checks, by name. In a business of three, accountability that is not attached to a person does not exist. One line saying who reviews generated customer communications before they go out is worth more than a paragraph of principles.
What to do when it goes wrong. A generated sentence made a promise you cannot keep, or an answer sent a customer the wrong information. Who is told, how fast, and who talks to the customer. Deciding that in advance takes ten minutes and saves a bad afternoon.
That is the whole document. It fits on one side of paper, it is understandable by somebody who started last week, and it is closer to the spirit of a literacy obligation than any external course.
What about the tools nobody told you about?
They are the real exposure, and no policy reaches them if it is written as a prohibition. Staff use AI tools that help them, and a rule that forbids it quietly produces the same behaviour with no record of it.
The pattern is predictable. Somebody drafts a customer reply in a free chatbot on their phone because the approved tool is slower, and a customer's complaint, including their name and order number, leaves your business without anyone deciding that it should. Nothing malicious happened. The approved path was worse than the unapproved one.
So the page should name at least one tool that is allowed and easy. A policy consisting only of prohibitions is a policy that gets routed around, and the version that works says here is what to use, here is what never goes into it, and here is who to ask if you are unsure. That last clause matters more than it looks, because the alternative is somebody guessing silently.
There is a second category worth an explicit sentence: AI features switched on inside software you already pay for. Those arrive without a decision being taken, frequently by default in an update, and they process whatever the host application can see. We went through how to audit them in a piece on the AI features your business tools turned on without asking.
Who counts as working on your behalf?
More people than a small business usually pictures. The duty reaches other persons dealing with the operation and use of AI systems on your behalf, and the Commission's guidance names contractors, service providers and clients as possible examples depending on the risk context.
For a shop that matters because the AI in your business is frequently not operated by you. A freelance copywriter drafting your listings, an agency running your ads, a virtual assistant answering your inbox, a bookkeeper reconciling your invoices: each may be using AI on work that carries your name, with your customers' data in it, under no instruction from you at all. If you have never asked, you do not know, and the honest first step is a single question in an email rather than a clause in a contract.
The practical version is to extend the same one page to anyone who touches your customer data, and to say so when you engage them. Two sentences in a scope of work covering what may be pasted where, and who signs off on customer facing text, does the job. It also surfaces a useful fact early, which is that some freelancers will tell you they use AI heavily and some will tell you they do not, and both answers change how much checking their output needs. The supplier side of that conversation is the same discipline we described for reading what supplier terms commit you to.
What help are you supposed to be getting?
Rather a lot, on paper. Article 4 does not only place a duty on businesses, it places one on the institutions, and small firms are named specifically as the intended beneficiaries.
Paragraph 2 of the article requires the Commission and member states to support and facilitate the efforts of providers and deployers, in particular small and medium enterprises, to meet the obligation, and requires the Commission to publish practical examples of compliance on its information platform. Paragraph 3 tasks the European AI Board with adopting recommendations for promoting AI literacy, taking account of European competence frameworks and common objectives. Both paragraphs were added by amendment.
Two things follow for a merchant. The first is that free official material exists and is the correct place to look before buying anything, because an example published by the body that wrote the rule is better evidence of compliance than a course sold by a company that did not. The second is a matter of proportion: enforcement by national market surveillance authorities is meant to be proportionate to the severity, nature and intentionality of an infringement, which is regulatory language for the fact that a small shop making a reasonable effort is in a very different position from one that made none and knew it.
That proportionality is worth taking seriously in both directions. It is a reason not to panic, and it is also the reason a dated page beats no page at all, because effort you cannot evidence looks identical to effort you did not make.
Is there an equivalent if you are not in Europe?
No binding one, and the main framework available is voluntary by design. That is a real difference in obligation and almost no difference in what a sensible shop does.
The United States reference point is the NIST AI Risk Management Framework. Its core is built from four functions: govern, map, measure and manage. Govern is described as cross cutting, establishing organisational structures, policies and accountability mechanisms, and infused through the other three. Map establishes the context and the potential impacts. Measure analyses and monitors risk. Manage allocates resources to the risks that were mapped and measured, and includes incident response planning.
Stripped of the vocabulary, that is the same one page. Govern is who checks. Map is which tasks you use AI for and what could go wrong in each. Measure is noticing when the output is wrong. Manage is what you do about it. A framework built for organisations with risk committees collapses neatly into a paragraph when the organisation is three people, and the collapse is legitimate rather than a shortcut.
If you sell into Europe from outside it, the broader obligation map is a separate question with a different answer, which we set out in our guide to which AI Act obligations actually reach a small seller.
How long should this take?
An hour to write and twenty minutes a year to revisit. Anything longer is a sign you are writing for an imagined auditor rather than for the people who work with you.
Start by listing the AI tools genuinely in use, which usually takes ten minutes and produces at least one surprise. For each, write down what it touches and who uses it. Then make the four decisions above and write them as sentences, not as policy language. Then, and this is the part that makes it real, send it to everyone and ask whether any of it describes something they already do differently. That question finds the gap between the document and the business faster than any review.
Revisit it when you add a tool, when a tool adds a feature, or once a year, whichever comes first. Date it, because an undated policy tells a customer, a regulator or an insurer nothing about whether it was ever true.
What this is really for
Not the regulator. The odds of a market surveillance authority examining a three person shop's AI literacy measures are low, and writing a document in anticipation of that is a poor use of an hour.
It is for the afternoon when a customer receives an answer that was confidently wrong, and you need to know who wrote it, what information the tool had, and whether anyone was supposed to read it before it went out. Every one of those questions is unanswerable without a decision you made earlier, and every one of them is answerable in a sentence if you did.
If you want the shape of a public facing version, the policy we publish about how AI is used in our own product is at our AI policy page, and it is deliberately short for the same reasons argued here. The customer facing side of this, meaning the terms and notices your shoppers actually read, is a different document with different rules, covered in our piece on the policies an online shop publishes about AI.
One page, four decisions, a date at the bottom. That is the whole obligation as it applies to a business your size, in the only sense that will ever matter to you.