The timeline: what is live now, and what comes next

The Act applies in stages, and the dates matter because many vendor pages still quote the old ones. Here is the current picture:

DateWhat it means for you
2 February 2025The AI literacy duty (Article 4) and the list of banned AI practices (Article 5) took effect.
August 2026National regulators began enforcing the Act. Each Member State sets its own penalties.
2 December 2027The full obligations for high-risk AI systems, including hiring and HR tools, apply. This date was moved from 2 August 2026 by the Digital Omnibus on AI, in force since 27 July 2026.

If a training vendor still quotes August 2026 for the high-risk deadline, their material has not caught up with the amendment. The bans and the training duty are live law today.

Three colleagues discussing AI compliance around a laptop in a bright office

What already applies: the AI literacy duty

Article 4 of the Act requires every provider and deployer of AI systems to take measures to ensure a sufficient level of AI literacy among their people. A deployer is simply a business that uses an AI system built by someone else, which covers almost every small business. The duty has applied since 2 February 2025, and there is no headcount or turnover threshold: a four-person practice using an AI tool is caught in the same way a bank is, as a Legal 500 briefing on AI literacy employer obligations explains.

It is broader than employees. Contractors, service providers, and anyone else operating AI systems on your behalf are in scope too. And it does not require a formal course. The EU AI Office, which coordinates AI policy, imposes no mandatory training format and no certificate. Its guidance suggests four practical steps: build a general understanding of AI in the business, clarify whether you provide AI systems or only deploy ones built by others, consider the risk level of the systems you use, and build training actions matched to each role. Simply pointing staff at a vendor manual written in technical language is unlikely to be enough.

One clarification arrived in July 2026. An amendment to Article 4 added an express note that you must take the measures, but you are not required to guarantee a specific level of literacy in any individual. The duty is about the steps you take, and you must be able to show them. Keep an internal record of training and awareness-raising as evidence in case of an inspection.

The part most small businesses miss: shadow AI

Before you train anyone, you need to know which AI tools are actually in use. The numbers on unapproved AI use are stark:

MeasureFigure
Employees using AI tools their organisation has not approved80%
Enterprises with any AI governance policy in place37%
Corporate-to-personal data transfers now flowing through generative AI32%
Additional breach cost when shadow AI is heavily involved$670,000
Median time to detect those breaches247 days
Drop in unauthorised use when approved tools are provided89%

Those figures, from a Cloud Security Alliance research note on shadow AI, matter for two reasons. First, you cannot document AI literacy for tools you do not know your staff use. Second, the fix that works is not a ban: when organisations provision approved AI tools, unauthorised use drops sharply. An inventory of every AI tool in the business is therefore the first compliance step, not an IT nicety.

What you must not do: the banned practices

Since 2 February 2025, the Act prohibits a list of AI practices outright. The one most relevant to small businesses is emotion recognition in the workplace or in education. A video interview tool that scores a candidate’s facial expressions or tone to infer enthusiasm falls under the ban, with only narrow medical or safety exceptions, as Belgian law firm Pitch sets out in its AI Act HR guide. If you use anything that claims to read how someone feels, switch it off and document that you did.

Hiring and HR tools: the December 2027 deadline

This is where most small businesses will feel the Act next. Annex III of the Act lists employment-related AI systems as high-risk: tools used for recruitment or selection, including placing targeted job advertisements, analysing and filtering applications, and evaluating candidates; systems affecting work terms, promotion, or termination; tools allocating tasks based on individual behaviour or personal traits; and systems monitoring or evaluating performance and behaviour.

The full high-risk obligations apply from 2 December 2027, after the Digital Omnibus on AI moved the date from 2 August 2026. The classification is what matters now, because you need the time to prepare.

If you use such a tool, you are a deployer, and the Act gives you specific duties. Use the system as its instructions say. Assign human oversight to people with real competence, training, and authority. Make sure the data you feed it is relevant and representative. Monitor it and keep the generated logs for at least six months. Inform workers’ representatives and affected workers before you put it into use, and tell the people subject to its decisions that the system is used. There is a narrow exception for tools that only perform a preparatory or procedural task with no significant risk, but the provider must document that assessment, and any tool that profiles people stays high-risk. No company is too small for these rules: there is no size threshold.

Hand writing notes on paper beside a laptop while documenting EU AI Act compliance requirements

What you do not need: a certificate

You will be sold one anyway. There is no EU AI Act certification scheme, no approved-provider list, no accreditation body for AI literacy training, and no certificate a regulator will ask to see, as learning provider Learnsignal explains in its guide to the certificate question. The European Commission has said so in plain language.

A certificate from a course is acceptable evidence that you took a measure, but it is not the measure itself. What a regulator wants is a documented, role-appropriate answer to four questions: who operates AI systems here, what does each role need to understand, what did you actually do and when, and how do you keep it current. Spend your budget on that, not on a framed PDF.

The three documents to have on file

No article of the Act names these three documents, but together they answer everything a regulator or inspector is likely to ask. Treat them as your compliance folder.

  1. An AI inventory. A simple list of every AI system the business uses or provides, who uses it, and for what. This is the foundation. You cannot train people on tools you have not identified, and an incomplete inventory becomes a legal gap once the high-risk obligations apply.
  2. A dated AI literacy record. For each role that operates AI, note what the people in it needed to understand, what training or awareness-raising you delivered, when it happened, and who delivered it. Keep it current as your tools change.
  3. A written AI use policy. Which tools are approved, what data must never be pasted into an AI tool, and who staff should ask before trying something new. Most businesses already have the makings of this in their staff handbook. It just needs the AI part written down.

What non-compliance can cost

BreachMaximum penalty
Banned practices (Article 5), such as emotion recognition at workUp to 35 million euros or 7% of global annual turnover
High-risk system obligationsUp to 15 million euros or 3% of global annual turnover
AI literacy duty (Article 4)No single EU-wide figure. Each Member State sets its own penalties, enforced by its national regulator.

These are headline maximums. Authorities must weigh the gravity of the infringement and the size and turnover of the business, so a small business is not treated like a multinational. Anyone, from a staff member to a competitor, can complain to the national regulator, so investigations can start from a complaint as well as from an inspection.

Frequently Asked Questions

Does the EU AI Act really apply to my small business?

Yes, the Act has no headcount or turnover threshold. If your business uses AI systems, even free consumer tools for drafting or customer support, the AI literacy duty applies. If you use AI in hiring, the high-risk rules apply from December 2027.

Do I need an EU AI Act certificate?

No, no official certificate exists, and no regulator will ask for one. What you need is a documented, role-appropriate training record showing who uses AI, what they were taught, and when. A course certificate can be part of that evidence, but it is never a substitute for it.

What is the minimum I should do in 2027?

List every AI tool in the business, decide what each role using AI needs to understand, deliver training proportionate to that need, and keep a dated record of what you did. Review the list whenever you adopt a new tool. That covers the duty that is already in force.

Our recruitment software only helps sort CVs. Is that high-risk?

If it analyses, filters, or ranks applications, or evaluates candidates, it generally falls into the high-risk employment category, and a tool that profiles people cannot use the narrow exception. The full obligations start on 2 December 2027. If it scores video interviews for emotion or enthusiasm, that is banned now, not in 2027.

Can the authorities fine a small business, or only large companies?

Any business can be sanctioned. Penalties are set nationally, and authorities must consider proportionality, including the size and turnover of the business, the gravity of the breach, and how cooperative you were. A documented training record and inventory are your best evidence of good faith.

Where to start

The AI literacy duty is the one obligation you can meet this month: map who uses AI in your business, teach each role what it needs to know, and write it down. If you would rather have the training built for you, Ailygence runs practical AI training for teams, matched to what each role actually does with AI. Book a call to discuss AI training for your team.