Blog
The EU AI Act for SMEs: what applies when, and what you need to sort out now
· 5 min read
You have put a chatbot on your website, you use AI to pre-sort incoming job applications, or you have a tool generating copy and images for marketing. Somewhere you heard that a major AI law came into force in August 2026. Somewhere else you read that those same rules have just been postponed. Both reports are correct, and that is what makes it confusing. The EU AI Act works with a timeline of obligations that take effect at different moments, and that timeline was also amended this year. Below, we set out what already applies, what has applied since recently and what has been postponed, with the official sources alongside.
What has been mandatory for a while: prohibited practices and AI literacy
The first two parts of the AI Act have applied since 2 February 2025: a list of prohibited AI practices and the obligation to ensure a sufficient level of AI literacy among people who work with AI. That is how it appears on the official timeline of the European Commission.
For most SMEs, the prohibited practices are not an everyday concern, but two of them directly affect HR-type applications. According to the explanation from the European Commission's official AI Act Service Desk, emotion recognition in the workplace and in education is prohibited, and the same goes for social scoring: classifying or assessing people on the basis of their behaviour or personal characteristics. If you use an AI tool that tries to measure employees' mood, motivation or engagement, it may fall under this ban, and phasing it out is not something to do "whenever it suits".
The AI literacy obligation is less sharply defined but broader: according to the European Commission's FAQ on the AI Act, it has applied since 2 February 2025 to everyone who provides or uses AI systems. There is no fixed number of hours or mandatory course attached to it, but there is an expectation that people who work with AI understand what the system can and cannot do, and what risks come with it.
Since August 2026: transparency for chatbots and AI content
A second layer came into force recently. According to the European Commission's FAQ on the AI Act, providers of chatbots and similar systems must make sure users know they are dealing with AI, AI-generated text, images, audio or video must be recognisable as artificially generated in a machine-readable format, and a deepfake must be visibly labelled as artificial. These transparency rules belong to the part of the European Commission's timeline that names August 2026 as its date.
What this means in concrete terms for a chatbot or an AI agent on your own channels, including the difference between the two, is set out in more detail in our earlier article on chatbots and AI agents. The main point here is this: it is no longer the distant future, it applies now.
The rules for providers of general-purpose AI models, the large language models that many applications run on, and the associated governance structure have also been in force since 2 August 2025, according to the official timeline of the European Commission. For an SME that uses such a model through a supplier, this is mainly useful to know: the obligation then lies primarily with the model maker, not with you as the user.
High-risk AI: postponed, but not postponed indefinitely
The most confusing news of this year is the shift in the strictest category: high-risk AI systems. Under the original European Commission timeline, these would have had to meet heavy requirements from 2 August 2026, such as a mandatory risk assessment, human oversight and documentation. That is no longer the case.
The European Commission reports that the Digital Omnibus on AI entered into force on 27 July 2026, and that the rules for the high-risk systems in Annex III, such as applications in recruitment and selection, education and biometrics, now apply from 2 December 2027. For high-risk AI embedded in physical products, such as lifts or toys, the date has moved to 2 August 2028. That is stated on the European Commission's news page about the Digital Omnibus and confirmed on the official timeline page. According to that same page, the reason for the postponement is that more time is needed to get standards, guidelines and supervisory authorities ready before companies are held to the requirements.
The Commission itself gives examples of what will soon count as high-risk: AI that places targeted job advertisements, filters applications or evaluates candidates, AI that assesses learning outcomes or steers the learning process, and AI for biometric identification or categorisation. If you already use such a system, the postponement gives you more time to get it in order, but the requirements are coming, and the date is now fixed in legislation that has already entered into force.
What does this mean for your business?
There are four concrete things you can do right now, regardless of how the rest of the timeline develops.
Find out whether there is an AI application running anywhere in your business that tries to measure or classify the emotions, mood or behaviour of employees or customers. If that counts as emotion recognition in the workplace or as social scoring, it is already prohibited and needs to be switched off, not scheduled for later.
Make sure the people in your business who use AI tools know what the system can do, where it goes wrong and when a human check is needed. That does not have to be formal training, but it should be something you can point to if someone asks about it.
Check your chatbot, your generated marketing copy and any AI images against the transparency requirements that apply now: does a visitor know they are talking to AI, and is generated content recognisable as such?
Look ahead to the systems that will count as high-risk in 2027 and 2028, particularly if you use AI in recruitment and selection or in assessing people. The postponement gives you room, not an exemption: start documenting in good time what a system does and why, so you do not have to catch up all at once just before the deadline.
Would you like to go through which AI applications in your business fall into which category and what that means in practice? Get in touch with us via the contact page. We are happy to help, drawing on hands-on experience with AI and OutSystems development, not just a legal checklist.