The EU AI Act in practice: what SMEs should really have in place by August 2026
There is probably an email in your inbox recommending an expensive compliance project because you use a transcription tool. My view, as someone who builds these systems: for a company with 50–500 employees that uses AI rather than developing it, that is, in my assessment, usually overkill. The AI Act gets serious in one place only, AI in personnel decisions, and only from December 2027.
The opposite position says you should treat everything as high-risk to be on the safe side. That spreads attention across cases the law does not cover. You end up with a thick file and still nobody overseeing how job applications are pre-screened. The audit disaster connects professional review, ownership and retrievable evidence. For SMEs, knowledge validation should define who checks information used in personnel decisions rather than substitute a general compliance file for oversight.
This article reflects the state of play in October 2026. It is a technical assessment, not legal advice.
What the Digital Omnibus changed
Regulation (EU) 2026/1744, the so-called Digital Omnibus on AI, entered into force on 27 July 2026. It amends the AI Act (Regulation (EU) 2024/1689). Three changes matter for SMEs.
- The Chapter III obligations for Annex III high-risk systems, such as candidate screening or performance assessment, apply from 2 December 2027 (Art. 113, point (c)). AI built into regulated products under Annex I follows on 2 August 2028.
- Article 4 on AI literacy has been rewritten. Providers and deployers must take measures to support the development of AI literacy among their staff. The text now states that they do not have to guarantee any specific level for any individual.
- The transparency obligations in Article 50 have applied since 2 August 2026. Only the machine-readable marking of AI-generated content (para. 2) gets extra time: systems already on the market have until 2 December 2026. That duty falls on the provider.
Deployer or provider
An SME that uses AI systems is a deployer under Article 3(4): it uses an AI system under its own authority. A provider (Art. 3(3)) develops a system, or has it developed, and places it on the market under its own name. Technical documentation (Art. 11), conformity assessment (Art. 43) and CE marking (Art. 48) for high-risk systems are the provider's job. Get the roles confirmed in the contract; a slide from the sales pitch does not count.
Three questions for every use case
The same tool can be harmless or high-risk. Minutes for a project stand-up fall into no high-risk category. Use the same technology to filter applications and you are in Annex III, point 4. So list use cases, not product names, and ask three questions about each.
- Does the system prepare decisions about employees or applicants, such as selection, promotion, dismissal, task allocation or performance evaluation? Then it is a candidate for Annex III, point 4, and from 2 December 2027 the deployer obligations in Article 26 apply to you.
- Does the system infer employees' emotions from voice, face or other biometric data? Then stop. Emotion recognition in the workplace has been prohibited under Article 5(1)(f) since 2 February 2025, except for medical or safety reasons.
- Do customers talk directly to an AI, or do you publish realistic AI-generated images, video or audio? Then Article 50 applies. The provider builds the AI notice into the chatbot (para. 1). Deepfakes, and AI text published to inform the public without editorial review, are yours to disclose (para. 4).
If a use case clears all three questions, Article 4 is essentially all that is left for it. I expect that to cover the large majority of use cases in a typical SME: drafting, translation, research, minutes.
Transcription: the notice comes from the GDPR
A meeting transcription tool is usually not an Article 50 matter. Everyone in the room still has to know they are being recorded, but that duty sits in the GDPR (the information obligations in Art. 13). If you have a works council, co-determination rules apply on top.
I recommend a fixed sentence at the start of every recorded meeting and a works agreement stating that recordings are not used to monitor performance. Our article on documenting employee reviews shows what the documentation can look like.
AI literacy: measures with evidence
Article 4 has applied since 2 February 2025. You owe measures, not a guaranteed outcome. You should still be able to show them. I would start with three building blocks:
- a training session built on real prompts and real mistakes from your own business
- a one-page handout per system
- a dated attendance list
Human oversight: Article 14 is for the provider, Article 26 is for you
Many guides cite Article 14 on human oversight. It is addressed to providers: high-risk systems must be built so that people can oversee them effectively, discard an output or stop the system. Your obligations as a deployer are in Article 26 and apply to Annex III systems from 2 December 2027. The ones that matter most for an SME:
- use the system in line with the provider's instructions for use (para. 1)
- assign oversight to people with the necessary competence, training and authority (para. 2)
- monitor operation and report risks or serious incidents (para. 5)
- keep automatically generated logs under your control for at least six months (para. 6)
- as an employer, inform workers' representatives and affected staff before the system goes live at the workplace (para. 7)
- inform people who are subject to decisions the system helps make (para. 11)
In practice: one named person per use case, a fixed review interval, a documented way to report errors. I have written separately about why I think human in the loop is the right default even where the law does not require it.
A roadmap in three meetings
- Collect every AI application actually in use, including the unofficial ones. Don't ask which tools people have; ask who used AI for what last week.
- Run the three questions over the list and name an owner for each use case.
- Set the training, the notice wording and the review intervals.
You end up with a document of a few pages, because the AI Act does not require deployers to run a certified AI management system or commission an external audit. Update it once a year, and whenever a use case drifts towards personnel decisions. That is the moment a minute-taking tool can turn into an Annex III case.
I don't see the biggest benefit in compliance at all. You see in writing where AI is at work in your company, and that is the best starting point for your next investment decision. Our article on responsible AI in practice shows how to build a framework from it, and the topic hub AI for mid-sized companies covers the budget questions.
Sources
- Regulation (EU) 2026/1744 (Digital Omnibus on AI), in force since 27 July 2026
- Regulation (EU) 2024/1689 (EU AI Act)
- Council of the EU, press release of 29 June 2026: final adoption of the Digital Omnibus on AI
- European Commission AI Act Service Desk: Article 4 (AI literacy), consolidated text
- AI Act Service Desk: Article 5 (prohibited practices)
- AI Act Service Desk: Article 26 (obligations of deployers of high-risk AI systems)
- AI Act Service Desk: Article 50 (transparency obligations)
- AI Act Service Desk: Article 113 (entry into force and application)
Frequently asked questions
Which EU AI Act obligations apply to SMEs?+
As of October 2026, an SME that only deploys AI is mainly bound by two articles: the AI literacy measures under Art. 4 and the prohibitions under Art. 5, both in force since February 2025. Art. 50 affects you directly if you publish deepfakes or AI text meant to inform the public. The broader deployer obligations in Art. 26 follow for HR-related Annex III systems on 2 December 2027.
Does an SME need certification for AI applications?+
No. The AI Act does not require deployers to be certified. Conformity assessment and CE marking are obligations of the provider of a high-risk system. If you plan to use a recruiting or assessment tool from 2027, ask the vendor now when it expects to have both, and don't lock in the rollout before then.
How much effort does implementation take?+
I would plan for the three meetings in the roadmap. How long they take depends mostly on how many use cases you find. The result is a table with a few columns per use case: system and vendor, your role, the outcome of the three questions, the owner, evidence of training, the review interval and the date of the last review.
Related articles
In a 30 minute software demo we show you how RECO secures and transfers knowledge in your company. Request a demo now and we'll get back to you with a proposed time.
