EU AI Act: what companies that use AI must do

Does the EU AI Act apply to SMEs that only use AI?

Yes. The AI Act attaches duties to roles and to uses, not to company size. A provider develops an AI system or has it developed and places it on the market or puts it into service under its own name. A deployer uses an AI system under its own authority in a professional context. A small firm that gives its sales team Copilot, or lets an agent pre-sort supplier invoices, is a deployer in the same way as a large group. Ask your lawyers whether any simplification applies to your firm. Here we assume that none does.

Many companies are still unsure where they stand. In Bitkom's 2026 survey (published 14 September 2026), 37% of companies believe they are affected by the AI Act as users of AI and 34% are still checking. Of the companies that are affected or still checking, 89% expect high implementation effort. In the BSI and TÜV-Verband AI security study 2026, reported by ZfK on 15 October 2026, 71% said they feel poorly informed about the AI Act's requirements.

It is different with an assistant of your own. If you build an assistant or agent, or have it built, and put it into use under your company's name, ask your lawyers whether you also count as its provider, because providers carry more duties than deployers. For the off-the-shelf tools most companies start with, the deployer role is the one that matters.

Which AI Act dates apply now, after the Digital Omnibus?

The Digital Omnibus on AI moved the high-risk dates and left the rest in place. According to the law firm Lewis Silkin (27 July 2026), the Omnibus was published in the Official Journal on 24 July 2026 and entered into force on 27 July 2026. Annex III high-risk obligations, which cover stand-alone systems in areas such as employment, move from 2 August 2026 to 2 December 2027. Annex I, which covers AI in regulated products, moves to 2 August 2028. AI embedded in products under the Machinery Regulation is taken out of the AI Act's direct high-risk rules. Article 50 transparency duties apply from 2 August 2026, and AI literacy under Article 4 remains as a duty to take measures.

Check the date of anything you read on the AI Act. Some German pages still online in November 2026 say that the full high-risk duties apply from August 2026, a date the Omnibus moved. In Germany, the Bundesnetzagentur supervises the AI Act. Its press release of 29 July 2026 announced that the German implementing act (KI-MIG) is in force, and it runs a KI-Service-Desk for companies with a self-check of a system's risk class. Banks and insurers should also expect BaFin to ask about AI: it published guidance on ICT risks of AI systems in financial firms on 18 December 2025.

AI literacy under Article 4: a training duty or a duty to take measures?

Article 4 asks deployers to take measures so that the people who use or run AI on their behalf have sufficient AI literacy. The rule has applied since 2 February 2025, as IHK Karlsruhe notes, and after the Digital Omnibus it remains a duty to take measures (Lewis Silkin, 27 July 2026). It does not name a course, a number of hours or a certificate. A lawyer puts the limit plainly: "Symbolische Maßnahmen genügen nicht" (Pöppel Rechtsanwälte, 5 May 2026), symbolic measures are not enough.

In practice, the measures that hold up fit the person's role and the tools they use day to day. Someone who drafts customer replies with an assistant needs to know where its answers come from, how to check a citation and which data must not go in. Someone who approves an agent's proposals needs to know what the agent cannot see and when to stop it. Keep a simple record per person or team: which tool, which content, when, and who delivered it. Our free AI usage review starts from the exports you choose to send and shows which tools are in use, which is the list your measures have to cover.

Many firms are behind on this. In Bitkom's 2026 survey, two thirds of companies rated their employees' AI skills as low and 25% offered no training at all. Our guide on change management for AI covers how to train staff on their own work.

Must I disclose that my chatbot is an AI?

From 2 August 2026, Article 50 requires that people are told when they interact with an AI system, and that certain AI-generated or manipulated content is marked as such. Some of these duties sit with the provider of the system and some with the deployer. For a chatbot your company configured and runs for customers, you are likely to carry both kinds in practice. Which of your outputs need a label, for example replies that an agent sends without a person reviewing them, is a question to settle with your lawyers.

Technically, a transparency notice is a small piece of work if it is planned in. The chat window shows a fixed sentence before the first message, so the notice is not hidden in the terms of use. Messages an agent sends carry a footer from a template, so nobody can remove it by editing a prompt. Each notice shown is logged with the version of its text, so you can prove what users saw on a given day. An internal assistant for staff should show the notice too: it costs nothing and makes the record complete.

Is our AI system high-risk? Recruiting, HR decisions and credit scoring

Most internal assistants are not high-risk. What makes a system high-risk is the decision it feeds. Annex III lists areas where AI is high-risk, and two of them reach ordinary companies. Employment: recruiting and selection, decisions on promotion and termination, and evaluating the performance of employees (Annex III point 4). Credit: assessing the creditworthiness or credit score of natural persons. Insurers should also note that risk assessment and pricing in life and health insurance is on the list. Gibson Dunn (27 May 2026) names "credit scoring" and "recruitment" among the Annex III examples. A tool that screens and ranks CVs is high-risk. An HR assistant that only answers questions about holiday rules from your works agreements does not match the high-risk examples in Grant Thornton's 2026 analysis. It still involves the works council when use can be traced to individuals.

For a deployer, high-risk means a set of duties from 2 December 2027. In outline, you use the system according to the provider's instructions for use. You give human oversight to people who have the competence, training and authority to overrule it. You make sure the input data you control is relevant for the purpose. You monitor the system in operation, report serious problems and keep the logs the system generates, as far as they are under your control. And you inform employees and their representatives before such a system is used at the workplace. Two rules apply already: Grant Thornton points out that emotion recognition at work is prohibited, and under Article 22 GDPR a fully automated rejection of an applicant without real human review is not allowed.

We do not build systems that decide about applicants, employees or credit without people. Where an assistant supports such work, it prepares information and a person makes the decision, with the reviewer, the reason and the outcome recorded. If you are unsure about a system's class, the KI-Service-Desk of the Bundesnetzagentur offers a self-check, and your lawyers make the call.

Is an AI officer (KI-Beauftragter) required by law?

The AI Act does not name such a role. Many companies appoint one to coordinate. Unlike the data protection officer under the GDPR, there is no mandatory AI officer: Article 4 asks for sufficient AI literacy of staff but names no role (skill-sprinters, 25 May 2026). In practice the coordination happens anyway. In the BSI and TÜV-Verband study reported by ZfK on 15 October 2026, 63% of AI-using firms had a designated AI responsibility.

Whoever takes the role, often the data protection officer or the head of IT, needs four things, and all of them should come from the systems: an inventory that updates itself from the gateway or licence exports, the risk class per use, the literacy record, and the notices and logs per system. With an LLM gateway, a single controlled access point to all models with keys, budgets and logs, most of that inventory comes out of a report and nobody has to fill in a survey.

What do we need to set up for each AI system we use?

Seven items, from inventory to supplier documentation. Work through the list once per system: the chat assistant, each agent, each tool a team bought on its own. Each line names what the system or the supplier should show you, so the answer is a setting or a document you can point to.

Data protection runs alongside: where personal data is involved, your data protection officer decides whether a data protection impact assessment is needed, and the works council has a say whenever logs make monitoring of performance or behaviour possible. Our guides on works council logging and on AI trace retention describe the settings in detail. If you want help with the technical side, talk to an engineer: the call is free, takes 30 minutes and needs no access to your systems or data.

AI system testingData protection officersWorks councils