The vendor of the recruitment system sends an email: their AI module falls under the AI Act, and they will take care of the documentation and the CE marking. The same week, the works council asks what the organisation itself has to arrange. HR looks at legal, legal looks at IT, and nobody is sure whether the answer is "nothing" or "a lot".
The answer sits in between, and it starts with your role. Under the AI Act almost every employer is a deployer: you use AI systems that someone else built. A concrete list comes with that role. Two obligations have applied to everyone since February 2025, one set applies from December 2027 to anyone using AI in recruitment, evaluation or dismissal, and everyday use of ChatGPT or Copilot mostly falls outside that set but not outside the GDPR.
Are you a provider or a deployer?
The AI Act distributes its obligations across roles. The provider develops an AI system or places it on the market under its own name. The deployer uses such a system under its own authority, for its own work. Personal, non-professional use does not count.
The heavy obligations sit with the provider: risk management, technical documentation, conformity assessment. That is what the vendor in the opening scene means, and it is correct. It does not mean the employer is done. The deployer has a shorter list of its own in the text of the law.
One exception worth knowing: you become a provider after all if you release an existing system under your own name, modify it substantially, or put it to a different purpose that makes it high-risk. An organisation that rebuilds a general language model into a screening tool for applicants moves from one role to the other. For anyone using a tool as delivered, this does not come up.
What already applies to every employer?
Two parts of the law have applied since 2 February 2025, whatever AI you use.
Article 4, AI literacy. Organisations take measures that support the AI literacy of the people who work with AI systems on their behalf. The Digital Omnibus softened this from an obligation of result to an obligation of effort, but the duty exists. What you can show for it is set out in AI literacy in 2026. It comes down to explaining what AI can and cannot do, which data does not belong in a prompt, and which tools the organisation has approved.
Article 5, the prohibited practices. Most of the prohibitions, such as social scoring by public authorities, do not touch an ordinary employer. One does: AI systems that infer the emotions of people in the workplace are prohibited, except for medical or safety reasons. That covers tools that score "enthusiasm" in a video interview, or software that reads an employee's mood from the tone of customer calls. If such a feature sits inside a package you use, you switch it off; the safety exception is narrow and meant for situations like fatigue detection for drivers.
Article 50 joined them on 2 August 2026: anyone running a chatbot for customers or citizens makes clear that they are talking to AI.
When is an HR application high-risk?
Annex III of the AI Act lists the applications that count as high-risk. Point 4 covers work. Two groups: AI for recruitment and selection (targeted job adverts, analysing and filtering applications, evaluating candidates), and AI for decisions within the employment relationship (promotion and termination, allocating tasks based on behaviour or personal traits, and monitoring and evaluating performance and behaviour).
The criterion is the intended purpose of the system, not the technology behind it. A simple scoring model that ranks CVs is high-risk. A language model that drafts a job advert is not. What most organisations actually encounter:
| Application | What the AI Act says | What you arrange |
|---|---|---|
| Recruitment tool that filters or ranks applications | High-risk (Annex III, point 4) | Article 26, from 2 December 2027 |
| Appraisal or scheduling system that scores performance | High-risk (Annex III, point 4) | Article 26, from 2 December 2027 |
| Emotion recognition in interviews or at the workplace | Prohibited (Article 5), since 2 February 2025 | Switch it off |
| Copilot or ChatGPT for text, summaries, translations | Not high-risk | GDPR, Article 4, your own policy |
| Chatbot for customers or citizens | Transparency (Article 50), since 2 August 2026 | Make clear it is AI |
The last two rows are where most employers live. What the GDPR asks there is covered in what the AI Act means alongside the GDPR.
What do you have to arrange as a deployer?
For a high-risk system the deployer's duties are in Article 26. In short:
- Use according to the instructions. The provider supplies instructions for use, and you use the system as they describe. A tool built to shortlist candidates is not used to reject them.
- Human oversight. You assign oversight to people who have the competence, the training and the authority to overrule an outcome. Someone who clicks "approve" without the option to decide otherwise is not oversight.
- Monitor the operation and keep the logs. You follow whether the system does what it should, report anomalies to the provider, and keep the logs the system generates insofar as they are under your control.
- Inform workers and their representatives. Before putting the system into use, you tell the affected workers and the works council that they will be subject to it.
- Use the DPIA. The information the provider supplies goes into the data protection impact assessment that the GDPR usually already requires in these situations.
For public bodies and a limited group of private organisations, a fundamental rights impact assessment comes on top (FRIA, Article 27). What ordinary privacy rules ask of HR beyond this is in working more safely with AI in HR.
Which dates count after the Digital Omnibus?
The Digital Omnibus on AI has been in force since 27 July 2026 and changed the timeline on one point: the obligations for high-risk systems in Annex III apply from 2 December 2027 instead of 2 August 2026. For AI embedded as a safety component in regulated products, the date is 2 August 2028.
Nothing else moved. The prohibited practices have applied since 2 February 2025, Article 4 too, and the transparency rules of Article 50 since 2 August 2026. The full timeline, with what did and did not shift, is in the AI Act timeline after the Digital Omnibus.
Three questions that get mixed up
The confusion in the opening conversation almost always comes from treating three different questions as one.
What is the system? The classification hangs on the intended purpose the provider has documented. It is in the vendor's documentation, and the vendor is responsible for it.
What do you do with it? The same product can be a writing aid in one organisation and a selection instrument in another. A recruiter who asks a general chatbot to put fifty CVs in order of suitability is not using a high-risk system in the sense of the law, but is making a decision to which the GDPR and the organisation's own policy apply.
Who is responsible? For use inside the organisation, that is always you, whatever the vendor has arranged. The vendor's email answers the first question and says nothing about the other two.
What does this look like in an organisation?
A mid-sized healthcare organisation has approved Copilot and circulated a short AI policy. An inventory shows ChatGPT, Perplexity and a meeting transcription tool in use alongside it. And the recruitment system HR has used for three years turns out, since an update, to have a feature that automatically ranks incoming applications. It was switched on, nobody had recognised it as AI, and it was described nowhere.
That is not an incident. It shows that "which AI do we use" is not a question you answer once, and that the high-risk application is not always the tool that was bought as AI. The organisation switches the ranking feature off until HR has decided what oversight looks like, adds the four other tools to the overview, and records a decision per tool. Whether such an overview is legally required is covered in is an AI register mandatory.
BeeSensible links a catalogue of AI tools to the use observed in the organisation. Every tool in the catalogue carries an EU AI Act classification and a risk score; the organisation sets the status itself, and everything starts at "no decision yet". What is observed is only the tool's domain, at most once per day per device, without a user id.
