In this article9 sections
Your employees are already using it
Does the EU AI Act apply to ChatGPT when your staff use it for daily tasks? The answer, for most European SMEs, is yes. The moment an employee pastes a customer complaint into ChatGPT to draft a response, or uses it to screen CVs, your organisation steps into a legal role defined by the Act itself.
This is not something sitting somewhere in 2030. The EU AI Act entered into force on 1 August 2024. The AI literacy duty in Article 4 and the prohibited practices have applied since 2 February 2025, and the Act applies in general, including the transparency rules, from 2 August 2026. The obligations for high-risk AI systems follow on 2 December 2027 (Annex III uses such as recruitment) and 2 August 2028 (AI in regulated products). The model behind ChatGPT has its own rules under Chapter V, but those fall on OpenAI as the model provider, not on you.
What the law actually calls you
The EU AI Act draws a sharp line between providers (companies that build AI systems) and deployers (organisations that put AI systems to work). OpenAI is the provider of ChatGPT. Your company, the moment it uses ChatGPT in a professional context, becomes a deployer.
Article 3(4) of the EU AI Act defines a deployer as "a natural or legal person, public authority, agency or other body using an AI system under its authority except where the AI system is used in the course of a personal non-professional activity." The phrase "personal non-professional activity" is the critical carve-out. Using ChatGPT at home to write a birthday card is personal. Using it at work, on company time, for business outputs, is not.
Size does not matter here. The Act applies to SMEs operating in the EU or whose AI outputs affect people in the EU. A 30-person logistics firm in Rotterdam using ChatGPT to summarise supplier contracts is a deployer under European law, the same as a 5,000-person corporation.
The risk level of your specific use case
Not every ChatGPT use case carries the same legal weight. The Act organises risk into tiers, and your obligations depend on which tier your actual use falls into.
Unacceptable risk practices have been banned outright since 2 February 2025. These include AI-based social scoring, emotion recognition in the workplace and certain real-time biometric identification. Two further bans, on AI that generates non-consensual sexual deepfakes or child sexual abuse material, apply from 2 December 2026. Standard ChatGPT use does not fall here.
High-risk uses are where most HR Directors need to pay close attention. Annex III of the Act lists employment-related AI applications as high-risk. These include systems used for recruitment, CV filtering, evaluating employees, promoting or terminating work relationships, and monitoring performance. If your team is using ChatGPT to help rank job applicants or assess performance reviews, you are moving into high-risk territory. Whether a general-purpose tool used this way is formally a high-risk system, and who then counts as its provider, is not settled; the safe course is to avoid using ChatGPT for these decisions or to treat that use as high-risk.
From 2 December 2027, deployers of Annex III high-risk systems must meet obligations including:
-
Using the system according to the provider's instructions for use
-
Ensuring human oversight is genuinely in place, not just nominally, by people with the competence and authority to act
-
Keeping the logs the system generates, for at least six months
-
Informing employees and their representatives before a high-risk system is used at work, and informing people that AI is used in decisions about them
-
A fundamental rights impact assessment only if you are a public body, a private body providing public services, or you use AI for credit scoring or life and health insurance pricing (Article 27)
Article 26 sets out these deployer obligations in full. It is worth reading. It is written in plain enough language to act on.
Limited-risk systems are those covered by the transparency rules in Article 50: chatbots, deepfakes and AI-generated text published to inform the public. Most internal ChatGPT use, such as summarising documents, drafting emails or translating, triggers no specific duty beyond Article 4.
Minimal-risk uses carry no mandatory obligations, though the Commission encourages voluntary codes of conduct.
What "human oversight" really means in practice
Human oversight is one of those phrases that sounds easy until you try to define it in a Monday morning meeting. As a legal duty under Article 26, it applies to high-risk systems from 2 December 2027: deployers must ensure that the humans assigned to oversee AI outputs actually have the competence, training and authority to intervene. It is also good practice for any AI output that affects a business decision.
In practice, this means three things for your office:
-
Someone must be named. You need a person, or a defined role, responsible for reviewing AI-assisted outputs before they affect a business decision.
-
That person must understand the tool. They cannot just rubber-stamp results. They need enough familiarity with how ChatGPT works and where it typically fails to catch meaningful errors.
-
They must have the authority to say no. If a manager is pressured to accept a ChatGPT-generated shortlist because "the AI picked them," that is not human oversight. That is human decoration.
For HR Directors specifically: if ChatGPT touches any part of your hiring pipeline, document the oversight process now. Who reviews the output? What criteria do they use? Where is that recorded?
When transparency applies to you
Article 50 has applied since 2 August 2026, and it splits the duties by role. The duty to tell people they are talking to an AI system (Article 50(1)) is a design duty of the provider: OpenAI for ChatGPT, or your company if you build your own customer chatbot. As a deployer, you have two duties: disclose deepfakes (realistic AI-generated or manipulated images, audio or video), and disclose AI-generated or manipulated text published to inform the public on matters of public interest, unless a human has reviewed it and someone holds editorial responsibility.
An employee who drafts a customer email with ChatGPT and sends it under their own name is not, by that alone, breaching Article 50. A voluntary disclosure policy can still be good practice for trust. Breaches of Article 50 fall in the fine tier of up to €15 million or 3% of global annual turnover; for SMEs and start-ups the lower of the two figures is the maximum.
Three practical steps to take this week
You do not need a legal team on retainer to start. These three steps cover the obligations that already apply, starting with Article 4, which requires you to take measures to support the AI literacy of staff who use AI tools.
Step 1: Map your AI use. Ask department heads to list every tool that uses AI, including ChatGPT, Copilot, Grammarly, and any AI features inside your existing software. You cannot manage risk you have not inventoried.
Step 2: Classify your use cases. Take that list and check each use case against Annex III. Any use touching recruitment, performance management, or credit decisions needs attention first. Everything else gets a lower-priority review.
Step 3: Write a one-page AI use policy. It does not need to be long. It needs to specify who can use AI tools, for what purposes, what must be disclosed, and who reviews outputs before they affect decisions. A short, enforced policy beats a 40-page document no one reads. Pair it with practical training on the tools your staff use, and keep a record of it: that is the core of your Article 4 measures.
Who supervises this
The EU AI Act is supervised at national level by market surveillance authorities, which supervise Article 4 from 2 August 2026. In the Netherlands, the Autoriteit Persoonsgegevens is the intended supervisor for most AI Act topics, alongside sector supervisors, but it will only get AI Act powers once the Dutch implementing law is in force. The AI Act sets no specific fine for Article 4; penalties are for national law, and the draft Dutch law foresees an order to fix shortcomings rather than a fine.
Beyond fines, the reputational damage from an AI-related employment dispute is harder to price. If a rejected job applicant argues that an AI system screened them out without adequate human review or transparency, you will want documentation showing how your process worked and who reviewed what. That record protects you under the GDPR today and under the AI Act later.
SMEs often assume the Act was written for big tech. It was not. Article 4(2), added by the Digital Omnibus (Regulation (EU) 2026/1744), obliges the Commission and Member States to support companies, in particular SMEs, in meeting the AI literacy obligation. Support is available. The obligations are real.
Start with a clear picture of where you stand
The single most useful thing you can do today is understand which of your current AI uses fall into which risk category. Once you know that, every subsequent decision becomes straightforward.
Khairos Comply offers a free 2-minute compliance check at comply.khairos.ai that maps your actual AI use cases against the Act's requirements and tells you where to focus first. No jargon, no lengthy intake form. Just a clear picture of your obligations so you can act on them.
References
Part of our guide to AI Literacy & Article 4
Not sure where your company stands?
Answer 15 questions about how your team uses AI. You get a compliance score and the gaps to close first, as a PDF.