The EU AI Act for companies that deploy AI agents and chatbots

If your company uses AI agents or chatbots built on someone else's model, the EU AI Act treats you mainly as a deployer. For typical business agents, two duties already apply today: AI literacy for the people who use them (Article 4, since 2 February 2025) and, for anything that talks to customers or staff, a clear disclosure that they are dealing with AI (Article 50, since 2 August 2026). Agents used in recruitment, staff evaluation or credit scoring are high risk, but those duties were postponed by the Digital Omnibus and now start on 2 December 2027.

This guide is written for owners, COOs and heads of IT who are rolling out agents, not training models. It covers which role you hold, where common agents fall, what changed in July 2026, and what to document before your first agent goes live.

#Provider or deployer: which one you are

The AI Act assigns duties by role. A provider develops an AI system, or has one developed, and places it on the market or puts it into service under its own name (Article 3(3)). A deployer uses an AI system under its authority in a professional context (Article 3(4)). The model vendor is the provider of the general-purpose model; the duties for those models have applied since 2 August 2025 and do not fall on you.

The grey zone is the agent itself. If you buy a finished product, such as a helpdesk tool with a built-in assistant, you are its deployer. If you or an agency build an agent on a model API and run it under your own brand, you are usually the provider of that AI system as well as its deployer. For a customer-facing chatbot this matters little in practice, because the disclosure duty falls on whoever builds the system, and that is you either way.

#The risk categories, and where typical business agents fall

The Act has no special category for agents. What counts is the purpose an agent is used for, not whether it chats, calls tools or acts on its own. There are four tiers: prohibited practices (Article 5), high-risk systems (Article 6 with Annexes I and III), systems with transparency duties (Article 50) and everything else, which carries only the general AI literacy duty.

Two details catch people out. First, Article 6(3) lets some Annex III systems escape the high-risk label if they only perform a narrow procedural or preparatory task, but never when the system profiles natural persons, and ranking applicants is profiling. Second, fraud detection is expressly carved out of the credit-scoring category, so a payments agent that flags suspicious transactions is not high risk on that ground.

#The timeline, and what the Digital Omnibus changed

On 19 November 2025 the European Commission proposed a Digital Omnibus on AI to simplify the AI Act and delay the high-risk rules, because harmonised standards were not ready. Parliament and Council reached a political agreement in May 2026 that replaced the Commission's flexible trigger with fixed dates. The text was adopted as Regulation (EU) 2026/1744, published in the Official Journal on 24 July 2026 and in force since 27 July 2026. It is law, not a proposal.

What the Omnibus did not do matters as much. It did not touch the Article 50 transparency date, it did not remove any high-risk duty (it moved the dates), and it kept AI literacy as an obligation on providers and deployers, although in softer wording. Do not confuse it with the separate Digital Omnibus proposal that would amend the GDPR: at the time of writing that one is still in the legislative process, so plan on the GDPR as it stands.

#Chatbot transparency under Article 50

Article 50(1) is the rule every customer-facing agent has to meet. It has applied since 2 August 2026, and fines for breaching it reach EUR 15 million or 3 percent of worldwide annual turnover, whichever is higher (for SMEs, whichever is lower) under Article 99(4).

The Commission's guidelines on Article 50, adopted on 20 July 2026, make this concrete. The "obvious" exception is to be read narrowly, so do not rely on it. The notice must come at the latest at the first interaction, be built into the system's design, and be clear: a line hidden in the terms and conditions is not enough. Text, audio or visual formats are all acceptable, so a voice agent can say it at the start of the call. The system should also answer truthfully whenever someone asks if it is a bot, and in sensitive settings, such as health topics or vulnerable users, periodic reminders may be needed.

Article 50 has three further parts. Providers of generative systems must mark output as machine readable (50(2)). Deployers of emotion recognition or biometric categorisation must inform the people exposed (50(3)). Deployers who publish deepfakes, or AI-generated text meant to inform the public on matters of public interest without human editorial responsibility, must disclose that (50(4)). An agent that drafts customer replies a human then reviews and sends is generally outside 50(4), because that provision targets public-interest publications, but check the guidelines for your exact use.

#AI literacy under Article 4

Article 4 has applied since 2 February 2025 to every provider and deployer, whatever the risk level. The Omnibus replaced the original duty to "ensure, to their best extent, a sufficient level of AI literacy" with a duty to support it, effective since 27 July 2026.

The new text adds that you are not required to guarantee any specific level of literacy for any individual. You still have to act. The Commission's AI literacy Q&A says there is no certificate requirement and no fixed format, that relying only on the vendor's instructions for use is generally not enough, and that an internal record of training and guidance is sensible. National authorities supervise this from August 2026, and the Act sets no fixed fine for Article 4 itself.

In practice, tailor it by role: a short session for everyone on what the agent does, where it fails and what must never go into it; deeper training for the people who configure it, review its output or handle escalations; and a named owner who keeps the material current when the agent changes.

#High risk in HR: what deployers will owe from December 2027

Recruitment and workforce management are where most mid-sized companies meet the high-risk rules. Annex III, point 4 covers systems for recruitment or selection (including targeted job adverts, filtering applications and evaluating candidates) and systems that make or support decisions on promotion, termination, task allocation based on behaviour or personal traits, and performance monitoring. From 2 December 2027, deployers of such systems must meet Article 26:

• Use the system in line with the provider's instructions for use.
• Assign human oversight to people with the competence, training and authority to overrule it.
• Make sure input data you control is relevant and sufficiently representative for the purpose.
• Monitor operation, report serious incidents and risks to the provider and authority, and suspend use where needed.
• Keep the automatically generated logs for at least six months.
• Inform workers' representatives and affected employees before the system is put into service at the workplace.
• Tell candidates and employees that a high-risk system is used in decisions about them; Article 86 adds a right to an explanation.

In Germany the works council is involved anyway: introducing technical systems suited to monitoring employee behaviour or performance requires co-determination under section 87(1) no. 6 BetrVG. The extra time to 2027 is best used to pick vendors that will deliver the documentation you need, not to wait.

#How the AI Act and the GDPR fit together

The AI Act regulates the system; the GDPR regulates the personal data flowing through it. An agent that reads customer emails, CVs or invoices with names on them is processing personal data, so all GDPR duties apply in parallel. The German data protection authorities set out their expectations in the DSK guidance on AI and data protection.

• Legal basis (Article 6 GDPR). The AI Act gives you none. Name one for each purpose, including any use of conversation logs to improve the agent.
• Processor contract and transfers (Articles 28 and 44 onwards). The model vendor and agent platform are usually processors. Sign a data processing agreement, check where data is stored and processed, and disable training on your data where the vendor offers that.
• Automated decisions (Article 22). Decisions with legal or similarly significant effects may not be based solely on automated processing without a specific exception. The Court of Justice held in SCHUFA (C-634/21) that a score can itself count as such a decision when it plays a determining role.
• DPIA (Article 35). Expect one for HR screening, scoring or large-scale monitoring. For high-risk systems, Article 26(9) tells you to use the provider's information for it, and a fundamental rights impact assessment under Article 27 can build on it.
• Transparency (Articles 13 and 14). Update privacy notices so they mention the agent, what it processes and who receives the data.

#What to document, and a checklist for your first agent

Take a hypothetical company with 300 staff rolling out a customer service agent that answers order questions and drafts refunds for human approval. Its file does not need to be long, but it should exist before go-live and be updated whenever the agent's scope changes. An authority, an auditor or a works council will ask for the same things: what the system is, why you classified it as you did, and how people stay in control.

#How we can help

Bespoke IT designs and builds AI agents for companies in Germany, Austria and Switzerland, and we treat the AI Act and GDPR points above as design requirements from the first workshop. If you want a second pair of eyes on which processes are worth automating and how they would be classified, start with our free AI audit. For legal sign-off we will tell you when you need your lawyer or data protection officer.