EU AI Act for AI Agents in Mid-Sized Companies: Which Duty Applies to Which Use?

Since 2 August 2026 the transparency duty of the EU AI Act applies. A chatbot that does not identify itself as AI has since then been a violation rather than a cosmetic flaw. At the same time, for the vast majority of use cases in mid-sized companies the regulation is far less demanding than the fine thresholds suggest: a phone assistant needs a notice at the start of the call, invoice processing needs trained staff and nothing else from the AI Act, and only those who pre-sort applicants or check creditworthiness operate a high-risk system with real documentation duties.

This article maps ten typical use cases to risk class, the concrete duty, the deadline and the responsible role. The depth on the two duties that affect everyone sits in separate articles: on the AI literacy duty under Article 4 and on the labelling duty under Article 50.

In brief: The EU AI Act (Regulation (EU) 2024/1689) sorts AI systems by risk. For chatbots, phone assistants and WhatsApp agents the transparency duty under Article 50 has applied since 2 August 2026. For back-office automation that does not decide about people, only the literacy duty under Article 4 applies, and that has applied since 2 February 2025. High-risk under Annex III in mid-sized companies is almost only applicant screening, performance evaluation and credit checks; their deployer duties under Article 26 apply, following the Digital Omnibus (Regulation (EU) 2026/1744, in force since 27 July 2026), only from 2 December 2027. Emotion recognition in the workplace has been banned since February 2025.

Which duty applies to which use?

The regulation does not ask which technology a system uses but what it does and about whom it decides. The same language-model interface can be minimal risk in one case and a high-risk system in the next. That is why the start is not a definition but a mapping:

Two things stand out in the table. First: the duty that affects everyone is not in the row with the highest risk but in every row: Article 4 has required every deployer since February 2025 to ensure staff know what they are using. Second: the three high-risk rows all concern decisions about people, not about transactions. Anyone who lets AI loose on invoices, enquiries and appointments stays in the lower classes. Anyone who lets it loose on applicants, employees or borrowers does not.

Who is a provider, who is a deployer?

The provider develops the system or places it on the market under its own name. The deployer uses it under its own responsibility. A mid-sized company with a chatbot is almost always the deployer, the chatbot vendor the provider.

The line shifts in two steps. Anyone who places a third-party system on the market under its own name or brand is a provider under Article 3 number 3, regardless of risk class. For high-risk systems Article 25 adds: anyone who substantially modifies such a system or changes the intended purpose of a system so that a high-risk system results takes on the provider obligations. Anyone who takes a standard chatbot and rebuilds it for applicant screening is no longer just a deployer.

For the division of roles in everyday operations that means: the provider owes the technical documentation, the labelling of the interaction under Art. 50 para. 1, the machine-readable labelling of generated content and, for high-risk, the conformity assessment. The deployer owes use according to the instructions, labelling kept switched on in operation, the training of its own staff and the inventory of the systems in use. None of these duties can be delegated to the other side.

What applies since when

The regulation entered into force on 1 August 2024 and takes effect in stages:

For practice in mid-sized companies a clear order follows: labelling and training are due and verifiable. Preparation for high-risk duties has gained time until December 2027, but anyone introducing applicant screening today should build it as if the duties already applied, because logging and oversight are hard to retrofit.

The six duties that count in mid-sized companies

1. Labelling (Art. 50)

Anyone interacting with an AI system must learn that, unless it is obvious from the circumstances to a reasonably well-informed person. In a chatbot the notice belongs in the first message, in a phone assistant at the start of the call, in a WhatsApp agent in the greeting. One sentence is enough: "You are speaking with the AI assistant of [company name]. On request I will connect you to a member of staff at any time." What additionally applies to generated images, audio and text is in the article on the labelling duty.

2. Literacy (Art. 4)

Providers and deployers must take measures that support the development of AI literacy among their staff. That has been the wording of Article 4 since the Digital Omnibus. The earlier formula of ensuring, to their best extent, a sufficient level has been dropped, and nobody has to guarantee a specific level. Certification is not prescribed; documented training with date, content and participants is the usual evidence. The duty has applied since 2 February 2025 and is the most frequently overlooked one. Details, including what does not threaten, are in the article on the AI literacy duty.

3. Inventory

The regulation does not literally prescribe an AI inventory, but without one none of the other duties can be evidenced. Per system: provider and model, purpose, types of data processed, risk class per the table above, data flows, responsible person. A table with six columns is enough.

4. Provider due diligence

Before the contract: is the provider's technical documentation available, is there a data processing agreement under Article 28 GDPR, where is the data processed, are inputs used for training. Providers of general-purpose AI models have been obliged to document and be transparent since August 2025; as a deployer you request those documents and file them with the inventory.

5. GDPR alongside the AI Act

The AI Act does not replace the GDPR and overlaps with it only in part. The GDPR governs personal data, the AI Act governs the system, even where it processes no personal data. Concretely, the AI Act adds to the GDPR: the labelling, the training and the inventory. What remains are data processing agreements, legal basis and information of data subjects, for instance for recorded calls.

6. Monitoring in operation

Deployers of high-risk systems must monitor operation and report serious incidents. For all other systems monitoring is not a duty under the regulation, but it is the precondition for labelling and escalation to work day to day: which conversations are stored for how long, who reviews them, how does a wrongly answered case reach a human. Anyone who can answer those three questions has the supervision that § 130 OWiG demands of every management anyway.

What does a violation cost?

Article 99 grades the fines by the severity of the breach; the higher amount applies, for small and medium-sized enterprises the lower one:

In Germany the implementing act (KI-MIG), in force since 29 July 2026, designates the Federal Network Agency as the central market surveillance authority. More important for practice than the authority is the counterpart in everyday business: a customer asking about the labelling, an applicant wanting to know whether a machine sorted their documents, a works council asking about the works agreement. Those questions come before the authority, and they come for certain.

Checklist: where do you stand?

Every no is an open task.

• The chatbot identifies itself as AI in its first message
• The phone assistant says so at the start of the call
• The WhatsApp agent says so in its greeting
• There is a list of all AI systems in use with risk class and responsible person
• Staff are trained, and the training is documented with date, content and participants
• A data processing agreement exists with every AI provider, and the technical documentation is on file
• The privacy policy names the systems and providers in use
• Recordings and logs have a defined and justified retention period
• The route from the assistant to a human is defined and tested
• No system in use evaluates applicants, employees or creditworthiness, or it is recorded and prepared as a high-risk system

Does the AI Act also apply in Austria and Switzerland?

Austria. Yes, directly and with the same dates. The AI Act is an EU regulation and needs no implementing statute to take effect. As a contact point Austria set up the AI Service Desk (KI-Servicestelle) at the regulator RTR in 2024, which provides information and advice on the Regulation. The obligations under the Regulation itself apply to a company in Vienna just as they do to one in Stuttgart. What differs are supervision and the accompanying national provisions, for example the Labour Constitution Act (Arbeitsverfassungsgesetz) instead of the Works Constitution Act.

Switzerland. Switzerland is not an EU member, so the Regulation does not apply there directly. Swiss companies should still know two things.

First, the scope reaches beyond the EU. Article 2(1) covers providers placing AI systems on the market in the Union irrespective of where they are established, and under point (c) also providers and deployers located in a third country where the output produced by the AI system is used in the Union. A Swiss company whose chatbot serves customers in Germany or Austria should therefore treat disclosure under Article 50 as its own obligation.

Second, Switzerland is working on its own rules. On 12 February 2025 the Federal Council decided to ratify the Council of Europe's AI Convention and to adapt Swiss law sector by sector. A consultation draft is due by the end of 2026. Until then the general statutes apply, above all the Federal Act on Data Protection, in force since 1 September 2023.

Frequently asked questions

We use a chatbot from an external provider. Am I affected by the AI Act? Yes, as a deployer. The provider supplies the labelling function and its technical documentation; you have to switch the labelling on, train your staff under Article 4 and list the system in your AI inventory. Anyone offering the chatbot under their own name is a provider themselves under Article 3 number 3; for high-risk systems Article 25 additionally governs the change of role, for instance on a change of purpose.

Does the AI Act apply if we only use ChatGPT for texts and emails? Yes. Anyone using an AI system professionally is a deployer and has had to take measures for the AI literacy of staff since 2 February 2025. A labelling duty for the texts only comes on top if they are published to inform the public on matters of public interest and nobody reviews them before publication. It does not apply to internal emails and quotes.

Is our AI phone assistant a high-risk system? No, as long as it answers calls, books appointments, records enquiries and forwards them. That is a case for the transparency duty under Article 50: the caller must learn at the start of the call that they are speaking with an AI. The assistant would only be high-risk if it classified emergency calls or decided on access to services listed in Annex III.

What applies if AI pre-sorts job applications? That is a high-risk system under Annex III number 4. Deployers then face the duties of Article 26: use according to the provider's instructions, oversight by trained persons, retention of logs for at least six months and informing employees and their representatives before putting it into service. The original 2 August 2026 deadline was postponed to 2 December 2027 by the Digital Omnibus (Regulation (EU) 2026/1744, in force since 27 July 2026). Co-determination of the works council under § 87 BetrVG applies today regardless.

Who enforces compliance in Germany? Under the German implementing act (KI-MIG), in force since 29 July 2026, the Federal Network Agency (Bundesnetzagentur) is the central market surveillance authority unless another body is responsible; the data protection authorities keep their competence for the GDPR. For most mid-sized companies the first contact in practice is not the authority but a customer or applicant asking about the labelling or the legal basis.

What does a violation cost? Article 99 grades the fines: up to EUR 35 million or 7 percent of worldwide annual turnover for prohibited practices, up to EUR 15 million or 3 percent for breaches of the other duties, up to EUR 7.5 million or 1 percent for false information to authorities. For small and medium-sized enterprises the lower of the two amounts applies.

Does the EU AI Act also apply to companies in Switzerland? Not directly, as Switzerland is not an EU member. However, Article 2 of the Regulation also covers providers and deployers in a third country where the output of their AI system is used in the EU. A Swiss company whose chatbot serves customers in Germany or Austria should therefore treat disclosure under Article 50 as its own obligation. In Austria the Regulation applies directly and with the same dates as in Germany.

Conclusion

For mid-sized companies the EU AI Act is not a conformity assessment but a sorting task: which system does what, and about whom does it decide. For the systems that handle enquiries, calls and documents the task ends at labelling, training and inventory. For the few that decide about people, it only begins there.

When choosing a provider, it pays to ask for exactly these points in writing beforehand: built-in labelling under Art. 50, information on where processing takes place and the documentation needed for inventory and provider due diligence. Whether a planned use falls into one of the high-risk rows should be settled before the contract is signed, not during the project.

This is not legal advice. Assigning a specific system to a risk class and interpreting individual provisions belong to a lawyer, especially for high-risk applications.

Sources

• Regulation (EU) 2024/1689 (EU AI Act) - EUR-Lex
• Regulation (EU) 2026/1744 (Digital Omnibus on AI) - EUR-Lex
• Article 4: AI literacy - artificialintelligenceact.eu
• Article 3: Definitions - artificialintelligenceact.eu
• Article 25: Responsibilities along the AI value chain - artificialintelligenceact.eu
• Article 26: Obligations of deployers of high-risk AI systems - artificialintelligenceact.eu
• Article 50: Transparency obligations - artificialintelligenceact.eu
• Annex III: High-risk AI systems - artificialintelligenceact.eu
• Article 99: Penalties - artificialintelligenceact.eu
• German Act on Market Surveillance and Innovation Promotion of Artificial Intelligence (KI-MIG), BGBl. 2026 I No. 223 - Gesetze im Internet
• Implementation guide to the AI Regulation (version 2.0) - Bitkom e.V.
• BSI - Security of AI systems - Federal Office for Information Security
• Article 2: Scope - artificialintelligenceact.eu
• AI Service Desk at RTR - Rundfunk und Telekom Regulierungs-GmbH, Austria
• Artificial intelligence: the Federal Council's regulatory approach of 12 February 2025 - Federal Office of Communications, Switzerland