# AI Act and GDPR: what to watch for in a business chatbot

Mesterséges Intelligencia Központ Kft. is a software development company based in Budapest, Hungary. Two rules matter most for a business chatbot: under the AI Act it must disclose that an AI is answering, and under the GDPR it must handle personal data lawfully.

> **In short:** Article 50(1) of the AI Act has applied since 2 August 2026: a chatbot must make clear, at the latest at the first interaction, that the user is talking to artificial intelligence, unless this is obvious. A typical customer-service chatbot is not a high-risk AI system, but if it handles personal data, the GDPR applies to it as well, among other things as to the lawful basis, information for users, data minimisation and a data processing agreement with the AI provider. This is information, not legal advice (checked on 17 October 2026).

## Which two laws are we talking about?

The AI Act is the European Union's regulation on artificial intelligence, officially [Regulation (EU) 2024/1689](https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32024R1689). As a regulation, it is directly applicable in every Member State. The official Hungarian text calls an AI system an „MI-rendszer”; in this article we use the everyday term “AI”.

Article 3 of the regulation distinguishes two roles. A **provider** is anyone who develops an AI system, or has one developed, and places it on the market or puts it into service under its own name or trademark. A **deployer** is anyone who uses an AI system under its authority, except for personal, non-professional use. If a company has a chatbot developed and runs it under its own name on its website, both definitions may fit, so the obligations are not something to leave to the developer alone.

The GDPR (the General Data Protection Regulation, officially [Regulation (EU) 2016/679](https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32016R0679)) applies to every processing of personal data, including what a chatbot does. Article 2(7) of the AI Act states that, apart from a few narrow exceptions that do not usually concern chatbots, the regulation does not affect the GDPR: the two sets of rules apply side by side, and neither replaces the other.

## What does Article 50 of the AI Act require?

Article 50 sets out transparency obligations for providers and deployers of certain AI systems. A business chatbot is mainly affected by paragraph 1, but the others are worth knowing too, because an assistant that also produces images, audio or published text can run into them.

| Paragraph | Who it binds | What it requires | When it affects a chatbot |
|---|---|---|---|
| (1) | the provider | for systems that interact directly with people, those concerned must be informed that they are interacting with an AI system, unless this is obvious | almost always: this is the basic chatbot case |
| (2) | the provider of a system generating synthetic text, images, audio or video | outputs must be marked in a machine-readable format so they are detectable as artificially generated | if the system generates content; the Commission's guidelines and code of practice give the details |
| (3) | the deployer of an emotion recognition or biometric categorisation system | the people exposed must be informed of the operation of the system | if the system infers emotions or intentions from biometric data |
| (4), first subparagraph | the deployer of a system generating deep fakes | it must be disclosed that the image, audio or video content is artificially generated | if the bot produces images, audio or video that deceptively resemble real people |
| (4), second subparagraph | the deployer of a system generating text published to inform the public on matters of public interest | it must be disclosed that the text is artificially generated, unless a person has reviewed it and someone holds editorial responsibility for it | rarely: an individual chatbot reply is usually not such published text |
| (5) | both | the information must be clear and distinguishable, given at the latest at the first interaction, and meet accessibility requirements | always, together with the disclosure under paragraph 1 |

### What does paragraph 1 mean in practice?

The disclosure has to appear at the start of the conversation, where the user can see it, not deep inside the terms and conditions. A sensible approach is for both the chat window header and the bot's first message to say: “An AI assistant is answering.” This matters even more if the bot has its own name or avatar, because a human name easily suggests that a member of staff is writing.

The “obvious” exception is narrow. Under the text of the regulation, it is judged from the point of view of a reasonably well-informed, observant and circumspect person, taking into account the circumstances and the context of use. The European Commission's questions and answers say the exception should be interpreted restrictively, because it deprives people of transparency, and they expressly name chatbots, AI agents and avatars as examples of systems that interact directly with people. In a customer-service chat it is therefore sensible always to disclose that an AI is answering.

## From when do the rules apply?

The AI Act entered into force on 1 August 2024, and its provisions became applicable in stages (Article 113). On 19 November 2025 the Commission proposed simplifying the regulation; this became [Regulation (EU) 2026/1744](https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32026R1744) of 8 July 2026, titled the “Digital Omnibus on AI”. It was published in the Official Journal on 24 July 2026 and entered into force on 27 July 2026. It did not postpone Article 50(1), but it changed several other dates (checked on 17 October 2026):

| Rule | Applies | Where it is laid down |
|---|---|---|
| Prohibited AI practices and AI literacy (Chapters I and II) | from 2 February 2025; the amendment's two new prohibitions from 2 December 2026 | AI Act Article 113, point (a) |
| The new, lighter wording of AI literacy (Article 4) | from 27 July 2026 | entry into force of Regulation (EU) 2026/1744 |
| Transparency obligations, including Article 50(1) | from 2 August 2026 | AI Act Article 113, general date of application |
| Machine-readable marking (Article 50(2)) for generative systems placed on the market before 2 August 2026 | compliance required by 2 December 2026 | AI Act Article 111(4), inserted by the amendment |
| High-risk systems under Annex III | from 2 December 2027 | AI Act Article 113, point (c), as amended |
| High-risk systems built into products (Annex I) | from 2 August 2028 | AI Act Article 113, point (c), as amended |

The Commission's AI Act page and its questions and answers on Article 50 give the same main dates. If a further amendment appears in future, it is worth checking on EUR-Lex, on the regulation's own page.

## AI literacy: what does the team need to know?

The original wording of Article 4 required providers and deployers to ensure, to their best extent, a sufficient level of AI literacy of their staff. Since the 2026/1744 amendment, the obligation is to take measures to support the development of AI literacy of staff and of other persons acting on their behalf; the text states that this does not require them to guarantee any specific level of AI literacy. The Commission and the Member States are tasked with supporting this effort, in particular for small and medium-sized enterprises.

For a company running a chatbot, this means short training tailored to each role in practice: what the bot can and cannot do, how to recognise a wrong answer, when and how to take over a conversation, and which data must never be typed into an AI tool. It is worth documenting.

## Is a customer-service chatbot high-risk?

The category is called a “high-risk AI system” in the regulation. Article 6 puts a system in it in two ways: if it is a safety component of a product covered by the EU product legislation listed in Annex I (or is itself such a product), and that product requires a third-party conformity assessment; or if its use falls into one of the areas in Annex III. The Annex III areas are biometrics, critical infrastructure, education, employment, access to essential private and public services, law enforcement, migration and border control, and the administration of justice and democratic processes.

A chatbot that answers questions about products, delivery, opening hours or order status falls into none of these areas, so it is not high-risk; the transparency rules of Article 50 apply to it. The same technology can, however, become high-risk depending on its intended purpose, for example if it:

- filters job applications or evaluates candidates (Annex III, point 4(a));
- evaluates the creditworthiness of natural persons or establishes their credit score (point 5(b));
- assesses risk or sets prices for natural persons in life or health insurance (point 5(c));
- evaluates eligibility for public benefits or services on behalf of a public authority (point 5(a)).

Under Article 6(3), an Annex III system is exceptionally not high-risk if it poses no significant risk, for example because it performs a narrow procedural task or only prepares a human decision, but it is always high-risk if it profiles natural persons. Anyone relying on this exception must document their assessment. The detailed requirements for high-risk systems apply from 2 December 2027 in Annex III cases. If a chatbot comes close to decisions like these, it is worth clarifying the classification with a lawyer before development starts.

## What does the GDPR require of a chatbot?

A chatbot conversation often contains personal data: a name, an email address, an order number, the description of a complaint. A company running the chatbot for its own customer service is then typically the controller, and the language model provider is typically a processor. On 19 November 2025 the Commission also proposed amending the GDPR ([COM(2025) 837](https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:52025PC0837)), but no act amending the GDPR has yet been published in the Official Journal (checked on 17 October 2026): what follows is based on the text in force.

### Lawful basis (Article 6)

Every purpose of processing needs a lawful basis from Article 6(1). For chatbots the most common are the performance of a contract (for example an answer about the status of an order), legitimate interests (with a balancing test) and consent (for example when signing up to a newsletter). If the conversations are later to be used for another purpose (for analysis or development, say), this has to be examined separately, because the original basis is not enough for further processing that is incompatible with the original purpose (Article 6(4)).

### Information (Article 13)

When the data is collected, the data subject must be told, among other things, who the controller is, the purposes and the lawful basis, the recipients, any transfer to a third country, the retention period, their rights and that they can lodge a complaint with a supervisory authority, which in Hungary is the National Authority for Data Protection and Freedom of Information (NAIH). The AI Act disclosure that “an AI is answering” does not replace this: the privacy notice must also be available in the chat window.

### Data minimisation and storage limitation (Article 5)

Under Article 5, only data necessary for the purpose may be processed, and only for as long as the purpose requires. For a chatbot, this comes down to three concrete decisions: the bot should not ask for more data than the answer needs; only the data needed for the task should reach the model, masked or pseudonymised where possible; and the conversation logs should have a retention period and a deletion routine. Data security (access control, encryption, logging) is part of the same article.

### Data processing agreement with the AI provider (Article 28)

If the language model provider processes personal data on the company's behalf, a written processing agreement is needed with it. Under Article 28(3), it must state, among other things, that the processor processes the data only on the controller's documented instructions, ensures confidentiality and security, engages further processors only with authorisation, helps respond to data subject requests, deletes or returns the data at the end of the service and allows audits. In practice, this means looking at the provider's business (API) terms and data processing addendum, not those of the consumer chat app: where the data is stored, for how long, and whether it is used for model training.

### Transfers outside the EU (Chapter V)

If the data is processed in a country outside the EU, a so-called third country, there must be a basis under Chapter V of the GDPR: an adequacy decision (Article 45) or appropriate safeguards such as standard data protection clauses (Article 46). For the United States, the Commission's [Implementing Decision (EU) 2023/1795](https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32023D1795) on the EU-US Data Privacy Framework is an adequacy decision, but only for US organisations included in the Data Privacy Framework List. The decision is in force; the action seeking its annulment was [dismissed by the General Court on 3 September 2025](https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:62023TJ0553), and an [appeal against that judgment is pending before the Court of Justice (C-703/25 P)](https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:62025CN0703) (checked on 17 October 2026). It is therefore also worth checking whether the provider offers storage or processing in the EU.

### Data protection impact assessment (Article 35)

An impact assessment is required if the processing, especially when using new technologies, is likely to result in a high risk to the rights of the people concerned. Article 35(3) expressly requires one in some cases, and the NAIH has published its own [DPIA list](https://naih.hu/hatasvizsgalati-lista), which includes, among others, profiling, automated decision-making with legal effects, using children's data for profiling or marketing, and combining data from different sources. A simple bot that answers product questions rarely needs one; if the bot combines customer data from the CRM, profiles people, or processes health data on a large scale, it probably does. It is worth recording the assessment in writing even when the result is “not required”.

### Automated decisions (Article 22)

Data subjects have the right not to be subject to a decision based solely on automated processing that produces legal effects or similarly significantly affects them. If a chatbot were to reject a complaint, a refund or an application on its own, this rule matters too: where such a decision is necessary for a contract or based on explicit consent, at least the right to human intervention, to express one's point of view and to contest the decision must still be ensured. The simplest approach is for the bot only to prepare such a decision and for a member of staff to make it.

## A practical checklist

| Task | Why | Legal source |
|---|---|---|
| State in the first message and in the chat window that an AI is answering | transparency | AI Act Article 50(1) and (5) |
| Clarify who is the provider, the deployer, the controller and the processor | sharing responsibility | AI Act Article 3, GDPR Article 28 |
| Record the bot's intended purpose and its risk classification | ruling out or managing high risk | AI Act Article 6, Annex III |
| Define the lawful basis for each purpose | lawful processing | GDPR Article 6 |
| Make the privacy notice available in the chat window | information | GDPR Article 13 |
| Ask for and pass to the model only the data needed | data minimisation | GDPR Article 5(1)(c) |
| Set a retention period for conversation logs | storage limitation | GDPR Article 5(1)(e) |
| Sign a data processing agreement with the AI provider | using a processor | GDPR Article 28 |
| Check the basis for transfers outside the EU | transfers to third countries | GDPR Chapter V |
| Assess and document whether an impact assessment is needed | high-risk processing | GDPR Article 35, NAIH list |
| Involve a person in significant decisions and hand over to staff on request | automated decisions | GDPR Article 22 |
| Train the staff who run the bot | AI literacy | AI Act Article 4 |

## Typical mistakes

- **Hidden disclosure:** “an AI is answering” appears only in the terms of use, not in the conversation.
- **A human persona without disclosure:** the bot gets a human name and photo, and nothing suggests it is a machine.
- **The full customer record in the prompt:** the model receives the whole CRM record when two fields would be enough for the answer.
- **A consumer tool with company data:** staff paste customer data into a free chat app without a processing agreement.
- **Logs kept forever:** conversations pile up with no deletion rule.
- **Automated decisions without human control:** the bot rejects significant requests on its own, and nobody reviews them.

## Frequently asked questions

### Do we have to disclose that an AI is answering?

As a rule, yes. Under Article 50(1) of the AI Act, an AI system that interacts directly with people must be designed so that those concerned know they are interacting with an AI system, unless this is obvious from the point of view of a reasonably well-informed person. Under paragraph 5, this must be communicated clearly and distinguishably at the latest at the first interaction. In a customer-service chat, the safest approach is therefore to say so in the first message.

### From when is it mandatory?

Article 50 has applied since 2 August 2026, and the amending Regulation (EU) 2026/1744 did not change this. For the machine-readable marking of generative outputs (Article 50(2)), providers of systems placed on the market before 2 August 2026 have until 2 December 2026 to comply. Article 4 on AI literacy has applied since 2 February 2025, with lighter wording since 27 July 2026.

### Does a customer-service chatbot count as high-risk?

Usually not. A chatbot that gives information about products, orders or services does not fall into the areas of Annex III of the AI Act, so the transparency rules apply to it. It can be high-risk, however, if it screens job applicants, assesses creditworthiness or prices life and health insurance risk, for example.

### Can we pass personal data to an AI provider?

Yes, if the processing has a lawful basis, the people concerned are informed about it, there is a data processing agreement with the provider (GDPR Article 28), and, if the data leaves the EU, the transfer has a basis too (Chapter V). It is worth passing on only as much data as the task needs, and checking whether the provider uses the data for model training.

### Do we need a data protection impact assessment?

You need one if the processing is likely to result in a high risk to the rights of the people concerned (GDPR Article 35). Rarely for a simple information bot; probably for profiling, automated decisions with legal effects, profiling of children's data, special categories of data processed on a large scale or combining data from several sources, all of which also appear on the NAIH's DPIA list. Either way, it is worth recording the decision in writing.

## How we can help

We build [AI chatbots and AI agents](https://www.mestersegesintelligenciakozpont.hu/en/services/ai-chatbots-and-agents), and a chatbot we build states in its very first message that an AI assistant is answering. When we connect language models (see our [AI integration](https://www.mestersegesintelligenciakozpont.hu/en/services/ai-integration) page), we make sure that only the data needed for the task reaches the model, we check the provider's data handling terms, the model sees only what the user asking the question may see, and conversations can be traced. For the technical side, read [Chatbot or AI agent?](https://www.mestersegesintelligenciakozpont.hu/en/knowledge-base/chatbot-vs-ai-agent) and [AI on company documents (RAG)](https://www.mestersegesintelligenciakozpont.hu/en/knowledge-base/ai-on-company-documents-rag).

This article is information, not legal advice: it is worth asking a lawyer or a data protection officer whether a specific processing operation complies. The first consultation is free; if you are planning a chatbot, write to us using the details on the [Contact](https://www.mestersegesintelligenciakozpont.hu/en/contact) page.

## Sources

- [EUR-Lex: Regulation (EU) 2024/1689 (AI Act)](https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32024R1689)
- [EUR-Lex: Regulation (EU) 2026/1744 (Digital Omnibus on AI)](https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32026R1744)
- [EUR-Lex: Regulation (EU) 2016/679 (GDPR)](https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32016R0679)
- [EUR-Lex: COM(2025) 837, the Digital Omnibus proposal](https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:52025PC0837)
- [EUR-Lex: Commission Implementing Decision (EU) 2023/1795 (EU-US Data Privacy Framework)](https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32023D1795)
- [EUR-Lex: General Court judgment in Case T-553/23 (Latombe v Commission)](https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:62023TJ0553)
- [EUR-Lex: appeal in Case C-703/25 P](https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:62025CN0703)
- [European Commission: AI Act](https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai)
- [European Commission: Transparency obligations under Article 50 of the AI Act (questions and answers)](https://digital-strategy.ec.europa.eu/en/faqs/transparency-obligations-under-article-50-ai-act)
- [NAIH: DPIA list (in Hungarian)](https://naih.hu/hatasvizsgalati-lista)