In short: Article 50 of the AI Act, Regulation (EU) 2024/1689, has applied since 2 August 2026. Anyone who chats with a bot or talks to an AI on the phone must be told so, at the latest at the first interaction, clearly and distinguishably. By its wording the duty falls on the providers of such systems. The Digital Omnibus, Regulation (EU) 2026/1744, did not postpone this disclosure duty. It postponed the technical marking of AI-generated content for older systems to 2 December 2026 and the high-risk obligations to December 2027 and August 2028. Breaches of Article 50 can cost up to €15 million or 3 percent of worldwide annual turnover. For small and medium-sized enterprises, the lower of the two amounts applies.
The first version of this article from 3 August 2026 contained errors. It named 2 August 2026 as the start of the high-risk obligations, promised small companies a 50 and 75 percent discount on fines and relied on statements about authorities and associations that we could not verify. We rewrote the text in October 2026 and checked every statement against the primary source: EUR-Lex, the European Commission's pages, the European Parliament and the German statute texts.
The deadlines at a glance, as of October 2026
The AI Act entered into force on 1 August 2024 and applies in stages. The Digital Omnibus was published in the Official Journal on 24 July 2026 and moved several dates. [2] [3] These are the dates that now apply:
| Date | What applies from this day |
|---|---|
| 2 February 2025 | Prohibited practices (Art. 5) and the AI literacy rules (Art. 4) |
| 2 August 2025 | Obligations for general-purpose AI models, governance, penalties |
| 2 August 2026 | The regulation applies generally, including the transparency duties in Art. 50. The disclosure duty has no transition period. |
| 2 December 2026 | Machine-readable marking of AI-generated content (Art. 50(2)) for systems placed on the market before 2 August 2026 |
| 2 August 2027 | Deadline for general-purpose AI models placed on the market before 2 August 2025 (Art. 111(3)) |
| 2 December 2027 | High-risk obligations for stand-alone systems under Annex III, for example in recruitment or creditworthiness |
| 2 August 2028 | High-risk obligations for AI in products under Annex I, such as machinery or medical devices |
For a company with AI on the phone or a chatbot, the bold row is the one that matters. It has applied since 2 August 2026, to new and to existing systems.
What Article 50(1) actually requires
The core fits in one sentence. Providers must ensure that AI systems intended to interact directly with people are designed and developed so that the people concerned "are informed that they are interacting with an AI system". [1] Four questions follow from that, and they come up again and again in practice.
Who does the duty fall on? By its wording, on providers. A provider is whoever 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. The Commission's questions and answers on Article 50 say the same: providers must design and develop their systems so that people are informed. [5] Whether a company that merely uses someone else's agent carries a duty of its own under paragraph 1, the text leaves open. More on that in the section on providers and deployers.
When? At the latest at the time of the first interaction (paragraph 5). According to the Commission's questions and answers, people must be informed from the start of the first interaction. [5] On the phone that means in the greeting, not after two minutes.
How? In a clear and distinguishable manner and in line with the applicable accessibility requirements (paragraph 5). The law does not prescribe any particular wording.
Are there exceptions? Yes, where the AI nature is obvious from the circumstances to a reasonably well-informed, observant and circumspect person. The Commission explicitly wants this exception read narrowly, because it takes transparency away from people. [5] A voice that sounds human is precisely not obviously an AI on the phone. Whoever relies on the exception carries the risk.
The rule is technology-neutral. It applies to text in a chat window just as much as to a voice on the phone, in the browser or on a tablet at reception. On 31 July 2026 the Commission announced that it would enforce the rules from 2 August. Chatbots and other interactive AI systems, it wrote, "will have to tell users they are dealing with AI, not a human". [6]
What a disclosure that holds up sounds like
The law says what has to be achieved, not which sentence to use. The wording below is taken from practice and is not an official template. All examples have two things in common: the disclosure comes first, and it calls the AI by its name.
| Situation | Example | What matters |
|---|---|---|
| Inbound call | "Dr Berger's practice, this is Lea, the practice's AI assistant. How can I help you?" | Disclosure in the first sentence, before any question |
| Chat on the website | "Hi, I'm the AI assistant of Weber Motors. I can answer questions about appointments and vehicles." | Visible before the first answer |
| Voice button in a chat | "You are about to talk to an AI assistant, not a human." | Before the microphone switches on |
| Outbound call | "Good afternoon, this is the AI assistant of Klein Property Management, calling about your Thursday appointment." | Sender, AI and reason within the first seconds |
| Handover to a human | "I'm putting you through to Mr Yilmaz from customer service now." | The switch to a real person is announced |
Five mistakes make a disclosure worthless:
- The disclosure comes too late. First the greeting, then three questions, and after a minute a "by the way, I'm an AI". Paragraph 5 requires the information at the latest at the first interaction.
- The disclosure only sits in the terms. Nobody reads them before a call. The information has to reach the person in the interaction itself.
- A human name with nothing added. "This is Lea" in a natural voice makes everyone think of a person. A name is allowed, but with "AI assistant" in the same sentence.
- The agent dodges the question. If someone asks "Am I talking to a human?", the answer has to be honest. One line in the agent's instructions is enough for that.
- The disclosure only exists in one language. If the agent speaks several languages, the disclosure belongs in the language of the conversation. A disclosure the caller does not understand is not clear.
The other paragraphs of Article 50
Article 50 covers more than the disclosure in a conversation. The paragraphs address different roles and are often confused.
- Paragraph 2, marking generated content: providers of systems that generate synthetic audio, image, video or text content must mark the outputs in a machine-readable format so that they can be detected as artificially generated. Exempt are, among others, purely assistive functions for standard editing. For systems that were on the market before 2 August 2026, the new Article 111(4) applies this duty only from 2 December 2026. [2] [5] A voluntary code of practice for implementation has existed since 10 June 2026. According to the Commission's questions and answers, the Commission and the AI Board have assessed it as adequate. [5] [7]
- Paragraph 3, emotion recognition and biometric categorisation: deployers of such systems must inform the people exposed to them. This becomes relevant as soon as a voice agent is meant to infer feelings from the voice. More on that in the high-risk section.
- Paragraph 4, deep fakes and texts of public interest: deployers must disclose when image, audio or video content has been artificially generated or manipulated and appears authentic. The same applies to AI-generated text published to inform the public on matters of public interest, unless a human has reviewed it and someone holds editorial responsibility. Audio with the cloned voice of a real person can be a deep fake within the meaning of Article 3(60).
The Digital Omnibus changed nothing about the disclosure in a conversation. In Article 50 it only replaces paragraph 7, the basis for codes of practice. [2] For the usual customer service setup with phone AI and a chatbot, paragraph 1 is therefore the one that counts today.
Provider or deployer: who has to do what?
The regulation distinguishes two roles:
- Provider is whoever 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, whether for payment or free of charge (Art. 3(3)).
- Deployer is whoever uses an AI system under its authority in the course of a professional activity (Art. 3(4)).
With platforms for voice and chat agents, the split is rarely perfectly clean. The platform supplies the system. The customer writes the greeting and the instructions and switches the agent on under its company name. Who is legally the provider of that specific agent depends on the individual case. That is why the division of tasks belongs in the contract.
In practice something else matters more: the caller hears your company. If the disclosure is missing, it reflects on you, no matter who should have built it in. Before going live, ask your provider three questions:
- Does the system show or speak the disclosure by itself, or do I have to set it up?
- Where in the contract does it say who is responsible for the disclosure?
- What happens to the disclosure when I change the greeting or the language?
White label and resale: whoever resells an AI system under its own brand can already be a provider under the definition in Article 3(3). For high-risk systems, Article 25(1) says so explicitly. Distributors, deployers and other third parties are considered providers there if they put their name or trademark on such a system (point a) or change the purpose of a system so that it becomes a high-risk system (point c). [1]
When a voice agent becomes a high-risk system
Most phone and chat agents are not high-risk systems. They take requests, answer questions and book appointments. Annex III does, however, list areas of use in which exactly such an agent can become high-risk: [1]
| Annex III | What is covered | Example from phone and chat |
|---|---|---|
| No. 1(c) | Emotion recognition systems | The agent analyses the voice to detect anger or uncertainty |
| No. 4(a) | Recruitment and selection, screening and filtering applications, evaluating candidates | An agent runs pre-screening calls with applicants and scores them |
| No. 5(b) | Creditworthiness and credit scores of natural persons, but not fraud detection | An agent pre-qualifies callers for financing |
| No. 5(c) | Risk assessment and pricing in life and health insurance | An agent asks for health details and calculates a premium |
| No. 5(d) | Evaluating emergency calls, dispatching emergency services, emergency patient triage | An agent decides which emergency call is handled first |
On emotion recognition: Article 3(39) means systems that identify or infer emotions or intentions on the basis of biometric data. An analysis of the voice can fall under that. An analysis of written text alone does not rely on biometric data.
The exception in Article 6(3): a system listed in Annex III is not high-risk if it poses no significant risk and meets one of these conditions. It performs a narrow procedural task, improves the result of a previously completed human activity, detects decision-making patterns without replacing the human assessment, or performs a preparatory task. The exception never applies if the system performs profiling of natural persons. Whoever relies on it must document the assessment before the system is placed on the market or put into service (Article 6(4)).
As a rule of thumb: if the agent only takes down details and a human decides, a preparatory task is likely. If it evaluates, sorts or prices people itself, it gets tight. For the obligations that then follow, the Digital Omnibus left a little more time: risk management, technical documentation, logging and human oversight apply to Annex III systems from 2 December 2027. [3]
What is never allowed
Since 2 February 2025, Article 5 has banned some practices outright. Three of them matter in practice for phone and chat: [1]
- Manipulation and deception (point a): techniques that materially distort a person's behaviour by appreciably impairing their ability to make an informed decision, causing or being likely to cause them significant harm.
- Exploiting vulnerabilities (point b): deliberately exploiting age, disability or a specific social or economic situation with the same effect.
- Emotion recognition in the workplace and in education (point f): banned, except for medical or safety reasons.
Two examples make the difference tangible. A sales script that deliberately pressures older people into signing a contract is moving towards points a and b. Software that analyses the mood of employees in your own call centre falls under point f. Analysing the mood of customers is not banned, but it is a case for Article 50(3) and Annex III.
Fines and competitor claims
Article 99 sets maximum amounts by the severity of the breach: [1]
| Breach | Maximum amount |
|---|---|
| Prohibited practices (Art. 5) | €35 million or 7 % of worldwide annual turnover |
| Other obligations, including Art. 50 | €15 million or 3 % of worldwide annual turnover |
| Incorrect or incomplete information to authorities | €7.5 million or 1 % of worldwide annual turnover |
For companies the higher of the two amounts generally applies. For small and medium-sized enterprises including start-ups, the lower one applies (Art. 99(6)). There is no flat discount. The member states lay down the rules on penalties, and enforcement runs through national market surveillance authorities (Art. 99(1) and Art. 70).
In Germany there is a second route besides the authority. Under § 3a of the Act against Unfair Competition (UWG), it is unfair to break a statutory provision that is also intended to regulate market conduct in the interest of market participants, if the breach is capable of appreciably harming the interests of consumers, competitors or other market participants. [8] Under § 8(3) UWG, competitors, registered trade associations, qualified consumer organisations and chambers can then demand that the conduct stop. [8] Whether Article 50 is such a market conduct rule is ultimately for the courts to decide. Once a court says yes, a competitor can become a risk too, not only the authority.
Outside the EU: the US in brief
If you call or message customers in the US, different rules apply:
- Federal: since the declaratory ruling FCC 24-17 of February 2024, AI voices in outbound calls count as an "artificial voice" under the Telephone Consumer Protection Act and generally need prior consent. [9] A duty to disclose AI at the start of every call was only proposed by the FCC in 2024 (FCC 24-84). The 2026 Unified Agenda lists the proceeding with no next action determined. [9] [10] According to the FCC, inbound calls are not covered by the TCPA.
- States: Maine requires a clear disclosure where consumers could otherwise believe they are communicating with a human, in text and in voice. Utah requires a disclosure when someone clearly asks, and in regulated occupations also unprompted for high-risk interactions. California has banned bots since 2019 that mislead online about their artificial identity to incentivise a purchase or influence a vote. California and New York also regulate so-called companion chatbots, meaning bots that simulate a sustained, human-like relationship. Bots used only for customer service are expressly excluded in California. [11]
Checklist: what to do now
- Put the disclosure in the first sentence. On the phone in the greeting, in chat visibly before the first answer.
- Have the agent answer the human question honestly. One line in the agent's instructions is enough.
- Check every language. The disclosure belongs in the language the conversation is held in.
- Make a test call and time it. Listen yourself for how many seconds pass before the disclosure.
- Clarify responsibility. Write down whether your provider enforces the disclosure or whether you set it up yourself.
- Write down the intended purpose. Note what the agent is for and what it is not for, for example no decisions on applicants, loans or insurance premiums.
- Treat recording as a separate question. It is a legal issue of its own. In Germany, anyone who records another person's privately spoken words without authorisation commits a criminal offence under § 201 of the Criminal Code. [12]
- Read sales scripts against Article 5. No pressure on vulnerable groups, no deception.
How EchoCall implements the disclosure
We describe this precisely because it is the same question you should ask every provider: what happens automatically, and what do you set up yourself?
- Chat widget: the widget of the AI chat agent shows an AI notice by itself before the first message, regardless of the greeting the customer writes. The text exists in German, English and French.
- Voice conversation in the chat widget: before the microphone starts, a notice card appears. It says that the visitor is about to talk to an AI, not a human.
- Chat greeting: in the editor, a button inserts a pre-written disclosure sentence before or after your own text, in 74 languages.
- Agent on the phone: here the customer writes the disclosure into the first message of the AI phone agent. The editor reminds you of Article 50 with a notice, and the help centre shows sample sentences. Our terms and conditions oblige customers to activate the disclosure in their conversation design and keep it activated. [13]
EchoCall runs on IONOS servers in Frankfurt and Karlsruhe. You will find the data processing agreement and the subprocessors on the GDPR page. Voice packages start at €75 net for 250 minutes and chat packages at €40, everything on the pricing page.
Frequently asked questions
Does the AI disclosure duty also apply to small businesses?
Yes. Article 50 has no exemption by company size. Small and medium-sized enterprises only benefit when it comes to the level of possible fines: for them, the lower of the two maximum values, fixed amount or share of turnover, applies (Art. 99(6)).
Is a notice in the terms or on the website enough?
No. The information has to reach the person at the latest at the first interaction, clearly and distinguishably. A sentence in the terms that nobody reads before calling does not achieve that. The disclosure belongs in the greeting on the phone or visibly in the chat window.
Does the disclosure have to come on every call?
The law speaks of the first interaction. Whether a second call from the same person is a new interaction, it does not say explicitly. In practice a disclosure on every call is the safe choice: the system often cannot reliably tell who is calling, and the sentence only takes a few seconds.
Who is liable if my provider does not build in the disclosure?
By its wording, Article 50(1) names the provider. Whether you as the company using the agent carry a duty of your own, the text leaves open. For your customers that is secondary, because they hear your company. So clarify before going live who sets up the disclosure, and put it in the contract.
Does the AI Act also apply to companies outside the EU?
In many cases, yes. It covers providers that place AI systems on the market or put them into service in the EU, wherever they are established. It also covers providers and deployers in third countries where the output of the system is used in the EU (Art. 2(1)).
May my AI agent have a human name?
The law does not prohibit it. What matters is that nobody believes they are talking to a human. "This is Lea, the practice's AI assistant" does both: the agent has a name, and the disclosure is in the same sentence.
Do I have to label AI texts or voice samples?
That depends on who you are and what you publish. The machine-readable marking under paragraph 2 falls on the provider of the system that generates the content, for older systems from 2 December 2026. As a deployer, paragraph 4 requires you to disclose deep fakes and AI texts on matters of public interest if no human holds editorial responsibility. A voice sample with a freely invented AI voice is usually not a deep fake, one with the cloned voice of a real person can be. Labelling voluntarily never hurts.
Sources
All sources are in English unless marked otherwise.
- Regulation (EU) 2024/1689 on artificial intelligence, Articles 2, 3, 5, 6, 25, 50, 70, 99, 111, 113 and Annex III: eur-lex.europa.eu
- Regulation (EU) 2026/1744, Digital Omnibus on AI, with the new Article 111(4): eur-lex.europa.eu
- European Parliament, Legislative Train on the Digital Omnibus on AI, with the postponed high-risk deadlines and the procedural steps: europarl.europa.eu
- European Commission, guidelines on the transparency obligations under Article 50 of 20 July 2026: digital-strategy.ec.europa.eu
- European Commission, questions and answers on Article 50: digital-strategy.ec.europa.eu
- European Commission, start of enforcement on 2 August 2026: digital-strategy.ec.europa.eu
- European Commission, code of practice on marking and labelling AI-generated content of 10 June 2026: digital-strategy.ec.europa.eu
- Act against Unfair Competition, § 3a and § 8 (in German): gesetze-im-internet.de, gesetze-im-internet.de
- FCC, declaratory ruling FCC 24-17 and proposal FCC 24-84: docs.fcc.gov, docs.fcc.gov
- Federal Register, Unified Agenda 2026: federalregister.gov
- Maine, Utah, California and New York: legislature.maine.gov, le.utah.gov, leginfo.legislature.ca.gov, leginfo.legislature.ca.gov, nysenate.gov
- German Criminal Code, § 201 (in German): gesetze-im-internet.de
- EchoCall, terms and conditions, section on the EU AI Act: echocall.de/en/terms-conditions
All links were checked on 6 October 2026. This article is not legal advice. For your specific use case, ask a law firm specialising in IT law.
