Do You Have to Tell People They Are Talking to AI? EU and UK Rules for Chatbots and Voice Agents

If a customer rings your business and an AI answers, should you tell them? Until recently that was a matter of good manners and brand risk. For anyone providing AI systems in the European Union, it is now also a legal question, because the AI Act's transparency rules began to apply on 2 August 2026. For a UK business, the picture is a mixture of EU rules that may or may not reach you, and UK data protection and telecoms rules that certainly do.

This article sets out what we found in the primary sources. It is general information, not legal advice, and it is not a substitute for a lawyer who knows your facts.

If you are still deciding whether a voice agent suits your business, start with AI voice agents for service businesses. This article assumes you are well along that road.

Does the EU AI Act reach a UK business?

The AI Act is an EU regulation, but its reach is not limited to businesses in the EU. The Commission's guidelines give an example: a company that provides a chatbot or AI agent on the Union market under its own name or trademark is a provider responsible for Article 50 compliance "regardless of whether the provider is established or located in the Union or in a third country". The guidelines also describe a company that develops a chatbot in-house and puts it into service in the Union for its own use as a provider.

So the practical questions for a UK business are:

• Do you have customers, callers or users in the EU?
• Are you providing the AI system, or using one built and supplied by someone else?
• If you build systems for clients, who is the provider in law: you, the client or the underlying vendor?

None of these has a one-line answer. If you only serve UK customers from the UK, the AI Act's transparency duty may not apply to you directly. If you sell to EU consumers, assume it needs to be examined.

What Article 50(1) requires

The Commission's guidelines break the interactive-system rule into parts.

1. It must be an AI system. Simple rule-based responders such as an out-of-office email are excluded.
2. It must be intended to interact, directly, with natural persons. The guidelines include voice assistants, customer-support chatbots and AI hotlines as examples. Backend machine-to-machine calls are excluded, as are transcription tools and recommender systems where nobody converses with the system.
3. The provider must design it so people are informed. The information has to be given, under Article 50(5), in a clear and distinguishable manner at the latest at the time of the first interaction, and in an accessible format.
4. The exception is "obvious" interaction. It is judged from the point of view of a reasonably well-informed, observant and circumspect person. The guidelines list factors including speech patterns, whether the voice is robotic or genuinely human-sounding, and the capability for advanced personalised interaction. A very natural voice makes the exception harder to rely on, not easier.

Two further points matter for businesses using AI to talk to the public.

• Agents are covered. The guidelines say AI agents that interact with people, for example by making bookings or managing correspondence, must disclose both that they are artificial and the person on whose behalf they act.
• Human oversight is not a loophole. According to the guidelines, the mere possibility that a human could intervene or review outputs should not be used to avoid disclosure.

The guidelines are also candid about their status. They state that they are non-binding and that authoritative interpretation of the AI Act rests ultimately with the Court of Justice of the European Union. The Commission's covering communication adds that the guidelines "will be formally adopted by the Commission at a later date, when all language versions are available". The version we read is the draft whose content the Commission approved on 20 July 2026.

What the recent amendment did and did not change

Regulation (EU) 2026/1744 amended the AI Act in July 2026. It is easy to read headlines about "delays" and assume transparency was among them. From the Official Journal text, it was not.

Selected AI Act dates after Regulation (EU) 2026/1744

Provision	Date	What the sources say
General application of the AI Act, including Article 50	2 August 2026	The regulation's recitals confirm the general date of application as 2 August 2026.
Article 50(2): marking of synthetic audio, image, video or text	2 December 2026 for systems already on the market before 2 August 2026	A four-month transitional period applies to providers of generative systems placed on the market before 2 August 2026. The Commission's FAQ says the grace period applies only to this marking and detection obligation.
Article 50(1): informing people they are interacting with AI	2 August 2026	The Commission's FAQ describes the grace period as limited to Article 50(2). We found no similar delay for Article 50(1).
High-risk system requirements under Chapter III, Sections 1 to 3	2 December 2027 or 2 August 2028 depending on the category	The recitals move these dates. This is separate from the transparency rules.

How disclosure works in practice

The AI Act does not prescribe a wording or a technique. The Commission's guidelines list examples, and these are the ones that matter most for customer-facing agents.

Disclosure techniques named in the Commission's guidelines

Channel	Example technique	Notes from the guidelines
Web chat	A plain-language label or banner and a first-turn greeting, possibly with a persistent badge.	Place disclosures close to the input field and use simple wording.
Voice and telephony	An explicit spoken statement at the start of the interaction.	Combine with periodic reminders in longer interactions, in particular after interruptions or a change in the AI's role. Audio cues such as tones are not sufficient on their own.
Email from an agent	An AI label at the top of the message.	Given as an example of a visual means.
Agents acting for a person	Disclose artificial nature and the person on whose behalf they act.	Where the provider cannot know in advance whether an agent will meet a person, the guidelines say the agent should be designed and instructed to disclose itself wherever that is reasonably likely.
Terms and conditions or a link	Complement only.	Insufficient when used alone.

Example opening lines for a phone agent follow. They are illustrative wording only, not tested against any regulator's view, so adapt them with advice.

Hello, you've reached Northgate Dental. I'm an AI assistant, so I'm not a person.
I can book appointments or take a message. If you'd rather speak to someone,
say "person" at any time.

A good opening does four things: it says it is AI plainly, says what the agent can do, offers a way to a human, and does not hide the disclosure inside a long greeting.

The UK side: data protection and PECR

The EU rule is about telling people it is AI. The UK rules we looked at are mainly about what you do with their information and how you contact them.

Call recordings and transcripts are personal data. A voice agent usually creates both. The ICO's transparency guidance says you must be "clear, open and honest with people from the start about who you are, how and why you use their personal information, and their rights". Its lawful basis guide says you must have a lawful basis, record which one you rely on for each purpose, and include it in your privacy information. In practice this means deciding, before launch:

• whether you record calls or keep transcripts, and why;
• which lawful basis you rely on, and whether legitimate interests has been properly assessed;
• how long you keep them;
• how a caller finds your privacy information during a phone call.

Outbound marketing calls follow PECR. The ICO's guide says an automated marketing call, meaning a call by an automated dialling system that plays a recorded message, needs the person's specific consent, and that general marketing consent or consent for live calls is not enough. The call must state who you are and allow your number to be displayed. Whether a particular AI voice system counts as an automated dialling system is a legal question we have not tried to answer. If you plan outbound marketing calls with an AI voice, get specific advice before you build anything. Inbound calls that people make to you are a different situation.

Suppliers. If call audio goes to a voice-AI vendor, that vendor is processing personal data for you. Check what its terms say about retention, location and use of audio for model training. We discuss the same question for text models in our model API comparison.

A practical checklist

1. 01Decide who the provider isWrite down whether you, your client or your vendor is the provider and deployer in law.
2. 02Script the disclosureSay it is AI at the first moment of contact, in words a child or older caller would understand.
3. 03Offer a human routeMake handover easy and test that it works.
4. 04Decide on recordingChoose a lawful basis, retention period and notice wording.
5. 05Check vulnerable callersConsider whether older people or people with disabilities may be callers and adapt the format.
6. 06Test and logCheck the disclosure plays on every call path and keep evidence that it does.

For how these parts fit into a build, see how AI voice agents work, and for budgeting see what an AI voice agent costs in the UK. Agents that act on callers' behalf also raise security questions, which we cover in prompt injection and AI agent security.

What we have not covered

This article does not cover high-risk AI systems, emotion recognition, deep fakes, penalties or enforcement, sector rules for regulated industries such as finance and healthcare, or consumer-protection law beyond the Commission's pointer to it. It also does not assess whether the UK has introduced AI-specific legislation since the sources above were published. Check current guidance from the ICO and the UK government before you rely on it.

Conclusion

The sensible default is simple: tell people early, tell them plainly and make it easy to reach a person. That is what the Commission's own examples describe for voice, and it costs a sentence. The harder questions are legal ones: who is the provider, whether EU rules reach you, and how recordings and outbound calls are handled under UK law. Settle those before launch.

If you are planning a voice or chat agent and want to talk through how it would be built and disclosed, our AI voice agents page explains how we approach it, and you are welcome to get in touch.