EU AI Act Article 50: What Your AI Phone Agent Must Say

EU AI Act Article 50: What Your AI Phone Agent Must Say From August 2

From 2 August 2026, an AI phone agent operating in the EU must disclose two things: that it is a machine — and on whose behalf it is calling. For a voice agent that means spoken, in plain language, at the latest at the first interaction. A line in your terms and conditions, a watermark in the metadata, or a vague "I'm your assistant" is expressly not enough.

That is not an interpretation. It is the outcome of the final guidelines the European Commission published on 20 July 2026 — thirteen days before the obligations apply. This article translates them into concrete configuration: which sentence belongs in your greeting prompt, when disclosure has to be repeated, who carries liability in white-label setups, and which artefacts you need when an authority asks. For the fundamentals, see our EU AI Act compliance checklist for voice AI — this piece covers what the guidelines added on top.

What was actually published on 20 July 2026

The Commission issued a 51-page set of guidelines on Article 50 of the AI Act. Formally they are non-binding — only the Court of Justice can give an authoritative interpretation. In practice they are the rulebook: market surveillance authorities across all 27 Member States will use them as the primary reference.

Article 50 builds four separate obligations that fall on different actors:

• Article 50(1) — systems interacting directly with people: chatbots, voice assistants, automated customer service. The provider must design the system so users know they are talking to a machine.
• Article 50(2) — generative systems: synthetic text, image, audio or video must be marked in a machine-readable format and be detectable as artificial.
• Article 50(3) — emotion recognition and biometric categorisation: the deployer must inform the people exposed to the system.
• Article 50(4) — deep fakes and AI-generated text on matters of public interest: the deployer must disclose the artificial nature of the content.

For a voice agent that answers or places calls, Article 50(1) is the core. If you also generate synthetic voice and store or distribute recordings, check Article 50(2). The addressees are not the narrow category of "AI producers" but the far broader group of businesses using AI in daily operations. A dental practice with an AI answering service is as much in scope as a corporation.

The obligation most teams miss: not just "I am an AI"

This is where the final guidelines add the most — and why a compliance configuration built in spring may be incomplete today.

The guidelines confirm that AI agents are covered, and they must disclose both their artificial nature and the person on whose behalf they are acting. An agent capable of making bookings, managing correspondence, negotiating, concluding contracts or executing purchases must state both. For an outbound agent, "this is an AI assistant" is not sufficient — the announcement must also make clear which company it is calling for.

Second point, on architecture: where a provider cannot determine in advance whether an agent will interact with a natural person, the agent must be designed at the architecture level to disclose itself in every situation where that is reasonably likely — including where the natural person is acting for a legal entity. The employee who picks up the phone for their employer is protected exactly like a private caller.

Third point, the most underestimated in practice: agents must also disclose themselves to the persons instructing them, at key steps such as authorisation, reporting and validation — and at every new interaction. A single announcement at the start of a long chain of contacts does not carry. The guidelines address multi-agent architectures explicitly.

What does not count as disclosure

The guidelines are most useful where they say what does not work. Expressly listed as insufficient when used alone:

• disclosures contained only in terms and conditions, URLs or documentation
• machine-readable markings such as metadata or watermarks that the user cannot perceive at the point of interaction
• unclear or ambiguous signals — a generic reference to an "assistant", or human-like representations that may mislead
• generalised statements such as "services on this website use AI"
• purely technical descriptions such as "this system uses LLMs", without explaining what that means for the user

What is required instead is a notice that is clear, in plain language, and placed at the moment the user comes into contact with the system. For voice systems that means an audible, spoken statement. The test is simple: a reasonably informed caller must understand, immediately and without effort, that there is no person on the other end.

In riskier contexts — financial or health advice, complaints handling, interactions with children or elderly users — a single notice at the start may not be enough. The guidelines consider periodic reminders likely necessary there.

When, how often, and with what evidence

From the guidelines you can derive a configuration rule per interaction scenario. This matrix is the practical core:

The last row deserves a warning: as soon as a human in the loop becomes reasonably likely, the obligation applies — and the burden of proof lies with you.

Provider or deployer? The role question decides who is liable

In most cases a company integrating a third-party AI agent into its telephony is a deployer: it uses the system under its own authority but did not build it. The qualification can shift. A business that substantially modifies the system, places it on the market under its own name or trademark, or changes its intended purpose can slide into the provider position — with considerably heavier obligations.

For agencies and IT integrators reselling voice agents as a white-label solution under their own brand, this is the defining question of the summer. If you market the platform under your own name, assume you fall into the provider role — and with it the duty to make disclosure technically possible and effective.

Two clarifications matter. "Authority" over an AI system does not require technical control — it is about deciding whether and how the system is used. And individual employees acting under company instructions are not separate deployers: the legal person is.

The "obvious" exception is narrower than most assume

Article 50(1) contains an exception where the artificial nature is obvious from the circumstances. The guidelines narrow it considerably: it should be limited to cases where almost no doubt remains. Code assistants used only by professional developers, internal tools for trained staff, or diagnostic aids for health professionals can qualify.

A customer-facing helpdesk agent on a public number does not. The burden of proof rests on whoever invokes the exception. For practically any voice agent taking external calls, it is not an option.

Deadlines: 2 August and the narrow December corridor

One misunderstanding is causing real damage: there is an extension to 2 December 2026 — but it is far narrower than usually described.

Three consequences. The extension covers only Article 50(2), so your phone agent's disclosure obligation lands on 2 August, unchanged. Systems that are partly interactive and partly generative benefit only in relation to the marking. And the decisive criterion is the date of placing on the market or putting into service of each individual system — without an inventory you cannot establish which deadline applies to you.

On technical feasibility the document is blunt: it is an objective notion, not dependent on the individual provider's resources. "Our team is too small" does not carry.

A Code of Practice on Transparency of AI-Generated Content has existed since 10 June 2026, confirmed by the Commission and the AI Board as an adequate voluntary tool. It is voluntary; the obligations are not. Non-signatories must demonstrate compliance by other adequate means and carry out a gap analysis against the Code — and should expect more requests for information.

Implementing it with Famulor: step by step

For an existing voice agent this is a configuration and documentation task, not a rebuild.

1. Write the disclosure sentence. Name artificial nature and principal in one line. Inbound: "Hello, you're speaking with an AI assistant at Dr. Becker Dental. How can I help?" Outbound: "Hello, this is an AI assistant calling on behalf of Kramer & Partners Tax Advisors." No "virtual colleague", no first name without qualification.
2. Anchor it in the prompt and the flow. Set the announcement in the Flow Builder as the first step of every entry path — not just the main one. Check secondary entries in particular: callbacks, IVR routes, campaign paths.
3. Cover every language. Disclosure must happen in the language of the conversation. Across 40+ supported languages that means one reviewed wording per language, not a machine translation that dilutes the meaning.
4. Set re-disclosure at handover points. On call transfer, name the switch. On callbacks and follow-ups, the obligation starts over.
5. Audit mid-call actions. Agents that book appointments, write data or trigger orders through mid-call tools fall squarely into the category for which the guidelines require naming the principal.
6. Automate the evidence. Use post-call analysis to verify per conversation that disclosure actually happened. That turns a claim into an artefact — exactly what the accountability principle demands.
7. Build a system inventory. Record per system: role (provider or deployer), applicable Article 50 category, supplier, and date of placing on the market or putting into service. Without this register everything else is blind.
8. Renegotiate supplier contracts. Role allocation, warranties on marking of generative outputs, provision of disclosure functionality, indemnities, flow-down of obligations along the supply chain.

Industry examples

Dr. Becker Dental, Münster, 12 staff. The agent takes calls outside opening hours and books appointments. The greeting names the AI nature and the practice. Because appointment booking is an acting capability, naming the principal is mandatory. When the call is routed to the hygiene assistant in the morning, the switch is announced.

Kramer & Partners Tax Advisors, 40 staff. Outbound campaign chasing missing documents. Naming the principal is especially sensitive: the recipient must know instantly that the firm is behind the call. Because deadlines and financial consequences are involved, repeating the notice briefly in longer conversations is the safe reading.

Nordlicht IT Services, 60 staff, white-label reseller. Sells voice agents under its own brand to trade businesses. Role assessment: very likely a provider. Nordlicht must therefore ensure disclosure is technically available and effective, and must instruct its customers. Those customers remain deployers and must ensure the notice genuinely reaches the caller.

Hansen Property Management, 850 units. A chat widget plus a phone agent for damage reports. Two channels, two formats: a visible text notice in the widget before the first message, a spoken announcement on the phone. If follow-up happens over WhatsApp, the obligation starts again with the first message of the thread.

Common mistakes

• Disclosure in the privacy policy instead of at the point of contact. The most common and most easily avoided error.
• Euphemisms. "Digital colleague", "smart service", a first name with no qualifier. Anything that does not state the artificial nature fails.
• Artificial nature only, no principal. Precisely the gap the final guidelines closed.
• Once at the start, never again. Callbacks, follow-ups and multi-agent steps trigger the obligation afresh.
• Only the main path checked. Secondary entries and campaign paths are what audits find.

What compliance costs — and what non-compliance costs

The cost side is unusually clear. An existing agent mostly needs configuration work: adjust the prompt, walk the flows, review language variants, switch on evidence reporting, build the inventory, review contracts. For an SME agent with a handful of entry paths this is a matter of hours, not months.

On the other side: breaches of Article 50 fall into the AI Act's second fine tier, up to EUR 15 million or 3% of total worldwide annual turnover, whichever is higher. For SMEs and start-ups the lower of the two applies — a proportionality mechanism, not an exemption. Add the supervisory powers: requests for information, on-site inspections, orders to take corrective measures, and in serious cases withdrawal of a non-compliant system.

To run the numbers for your own case — call volume, time saved, cost per conversation — you can estimate it in a couple of minutes.

Conclusion

The final guidelines removed the last excuse for waiting. For voice agents the obligation reduces to one sentence: disclose out loud that this is an AI and who it works for — at every new interaction, in the language of the conversation, documented so it can be proven. 2 August 2026 arrives with no transition period.

Three things worth finishing this week. Update the disclosure sentence for every active agent — artificial nature and principal. Click through every entry path, not just the main one. Automate the evidence so a request for information is a matter of minutes rather than weeks.

That is what Famulor is built for: disclosure as the first step of every path in the Flow Builder, reviewed wordings across 40+ languages, re-disclosure at transfer points, and post-call analysis as automatic per-conversation evidence.

Concrete next step: open your active agent, play back the first ten seconds of a real call, and check whether it names both the artificial nature and the principal. If not, that is the change due before 2 August — and it takes minutes.

FAQ

Does my AI phone agent have to say it is an AI from 2 August 2026?

Yes. Under Article 50(1) of the AI Act, a voice agent must disclose its artificial nature audibly and in plain language, at the latest at the first interaction. The final guidelines of 20 July 2026 additionally require naming the person or company on whose behalf the agent is acting.

Is a notice in the terms and conditions or privacy policy enough?

No. The guidelines expressly list disclosures contained only in terms and conditions, URLs or documentation as insufficient. The notice must appear at the point where the user comes into contact with the system — spoken, in the case of voice agents.

Is it enough if the agent introduces itself once at the start of the call?

Not in every case. The obligation applies afresh at every new interaction, including callbacks and follow-up calls. In riskier contexts such as financial or health topics, or conversations with children and elderly users, the guidelines consider periodic reminders likely to be necessary.

What does the duty to name the principal mean in practice?

The agent must make clear which company it is acting for — not merely that it is a machine. This applies especially to agents that can book appointments, manage correspondence, negotiate, conclude contracts or execute purchases.

Does the extension to 2 December 2026 apply to my phone agent?

Not for disclosure. The transitional period to 2 December 2026 covers only the machine-readable marking obligation under Article 50(2), and only for systems placed on the market or put into service before 2 August 2026. The Article 50(1) disclosure obligation applies from 2 August with no transition.

When is the artificial nature "obvious" so disclosure is not needed?

Only where almost no doubt remains — for instance internal tools for trained staff or code assistants for professional developers. A customer-facing agent on a public number does not meet the exception, and the burden of proof rests on whoever invokes it.

How high are the fines for breaching Article 50?

Up to EUR 15 million or 3% of total worldwide annual turnover, whichever is higher. For SMEs and start-ups the lower of the two applies. There is no exemption based on company size.

What evidence should I keep for a regulatory inspection?

A system register with role, Article 50 category, supplier and date of putting into service, plus disclosure evidence such as prompt versions and transcript samples, the supplier's technical documentation, and the supplier contracts. Corrective measures taken after the fact are generally not considered sufficient to avoid a sanction.