# Must you tell people they are talking to an AI?

> Justas Butkus is a fractional AI officer based in Vilnius, Lithuania, founder of AINORA, MB and of Impetora, and a graduate of ISM University of Management and Economics.

Justas Butkus is a fractional AI officer based in Vilnius, Lithuania, who builds and operates production AI voice systems and advises companies deploying AI in customer-facing operations across the UK and EU. His applied work covers AI disclosure and cross-border outreach rules across EU member states, the UK and the US.

**Yes, but the duty is split. Article 50(1) requires the **provider** to design the system so people are informed they are interacting with an AI, unless that is obvious. Deployers carry Articles 50(3) and 50(4). Article 50 has applied since 2 August 2026 under Article 113, and it covers AI systems that call or chat with your customers.**

Canonical: https://justasbutkus.com/eu-ai-act/voice-agent-disclosure/
Last updated: 2026-08-01

---

## What Article 50 actually requires

Article 50 is the transparency provision. In plain terms it does three things relevant to anyone running an AI system that talks to customers. Note who each one binds — the paragraphs do not all point at the same party:

- **Disclosure of interaction — Article 50(1), on the provider.** Systems intended to interact directly with people must be designed so those people are informed they are interacting with an AI system, unless that is obvious to a person who is reasonably well-informed, observant and circumspect, taking into account the circumstances and the context of use.
- **Labelling of synthetic content — Article 50(2), on the provider.** Audio, image, video and text generated or manipulated by AI must be marked in a machine-readable way as artificially generated, with narrow exemptions including assistive or standard editing that does not substantially alter the input data.
- **Deepfake disclosure — Article 50(4), on the deployer.** Content that convincingly resembles real people, places or events must be disclosed as artificially generated, with narrow exemptions for artistic, satirical and fictional works, and for published text that has undergone human editorial review.

For a voice agent making or answering calls, the first is the one that bites. The exemption for the obvious is narrower than people assume: modern speech synthesis is precisely the case where it is no longer obvious.

## The dates that matter

**EU AI Act application timeline, as amended by Regulation (EU) 2026/1744 (Digital Omnibus on AI), in force 27 July 2026**

| Date | What applies |
| --- | --- |
| 2 February 2025 | Prohibited practices under Article 5, and Article 4 AI literacy obligations for providers and deployers |
| 2 August 2025 | Baseline obligations for general-purpose AI model providers; governance and the penalties regime |
| 2 August 2026 | Article 50 transparency applies under Article 113, and the enforcement regime bites for the obligations already in effect |
| 2 December 2026 | Grace period ends for synthetic-content systems already on the market; two new Article 5 prohibitions take effect, covering non-consensual intimate imagery and AI-generated child sexual abuse material |
| 2 August 2027 | Deferred regulatory sandbox obligations |
| 2 December 2027 | Standalone high-risk systems under Annex III, deferred by the Digital Omnibus |
| 2 August 2028 | High-risk AI embedded in already-regulated products, deferred by the Digital Omnibus |

Penalties for breaching the transparency obligations sit in the middle tier of the Act's sanctions regime, set out in [Article 99](https://artificialintelligenceact.eu/article/99/): up to €15,000,000 or 3% of total worldwide annual turnover, whichever is higher. The top tier, reserved for the prohibited practices in Article 5, runs to €35,000,000 or 7%.

> A detail that is routinely reported incorrectly: for SMEs and startups, Article 99 caps each fine at whichever of the amount or the percentage is LOWER, not higher. So a smaller company does not face the headline number that gets quoted in most coverage of the Act. Worth knowing before anyone sells you compliance work on the strength of a €35 million figure that would never apply to you.

## Does the Act require you to appoint an AI officer?

No. This is worth stating plainly because it is widely misrepresented.

Unlike the GDPR, which mandates a Data Protection Officer in defined circumstances, the AI Act requires no specific role, no named title and no particular governance structure. Article 4 requires providers and deployers to take measures that support the development of AI literacy among the people operating these systems. There is no mandated curriculum, examination or certificate.

What the Act does create is accountability that has to land somewhere. Appointing a responsible person is the common way organisations demonstrate that, and it is what auditors, insurers and enterprise customers increasingly ask to see. That is a market practice responding to a real obligation, not a legal requirement, and anyone telling you otherwise is selling from fear.

## What disclosure looks like on a real call

The provision is short. Implementing it without ruining the call is where the actual work is.

- **Say it early, and in the flow.** Disclosure at the top of the call, in the natural greeting, rather than buried in a legal preamble nobody listens to.
- **Say it in the language being spoken.** A disclosure in English on a call conducted in Lithuanian, German or Polish does not inform anyone.
- **Handle the direct question.** If a caller asks whether they are speaking to a person, the system must answer honestly and immediately. This needs to be a hard rule, not a matter of the model's discretion.
- **Log it.** Keep a record showing disclosure occurred on each call. When a regulator or a customer asks, the recording or transcript is the evidence.
- **Keep a route to a human.** Not strictly required by Article 50, but it is the thing that turns a complaint into a resolved call.

Member states also layer their own telecoms and consent rules on top of this, and they differ considerably on calling hours, consent basis and caller identification. The Act is a floor, not the whole picture.

## Sources, and a necessary caveat

- Regulation (EU) 2024/1689 (the AI Act) — [Article 50](https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX%3A32024R1689), [Article 99 penalties](https://artificialintelligenceact.eu/article/99/), [Article 113 application dates](https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX%3A32024R1689).
- Regulation (EU) 2026/1744, the Digital Omnibus on AI, in force 27 July 2026, which amended Article 4 and deferred several high-risk deadlines.

> This page is general information about the Regulation as it stands on the date above. It is not legal advice, it does not account for national implementing rules, and it does not replace advice on your specific circumstances.

## Who carries the obligation

The obligations are split, and this is the part most summaries get wrong. **Article 50(1) — designing the system so people are told they are dealing with an AI — sits with the provider**, not with you as the buyer. Deployers carry Article 50(3) where emotion recognition or biometric categorisation is involved, and Article 50(4) for deepfakes and AI-generated text. Separately, Article 26 requires you to operate the system in accordance with the provider's instructions.

So buying the system does not make the Article 50(1) design duty yours. It does not leave you clean either: the customer complaint, the procurement questionnaire and any national consumer, telecoms or data-protection breach land on you regardless. And if you put the system on the market under your own name or substantially modify it, you can become the provider yourself.

Procurement packs and insurer renewal forms have started carrying AI governance questions, and that is where most mid-market companies meet the Act first, rather than through a regulator.

## Frequently asked questions

### Does the EU AI Act require disclosing that a caller is an AI?

Yes. Article 50 requires that people are informed they are interacting with an AI system unless it would be obvious to a reasonably observant person. For a synthetic voice on a phone call, that exemption is narrow, so disclosure is the safe and expected position.

### When did Article 50 become enforceable?

2 August 2026, which is the Act's main compliance deadline and the point at which supervisory authorities gained full enforcement powers. Systems generating synthetic content that were already on the market before that date have until 2 December 2026.

### What are the penalties for getting transparency wrong?

Breaches of the transparency obligations fall in the middle sanctions tier: up to €15,000,000 or 3% of total worldwide annual turnover, whichever is higher. The most common practical cost, though, is failed procurement diligence rather than a fine.

### Does the AI Act require us to appoint an AI officer?

No. It mandates no role, title or governance structure, unlike GDPR's Data Protection Officer. It does create accountability obligations that have to sit with someone, which is why many organisations name a responsible person, but that is practice rather than law.

### Does this apply to us if we are outside the EU?

It can. The Act has extraterritorial reach: if your AI system is used in the EU, or its output affects people in the EU, the obligations can apply regardless of where your company is established.

### We bought our AI calling system from a vendor. Is it their problem?

No. Providers and deployers each carry duties, and buying the system does not transfer yours. If you point an AI caller at your customers you are the deployer, and informing those customers is your obligation.

## Related

- [What a fractional AI officer does](/fractional-ai-officer/) — Who owns this inside a company, and when the role is worth filling.
- [EU AI Act, in depth](https://impetora.com/eu-ai-act/overview) — Risk classification, conformity assessment and ISO 42001 mapping.

## If AI is already talking to your customers

The disclosure question is usually the visible edge of a larger one: nobody is quite sure what AI is running, who approved it, or who answers for it. That is worth half an hour.

Contact: justas@ainora.lt · [LinkedIn](https://www.linkedin.com/in/justas-butkus/)