AI Act
The caller has to be told. Today, only you can tell them
Article 50 of the AI Act requires that a person interacting with an AI system be informed of it. This product does not yet enforce that, the setting that looks like it would is about recording rather than about AI, and every assistant we ship greets a caller as a human would.
Nothing in CallAgent makes an assistant disclose that it is an AI. If your assistant does not say so in its own first message, nobody says it — and since 2 August 2026 that is a live obligation, not a future one.
What Article 50 actually asks for
Article 50(1) puts the duty on the provider of an AI system intended to interact directly with natural persons: the system must be designed and developed so that the people concerned are informed that they are interacting with an AI system, unless that is obvious to a reasonably well-informed person in the circumstances. The information has to be given at the latest at the time of the first interaction.
A voice assistant on a telephone line is the case the Article was written for, and the "obvious" exemption does not save it. The whole point of a modern speech pipeline is that it does not sound like a machine; a caller who cannot tell is precisely the caller the Article protects.
The transparency obligations in Chapter IV apply from 2 August 2026. There is no grace period left to plan around.
| Who it binds | The provider of the system, and — through Article 50(4) and the deployer duties — the business running it. |
|---|---|
| When | At the latest at the first interaction. On a telephone call, that is the greeting. |
| The exemption | "Obvious to a reasonably well-informed person." A natural-sounding voice does not qualify. |
| In force | Since 2 August 2026. |
What the product does about it, which is nothing
The greeting an assistant speaks is a free-text field. It is nullable, it is validated for length and for template variables, and nothing anywhere checks what it says. An assistant with an empty greeting simply says nothing and waits, and an assistant whose greeting is "Good morning, how can I help?" says that.
There is a consent mechanism in the product and it is easy to mistake for this one. It is not: it announces that a call is being recorded, its own source says so, its default is to announce nothing, and the wording it falls back to is "This call is being recorded" — which tells a caller about a microphone, not about a machine.
There is also a field on an assistant called identity disclosure. It is stored, it is returned by the API, and it is never read: it is not included in the configuration sent to the voice gateway, so it cannot affect a call. And there is a post-call scoring field that asks a model whether the assistant identified itself, which measures the problem rather than preventing it.
| Greeting validation | Length and template variables only. No content rule. |
|---|---|
| Consent policy | About recording, not about AI. Defaults to announcing nothing. |
| Identity disclosure flag | Stored, never sent to the gateway, never acted on. |
| Post-call scoring | Asks a model whether disclosure happened. After the call. |
| Shipped greetings | None of the demo assistants or templates discloses. They read as receptionists. |
The announcement path does not fire at all
While writing this page we went looking for the recording announcement in the running configuration and could not find it arriving. The reason is a key mismatch: the service that plans the announcement reads the assistant's plan from one place in the call snapshot, and the resolver writes it to another, under different field names again. The lookup therefore always comes back empty, the policy falls through to the organization default, and that default is to announce nothing.
Which means the honest statement about this build is stronger than "AI disclosure is not enforced". No announcement of any kind is spoken on a call today. Not the AI disclosure, which was never built, and not the recording notice, which was.
That is a bug rather than a policy, it is on the record here, and it is on the list ahead of any new feature. It is written on this page rather than only in a tracker because a compliance page that omits the compliance bug is the thing this site exists not to be.
Put it in the first message
Until the platform enforces it, the disclosure has to be the first thing your assistant says, written into the greeting by you. It is one line, it costs you about two seconds of call time, and it is the difference between a compliant deployment and one that is not.
Two things make it work rather than merely exist. Say it in the first sentence, before the caller starts explaining their problem — a disclosure at the end of a paragraph has already lost the point. And say it in a way a person actually parses: "an automated assistant" and "a virtual assistant" are both weaker than saying the word.
If your assistant transfers to a human, say that too when it happens. A caller who was told they were talking to an AI and then reaches a person should be told the switch has happened, or the first disclosure starts working against you.
-
ro
Bună ziua, sunt asistentul virtual — un sistem de inteligență artificială — al companiei :company. Cu ce te pot ajuta?
-
en
Hello, I am :company's AI assistant — an artificial intelligence, not a person. How can I help?
-
combined
Bună ziua, sunt asistentul cu inteligență artificială al companiei :company. Apelul este înregistrat. Cu ce te pot ajuta?
The third combines the AI disclosure with the recording notice, which is the shape most deployments end up needing, and gets both out of the way in one breath.
Where the enforcement would have to live
Published rather than kept internal, because a customer buying on the strength of this page should be able to check whether the change landed.
The one place with authority over the first thing a caller hears is the voice gateway's first-turn path — the same choke point the recording notice already uses, which is what makes "the caller was told before the assistant said anything" true rather than usually true. A disclosure added there covers the inbound greeting, the held outbound greeting, and the case where an assistant has no greeting at all and would otherwise say nothing before the caller speaks.
Doing it only in the control plane, by prefixing the greeting before it goes on the wire, is cheaper and is not enough: a phone-number hook can override the greeting after that point, and a model-generated first message never passes through the field at all.
The second half is the default. A disclosure setting that ships off has the same effect as no setting, so the default has to be to disclose, and turning it off has to be a deliberate act with the reason recorded.
| Where | The voice gateway's first-turn announcement path, beside the recording notice. |
|---|---|
| Not only where | Prefixing the greeting in the control plane is bypassed by number hooks and by model-generated first messages. |
| Default | On. A transparency control that ships off is not a control. |
| Also | Every shipped demo and template greeting has to disclose, or the examples teach the opposite. |
| Test | No test asserts anything about greeting content today. One has to. |
What else Article 50 covers
The rest of the Article is about generated content: synthetic audio, image, video and text have to be marked machine-readably, and deep fakes and AI-generated text published to inform the public have to be disclosed. Emotion recognition and biometric categorisation carry their own duty to inform the people exposed to them.
This product synthesises speech, which is generated audio content. Nothing in the pipeline watermarks it, and the practical exemption for assistive and editing functions is doing a lot of work in an argument that it need not be. If you are recording and republishing an assistant's voice rather than speaking it down a line, that is a separate question and it is one to take advice on.
The product does not do emotion recognition, and the analysis it runs after a call is a summary and a set of structured fields rather than an inference about a caller's emotional state. If that changes, it belongs on this page before it ships.
These pages describe what this software does. They are not legal advice and they do not decide whether what you are doing with it is lawful. That decision is yours, and in most of the interesting cases it depends on facts about your business that no software can see.
If you are deploying this in the Union
The disclosure has to be in your assistant's first message until the platform enforces it. We will help you write it, in Romanian or English, and check it against the rest of the greeting.
The other European documents
None of the seven documents in this set has been through a lawyer. They are engineering statements about a running system, published because a wrong description is worse than a missing one, and they will be reviewed before anybody is asked to sign anything.
Notes
-
1
The greeting field is
assistants.first_message; its API validation is length and type only. -
2
The consent policy enum covers recording announcements and defaults to
nonein the migration, in the model fallback and in the resolving service. -
3
The unread assistant field is
compliance_plan.identity_disclosure_required; it is not present in the snapshot sent to the voice gateway.