Your AI SDR now has to say it is an AI

What changed on 2 August

Article 50 of the EU AI Act took effect four weeks ago. The core obligation is short:

Providers must design AI systems that interact directly with people so that those people are informed they are interacting with an AI system, unless it is obvious to a reasonably informed and observant person. The notice has to come at the point of first interaction, in a clear and distinguishable way.

Read the text at Article 50 of the AI Act, and the European Commission's own FAQ on the transparency obligations.

Three things about it matter commercially.

It applies to genuine two-way interaction with a natural person. A form that collects data is not caught. An agent that holds a conversation is.

Synthetic content has to be marked. Providers must mark AI-generated audio, video, images and text with machine-readable markers, and deployers must clearly label deepfakes at first exposure. If your outbound uses a cloned voice or a generated face, this is the clause that governs it.

The penalties are not nominal. Non-compliance with Article 50 carries fines of up to 15 million euros or 3 percent of global annual turnover, whichever is higher.

There is a limited grace period to 2 December 2026 for systems already placed on the market before 2 August. If you launch an AI agent into the EU market today, you do not get it.

The EU has three layers, not one

This is where most guidance goes wrong. The AI Act tells you how your agent must introduce itself. It says nothing about whether you were allowed to contact that person in the first place. Two other regimes answer that.

Layer one: can you process their data at all

Cold outreach to a named person is processing personal data, so you need a legal basis under Article 6 of the GDPR. For B2B prospecting the usual basis is legitimate interest, Article 6(1)(f).

Legitimate interest is not a free pass. It requires a documented Legitimate Interest Assessment covering three tests: that your purpose is legitimate, that the processing is necessary to achieve it, and that your interest is not overridden by the individual's rights. For B2B outreach to a corporate address, about a role that person actually holds, the balance usually comes out in your favour. To a personal webmail address, it usually does not.

Write the assessment down. An LIA you cannot produce is, for regulatory purposes, an LIA you never did.

Layer two: you have to tell them where you got their details

This is the obligation most outbound programmes miss entirely. Article 14 of the GDPR applies whenever you obtain personal data from somewhere other than the person themselves, which is exactly what prospecting is. You must tell them who you are, why you are processing their data, your legal basis, how long you keep it, their rights, and the source the data came from.

In practice this is a line in the footer and a link to a privacy notice, not an essay. But the source disclosure has to be there, and "we found you on the internet" is not a source.

They also have an unqualified right to object to direct marketing under Article 21. Unqualified means there is no balancing test and no discretion. They object, you stop, permanently, across every campaign and every channel.

Layer three: national law, which is where it stops being harmonised

The ePrivacy Directive governs electronic marketing separately from the GDPR, and because it is a directive rather than a regulation, each member state implemented it differently. There is no single European answer to "can I send this email."

Germany is the strict end. Section 7 of the UWG, the unfair competition act, generally requires prior express consent for commercial email, including B2B. Legitimate interest under the GDPR does not rescue you from it, because it is a different statute doing different work. Germany is also an unusually litigious market for this, with competitors and trade associations able to act, not only regulators.

France is the permissive end for B2B. The CNIL's position is that B2B prospecting to a professional address is permissible on a legitimate interest basis where the message relates to the recipient's professional role. See the CNIL guidance on commercial prospecting.

Most other member states sit between the two, generally closer to France for genuine B2B to corporate addresses.

The practical consequence: if you run one European sequence with one set of rules, you are compliant in some countries and not others. Segment by country before you segment by persona.

The US took a different route to a similar place

There is no US equivalent of the AI Act. But telephone law caught AI voices two years before Europe did, and it is considerably less forgiving than most people running outbound expect.

AI voices are artificial voices

On 8 February 2024 the FCC issued a declaratory ruling that AI-generated voices are "artificial" for the purposes of the Telephone Consumer Protection Act. See the FCC announcement and the full ruling, FCC 24-17.

That single classification does the work. Once your agent's voice is "artificial," the TCPA's consent requirements attach whether or not you used an autodialer, and whether the speech is pre-recorded or generated live in the conversation.

The exposure is what makes this serious. TCPA damages are statutory, at 500 dollars per violation and up to 1,500 dollars for a willful one, with a private right of action and no cap. Per call. A campaign of ten thousand non-compliant calls is an arithmetic problem, not a risk assessment.

Email is much more permissive, and people over-correct

CAN-SPAM is an opt-out regime, not an opt-in one. You do not need prior consent to send a commercial email in the United States. You do need accurate header information and a non-deceptive subject line, identification of the message as an advertisement, a valid physical postal address, a clear opt-out mechanism, and you must honour opt-outs within ten business days.

Civil penalties are assessed per individual email, which is how a modest campaign becomes an existential number.

The gap between US email rules and German email rules is the single widest divergence in this whole area. One sequence cannot serve both.

Some states legislate the bot question directly

California's bot disclosure law, at Business and Professions Code section 17941, makes it unlawful to use a bot to communicate with a person in California with the intent to mislead them about its artificial identity, where the purpose is to incentivise a sale or influence a vote. It has been in force since 2019, and it lands close to where the EU AI Act arrived seven years later.

And recording is its own problem

AI voice agents record by default, and roughly a dozen US states require all-party consent to record a call, including California, Florida, Illinois, Pennsylvania and Washington. The safe operating rule for a multi-state programme is to disclose recording at the start of every call regardless of where the number sits, because your agent does not reliably know where the person actually is.

Side by side

What a compliant programme actually looks like

Nine things. None of them are difficult, and most outbound programmes are missing at least four.

1. Your agent identifies itself as an AI at the start of every conversation. Required in the EU now, close to required in California, and good practice everywhere. It also converts better than people expect, because it removes the moment where the prospect starts to suspect rather than know.
2. You segment by country before anything else. Germany does not receive the same sequence as France, and neither receives the US sequence.
3. You have a written Legitimate Interest Assessment. One document, three tests, kept current.
4. Your first message discloses where the data came from and links to a privacy notice.
5. Every message and call carries a working opt-out, and opt-outs propagate to a global suppression list across every channel, not just the one they replied to.
6. You have prior express consent before any AI voice call to a US number. This is the single largest source of exposure in an AI outbound programme.
7. You disclose recording at the start of every call, regardless of state.
8. A cloned voice or face has documented consent from the person it belongs to, and synthetic media is marked.
9. You keep the records. Consent, source, assessment, opt-out timestamps. Compliance you cannot evidence is compliance you do not have.

What is still unsettled

Anyone who tells you this area is stable is selling something.

The FCC has proposed further rules that would define AI-generated calls and require in-call disclosure, but those remain proposals rather than law. Consent standards under the TCPA are being actively litigated. State-level AI legislation in the US is moving quickly and inconsistently, and several frameworks have shifted or been redrafted in the past year. The ePrivacy Regulation intended to replace the current directive has been in negotiation for years without landing.

The direction of travel, though, has been consistent for a decade and points one way: more disclosure, not less. A programme built on being upfront about what it is does not need rewriting every time a rule changes. One built on the prospect not noticing does.