EU AI Act and cold email: when AI-written outreach needs a disclosure

Short answer: does the EU AI Act apply to cold email?

The EU AI Act does not ban AI-written cold email and, on its plain text, does not require the person who sends a sales email to label it, but an AI agent that carries on an email conversation with a recipient is covered by Article 50(1).

Article 50 puts the duties that can reach a sales email on the providers of AI systems: design systems so people are told when they are talking to an AI, and mark generated text so a machine can detect it. Whether a particular email falls under them depends on who is really talking to the recipient, a person or the system.

This guide is not legal advice.

What Article 50 says about AI-generated emails

Article 50 of Regulation (EU) 2024/1689 sets four transparency duties and a timing rule. Only some reach email.

Two points are easy to miss.

First, a company that uses an AI tool in its business is a deployer, and the deployer duties in 50(3) and 50(4) do not reach a private business message. Article 50(4) needs text that is published to inform the public on a matter of public interest, and the Commission's guidelines list "private, interpersonal correspondence (for professional purposes)" as text that is not published.

Second, the provider duties are about how a system is built. A sales team using a vendor's tool does not build the marking or the disclosure. A team that builds its own agent in house and puts it into service under its own name can be the provider, according to the guidelines.

AI SDR compliance in Europe: case 1, an AI agent that talks back

The guidelines, published on 20 July 2026, split 50(1) into four elements: an AI system, intended to interact, directly, with natural persons. Their reading is wide in three ways that matter for sales.

• Email counts. Interaction can happen through a phone, e-mail or an online interface, and a system that can write and send messages can interact directly. A human does not have to start the exchange.
• AI agents are named. Agents are covered when they can interact with people in tasks such as managing correspondence or negotiating, and they should disclose both that they are artificial and who they act for.
• B2B is not an escape. The people concerned may be professionals, including a person acting for a legal entity.

Two limits are worth knowing. Simple rule-based auto-responders are not AI systems for this purpose, and the guidelines name traditional out-of-office emails as an example. The guidelines also say direct interaction typically involves real-time or near real-time exchange, which fits email less neatly than chat. They still name e-mail, so that is a weak reason to assume a conversing agent is outside the rule.

The "obvious" exception is narrow. The guidelines limit it to cases where there is almost no doubt about the nature of the interaction, and they warn that human-like writing makes this harder. On that standard, a fluent email from a first-name sender is unlikely to count as obviously machine-written.

On format, the guidelines give a plain example: an email generated by an AI agent that features an AI label at the top. Disclosure only in terms and conditions, or a generic notice such as "Services on this website use AI", is not enough.

The marking duty under 50(2) and short emails

Article 50(2) asks the provider to mark generated text in a machine-readable way. The guidelines list AI-generated summaries, and rewriting that changes style, structure and meaning beyond grammatical and minor stylistic correction, among the changes that need marking. Grammar correction, spell-checking and AI translation of existing text fall under the standard-editing exception. An AI system that writes a first draft is generating text, not editing it.

The guidelines also describe a narrow business-to-business exemption. It needs output that is strictly technical, meant for a limited number of professionals inside the organization and not intended to be shared outside the company, so an outbound sales email does not qualify.

The voluntary Code of Practice, which the Commission assessed as adequate on 8 July 2026, exempts "very short text" from its watermarking measure. For now that means text under 200 tokens. A short first email is often below that, though counts vary by tokenizer and language. The Act sets no length threshold, and the Code applies only to those who sign it. The Commission's opinion adds that adherence to the Code is not conclusive evidence of compliance.

Disclose AI in sales emails? Case 2: AI drafts, a person approves, a tool sends

The guidelines point two ways. Direct interaction excludes mediated cases, for example customer service staff who use AI assistance tools to help them communicate, or output that reaches the person through someone else who disseminates it. But the mere possibility of human review must not be used to get around 50(1), and products that blend AI replies with human-curated content need disclosure unless the AI outputs were "properly reviewed and sent by humans as the main interlocutors".

That gives a reading at each end. A person who reads the draft and sends it is on the mediated side. A person who could review but does not is on the other.

The middle is open. If a person clicks approve in a queue and the platform sends from their mailbox on a schedule, is that "sent by humans"? Is a glance at the draft "properly reviewed"? The guidelines do not address approval queues, they are non-binding, and they state that only the Court of Justice of the European Union can give an authoritative interpretation of the Act.

We know of no published enforcement decision or regulator statement applying Article 50 to sales email as of 17 October 2026, but we did not check every member state's authority.

Does human review exempt AI-written emails from the AI Act?

Not in the way many summaries suggest. They cite the exception in Article 50(4): disclosure is not needed where AI-generated text has undergone human review or editorial control and a person holds editorial responsibility. It applies only to text published to inform the public on matters of public interest, such as news articles or public safety warnings. Private business email never enters 50(4), so the exception neither frees cold email nor puts it at risk. It does not appear in 50(1) or 50(2).

The guidelines do define review there: a deliberate examination of the substance by someone with relevant knowledge, not spell-checking or a cursory sign-off. That is guidance for 50(4), not a legal test for email. It is still a sensible reference point if you want to document that a person really looked.

AI Act timeline for 2026: dates, enforcement and fines

The Commission proposed the Digital Omnibus on AI on 19 November 2025, with a political agreement on 7 May 2026. It was adopted as Regulation (EU) 2026/1744 of 8 July 2026, published on 24 July and in force from 27 July. For Article 50 its main effect is a new Article 111(4) with the 2 December date. The proposal offered six months for marking and the adopted text gives four, so older summaries may show a later date.

The guidelines add that a system that is partly interactive and partly generative gets the extra time for marking only: the duty to tell people they are dealing with an AI applied from 2 August.

Who enforces Article 50, and the fines

National market surveillance authorities enforce Article 50 for most systems. The AI Office is competent for AI systems built on a general-purpose model by the same provider as the model. Any person with grounds to think the rules were broken can lodge a complaint with a market surveillance authority.

Fines go up to EUR 15 million or 3% of worldwide annual turnover, whichever is higher. For SMEs, start-ups and small mid-cap companies the lower of the two applies (Articles 99(4)(g), 99(6) and 99(6a)). Other law keeps its own enforcers: the guidelines say the AI Act's market surveillance role does not affect the powers of, for example, data protection and consumer protection authorities.

How to disclose AI in sales emails: a practical policy

1. Classify each workflow as an agent that converses, a draft a person reads and sends, or an approved draft that a tool sends. Write the answer down with the date.
2. If an agent replies on its own, say in the visible body of the first message that it is an AI and who it acts for. A link to a policy page can add to that but, per the guidelines, does not replace it.
3. If a person reviews and sends, keep evidence: who approved, what they edited, how many drafts they rejected.
4. If a tool sends after a quick approval, or on its own, consider a plain disclosure line. It costs little and covers the unclear case. We found no reliable published data on how it affects replies, so test it on your own list.
5. Do not invent a human experience. A made-up memory such as "I enjoyed your talk" can mislead the reader whatever Article 50 says. See how to keep made-up details out of AI personalization.
6. Keep the other rules separate. Article 50(6) leaves other transparency obligations for deployers in Union or national law in place. Whether you may send at all is a question for data protection and marketing law: see the GDPR cold email guide and the German UWG guide.
7. Ask your vendor whether it has signed the Code of Practice, how it marks text and whether disclosure can be switched off. The AI SDR buyer's checklist has more questions.

How FoxKnock handles this

FoxKnock drafts, you approve, and your own mailbox sends. The default approval level is Review each, so every email waits for a person. Automatic sending is opt-in and earned per sequence: it unlocks only after at least 50 emails approved one by one, with under 10% edited and under 5% rejected, and never because a timer ran out. Even then, only drafts that pass the quality gate send on their own. See approval-first outreach.

Mapped to the cases above:

There is an optional AI disclosure footer, off by default. Its text reads "This message was drafted with AI assistance and reviewed under our sending rules", and it is added only to AI-written messages, after your signature in small print. The AI inbox drafts suggested replies that you send yourself. Apart from earned automatic sending, the one message FoxKnock can send without anyone approving it first is the opt-in booking-link reply to a meeting request, a fixed-wording reply with your own booking link, off by default.

The footer is a visible line, not the machine-readable marking that 50(2) describes, and it sits at the bottom of the email where the guidelines' example puts an AI label at the top. FoxKnock is itself a vendor of a text-generating system, so the provider duties above are ours to assess too, and this guide is not a statement that every Article 50 duty is met. We are not neutral here. If you switch on automatic sending, switching the footer on as well is the conservative choice. AI features send content to Claude, by Anthropic.

If you want to see the approval flow on your own sequences, the Free plan has 500 one-time credits and needs no card. You can create a free account and leave the approval level on Review each.

Not legal advice

This guide reads the two regulations and the Commission's guidance as published on 17 October 2026. The guidelines are non-binding, and we know of no court ruling on Article 50 yet. It is not legal advice. For a real sending program, ask a lawyer in the member states you send into.