The EU AI Act and AI-written cold email: what applies

Using an AI model to draft cold email copy does not, by itself, put you inside the EU AI Act's high-risk regime. The Act is product-safety legislation organised by risk category, and marketing copy generation is not one of the high-risk uses it lists. The law that actually governs your outbound — lawful basis, notice, objection rights, profiling — is still the GDPR, plus the national laws implementing the ePrivacy Directive on unsolicited electronic marketing.

That is the short answer, and it is worth stating plainly because this topic attracts a lot of alarming and inaccurate marketing. What follows is the division as best it can be stated. None of this is legal advice, the Act's implementation is still producing guidance and standards, and its application timeline has itself been the subject of amendment proposals — so verify the current position with counsel before relying on any of it for a compliance decision.

What the Act is, in one paragraph

Regulation (EU) 2024/1689 entered into force in August 2024 and applies in stages rather than all at once. The prohibitions on a short list of unacceptable practices and the AI-literacy duty came first, in early 2025. Obligations on general-purpose AI models followed later in 2025. The bulk of the regime — including the high-risk classification system and the Article 50 transparency duties — was legislated to apply from August 2026, with a further tranche for AI embedded in already-regulated products after that. Those dates have been subject to proposed adjustments, so treat the schedule as something to check rather than something to quote.

Which parts could touch outbound at all

The transparency provisions are where most of the confusion comes from. They include a duty to tell a person when they are interacting with an AI system unless it is obvious, and duties around marking machine-generated or manipulated content. Neither maps cleanly onto "I used a model to draft an email that I then reviewed and sent under my own name". The chatbot-style disclosure concerns interactive systems; the content-marking duties are framed around providers of generating systems and around certain publication contexts, with carve-outs where AI performs an assistive editing function and does not substantially alter the human input.

Where this *does* get closer: a fully automated agent that conducts the back-and-forth of a conversation with a prospect without a human in the loop is a different thing from a drafting tool, and the interaction-disclosure question becomes live. If that is your setup, get advice.

What actually binds your outbound

This is the part worth your attention, because it applies now and is enforced.

• Lawful basis (GDPR Article 6). For B2B outreach you can often rely on legitimate interest, but you have to perform the balancing assessment and write it down. Using AI to generate volume does not change the basis; it just makes the balancing harder to argue if the volume is indiscriminate.
• Notice (Articles 13–14). When you collect personal data from a third-party source rather than from the person, you owe them information about it — including where you got it. A link in the email to a short, honest privacy notice is the practical answer.
• Objection and erasure. An objection to processing for direct marketing must be honoured. Suppress globally, immediately, across every campaign and mailbox.
• Automated decision-making and profiling (Article 22). Scoring and ranking prospects is profiling. Article 22's restrictions bite when a decision is solely automated and has legal or similarly significant effects, which sending an email generally does not — but if your pipeline makes automated decisions about people beyond who receives a message, look at it.
• ePrivacy national laws. Rules on unsolicited electronic marketing are set nationally and differ meaningfully across member states, including on whether B2B addresses need consent. Check the countries you send into.
• CAN-SPAM, if you send into the US: no deceptive headers or subject lines, a clear opt-out, a valid physical postal address, opt-outs honoured within 10 business days.
• Bulk-sender mechanics. At volume, List-Unsubscribe with an HTTPS URL and a mailto, plus List-Unsubscribe-Post: List-Unsubscribe=One-Click as specified in RFC 8058. Gmail and Yahoo expect one-click unsubscribe from bulk senders, and Google asks that reported spam stays below 0.3%.

Warm-up and the Act

Email warm-up is not an AI system in the Act's sense, and nothing about scheduling sends, measuring placement or tracking reputation falls into its risk categories. Where a warm-up or outreach product advertises "AI", the relevant question for you is a GDPR one — what personal data it processes, where, and under what agreement — not an AI Act one.

A short practical checklist

1. Keep a human reviewing and sending. Drafting assistance is a very different posture from an autonomous agent conversing with prospects.
2. Write down your legitimate-interest assessment, and make it reflect what you actually do.
3. Publish a short privacy notice and link it from your outreach, saying where data came from.
4. Suppress objections and unsubscribes globally and instantly.
5. Keep segments narrow. Indiscriminate volume is the fact that makes every legal argument harder, and it is also what damages your deliverability.
6. Record which AI tools you use for what, and give the people using them basic training. That is cheap and satisfies the literacy duty.
7. Do not let AI write claims you cannot substantiate. Consumer-protection and advertising law applies to your copy regardless of who drafted it.

What not to do

Do not assume the Act bans AI-written outreach. It does not, and vendors implying otherwise are selling something.

Do not assume the Act replaces GDPR. They operate alongside each other, and GDPR is the one that governs your prospect data.

Do not treat a model's output as fact-checked. Invented statistics and invented customer claims are your liability.

Do not use AI to scale a list you have no lawful basis for. Volume is the aggravating factor in every version of this.

Do not cite the Act's deadlines from a blog post, including this one. Check the current text.

Where MailPilot fits

MailPilot deals with the sending side rather than the copy side: it warms mailboxes over OAuth on Gmail and Outlook, or over SMTP/IMAP for SendGrid, Amazon SES, Mailgun, Postmark and custom SMTP, checks SPF, DKIM and DMARC on each sending domain, measures real inbox placement per provider through a seed-inbox network, and tracks per-mailbox reputation daily. Whoever or whatever drafts the email, those are the signals that decide whether it arrives.