Email is a covered channel. Article 50(1) applies to AI systems that interact directly with natural persons, and an AI agent that reads and answers customer email on its own is exactly that. The medium being email rather than a chat widget changes nothing.
The dividing line is autonomy. If your AI drafts replies that a human reads, edits, and sends, that's human agent-assist, and the Commission's draft guidelines place it outside the chatbot disclosure duty; the human sending the message is the interaction. If the AI sends replies itself, with nobody reviewing each one, the duty applies, and the disclosure has to reach the recipient in the email itself: a clear line identifying the reply as coming from an AI assistant, not a note buried in your terms or an email signature nobody parses as a disclosure.
A separate worry people bring to this question can usually be dismissed. The AI-text labelling duty in Article 50(4) covers text published to inform the public on matters of public interest. One-to-one correspondence isn't "published" (it doesn't reach an indeterminate public audience), so routine AI-written customer-service email doesn't trigger that duty regardless of review workflow.
Practical pattern for autonomous email agents: open with the identification line ("You're receiving an automated reply from our AI assistant..."), offer a human path ("reply 'agent' to reach a colleague"), and keep the template on file, dated, as evidence. For mixed pipelines where some mail is auto-sent and some human-sent, apply the notice to the automated templates rather than trying to badge every message.
If the honest description of your setup is "the AI answers, a human spot-checks occasionally," treat it as autonomous. Spot-checking a sample later is not a human reviewing and sending each message.
Not sure how this applies to your specific setup? The free 2-minute check tells you which of the four Article 50 duty areas likely apply to your company. No email required.
Take the free check Generate disclosure text