Do You Have to Disclose AI in Sales Emails? EU and US Rules

Usually no. Say AI helps draft a sales email, and a named person on your team checks it and sends it from their own mailbox. Then none of the three regimes covered here (the EU AI Act, California's bot chapter and the FTC's rules) makes you tell the prospect that a tool was involved.

A disclosure duty shows up in two situations:

• The prospect is interacting with an AI system, for example a bot answering their reply or a voice agent on a call.
• AI is used to mislead someone about who they are dealing with.

The EU AI Act put the question on the table. Its transparency rules sit in Article 50 of Regulation (EU) 2024/1689, and the European Commission's guidelines for providers and deployers confirm they apply from August 2, 2026. You do not need an EU office for this to matter: under Article 2, the Act also covers providers and deployers outside the EU when the output of their AI system is used in the EU.

The parts relevant to sales deal with direct interaction with AI, machine-readable marking by the companies that build AI tools, deepfakes and some published public-interest text. A private sales email that a person checked falls under none of them by default.

So one test settles most cases: is a person standing behind the message, or is the prospect talking to software?

In outbound, that test mostly bites after the prospect hits reply. Drafting happens before anyone answers. The rules switch on only when a machine writes back.

I cover disclosure only, under those three regimes. Other US states are outside that scope, so check any state you sell into separately.

Consent, opt-outs and spam rules are a topic of their own, and getting them wrong has led to regulator fines. Nothing here is legal advice.

What kind of AI use changes the answer?

AI enters outbound at three points, and the law treats each one differently.

AI drafts, a person sends. A rep asks a tool for a first draft or a subject line, rewrites whatever sounds off and sends it from their own mailbox. The prospect hears from exactly the person they think they hear from. That is ordinary writing help, in the same family as a spell checker.

AI writes, nobody reads. Picture a tool generating opening lines for 3,000 contacts and pushing them straight into a sequence. It is still your message in your name, so sending it triggers no label. The exposure is in what it says: unchecked AI happily invents a mutual connection, compliments a post the prospect never wrote or claims a result you never had.

IBM's guide to AI hallucinations defines this as a mistake where the system confidently generates false or fabricated information. IBM also notes that hallucinated output can "sound right" even when it is wrong or unverifiable. That is the problem with 3,000 openers nobody read. The fake ones look just like the real ones.

AI replies. Software reads the prospect's answer, writes back and books the meeting. At that point the prospect is in a conversation with an AI system, which is exactly what the rules below were written for.

The middle level is the risky one. A review counts only if someone reads for meaning and would sign the result with their own name.

Which parts of outbound to hand to software in the first place is a separate decision.

When does the EU AI Act require disclosure?

Article 50 contains several separate duties. Four of them touch sales. A fifth, on emotion recognition and biometric categorization, rarely does, and its one sales edge case is in the questions at the end. Only one of the four concerns conversations, and none of them puts a label on every AI-assisted email. The full Article 50 text takes five minutes to read.

In sales terms, a deepfake is a synthetic video, audio clip or image of a real person that looks or sounds authentic, and public-interest text is published content about public matters.

A required notice has to reach the person at the latest at their first interaction with the system or first exposure to the content.

The Commission's Article 50 FAQ describes one limited grace period. It covers only AI systems placed on the market before August 2, 2026, and only the machine-readable marking duty in Article 50(2); providers of those systems must comply from December 2, 2026. For a sales team the practical effect is small, because marking is the tool vendor's job either way.

What this means in day-to-day outbound:

• A reply bot talking to EU prospects. People must be told they are talking to AI unless that is obvious. The design duty sits with the tool's provider, so choose a tool that announces itself and never strip the notice out. An email thread signed by a named rep is built to look human, so "it was obvious" is a weak defense.
• An AI-drafted email. Machine-readable marking (a hidden tag in the file that software can detect) happens inside the tool, and assistive editing is carved out anyway. Nothing has to appear in front of the recipient.
• Public-interest text. The Commission's FAQ limits labeling to text that is both published and about public-interest topics. Human review or editorial control removes the duty, while a surface check such as spell-checking does not count. A one-to-one email is unpublished.
• An AI video of you. A synthetic clone of the founder in a video sent to prospects may count as a deepfake, and then you, as the deployer, must disclose it.

A legal commentary on Lexology reads ordinary commercial advertising and most marketing copy as generally outside the public-interest text rule. The same piece warns that AI text making claims about health, safety or sustainability could fall inside it.

Does California make you say a bot is a bot?

California's Business and Professions Code has had a chapter on bots since July 1, 2019, and the definitions in section 17940 set how far it reaches. They are narrow:

• A bot is an automated online account where software produces all or nearly all of the actions or posts.
• Online means appearing on a public-facing website, web application or digital application.

Taken together, the chapter targets automated accounts in public digital spaces. A rep's email to one prospect is no such account, so it falls outside those definitions.

Section 17941 of the same chapter narrows things further. Using a bot to talk to someone in California online is unlawful only in one case. The aim must be to mislead them about the bot's artificial identity, so as to knowingly deceive them about the content of the communication to drive a purchase or sale, or to sway a vote.

Whoever runs the bot avoids liability under that section by disclosing that it is a bot, clearly and conspicuously.

Two outbound setups come much closer: a chat assistant on your public website, and a LinkedIn profile where software sends substantially all of the messages. If either one talks to people in California, treat it as a bot question and have a lawyer look at it.

What does the FTC look for in AI outreach?

The FTC has no rule requiring a label on an AI-assisted sales email. What it pursues is deception about who is speaking.

In February 2024, the FTC finalized a rule against impersonating governments and businesses, and proposed extending it to individuals. The rule gives the agency a way to go after scammers who fake official or company email addresses and websites, or who use borrowed seals and logos to reach consumers by mail or online.

The same release warned that AI deepfakes threaten to "turbocharge" that kind of fraud. It also asked whether an AI platform that knowingly supplies tools used for impersonation should be liable.

The FTC already acts on AI deception. When it announced Operation AI Comply in September 2024, then-Chair Lina Khan said there is "no AI exemption from the laws on the books." One case targeted DoNotPay over claims about its AI chatbot service. The complaint alleged the company never tested whether the chatbot's output matched a human lawyer's.

The release is written about consumers, and you sell to businesses. The logic carries over to outbound anyway:

• Send as a real person only when that person wrote or approved the message.
• Use sending domains that belong to you, never ones dressed up as another company's.
• When a prospect asks whether they are talking to a human, a bot has to say it is a bot.

The first rule covers invented personas too: a fictional rep with an AI-generated headshot impersonates nobody real. The prospect is still misled about who they are dealing with.

Domains are where my team is most careful. We set up each client's sending domains and authenticate them with SPF, DKIM and DMARC, which tells receiving servers the mail really comes from the company that owns them. The setup itself is explained in our deliverability and infrastructure FAQ.

Should you disclose AI anyway when no law requires it?

For a reviewed email, I would leave the AI line out. Your prospects do not care which AI tool you pay for, just as they never cared which CRM or automation platform sits behind your outreach.

A footer reading "written with AI" answers a question nobody asked, and it moves the reader's attention from your offer to your tooling. The purchase decision comes after an audit of you anyway: your profile, your site, your proof, never your tool stack.

Short messages that sound like one person writing to another still beat polished templates that sound generated. If a reader already suspects a machine, a label will not fix that, and an edit will.

Even where no law forces it, act in these cases:

1. A prospect asks whether a person is on the other end. Tell the truth, every time.
2. A bot runs the conversation, even in a market no law reaches.
3. A draft contains a personal detail nobody checked, such as praise for a talk the prospect never gave. Delete the line instead of labeling it.

What does a compliant disclosure workflow look like?

Check every AI touchpoint against four gates, in this order:

1. Is it a drafted email? A named person approves the template and reads every generated personal line before each batch goes out. If it passes, no label is needed.
2. Is a bot answering in the thread? If software replies without a person, its first message says so. For example: "This is the AI assistant for our sales team. Reply 'person' and a colleague will take over."
3. Is it a public AI asset? Synthetic video or audio of a real person gets a visible label. So does published text on public-interest topics, unless a person holds editorial control over it.
4. Is it a live AI interaction? Voice agents and website chat announce that they are AI at the very start, before the first question.

Place the notice where the prospect first meets the AI: the opening line of the bot's first message, the first sentence of the call, beside the chat window. A footer under the signature is the spot nobody reads.

The simplest setup never reaches gates 2 and 4. In outbound, the first message is there to start a conversation, and a pitch in it gets you ignored. The reply is where a deal actually begins. Conversations come first. Calls come out of conversations. Clients come out of calls. That is also the exact stage where a bot answering would trigger the duty to disclose.

In our campaigns, senior people on my team read and answer the replies, so a prospect deals with a person from the first reply onward. If you are weighing an autonomous AI SDR against people for that job, compare the cost per meeting first.

Which outbound setups need an AI notice?

The table describes a standard B2B setup. The conditions sit in the sections above. When your setup straddles two rows, follow the stricter one.

The rule I would pin above every outbound desk: if a prospect could reply "am I talking to a real person?" and the honest answer is no, tell them before they need to ask.

Unsure where AI sits in your outbound setup? Book a call and walk me through it, and I will tell you which parts to raise with your lawyer.