AI Outbound Compliance Handbook: email, WhatsApp, LinkedIn and voice, one page
Last verified: 2026-09-11
An AI sales agent is a machine that writes to strangers on your behalf, at volume, in your name. Every rule that applied to a human rep still applies, plus one new one about being honest that it is a machine. This page is all of them in one place, in the order a message meets them: may I contact this person, on this channel, at this hour, saying what, and what happens when they say stop.
We build BOSRAI, an AI sales rep, so read what follows knowing that. Where the product does something about a rule, it says so in a box. Where we could not verify a position to our own standard, it says that too. We are not lawyers and this is not legal advice.
Summary
- Being an AI does not change whether you may send. GDPR, ePrivacy, CAN-SPAM, PECR and CASL decide that, exactly as they did for a person with a mail-merge.
- Being an AI adds one duty: say so. In the EU that is Article 50 of the AI Act, applicable since 2 August 2026, and the duty sits with the provider of the agent. For voice in the US, the FCC put AI-generated voices inside the TCPA in February 2024.
- The channel rules are stricter than the law in two places. Meta's WhatsApp policy and LinkedIn's terms will end your access faster than any regulator will fine you.
- Suppression is the part that gets audited. Unsubscribe, STOP and do-not-contact have to work in seconds, across every channel, forever.
1. May I contact this person at all
EU: GDPR and the ePrivacy Directive
Two questions, in this order. First, do you have a lawful basis to hold and use the person's data (GDPR, Article 6). For B2B prospecting the usual answer is legitimate interest (Article 6(1)(f)), which needs a documented balancing test: your interest in reaching them, against their reasonable expectation of being contacted at work about their work. Second, does the ePrivacy Directive (2002/58/EC, Article 13) allow the channel. Article 13 requires prior consent for electronic direct marketing to natural persons, with a "soft opt-in" for existing customers, and leaves it to each Member State how far that protection extends to business addresses.
That last clause is why the answer differs by country:
- Germany treats email advertising to businesses as needing consent under the Act Against Unfair Competition (UWG §7). Cold B2B email is the exception, not the rule, and the regulators have said so repeatedly.
- France allows B2B prospecting to a professional address on legitimate interest, if the message is about the person's professional function and there is a clear opt-out (CNIL guidance on prospection B2B).
- Netherlands, Italy, Spain, Nordics: broadly legitimate-interest territory for role-relevant B2B email with an opt-out, with Italy's Garante the most active enforcer in the group.
Whichever basis you use, Article 14 of GDPR applies whenever the data did not come from the person: you owe them a privacy notice at the latest at the first communication, saying who you are, where the data came from and how to object. A link to a privacy page in the footer of the first email is how most compliant programmes meet it.
Open-tracking pixels. The French and Italian regulators have both said that a pixel that reports when an email is opened is tracking that needs consent, the same as a cookie. A cold email cannot have that consent.
What BOSRAI does. Every contact carries its source and the reason it matched your audience, so the balancing test has a record behind it. The first email carries your legal name, postal address, an unsubscribe link and a privacy link. Contacts we can place in France or Italy receive plain-text email with no HTML part, so there is no pixel and no rewritten link.
United Kingdom: PECR and UK GDPR
The Privacy and Electronic Communications Regulations distinguish corporate subscribers (a limited company, an LLP) from individual subscribers (sole traders, partnerships in most of the UK, and anyone on a personal address). You may email a corporate subscriber's employees about their work without consent, with an opt-out; individual subscribers need consent. UK GDPR still needs a lawful basis and an Article 14 notice. The ICO's direct marketing guidance is the reference. What the rules require when the sender is an AI rather than a person is a question we have with counsel and will publish when it clears.
United States: CAN-SPAM
CAN-SPAM does not require consent for B2B email. It requires: an accurate From, no deceptive subject line, identification as an advertisement where that applies, a valid physical postal address, a working opt-out that is honoured within ten business days, and no further sending after an opt-out. Since February 2024, Gmail and Yahoo also require senders over 5,000 messages a day to carry one-click List-Unsubscribe headers and to keep spam complaints under 0.3%. Those are not law, but they decide whether anything gets delivered.
What BOSRAI does. Legal name, postal address and one-click unsubscribe on every email, plus the List-Unsubscribe headers regardless of volume. An unsubscribe marks the person do-not-contact within seconds, stops every sequence they are in and deletes their waiting drafts.
Canada: CASL
The strictest of the four. Commercial electronic messages need express or implied consent before sending. Implied consent for B2B exists where the address was "conspicuously published" without a statement refusing unsolicited messages, and the message relates to the person's role. Every message must identify the sender, give contact details and carry an unsubscribe that works for at least 60 days. Penalties reach CAD 10 million per violation for organisations.
2. The new duty: say it is an AI
EU AI Act, Article 50
Applicable since 2 August 2026. The transparency duty in Article 50(1) is on the provider of an AI system that interacts directly with people: the system must be designed so people know they are dealing with an AI, and the Commission's guidelines (C(2026) 5054 final) add that an agent must disclose both its artificial nature and the party on whose behalf it acts. The guidelines name email and "managing correspondence" as in scope, and say that a person representing a company is still a natural person. A human who genuinely reviews and sends a draft is treated as the sender; an approval queue nobody opens is not.
What Article 50 does not do: it does not require an AI footer on every cold email (that duty covers text published to inform the public), and it does not make unlawful outreach lawful. Fines reach €15 million or 3% of worldwide turnover, whichever is higher, and for SMEs whichever is lower. Our full reading of the guidelines, with paragraph numbers, is in EU AI Act Article 50 and AI sales agents.
What BOSRAI does. We are the provider, so the duty is ours. Any message the rep writes and sends with no person approving it ends with one line: "Written and sent by [company]'s AI sales assistant. A person at [company] reads every reply." Both required facts are in it. On WhatsApp and LinkedIn the line appears in brackets; the phone rep says it in its opening sentence. Drafts a person approves in the Inbox are that person's message and carry no line. The line is on by default and a customer who turns it off takes the responsibility with it.
United States: voice
On 8 February 2024 the FCC declared that AI-generated voices are "artificial or prerecorded" voices under the Telephone Consumer Protection Act, so a call made with one needs the prior express consent the TCPA requires, and for marketing calls to mobile numbers prior express written consent. Whether that reaches B2B calls to a business mobile is the point we could not verify to our own standard. The safe reading is that an AI should only call people who asked for the call.
What BOSRAI does. The phone rep calls only people who requested a call from the booking page, opens by saying it is your company's AI assistant, and ends the call the moment someone says it is a bad time.
3. The channel rules, which bite first
Two layers. Meta's policy: on the WhatsApp Business Platform a business may only start a conversation with a person who has opted in, using an approved template; outside a 24-hour customer-service window everything is a template. Numbers are banned on volume and on reports, and a ban on a number a business has used for years is the real cost. Regional law: the UAE's telecom regulator treats unsolicited marketing messages as spam and the Personal Data Protection Law (Federal Decree-Law 45/2021) is consent-first; Saudi Arabia's PDPL takes the same line. In practice, in the markets where WhatsApp is the business channel, cold WhatsApp is the fastest way to lose the number.
What BOSRAI does. A sequence cannot open on WhatsApp; the editors and the API refuse it. A WhatsApp step goes only to someone who has already engaged: replied on any channel, accepted your LinkedIn invitation, clicked, visited, booked a call, or gave you their number. Every message names the sender in the first line, ends with "Reply STOP and I will not message again", and is sent between 07:00 and 21:00 where the person is. A STOP is honoured without a model in the loop. A new number starts on a small daily cap that grows with use.
LinkedIn's User Agreement prohibits automation that accesses the service other than through the interfaces LinkedIn provides, and the platform restricts accounts that send invitations or messages at a rate a person would not. There is no law here, only a terms-of-service risk that ends in a restricted profile. Working limits in 2026 are around 100 invitations a week per account, lower for new or low-acceptance accounts.
What BOSRAI does. Messages go from your own account, paced across working hours, with a hard cap of 40 invitations a day and lower caps while an account is new. When acceptance falls below the line LinkedIn watches, the rep rests the account. Invitations are blank by default.
4. Timing
No law sets sending hours for email. For calls, the TCPA's window is 8 a.m. to 9 p.m. in the recipient's time zone, and several US states are narrower. For WhatsApp there is no statutory window, but a message at 02:00 is the kind that gets reported.
What BOSRAI does. Email is spread across the recipient's working day where we can tell it. WhatsApp keeps to 07:00 to 21:00 local time; a draft approved at night says "Sends at 07:00 their time".
5. Suppression: the part that gets audited
Every regime above ends the same way: when someone says stop, it has to stop, on every channel, and stay stopped. The failure that reaches regulators is rarely the first message. It is the fourth message to someone who said no after the second.
What BOSRAI does. Unsubscribe links, STOP replies in any wording, and "not interested" replies the rep classifies all mark the person do-not-contact, end every sequence they are in and delete their waiting drafts. A do-not-contact mark survives later hunts and imports.
6. Deliverability is compliance by another name
A burned domain is the fine nobody sends you. Bounce rates above 2 to 3% and complaint rates above 0.3% get a domain throttled or blocked, and a domain that cannot deliver cannot comply with anything.
What BOSRAI does. Sends are capped per mailbox and spread through the day. When bounces cross the line that hurts reputation the rep pauses and resumes itself when the rate recovers. Catch-all addresses are parked rather than sprayed. Emails go out as plain text with a plain HTML twin, the way a person's mail looks.
7. The checklist to run against any vendor
- Where in the product is the Article 50 disclosure, and does it name the party the agent acts for? Ask to see a sent message.
- What happens in the first sixty seconds after an unsubscribe, on each channel? Ask for the sequence of events, not the policy.
- Can the product open a conversation on WhatsApp cold? If yes, whose number gets banned?
- What does the first email carry: legal name, postal address, opt-out, privacy notice?
- Does the vendor track opens in France and Italy?
- What are the LinkedIn caps, and what happens when acceptance drops?
- Does the voice agent say it is an AI in its first sentence, in words?
- Where is the record of where each contact came from and why they were selected?
The answers to those eight questions for BOSRAI are on the Compliance page.
What we could not verify
The UK position on AI-sent B2B outreach under PECR; whether the FCC's AI-voice ruling reaches B2B calls to business mobiles; and the German regulators' view of an AI-drafted, human-approved email under UWG §7. Each is stated above as an open question rather than filled in. We will update this page when any of them clears.
Primary sources. Regulation (EU) 2024/1689, Article 50 · Commission guidelines on Article 50, C(2026) 5054 final · GDPR Articles 6 and 14 · ePrivacy Directive 2002/58/EC, Article 13 · ICO guide to PECR · FTC CAN-SPAM compliance guide · CASL, Government of Canada · FCC declaratory ruling on AI-generated voices, 8 February 2024 · WhatsApp Business Messaging Policy · LinkedIn User Agreement · Gmail sender guidelines.
Last verified: 11 September 2026.