AI Sales Agent Disclosure: What the EU AI Act Actually Requires

Yes. If your AI sales agent talks to anyone in the EU, you are legally required to tell them they are talking to an AI. That obligation has been enforceable since 2 August 2026 under Article 50 of the EU AI Act. If the agent runs on WhatsApp, Meta's messaging policy adds a second layer: the customer must always be able to reach a human. Neither requirement is technically difficult, but both need to be built into the product from the start, not added after a compliance review flags them.

What Article 50 of the EU AI Act actually says

The relevant regulation is EU 2024/1689, and the transparency obligations are unusually readable for EU legislation. For a sales or support agent, three clauses matter:

• Article 50(1): AI systems designed to interact directly with people must inform those people they are talking to an AI, unless it is obvious to a reasonably well-informed, observant person.
• Article 50(2): Systems that generate synthetic text, audio, images or video must mark that output as artificially generated in a machine-readable format, where technically feasible.
• Article 50(4): AI-generated text published on matters of public interest must be disclosed as such, unless a human editorially reviewed it and takes responsibility.

For a sales agent, 50(1) is the clause that applies. The European Commission's AI Act overview confirms the phased timeline: prohibited practices from February 2025, general-purpose model obligations from August 2025, and the remaining provisions including Article 50 from 2 August 2026. The penalty ceiling under Article 99 is the higher of €15 million or 3 percent of worldwide annual turnover.

Two things founders consistently get wrong:

Scope follows the user, not the vendor. The Act applies wherever the user is located. A clinic in Dubai running an Arabic-language WhatsApp agent is in scope the moment a patient messages from Paris. The test is where the customer sits, not where you are incorporated.

"Obvious" is a weak defence. A chat window that opens with "Hi, how can I help?" in a warm tone, responds in under two seconds, and never mentions AI is not obviously a machine to an ordinary user. The regulator applies the customer's perception, not yours.

WhatsApp's own policy, and why it matters more immediately

If your agent runs on the WhatsApp Business Platform, the WhatsApp Business Messaging Policy requires you to disclose that the person is interacting with an automated experience, and to provide a clear way to request a human agent. Meta can throttle or shut a business number for violations. Appeals move at Meta's pace. A regulator might take a year to notice a non-compliant deployment; a number ban lands on a Tuesday afternoon and takes your revenue channel offline immediately.

This is why the escalation path is not optional scope when you are pricing a deployment. It should already appear in the estimate for what an AI WhatsApp sales agent costs. If it does not, ask why.

The geographic picture beyond the EU

Other markets are converging on the same disclosure pattern. California has required since 2019 that a bot used to sell goods or services identify itself as a bot. Gulf regulators publish AI ethics principles that name transparency toward end users, and public-sector and healthcare tenders in Saudi Arabia and the UAE increasingly ask for it in writing, a pattern covered in more detail in AI data residency considerations for Saudi Arabia and UAE. Return to the Dubai clinic: one patient messaging from Paris triggers Article 50, a local hospital tender asks for the same disclosure in the contract, and a customer with a California number is covered by the 2019 state law. Three rulebooks, one first message. Building to the EU standard once is cheaper than tracking each jurisdiction separately.

How to disclose without damaging conversion

The concern is legitimate. You built a sales agent because it converts, and now you are opening every conversation with a notice. Here is what actually works.

One line, at the start, in plain language

Article 50(1) asks for a clear notice before or at the first interaction. That does not require a paragraph of legal text. One short message is enough:

``text Hi, I'm Layla, the AI assistant for Northwind Clinic. I can book appointments and answer questions. Type "human" any time to reach our team. ``

That single message satisfies the AI Act and WhatsApp. It also sets expectations usefully: when the agent asks a qualifying question, the customer understands the context and does not feel interrogated.

A name is fine. A fake human identity is not.

Giving the agent a persona is entirely reasonable. Pairing that persona with a profile photo of a real-looking person and no disclosure is exactly the pattern the law was written to stop. If you want a named character, pair it with the words "AI assistant" in the first message and in the WhatsApp business profile description.

Make handoffs explicit in both directions

When the conversation moves to a person, say so: "Passing you to Omar from our team." When a human hands back to the agent after hours, say that too. Customers tolerate automation. What they do not tolerate is uncertainty about who they are dealing with. The mechanics of this are covered in AI agent escalation paths.

Machine-readable markers for generated files

If your agent sends generated quotes, reports or images, Article 50(2) expects a machine-readable marker such as content metadata or a provenance credential. Most teams can handle this by routing generated PDFs and images through a templating pipeline that stamps metadata at creation. Ask any vendor you are evaluating how they handle it. If they do not know, that is a signal worth acting on. See the AI vendor evaluation checklist for a fuller set of questions.

What we found in our own outreach

We run our own outbound engine. It pulls facts from a prospect's public site, uses a smaller local model to extract specific signals, and drafts one tailored message per prospect in a single call. The relevant boundary here is where the automation stops: a person reviews and sends, and a person handles every reply.

That boundary turned out to be commercially useful, not just compliant. In testing, asking for a simple "reply YES" consistently outperformed linking to a calendar. The reason is friction reduction, but the disclosure benefit runs alongside: the moment someone replies, they are talking to a human, and we say so honestly. We wrote about the CTA finding in AI sales agent CTA design. If your funnel has a natural point where a human takes over, put the handoff there and be explicit about it. You get better reply rates and a cleaner compliance position at the same time.

Checklist before going live

• First message discloses AI in the user's language, and it renders correctly in the WhatsApp preview.
• A human escape hatch exists, tested on a real device, and it works outside business hours (even if the reply is "a person will respond at 9am").
• WhatsApp business profile description mentions automated assistance.
• Handoff messages are explicit in both directions, AI to human and human back to AI.
• Generated files carry provenance metadata, or you have a documented technical reason why it is not feasible.
• Logs record which turns were AI and which were human, so you can demonstrate compliance if asked.
• Your contract with the vendor names who is the "provider" and who is the "deployer" under the Act, because the obligations differ between them.

For a broader view of what to verify before signing with any AI vendor, the AI agent vendor evaluation checklist is worth going through before you commit.

The real risk is not the fine

The €15 million ceiling in Article 99 is the number everyone quotes. It is not the first thing that will hurt you. An agent that pretends to be a person, gets caught, and gets screenshotted is a brand problem before it is a legal one. A WhatsApp number ban lands months before any regulator sends a letter. Everything Article 50 asks for fits into the Layla message above: who the customer is dealing with, what it can do, and the one word to type to reach a person. Build the disclosure, the escape hatch and the AI-versus-human logs into the first sprint and you will not think about Article 50 again.

If you are scoping a WhatsApp or web sales agent for a market that includes the EU or the Gulf and want disclosure, escalation and logging handled properly from day one, let's talk.