EU AI Act and sales calls: what applies now

This is an overview, not legal advice. Most of what has been written about the AI Act deals with high-risk systems, fine levels and deadlines that keep moving, and almost none of it matters to an ordinary sales team. What matters fits in one article, Article 50, and it applies now. In July 2026 the EU did push the heavy high-risk obligations back by more than a year, which has persuaded plenty of people that the whole regulation is on ice. It is not. The transparency rule did not move by a day.

What does Article 50 say, verbatim?

Article 50(1) places a design duty on the provider. AI systems meant to interact directly with natural persons have to be built so that those persons 'are informed that they are interacting with an AI system, unless this is obvious from the point of view of a natural person who is reasonably well-informed, observant and circumspect'. Article 50(5) sets the timing: the information must be given 'in a clear and distinguishable manner at the latest at the time of the first interaction or exposure'. Article 50(4) then reaches beyond the builder to whoever uses the system: if you deploy an AI that generates audio, images or video resembling real people, a deep fake, you must disclose that the content was artificially generated.

Three things here are worth dwelling on. 'Obvious' is judged from where the customer sits, not from where you sit: a well-made AI voice on the phone is by definition not obvious. 'At the latest at the first interaction' means, in practice, the opening of the call, not a terms document emailed afterwards. And 'deep fake', as defined in Article 3(60), includes audio that 'resembles existing persons', so a cloned version of your best rep's voice in outreach falls under Article 50(4), no matter how good the idea sounded in the sales meeting.

Who carries the responsibility: the provider or you?

On paper, the design duty in 50(1) belongs to the provider, meaning the company that develops the system. You, as the buyer and user, are the 'deployer' in the regulation's vocabulary, defined in Article 3(4) as whoever uses an AI system 'under its authority'. In practice, though, you are the one choosing the settings, writing the script the AI voice reads, and deciding whether the line 'hi, I'm an AI assistant from company X' stays in or gets cut because it 'hurt conversion'. Article 25 adds that you can become a provider yourself if you rebuild the system or put your own name on it. The sensible approach is therefore to treat the requirement as your own, confirm that the tool meets it, and keep a record of the check. Bear in mind that the practical exposure landing on the user is a reading of how the rules fit together, not settled case law. The regulation is new and has barely been tested.

Four situations in a sales team

1. An AI voice agent calls prospects. Article 50 applies in full here: the agent has to introduce itself as an AI at the start of the call. Swedish law then adds a layer. Section 19 of the Marketing Act (marknadsföringslagen) requires prior consent for 'automatic calling machines or other similar automatic systems for individual communication not operated by a person' when the recipient is a natural person. A fully autonomous calling AI is arguably exactly that, while a human working a dialer is not, but no authority has ruled on the question, so talk to a lawyer before the agent calls consumers or sole traders. Consumer calls also trigger the written-confirmation rule in the Distance Contracts Act, as they always have.
2. A chatbot on the website or in the portal. The same rule applies: the customer has to understand it is an AI, no later than the first message. The Commission's guidelines of 20 July 2026 use exactly this example: a chatbot that opens by saying it runs on AI technology, or a voice assistant that says so at the start of the session.
3. AI-written sales emails and proposals. The answer here is reassuring. The labelling duty in Article 50(4) covers AI-generated text published 'with the purpose of informing the public on matters of public interest'. A sales email that a human writes with help from an AI tool is nothing of the sort, and you do not need to add 'this email was written by AI'. The line is drawn at autonomous agents. The Commission's guidelines cite, as an example of correct disclosure, 'an email generated by an AI agent sent to a natural person that features an AI label at the top', so an agent that handles the email conversation with a customer on its own counts as a system interacting directly with people. The Marketing Act's ordinary ban on misleading marketing still applies regardless, and a cloned voice or a synthetic video message from a named rep is a different question altogether, see point one.
4. Internal AI training, such as roleplay against an AI customer. No real customer is involved, and the rep knows the other party is an AI, so the purpose of Article 50, that nobody is tricked into thinking they are talking to a human, is satisfied by the context itself. The provider still has to make the AI nature clear in the interface, and the Commission's guidelines describe the obviousness exception as a narrow one: an average user should be left with almost no doubt, which the guidelines take to be the case for internal assistants used by trained staff. One question worth putting to coaching-tool vendors: since 2 February 2025, Article 5(1)(f) has banned AI systems that infer employees' emotions in the workplace, with exceptions for medical and safety reasons. Judging how a message is delivered, its pace, clarity and energy, is not the same as classifying what a person feels, but the vendor should be able to show in its documentation where that line sits. Ask.

What was postponed in July 2026, and what was not?

The Digital Omnibus on AI, Regulation (EU) 2026/1744, appeared in the Official Journal on 24 July 2026 and took effect on 27 July. It pushes the high-risk obligations for Annex III systems to 2 December 2027 and those for Annex I systems to 2 August 2028. Article 50 is left exactly where it was. The Commission's own FAQ of 24 July 2026 could not be clearer: Article 50 applies from 2 August 2026. The one concession concerns the marking of AI-generated content under 50(2), where systems already on the market before 2 August get until 2 December 2026. The Commission's guidelines on the transparency obligations arrived on 20 July 2026 and set out, among other things, how 'obvious' should be read: the bar rises when the audience is children or vulnerable groups.

What happens if you ignore it?

Article 99(4) caps the fine for breaching Article 50 and several other provisions at 15 million euro or 3 percent of worldwide annual turnover, whichever is higher. Enforcement rests with national market surveillance authorities. On 12 June 2026 Sweden named five authorities for the AI Act: the Swedish Post and Telecom Authority, the Swedish Authority for Privacy Protection, Finansinspektionen, the Medical Products Agency and Swedac. Which one will end up examining an ordinary sales chatbot had not been decided as of September 2026. Layered on top of all this is ordinary consumer and marketing law. An AI voice that fails to say who is calling and why breaks the same rules as a human who hides it, and the Swedish Consumer Agency has brought several such cases against telemarketers in 2025 and 2026.

Five steps before next quarter

For an ordinary B2B team this is not a big compliance project. It is an inventory plus a handful of decisions, taken in the order of the steps below: map which tools actually talk to customers, confirm the introduction is still in place, rewrite the scripts, decide on automated dialing, and write it all down. One more duty applies. Since February 2025 Article 4 of the regulation has required anyone using AI systems to support their staff's AI literacy. The July 2026 omnibus rewrote that article and made clear that no particular level has to be guaranteed for any individual, but the obligation to do something stands, so an hour with the team is more than just good practice.

Train without touching a single customer

The logic is hard to argue with: the best way to get good at real customer calls is to stop practising on real customers. Salesprep is an internal training tool. The AI plays the customer, nobody real is on the other end, and the person practising is in no doubt that they are talking to an AI. Every call is scored on six or seven metrics from a library of ten, always with a written comment, and three calls are free when you create an account. When your AI agent later makes a real call, you will be the one who decided what it says in its first sentence, and these days that sentence starts with what it is.