The short version: probably much less than you fear, and one thing that already applies right now.
Regulation (EU) 2024/1689 - the AI Act - has produced a great deal of anxiety among people whose actual exposure is one sentence of disclosure. It is a risk-tiered regulation, and which tier you land in matters far more than the length of the text.
Prohibited practices - social scoring, certain biometric categorisation, manipulation exploiting vulnerabilities. Banned outright since February 2025. A sales agent does not go here unless you have designed something genuinely alarming.
High-risk - the Annex III list: employment and worker management, education, creditworthiness assessment, essential public services, law enforcement, migration, biometrics. This tier carries the heavy obligations - risk management systems, data governance, technical documentation, logging, human oversight, conformity assessment.
Transparency - systems that interact directly with people, or generate synthetic content. Light obligations, centred on people knowing what they are dealing with.
Minimal - everything else. No specific obligations.
An agent that answers product questions, recommends items, books a meeting and captures a lead is a transparency-tier system. It interacts with people directly, so Article 50 applies. It is not in Annex III, so the high-risk regime does not.
The obligation is genuinely modest: people must be informed they are interacting with an AI system, unless that is already obvious to a reasonably well-informed person in the circumstances. That is a label, done well.
This is not a future deadline. The transparency obligations became applicable on 2 August 2026. If you are reading this the week it was published, they are already in force.
Worth checking rather than assuming.
Creditworthiness. Annex III covers AI used to evaluate the creditworthiness of natural persons or establish their credit score. If your agent moves from "here are our finance options" to anything that scores or screens an individual for credit, the tier changes and the obligations get serious. Financing, instalments and lease qualification are the places this quietly happens.
Recruitment. Also Annex III. An agent doing customer sales is fine; the same agent pointed at your careers page to screen applicants is not the same system in regulatory terms.
If either applies to you, stop reading blog posts and talk to counsel.
Voice raises the stakes on disclosure, because a good synthetic voice is exactly the case where a person might reasonably not realise. The safe reading is that the disclosure should come early and audibly - in the agent's own opening, not buried in a policy page - and that a person who asks whether they are talking to a human should get a straight answer.
There is also a separate strand of Article 50 on marking synthetic audio and video content. Whether and how it bites on a live assistant that plainly announces itself is exactly the kind of question to put to a lawyer rather than settle from a vendor blog.
Our agents disclose that they are AI, in text and in voice. The disclosure wording is configurable under Appearance, because it has to sit in your brand voice and your language - but the mechanism is not something you can switch off, which we think is the right default for a shared obligation.
Beyond the label, conversations are logged so there is a record of what was said, and a human override exists for live conversations. Neither is required of a transparency-tier system. Both are cheap, both help if anyone ever asks, and one of them is simply good operational sense.
For most people deploying a sales agent, AI Act compliance is: say it is an AI, mean it, keep records, and check you have not wandered into credit scoring. The regulation is long. Your slice of it is short.