What changed for providers on that day.
MiCA
MiCA: an AI agent talking to a client is a regulated communication.
Last reviewed: 5 October 2026
MiCA sets a single EU regime for crypto-assets and the firms that offer services in them. It regulates the communication, not the channel. The moment an AI agent onboards a client, answers a suitability question, explains a product or prompts an action, that exchange carries the same fair-clear-and-not-misleading duty, the same risk-warning expectations and the same record-keeping consequences as one held by a person.
What the law requires
- Providing crypto-asset services in the Union requires authorisation as a crypto-asset service provider.
- Regulation (EU) 2023/1114, Art. 59
- A provider must act honestly, fairly and professionally in the best interests of its clients, and all information addressed to them (marketing communications included) must be fair, clear and not misleading and identifiable as marketing.
- Regulation (EU) 2023/1114, Art. 66(1) and 66(3)
- Client crypto-assets and funds must be held safely and kept separate from the provider's own.
- Regulation (EU) 2023/1114, Art. 70
- A complaints-handling procedure must be in place, published, and complaints must be recorded together with the measures taken.
- Regulation (EU) 2023/1114, Art. 71
- An offer of a crypto-asset other than an asset-referenced or e-money token requires a published white paper.
- Regulation (EU) 2023/1114, Art. 6
- The asset-referenced and e-money token titles have applied since 30 June 2024; the rest, including provider authorisation, since 30 December 2024.
- Regulation (EU) 2023/1114, Art. 149
- Member States set the administrative penalties. For crypto-asset service providers the Regulation fixes the minimum maximum for legal persons at at least 5% of total annual turnover.
- Regulation (EU) 2023/1114, Art. 111(5)
What Audact does about it
- The approved risk-warning wording is held to, in the client's language; each sentence segment the model produces is checked before it is sent to speech, and a segment already spoken is not withdrawn.
- In controlled testing
- Every warning, consent and policy verdict leaves a receipt carrying its position in the conversation: which warning, at which second, before or after which question.
- Eligibility resolved before the line connects: jurisdiction, retail or professional classification, and whether the asset class may be discussed at all.
- In development
- One export shaped for a supervisory request, covering the interactions, the consents and the verdicts behind them.
- In development
Channel
The duty attaches to what was said, not to who said it.
MiCA never mentions voice agents, and it does not need to. Article 66 is written around the information addressed to clients; the medium is not a qualifier. An agent that describes a product, answers a risk question or nudges towards a trade is producing exactly the material the article governs.
That has a practical edge that human teams already know: the compliance question is rarely what the script says, it is what was actually said on the day. A per-conversation record carries a hash of the output the gate released, the decision and the recorded policy version, and it is the same record the complaints procedure of Article 71 needs when a client disputes what they were told.
Marketing
Identifiable as marketing, in the moment.
Article 66(3) asks for two things at once: the information must be fair, clear and not misleading, and marketing communications must be identifiable as such. In a written channel that is a label. In a conversation it is a sentence, spoken at a point in the exchange where it still frames what follows.
Placing it is a policy decision, not a scripting one: it varies with the jurisdiction and with what the client has already been told. That is the same mechanism the Article 50 disclosure runs on, which is why both end up in one policy pack rather than two.
Note
Information, not legal advice.
This page sets out the rules as we read them, with the article numbers so you can check us. It is not legal advice, and your own position depends on facts this page cannot know. Take advice from qualified counsel before you rely on it.
Questions
- Does MiCA cover AI voice and chat?
- Yes, by covering the communication. The duties in Article 66 attach to information addressed to clients regardless of the channel it travels through, so AI-mediated onboarding, support and advice raise the same suitability, risk-warning and record questions as a human conversation.
- Since when does MiCA apply?
- The asset-referenced and e-money token titles have applied since 30 June 2024, and the rest of the regime (including authorisation of crypto-asset service providers) since 30 December 2024, under Article 149.
- What are the penalties?
- Member States set them. Article 111 obliges them to provide administrative penalties and sets minimum maxima; for crypto-asset service providers that minimum maximum for legal persons is at least 5% of total annual turnover. The figure that actually applies to you is the one in your national implementing law.
- Do we need MiCA and the AI Act at the same time?
- Yes. MiCA governs what may be said and what must be recorded; Article 50 of the AI Act governs telling the person they are speaking to an AI. They apply to the same conversation and neither covers the other.
Approved wording, held to in the moment
Control shows how a policy pack decides what may be said, in which jurisdiction, and what happens to wording that falls outside it.
Related
The 30 December 2024 date DORA The operational resilience side of the same firm.
EU AI Act Article 50 The disclosure duty on the same call.
Financial services The flows this shows up in.
All frameworks
