Disclosure, vulnerable callers, a route to a person, recording and payments: what the rules expect of an AI voice agent in a regulated firm, and a one-page sign-off sheet to complete before go-live.
When one of your people says the wrong thing on a call, you know what happens next. There's a script they should have followed, a QA form, a supervisor, a training record and a complaints process. Someone answers for it.
When a voice agent says the wrong thing, the caller hears your firm, not your vendor. So the question for 2027 is simple. Who signed off what it says, and what evidence do you keep?
PolyAI is clear about where that sits. Writing about financial services in July 2026, it said it has not sought direct regulatory approval itself: "The approval sits with the bank." That is the right answer, and it holds for every platform. The vendor supplies the agent. The firm owns what it says.
Here is what the rules expect, in summary. This is not legal advice, so check the detail with your own counsel.
Tell callers they are talking to AI
Article 50(1) of the EU AI Act has applied since 2 August 2026. People must be told they are interacting with an AI system, at the latest at the first interaction. It also reaches providers and deployers outside the EU when the AI's output is used in the EU. A UK or North American contact centre answering EU callers is in scope. Under Article 99 of the Act, breaches can cost up to €15m or 3% of worldwide annual turnover. The Commission's July 2026 guidelines take a narrow view of when AI is "obvious", and expect more disclosure where older or vulnerable users are likely (Ropes & Gray, August 2026).
In the US, Maine has since September 2025 barred using an AI chatbot, voice included, in a way that may mislead consumers into thinking they are dealing with a human. The exception is clear notice. Utah requires disclosure when a consumer asks, and at the outset of high-risk interactions in regulated occupations. California's AB 3030 covers healthcare providers using generative AI for clinical patient communications. They must give a disclaimer, spoken at the start and end of an audio interaction, and instructions for reaching a human.
The sign-off: who approved the exact opening words, for every region you take calls from?
Spot vulnerability, and act on it
The FCA has said it does not plan AI-specific regulation. It relies on the Consumer Duty and the Senior Managers and Certification Regime, so the obligations you already have still apply when an agent takes the call. Its vulnerable customer guidance, FG21/1, expects firms to spot signs of vulnerability, make it easy for customers to tell them about their needs, and respond flexibly.
It can be done in the most sensitive calls. PolyAI reports that NatWest's deployment handles inbound collections calls from customers in financial difficulty "under full FCA conduct and Consumer Duty obligations". No outcome figures have been published, but the point stands: the obligations do not relax because a machine answers the phone.
The risk is real, too. The Ministry of Justice's GOV.UK record for its own voice assistant lists "failure to detect safeguarding related statements" among the risks it manages.
The sign-off: who approved the list of vulnerability cues, and the route each one takes?
Let callers reach a person
Customers expect it. Gartner found 87% say it is "essential for companies to provide an option to reach a human agent when using genAI" (August 2026).
The law is moving the same way where decisions are involved. In the UK, the Data (Use and Access) Act changes in force since 5 February 2026 allow significant automated decisions only with safeguards. One is the right to obtain human intervention. Colorado's SB 26-189 takes effect on 1 January 2027. It covers automated decision-making that materially influences consequential decisions in areas including lending, insurance and healthcare. From that date, those decisions need notice before use, notice after an adverse outcome and, on request, human review and reconsideration to the extent commercially reasonable. Quebec's Law 25 requires firms to tell people when a decision is based exclusively on automated processing and to let them submit observations.
Most voice agents do not make decisions like these. They answer questions and complete simple tasks. If yours shapes a decision about credit, cover or care, these rules apply to it.
The sign-off: who approved the handover rules, and who receives each route?
Recording and payments
An AI agent records and transcribes by design, so the notice matters more. In the UK, lawful business recording needs "all reasonable efforts" to tell callers the call may be recorded (SI 2018/356). In the US, 13 states require all-party consent (Kilpatrick, July 2024).
Card payments need their own answer. Under PCI DSS v4.0.1, an agent that hears or transcribes card numbers brings its audio, transcripts and logs into PCI scope. Masking or payment links keep the numbers out.
The sign-off: who approved what is recorded, how long it is kept, and how card numbers stay out?
The one-page sign-off sheet
Put this on one page, with a name and a date against every line. Re-sign a line whenever it changes.
- Opening words. Disclosure, recording notice and how to reach a person, for each region. Signed by compliance. Evidence: the approved script version, and sampled calls showing it was said.
- Vulnerability cues and routes. Signed by your vulnerable customer lead or conduct risk. Evidence: the cue list, test calls and handover logs.
- Handover rules and owners. Signed by the operations lead. Evidence: routing tests in and out of hours, and proof the context arrived.
- The knowledge the agent answers from. Signed by the policy or product owner. Evidence: a versioned knowledge base, and answers checked against source documents, claim by claim.
- Data, recording and payments. Signed by the data protection officer and the PCI owner. Evidence: the data flow, retention settings and a masking test.
- Decisions the agent may make, if any. Signed by risk. Evidence: the notice wording and the human review route.
- Changes. Signed by a named approver for each release. Evidence: the change log and test results against reference calls.
- Monthly review. Signed by the business owner. Evidence: the scorecard against baseline, complaints, handovers and accuracy checks.
Line 7 is the one that slips. PolyAI says its improvement agent proposes changes on a branch and never publishes without human approval. Good. Someone in your firm still has to be that human, with their name on the change.
In Canada, model risk teams will ask a further question. OSFI Guideline E-23 applies to all federally regulated financial institutions from 1 May 2027. It counts AI methods as models, and expects an inventory, a risk rating per model and lifecycle governance. If you are a Canadian bank or insurer, expect your voice agent to be asked about.
Evidence, not assurance
None of this makes a firm compliant on its own. What it does is give your compliance team something to review, and your regulator something to read if they ask. A voice agent that can't show who approved what it said is a risk, however good it sounds.
At ETT, the sign-off sheet is part of a voice agent's design from the first week. For the disclosure rules in detail, see our EU AI Act Article 50 checklist, and for the records to keep, the AI evidence pack. If you are planning a voice agent for a regulated contact centre, start with our conversational AI and voice automation service.
Sources
- PolyAI blog, "Voice AI compliance in financial services", 28 July 2026
- EU AI Act Articles 2, 50 and 99; European Commission
- Ropes & Gray, August 2026
- Maine PL 2025 c.294; Utah SB 226 (Perkins Coie); California AB 3030 (ArentFox Schiff)
- FCA, 6 July 2026, and FG21/1
- GOV.UK Algorithmic Transparency Record, Ministry of Justice (9 September 2026)
- Gartner, 4 August 2026
- Data (Use and Access) Act 2025 (HSF Kramer); Colorado SB 26-189 (Finnegan); Quebec Law 25
- SI 2018/356; Kilpatrick, July 2024; PCI SSC; OSFI Guideline E-23
- PolyAI product documentation, docs.poly.ai (vendor description)