Skip to main content
Audact
Explore a workflow

Sector · Insurance

Insurance: let AI take the call, not give the advice.

By application

Last reviewed: 25 October 2026

On an insurance line one sentence quickly becomes a promise: am I covered for this, which policy should I take, what will it cost, will my claim be paid? A voice agent that answers freely is stating cover, giving advice and getting ahead of the claims handler, and under the Insurance Distribution Directive a personal recommendation is advice, with duties of its own and, nationally, a licence. Audact puts your business rules between the AI and the caller: supported replies are checked before they are spoken, product information stays within what you approved, and cover, recommendations and claim outcomes go to the people authorised to give them. Each decision leaves a tamper-evident record. Insurance is by application: each deployment is scoped before it goes live, because the line between information and advice is where the licence starts.

What the law requires

A person interacting with an AI system must be told so, unless it is obvious from the circumstances.
Regulation (EU) 2024/1689, Art. 50(1)
Insurance distribution covers advising on, proposing and preparing insurance contracts, and assisting with them, in particular in the event of a claim. Advice means a personal recommendation to a customer.
Directive (EU) 2016/97, Art. 2(1), points (1) and (15)
Before a contract is concluded, the distributor establishes the customer's demands and needs. Any contract proposed must be consistent with them, and advice comes with a personalised recommendation explaining why the product fits.
Directive (EU) 2016/97, Art. 20(1)
In Germany, brokering insurance as a business needs a permit from the chamber of industry and commerce, and the intermediary asks about the customer's wishes and needs, advises, gives the reasons for each recommendation and documents it.
§ 34d(1) GewO and § 61(1) VVG (DE)
In the Netherlands, advising on insurance needs a licence, and before advising the adviser gathers information on the customer's financial position, knowledge, experience, objectives and risk appetite, as far as relevant.
Wet op het financieel toezicht, art. 2:75 and art. 4:23 (NL)
AI systems intended for risk assessment and pricing in relation to natural persons in life and health insurance are high-risk.
Regulation (EU) 2024/1689, Annex III, point 5(c)
A person has the right not to be subject to a decision based solely on automated processing that produces legal or similarly significant effects. Where such a decision is allowed, the controller must offer human intervention, the right to express a view and the right to contest it.
Regulation (EU) 2016/679, Art. 22(1) and 22(3)

What Audact does about it

Scope limits checked while the call runs: a supported reply that states what is covered, or which product suits the caller, is held back, and the question goes to an authorised person, patent-pending (UK IPO).
In controlled testing
Product information within what you approved: features, documents, payment dates and premiums are checked against your approved set before they are spoken. A supported reply outside it is replaced by a fallback reply before it is spoken.
In controlled testing
The check covers supported replies and runs within an evaluation window. If it does not answer in time, the original reply is released unchanged and a FAIL_OPEN event records the miss.
In controlled testing
First notice of loss taken down (what happened, when, where, who is involved, the policy number) and passed to a named claims handler, with no statement on whether the claim is accepted or paid.
In controlled testing
A caller in distress after an accident or a bereavement passed to a person rather than kept in the flow.
In development
The Article 50 disclosure spoken at the start of the call, in the caller's language.
In controlled testing
A tamper-evident receipt per governed interaction: whether the reply was allowed, modified or blocked, with the input and the output as hashes. It carries an authentication code, not a signature.
Working today
Call-backs to web leads placed only to people who asked for one, within the contact rules of their country.
In controlled testing
Each insurance deployment scoped before it goes live: which questions the agent may answer, which go to a person, and which data it may take.
By application
The same rules for AI chat on your website and in your customer portal.
In controlled testing
AI-written email checked for what it commits to, such as a premium, a date or a promise, before it is sent.
In controlled testing
Custom function and tool calling during the call: which of your systems the agent reaches, such as your policy administration, is scoped per deployment.
In controlled testing

Where it binds you

AI can bind your business.

Five things an insurance line hears every week. Each has a sentence that sounds helpful and commits you to something, and a reply your rules can let through instead.

  • The caller asks
    Am I covered if my basement floods?
    The sentence that binds you
    Yes, water damage is covered on your policy.
    What your rules let through
    No statement on cover. The question goes to an authorised person, with the policy number taken down.
  • The caller asks
    Which policy should I take?
    The sentence that binds you
    The comprehensive plan is the best fit for you.
    What your rules let through
    Information on the products you approved. A recommendation comes from a licensed adviser.
  • The caller asks
    What would it cost me?
    The sentence that binds you
    About thirty euros a month, I'd say.
    What your rules let through
    A premium from a quote you approved, or a call-back with one.
  • The caller asks
    Will my claim be paid?
    The sentence that binds you
    Don't worry, that will be paid out.
    What your rules let through
    The claim is taken down for a claims handler. The decision stays with them.
  • The caller asks
    Can you cancel my policy today?
    The sentence that binds you
    Done, it's cancelled.
    What your rules let through
    The request is taken down and passed on. The agent does not say it is done.

Illustrative examples, not recordings. You set the rules; Audact checks the supported reply against them before it is spoken.

Where it starts

Five calls to start with.

  • Service calls. Policy documents, payment dates, how to report a claim and opening hours, answered within what you approved.
  • After hours. A first notice of loss at night or at the weekend taken down for a claims handler in the morning, with urgent cases routed the way you set.
  • Web-lead call-back. Someone who asked for a quote is called back within the contact rules of their country and passed to an adviser. The agent takes the details; the adviser gives the advice.
  • First notice of loss. What happened, when, where, who is involved and the policy number, passed to a named claims handler.
  • Broker offices. The calls an intermediary misses while advising, answered and routed to the right adviser.

The line

Information is allowed. A recommendation is advice.

Telling a caller what a policy document says, when the premium is due or how to report a claim is information. Telling them which policy suits them is a personal recommendation, and under the IDD that is advice: it needs the customer's demands and needs established first, a recommendation with reasons, and people with the knowledge and ability to give it. In Germany and the Netherlands it also needs a licence.

A general-purpose assistant crosses that line within a sentence, because recommending is what it was trained to do. In Audact the line is a rule checked while the call runs: when a supported reply would make a recommendation or a statement on cover, it is held back, and the agent moves to what it may do: give the approved information, take the details, pass the caller to an authorised person. The refusal is recorded on the receipt for that call.

Claims

A claim is a decision. The agent takes the notice; a person decides.

First notice of loss is where a voice agent helps most, and where it can do the most harm. The caller wants to hear that it will be fine, and a reassuring sentence about cover or payment is exactly what a claims handler may later have to take back. The agent takes down what happened, when and where, and passes it to a named handler. Whether the claim is accepted stays with them.

Two rules point the same way. A decision on a claim based solely on automated processing, with legal or similarly significant effects, is restricted by Article 22 GDPR. And an injury or an illness brings health data, a special category under Article 9 that pulls in the impact assessment of Article 35; that assessment is part of scoping a claims line, not paperwork afterwards.

The record

A record of each decision, not a transcript.

Every check that answers in time leaves a receipt: whether the reply was allowed, modified or blocked, with the input and the output stored as hashes rather than words.

The receipt is tamper-evident. It shows that the record is intact: a later change or reordering is detectable by recomputation against a retained commitment. It carries an HMAC-SHA256 code under a symmetric key held by Audact, so whoever can check it can also produce it: an authentication code, not a signature. It does not establish who issued it, and whether the underlying event is true. Three of its values can be recomputed without a key: the input hash, the output hash and the hash of the preceding evidence event.

Whether a claim was then registered in your system is for that system to confirm, not for the receipt. You can recompute a sample receipt in your own browser.

The law, in general terms

Article 50 already applies. The distribution rules bite harder.

Article 50 of the AI Act has applied since 2 August 2026: the caller has to be told they are speaking to an AI system. It is a transparency duty, not a ban. Nothing about running a voice agent on an insurance line became unlawful that day.

The exposure sits in what the agent says, and the distribution rules decide it: honest, fair and professional conduct in the customer's best interests, people with the appropriate knowledge and ability, and nationally the licence and the duty to ask, advise, give reasons and document. Pricing is a separate line: AI used for risk assessment and pricing of people in life and health insurance is high-risk under Annex III of the AI Act. An agent that answers calls and passes them on is not doing that; one that prices a policy would be.

General information, not legal advice.

Questions

Can the agent tell a caller whether something is covered?
Not under the rules on this page. A supported reply that states cover is held back at run time, and the question goes to an authorised person, with the policy number and the question taken down. The agent can give the information you approved, such as where to find the policy document or how to report a claim.
Can the agent recommend a product?
Not under the rules on this page. A personal recommendation is advice under the IDD, with its own duties and, in Germany and the Netherlands, a licence. A supported reply that recommends a product is held back; the agent informs within what you approved and passes the request to a licensed adviser. The refusal is recorded on the receipt for that call.
What does the agent do with a claim?
It takes the first notice: what happened, when, where, who is involved and the policy number, and passes it to a named claims handler. A supported reply that says whether the claim is accepted or paid is held back. That decision stays with a person, which is also where Article 22 GDPR points.
Why is insurance by application?
Because the line between information and advice is the one that carries the licence, and it runs differently per product and per country. Each insurance deployment is scoped before it goes live: which questions the agent may answer, which go to a person, and which data it may take. For a claims line that handles health data, the impact assessment of Article 35 GDPR comes first.
Does Audact take care of the IDD for us?
No. Those duties sit with the distributor, and software cannot hold a licence. What Audact does is narrower and checkable: it tests each supported reply against the rules you set, before the caller hears it, and records the decision. This page is general information, not legal advice.

Let AI take the call. Keep control of what it says.

See how Governed Voice AI checks a supported reply against your rules before it is spoken, and what each call leaves behind.

See Governed Voice AI

Related