How to answer insurance policy questions with an AI agent

Quote the policy; do not rule on it. The agent can explain what a customer's own schedule and wording say about limits, excesses, exclusions and endorsements, citing the section. It must not tell anyone whether a particular loss will be paid. "Am I covered for this?" about something that has happened is a claim, and goes to the claims team.

What is the policy query process, and who owns it?

Policyholders ask about their cover constantly. What is my excess? Does my home policy cover accidental damage? Is my partner insured to drive? What happens if I travel abroad for longer than a month? The answers sit in two documents: the schedule, which is specific to the customer, and the policy wording, which is shared across many customers but changes between versions.

The underwriting or product team owns the wording. Customer service owns the conversation. Claims owns every decision about whether a loss is paid. Mid-term changes to the policy belong to the policy administration team. The agent answers the first kind of question and stops short of the others, the boundary our guide to the AI support agent that knows when to say no draws for every support agent.

What are the steps, and who acts at each one?

1. Agent, depth 1: authenticates the customer and identifies the policy, including the wording version that applies to this policy term.
2. Agent, depth 1: reads the schedule and wording. Sums insured, limits, excesses, named drivers or insured persons, endorsements, and the applicable wording version.
3. Agent, depth 3: answers questions about what the policy says, quoting the section and noting any endorsement that changes it.
4. Agent, depth 1: recognises hypothetical cover questions. "Would I be covered if…" is answered by quoting the relevant sections and saying that cover is decided when a claim is assessed, on its facts.
5. Agent, depth 1: recognises loss questions. "My kitchen flooded last night, am I covered?" is a claim. The agent offers to start the claim and does not comment on cover.
6. Agent, depth 2: drafts mid-term change requests, such as adding a driver or changing an address, with any premium change taken from the rating system, for the policy administration team.
7. Person, policy administration: approves changes.
8. Person, claims handler: decides cover for any loss.

Which controls matter, and what evidence does each leave?

The right wording version. Customers renewed under one wording may have an older one for an earlier period. Evidence: wording version cited in each answer.

Quotes, not paraphrase, for limits and exclusions. Small changes in words change cover. Evidence: the quoted passage stored with the answer.

No cover decisions. "Yes, that's covered" about a real loss is outside scope, and certification tests for it in many phrasings. Evidence: scope-gate results and sampled conversations.

Premiums come from the rating system. The agent never estimates a premium change. Evidence: the rating reference on each change request.

Ambiguous wording goes to a person. Where the text could reasonably be read two ways, the agent says so and routes to the product team. Evidence: routing records and the product team's answer.

What should you measure?

• Questions answered from the policy, by topic. Repeat topics show where the wording is hard to understand.
• Loss questions converted into claims, and how quickly.
• Ambiguities routed to the product team, and whether the wording was later clarified.
• Complaints where the customer says they were told they were covered. Each one is reviewed against the filed conversation.
• Change requests drafted and approved.

Where does it usually go wrong?

Hypotheticals become rulings. A customer describes an event in the past tense as a hypothetical. The agent must notice and treat it as a claim.

Endorsements are missed. A policy may exclude a named item or add a higher excess. Answers that read only the general wording are wrong for that customer.

Documents are paraphrased into the customer's words. Customers ask "am I fully comprehensive?" or "is my phone covered outside the house?". The agent must map the question to the policy's own terms and quote them, not answer in the customer's vocabulary, which may not match how the policy defines things.

General product information is mixed with the customer's policy. The marketing page says "covers accidental damage" but this customer did not buy that option. The schedule wins.

Renewal changes are explained from the wrong year. Customers asking why cover changed at renewal need both versions compared, which is where handling policy renewal questions with an AI agent picks up.

Premium questions slide into billing. "Why did my direct debit go up?" is partly a policy change and partly a billing question. The billing part is covered by the same rules as can an AI agent answer questions about a customer's bill?

Where Pryme fits

The schedule, wording and endorsements reach the agent as sealed sources from the policy administration system through the Pryme Context Protocol, each with its version. Attest's scope gate is run on your own past conversations where customers described real losses, to prove the agent routes them rather than ruling on cover.