It depends on what "set" means. An AI agent can calculate a quote by running the approved rating model on a customer's details, at depth 3, much as rating engines already do. It should not decide the rates, change the rating factors or override a price for an individual. Those are pricing decisions, and named people own them.
Why is a premium hard to take back?
A quoted premium that the customer accepts becomes a contract for the policy term. If it was too high, the insurer can refund the difference, but only after someone notices. If it was too low, the insurer usually has to honour it.
The bigger risk is scale. A mistake in a single quote costs one refund. A mistake in a rating factor reaches every policy priced while it was live, across renewals as well as new business. Finding the affected customers, recalculating and refunding them is a remediation project, not an undo. That is the shape of risk described in why autonomy in regulated industries is a question about reversal.
Pricing also carries fairness risk. A factor that tracks a protected characteristic indirectly can disadvantage a group of customers for months before anyone sees it. That harm is not repaired by a refund.
What can the agent do, and what stays with pricing staff?
| Task | The agent | Stays with a person |
|---|---|---|
| Producing a quote | Runs the approved rating model at depth 3 | Nothing routine |
| Applying published discounts | Applies them when the rule's conditions are met | Discounts outside the published rules |
| Explaining a premium | Explains the main factors to the customer in plain words | Disputes about whether a factor is fair |
| Renewal pricing | Calculates the renewal and flags large changes | Approving renewals above the insurer's threshold |
| Price override requests | Drafts the request with the case for it | The underwriter or pricing lead decides |
| Changing rates or factors | Never | The pricing committee, with sign-off |
| Monitoring outcomes | Compares prices across customer groups and flags patterns | Investigating and deciding what to change |
The quote itself is the safe part, because the model doing the pricing has been approved and the agent only runs it. The limits are in the rows below it. Underwriting decisions sit next to pricing ones; see whether an AI agent can underwrite an insurance policy.
A worked example
Take a pet insurer. A customer with two dogs asks for a renewal quote and mentions that a competitor is cheaper. The agent runs the rating model, applies the published multi-pet discount, and returns a renewal premium of 640, up from 520.
The customer asks for a better price. The agent cannot offer one. It explains the two factors that moved most, the older dog's age band and a claim during the year, and offers to pass the request to the retention team. It drafts the request with the quote breakdown.
The agent also notices that a new customer with the same two dogs would be quoted 540. It flags the gap to the pricing team as a pattern to review, without changing anything for this customer. The pricing team decides whether the renewal price is fair value.
What do the rules say about AI in pricing?
The EU AI Act lists risk assessment and pricing in life and health insurance among its high-risk uses, which brings risk management, logging and human oversight duties for those products. In the UK, the FCA Consumer Duty expects firms to deliver good outcomes for retail customers, including on price and value. In Singapore, the MAS FEAT principles set expectations on fairness and accountability; see our brief on AI agents for compliance teams in insurance in Singapore. In Australia, APRA and ASIC supervise insurers, as covered in AI agents for compliance teams in insurance in Australia.
This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
What controls make it safe?
- Read-only access to rates. The agent can call the rating model but cannot write to rate tables or factor files.
- Version stamps. Every quote records which version of the rating model produced it, so a faulty version can be traced to every affected customer.
- Thresholds on change. Renewals that move by more than the insurer's set amount are held for review.
- Overrides are drafts. Any request to depart from the model goes to a named person with authority.
- Outcome monitoring. The agent watches for price patterns across customer groups and reports them; it never adjusts for them on its own.
Questions people ask
Can the agent negotiate on price?
No. Negotiation is a discretionary pricing decision. The agent can explain the price and route a request to someone who has authority to change it.
Can the agent suggest cheaper cover?
It can show the options and their prices from the approved range. Recommending one over another based on the customer's needs moves towards advice, which needs its own controls.
Who owns the rating model?
The pricing function, under the insurer's model governance. The agent is a user of the model, not its owner, and that separation is a control in its own right.