Why underwriters are the real control in an AI agent deployment

Accepting a risk and setting its price commits the insurer for the life of the policy. AI agents can extract submissions, check appetite and draft quotes, but the underwriter decides whether to accept, at what price and on what terms, which makes underwriters the real control and their training a priority.

What does an agent take from an underwriter's desk?

The extraction and the first pass. At depth 1 the agent reads the submission, the proposal form, loss history, third-party data and underwriting guidelines. At depth 2 it drafts: a structured risk summary, a check against appetite with any breaches flagged, a proposed rating and a quote with standard terms. For small, standard risks within tight parameters, quoting may later be certified at depth 3.

Binding cover is a different matter. Once cover is bound, the insurer is on risk, and unwinding that is hard. Binding, and any decision outside the underwriting guidelines, belongs at depth 4 or with a named underwriter. Most insurers keep it with a person.

Why does underwriting authority map so well to depth?

Insurers already grade underwriters by authority: who may write which classes, up to which limits, and what must be referred. Agents fit this structure naturally. Each agent capability has a depth, and the organisation's own limits work like an authority letter. The verdict on each action names the bound that decided it.

The underwriter's training should make that link explicit. An agent's quote that exceeds its limit is held, exactly as a junior underwriter's quote would be referred. The underwriter who receives it is exercising their own authority, and the record shows it. Every verdict and every human decision is filed in an append-only record, which makes underwriting audits far more precise than file reviews used to be.

Where must the underwriter look hardest?

On the facts the proposer supplied. Under the Pryme Context Protocol, every fact the agent uses carries its provenance. Claims history from the insurer's own system is sealed. A broker's presentation is signed, with a named party answering for it. The proposer's answers on the proposal form are declared. A survey report attached to an email is external until checked.

Agent quotes are built on all of these at once. An underwriter who reads the provenance before the price will spot when a rating rests on declared answers that should have been verified, or on an external document that has not been confirmed. That is where the underwriting judgement lives now.

Brokers add a further layer. A broker's presentation is signed, but it is also written to win terms. Agents read it literally. Experienced underwriters read it as an argument, noticing what is emphasised, what is played down and what is missing. Teaching newer underwriters to read submissions that way, with an agent's summary beside them, is one of the most valuable things a senior underwriter can do.

What stays with the underwriter?

The decisions that define the risk the insurer carries:

1. Accepting unusual or large risks. Anything outside standard parameters or guidelines.
2. Pricing judgement. Loads and discounts that reflect what the data cannot fully show.
3. Terms and exclusions. Tailoring cover to the risk, which an agent can propose but not decide.
4. Referral. Recognising when a risk exceeds their authority and sending it upward.
5. Fair treatment. Making sure similar customers are treated alike and that pricing is defensible.

The fuller list is in what underwriters should never let an AI agent do.

How do regulation and fairness raise the stakes?

Risk assessment and pricing for life and health insurance is listed as high-risk in Annex III of the EU AI Act, which brings human oversight under Article 14 and AI literacy duties under Article 4 for those using such systems. GDPR Article 22 restricts decisions based solely on automated processing that have legal or similarly significant effects. Insurance supervisors across markets expect fair treatment and explainable pricing.

None of that rules out agents. It does mean the underwriter must be able to explain a decision the agent drafted, in terms a customer and a supervisor would accept. An underwriter who cannot say why a risk was loaded, beyond "the agent rated it that way", is not providing oversight. This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.

What does an underwriter need to learn?

Depth and how it is earned: the role sets a ceiling, evidence from certification against the insurer's own cases sets the depth, the lower wins. Verdicts: held, refused, and withheld, which is a platform floor for every agent. Provenance, as above. Certificates: they lapse after 90 days, and a run that clears no depth drops the agent to depth 1 the same day, so automated quotes become drafts until it recovers. And the review method, set out in how underwriters review an AI agent's decisions.

Underwriters sit beside operations managers, who run the queues, and collections agents, who deal with the money end of the policy. See operations managers as the real control and the overview in AI agent skills by role. The brief on underwriters in insurance in France covers that market.

Where the Academy fits

Pryme Intelligence Academy's Operator track trains underwriters who review agent quotes every day, with module tests set in insurance. Senior underwriters who design and certify the agents themselves may take the Builder track, which ends in a capstone agent certified in Attest to depth 2 or higher. An insurer can make binding authority depend on a current certificate.