The case for certifying underwriters before AI agents go live

Underwriting authority has always been personal: a named underwriter holds a written limit and answers for every risk they accept. Agents do not change that. They read the submission and draft the risk view; the underwriter still binds. Certifying underwriters before go-live makes sure that authority is exercised over the agent's work with full care.

How does an agent fit into underwriting?

An underwriter receives a submission, often from a broker. It contains proposal forms, schedules, loss history and supporting documents in many formats. The underwriter reads it, checks it against appetite, assesses the risk, prices it, sets terms and decides whether to quote, refer or decline.

A governed agent takes the reading and drafting. At depth 1 it reads the submission and the insurer's own data. At depth 2 it drafts a risk summary, flags gaps and proposes terms. For simple, well-understood risks within appetite, an insurer may let it issue quotes at depth 3 within limits the underwriting function sets, once Attest has shown it does so correctly on the insurer's own book. Anything outside appetite, any referral and every decline is held for a named underwriter.

Why is a broker submission a provenance problem?

Almost everything in a submission is declared or external. The proposal form states what the insured says about themselves. The loss history may come from the insured, a previous insurer, or a spreadsheet from the broker. Surveys and valuations may be signed by an identifiable professional, or may be an attachment with no clear author.

The agent records each fact's class. But a draft risk summary reads with the same confidence whether its key facts are signed or merely declared. A certified underwriter checks: is the sum insured supported by a signed valuation, or declared on the form? Is the five-year loss history from the previous insurer, or a broker's summary? Does the risk summary rely on anything the insurer itself holds as sealed?

This matters more for underwriting than many other roles because the whole contract rests on what was disclosed. An underwriter who accepts a risk on declared facts without realising it has made a different decision from one who knew.

Where does regulation sharpen the case?

For some lines it is explicit. The EU AI Act lists risk assessment and pricing in relation to natural persons for life and health insurance as high-risk in Annex III. Deployers of high-risk systems must provide human oversight under Article 14, and the people providing it need the competence to do so meaningfully. An underwriter who approves an agent's pricing for a life policy is that oversight.

For commercial lines the rules differ, but the principle is the same. Supervisors expect insurers to show who made each underwriting decision and on what basis. The verdict record shows the first. Certification is how you show the second: that the person was equipped to judge what the agent prepared.

What should an underwriter's certificate prove?

That they can trace every material fact in a risk summary to its source and provenance class; read why a case was held or referred; tell when an agent's proposed terms rely on a gap it could not see; and decline or refer with reasons that are their own. A practicum of real underwriting decisions, observed and scored, shows whether they do this on live submissions or only in theory.

How should underwriters take part in pre-launch testing?

Before the agent quotes anything, insurers should red-team it before go-live: submissions with misstated sums insured, loss histories that contradict other evidence, and risks just outside appetite. Underwriters should review the agent's drafts on those cases blind. Where its draft would have convinced them, the gap is in the training, and it should be closed before certification is signed off.

The AI agent go-live approval checklist should then include a line confirming that every underwriter with authority over agent-prepared risks holds a current certificate.

What is the trade-off?

Underwriters are experienced, expensive and often sceptical of new controls. Asking them to sit an exam and a practicum can feel like questioning their judgement, and pulling them off live submissions costs business.

The response is that certification tests something new: reading an agent's work, not underwriting. A senior underwriter's risk judgement is not in question. What is in question is whether they will spot that the agent's tidy summary rests on a declared figure, under the time pressure of a busy renewal season.

Sequencing helps. Run the agent at depth 2 first, with every quote and referral decided by an underwriter. That is practice on real submissions. Only allow depth 3 quotes on simple risks once the underwriters who sample them are certified.

The record makes the rest plain. Each bound risk will carry an underwriter's name, and the question of whether underwriters are accountable for an agent's mistakes depends on whether they were equipped to see what the agent missed. The same reasoning drives the case in collections and credit control: whoever's name is on the decision should be tested on the system that prepared it.

What should a chief underwriting officer do now?

Link underwriting authority over agent-prepared risks to a current certificate. Keep referrals and declines held. Use red-team cases as training material. Review the agent's appetite rules at each of its 90-day recertifications, with an underwriter involved. The guide to AI literacy and certification for people who work with agents sets out which level fits which job.

Where the Academy fits

Underwriters who review agent-prepared risks take Pryme Intelligence Academy's Operator track, with module tests set in insurance and in their own role. Those who design underwriting agents and their appetite rules can go on to the Builder track. Both certificates are valid for 12 months and listed on a public register by name, credential and expiry.