How to teach underwriters to bank corrections that improve an AI agent

Underwriters bank a correction by recording why they changed the agent's risk summary, rating inputs or referral, then routing that reason to the agent owner as a brief, source or bound change, re-certified before go-live. The underwriting rule: no banked change may give the agent more room than written authority and the insurer's appetite allow.

What does the agent do for underwriters?

An underwriting agent gathers the submission, extracts risk information from proposal forms, surveys and loss histories, checks the risk against appetite, prepares rating inputs and drafts a recommendation or a referral note. It works at depth 1 and 2. Binding cover, setting terms outside guidelines and declining risks stay with underwriters who hold the authority to do so.

Underwriters correct it in recognisable ways. It reads a five-year loss history and misses that the largest claim is still open and reserved. It treats a risk as within appetite because the trade is listed, when the appetite statement excludes that trade above a certain turnover. It applies a rating factor from an old rating guide. It writes a referral note that says what the risk is but not why it needs referral.

The cross-role skills map is in AI agent skills by role.

Why does written authority set the boundary?

Underwriting runs on delegated authority. Each underwriter has written limits: classes, sums insured, territories, deviations from the rating guide. An agent working for an underwriter cannot hold more authority than the person it works for, and in practice it holds much less, because it drafts and the underwriter decides.

This shapes banking. Suppose an experienced underwriter regularly accepts a certain risk feature that the guide treats as a referral trigger, because their authority allows it. If they correct the agent's referrals and someone banks the change, the agent now stops referring that feature for every underwriter, including juniors whose authority does not cover it. A banked change must be safe for the least-authorised person who relies on the agent. Changes that widen what the agent treats as acceptable go through the underwriting manager and the authority framework.

What are the steps?

1. Agent, depth 1: extracts risk information from the submission, each fact stamped with provenance.
2. Agent, depth 2: prepares rating inputs and drafts a recommendation or referral, citing the appetite statement and rating guide sections used.
3. Person (underwriter): corrects and records a reason: missed fact, appetite misread, stale guide, referral unclear, authority-dependent, or one-off.
4. Agent, depth 1: groups reasons fortnightly by class and reason.
5. Person (underwriting manager): decides what to bank. Anything marked authority-dependent stays out, or is banked only as a rule tied to the reviewing underwriter's authority.
6. Person (agent owner): makes the change, such as a rule that open claims with reserves are always summarised separately, or the current rating guide replacing the old one as a sealed source.
7. Attest replays the agent against past submissions, including risks that went on to claim, before go-live.

What else must underwriters learn?

How to judge provenance in a submission. A proposal form is declared by the insured or broker. A survey report is signed by the surveyor. Claims data from the insurer's own system is sealed. An agent that weighs a declared statement above a contradicting survey is making an error, and the banked fix is a rule about how conflicts between classes are resolved.

How pricing rules interact with regulation. In the EU, risk assessment and pricing for life and health insurance of individuals is listed as high-risk under the AI Act, which brings human oversight duties. Underwriters in those lines should expect more scrutiny of any change that affects pricing, and keep the agent's role to drafting. The everyday review checks are covered in how underwriters review an AI agent's decisions.

Which controls and evidence matter?

• Authority-dependent corrections tagged. Evidence: the reason code and the underwriter's authority level.
• Rating guide and appetite from the system of record. Evidence: provenance and version on every citation.
• Manager approval for widening changes. Evidence: approver name on each.
• Replay on risks that claimed. Evidence: harm and escalation gate scores.

What should underwriting measure?

Repeat corrections per reason code by class. Referrals the underwriter found unnecessary, and referrals the agent should have made but did not; the second matters more. Time from a rating guide or appetite change to the agent's sources. And corrections marked authority-dependent, as a guide to how varied practice is across the team.

Where does it usually go wrong?

Senior underwriters' authority leaks into the agent through banked corrections, so juniors receive drafts that treat referrable risks as routine.

Guides change and the agent keeps the old one. Rating guides are revised often; tie each revision to a source update.

Referral notes get banked as format rather than reasoning. A tidy note that does not say why a risk needs referral is no help to the person receiving it. Collections teams face a related problem with arrangements that depend on who approved them, as teaching collections agents to bank corrections explains, and insurers in different markets add their own supervisory expectations, as AI agent training for underwriters in insurance in Ireland describes.