What this guide is for
O*NET describes insurance underwriters as evaluating applications and deciding whether to provide coverage and on what terms. Those are consequential insurance decisions, not tasks to hand to a general-purpose model. NAIC says insurers remain responsible for applicable requirements when AI supports insurance decisions, including fairness, accuracy, and avoiding unfair discrimination.
Make one review packet, not a synthetic underwriting decision
needs_review.Bound the scenario and authority.
Name the training or approved purpose, line-of-business/jurisdiction owner, authorized underwriter, procedure owner, and prohibited decisions. Stop for private data, unapproved sources, a requested conclusion, or an instruction to communicate or update a system.
Inventory only approved material.
AI may capture labels, owner, version, date, stated permitted use, and apparent missing/duplicate/conflicting status. It must not evaluate an application or infer risk from the material.
Make unknowns visible.
Where source, version, authority, permitted use, or context is missing or conflicts, label it
unknown. Do not fill the gap with model memory or a plausible assumption.Draft questions, never decisions.
Prepare neutral questions for the named underwriting, product, legal/compliance, actuarial/model-risk, and privacy/security reviewers. Do not draft acceptance, declination, pricing, terms, notices, or messages.
Hand off and stop.
Create the review packet and name the people who can approve, revise, reject, or escalate it. The packet never authorizes an external action.
Give AI a clerk job, never an underwriter job
Organizes approved source labels, dates, owners, and stated permitted uses.
Flags missing, stale, duplicate, or conflicting support without resolving it.
Drafts neutral reviewer questions tied to named human authority.
Safe preparation vs. approval-required work
| AI may prepare | Human authority must decide or do |
|---|---|
| Inventory approved source labels and metadata | Evaluate an application, choose inputs, or determine eligibility/risk |
| Flag missing, duplicate, stale, or conflicting support | Accept, decline, classify, set coverage, limits, deductibles, terms, or exceptions |
| Produce an unknown list and neutral review questions | Set rate, premium, price, notice, cancellation/nonrenewal, or policy action |
| Assemble a local review-only receipt | Validate a model; make fairness, legal, privacy/security, or compliance conclusions; communicate or write a system |
Insurance AI governance varies by state, insurer, product, facts, and data. NAIC material is not a universal self-executing rule. Obtain current legal/compliance, underwriting, product, model-risk, privacy/security, and procedure review.
Five ways a preparation tool becomes unsafe
- Calling a summary a risk evaluation. A tidy source list is not an application evaluation, class, or underwriting conclusion.
- Turning an unknown into a plausible term. Coverage, limits, rates, premium, and exceptions require authorized sources and people.
- Combining sources without preserving their version. Preserve owner, date, permitted purpose, and conflict; do not silently normalize them.
- Using more data “just in case.” Minimize inputs and do not upload private applicant, insured, producer, policy, medical, credit, financial, or rating data without separately authorized review.
- Letting a draft leak into an insurance action. The packet ends at
needs_review; it cannot quote, bind, notify, message, upload, or update a system.
Underwriting Evidence & Review Desk
Enter only fictional or organization-approved, minimum-necessary text. This form keeps data in your browser and produces a review-only receipt; it does not send or save anything.
What this manual is built on
- O*NET: Insurance UnderwritersOccupation and task context; not an underwriting procedure.
- NAIC: Artificial IntelligenceCurrent insurer-AI governance context; check state and carrier applicability.
- NAIC: AI Model Bulletin announcementGuidance and governance context, not a universal legal rule.
- NIST AI 600-1Voluntary generative-AI risk-management framing.
- FTC: Gramm-Leach-Bliley Act guidancePrivacy and safeguarding context for covered organizations.
The cited sources and the guide boundaries should be rechecked before real-world use.
Connect this desk without blurring authority
Start at the Insurance AI Guide Shop. For intake/service preparation, use the Insurance Claims & Policy Clerks guide. For post-loss vehicle evidence, use the Auto Damage Appraisers guide. For least-privilege role design, use the AI Agent Workflow Builder.