How to Build an AI Underwriting Assistant (Playbook)

Underwriters spend most of their time assembling and reading files rather than judging risk. An AI underwriting assistant absorbs the assembly: extracting data from submissions, checking completeness and appetite, retrieving guidelines, and drafting a cited risk summary in the underwriter's own structure, while the underwriter keeps the decision. This playbook covers the build, following FISTA's AI agents practice and the controls in the AI controls for financial services whitepaper. Context is in ai underwriting in insurance. This is general guidance, not legal advice.

What does the assistant do?

Stage	Function	Control
Intake	Submissions from email, broker portals, APIs	Deduplication; case linking
Extraction	Application, loss runs, financials, schedules, inspections	Confidence; validation; review
Completeness and appetite	Missing information; out-of-appetite flags	Rules from underwriting guidelines
Enrichment	Third-party data with provenance	Approved sources only
Guideline retrieval	Applicable rules and referrals	Citations
Risk summary	Structured, cited draft	Underwriter review
Decision documentation	Underwriter decision, rationale, referrals	Audit trail

Step 1: Specify with underwriters

Work with underwriting leadership to define lines and classes in scope, the risk summary structure underwriters use, completeness rules, appetite rules, referral triggers, prohibited behavior (no recommendations on regulated decisions unless explicitly designed, no unsupported statements), and quality thresholds. The specification is the evaluation basis. See how to write an ai spec.

Step 2: Build intake and extraction

Consolidate submission channels; classify document types; extract structured fields per document type with confidence and positions; validate with rules (totals, dates, identifier formats) and cross-document consistency (application versus loss runs). Low-confidence fields route to review. The pipeline pattern is in how to build an ai data extraction pipeline.

Step 3: Check completeness and appetite

Apply deterministic rules from the guidelines: required documents and fields per class, appetite exclusions, and referral triggers. Generate broker follow-up requests for missing items as drafts for underwriter or assistant approval. Rules are code, versioned with the guidelines. See rules engine vs llm.

Step 4: Retrieve guidelines with citations

Index underwriting guidelines, bulletins, and appetite documents with versions and effective dates; retrieve the sections relevant to the class, exposures, and flagged conditions; and attach citations. Guideline retrieval turns experienced underwriters' tacit knowledge into checkable references for less experienced ones. The retrieval design is in the enterprise RAG reference architecture whitepaper.

Step 5: Draft the risk summary

Generate the summary in the underwriter's structure: exposure overview, loss history analysis, financial indicators, controls and mitigations, guideline considerations and referrals, and open questions. Every statement cites an extracted field or guideline section; missing data is flagged, not filled. Validation checks citations and blocks recommendations outside scope. See llm output validation.

Step 6: Preserve decision rights

The underwriter reviews the summary with source documents one click away, edits, records the decision and rationale, and applies referrals. The system logs the summary version, edits, decision, and approver. Where straight-through processing is considered for simple classes, it is a separate specification with rules, validation, fairness testing, and oversight. Design is in what is a human approval gate and ai human oversight requirements.

Step 7: Evaluate

Build a labeled set of past submissions with underwriter-verified extractions, expected completeness and appetite outcomes, and reference summaries. Measure field-level accuracy, rule correctness, guideline retrieval relevance, summary faithfulness and completeness, and adversarial resistance to documents containing manipulative text. Wire the suite into CI. Method is in the AI evaluation and testing whitepaper.

Step 8: Launch and monitor

Launch with a pilot underwriting team; measure time from submission to decision, edit rates by summary section, referral accuracy, and underwriter satisfaction; sample summaries for faithfulness weekly. Track quote-to-bind and loss outcomes over time as lagging indicators. Observability design is in the AI observability whitepaper.

Worked example: small commercial property

A carrier pilots the assistant on small commercial property submissions. Applications, loss runs, and inspection reports arrive from brokers by email. Extraction captures building attributes, occupancy, protection class, values, and five-year loss history with confidence; validation cross-checks values against schedules and flags a mismatch between application and loss-run policy periods. Appetite rules flag an occupancy outside appetite and generate a decline draft; for in-appetite risks, guideline retrieval surfaces the sections on roof age and protection class. The summary presents exposures, loss analysis, and guideline considerations with citations, and lists two open questions for the broker. The underwriter reviews, edits the loss analysis, decides, and documents. Pilot metrics show time to decision falling and edit rates concentrated in the loss analysis section, which drives a prompt and evaluation refinement.

What governance applies?

Model risk management for in-scope components, fairness testing of outcomes where decisions affect protected groups, adverse-action reasoning where required, audit trails, access control over submission data, and change control. See ai model risk management and ai explainability requirements.

What does it cost to run?

Cost scales with submission volume and document pages; value is measured in underwriter time per file, turnaround, consistency, and downstream quote-to-bind and loss performance. Drivers are in document ai cost.

What are the common mistakes?

• Summaries without citations that underwriters cannot check.
• Extraction without validation feeding confident wrong summaries.
• Guidelines indexed without versions, citing superseded rules.
• Blurring decision rights without the controls straight-through processing requires.
• Evaluating fluency instead of faithfulness and field accuracy.
• Ignoring fairness and adverse-action requirements until audit.

What team does the assistant need?

The build needs senior underwriters to define the summary structure and label submissions, a compliance partner for decision-rights and fairness controls, a document and retrieval engineer, and a product owner from underwriting operations who prioritizes by turnaround and quality rather than by novelty.

How FISTA Solutions builds underwriting assistants

FISTA Solutions builds underwriting assistants to this playbook: underwriter-defined specifications, confidence-scored extraction with validation, rule-based completeness and appetite checks, versioned guideline retrieval with citations, cited risk summaries in the underwriter's structure, enforced decision rights, and labeled-set evaluation under model risk governance. The AI agents practice delivers the assistant, AI enablement the document and retrieval platform, and forward deployed engineers embed with your underwriting team. The record behind the work is 150+ projects with 99.9% uptime.

This playbook is general guidance, not legal advice. To scope an underwriting assistant, message FISTA on WhatsApp, or read ai in commercial insurance for the sector context.