Learn . Use cases
AI use cases in insurance.
Seven patterns where AI is paying back inside US carriers and MGAs right now — and what each one costs to ship, what breaks in production, and what state DOIs care about.
Updated . 2026-05-17 . 9 min read
Insurance has the rare combination that makes AI pay back fast: document-heavy workflows, repeatable decisions, structured-data shortages, and operational metrics that move money directly. The discipline is making sure every AI touchpoint is auditable back to source, that no decision authority shifts to a model without explicit human-in-the-loop, and that the compliance posture survives a DOI market-conduct exam. We've shipped these patterns across P&C, life, and specialty lines.
Below: seven production patterns ordered roughly by payback speed for a typical mid-market carrier. Numbers are observed ranges across AISD/inVerita engagements — not vendor claims. See our insurance industry hub for engagement structure and the claims triage case study for end-to-end detail.
Use case 01
FNOL & first-touch triage agents
45 min → 8 min handle time
First Notice of Loss is where most claim leakage starts — slow intake, missing data, wrong adjuster routing. An FNOL agent takes inbound calls, emails, or portal submissions, asks the right follow-up questions in plain language, extracts the structured fields (policy ID, loss date, peril, severity signals), pulls the policy from PAS, runs initial coverage validation, and routes to the right adjuster bucket. We typically see end-to-end handle time drop from 45 minutes to 8 minutes for property and auto first notices.
Failure mode + mitigation
Agent over-promises coverage in the customer-facing transcript. Mitigation: strict prompt boundaries, deterministic policy-language inserts (never let the LLM paraphrase coverage), and human review for any line that quotes the policy. Eval set covers 200+ recorded calls with adversarial scenarios.
Use case 02
Claims document processing
60-80% ↓ adjuster review time on document-heavy claims
Claims pull medical records, repair estimates, police reports, prior claims, adjuster notes, recorded statements. Document agents extract the structured signals (diagnosis codes, treatment dates, repair line items, fault indicators) and surface contradictions across documents. Adjusters get a structured summary and a pile of original docs with citation links instead of a 200-page PDF to read.
Failure mode + mitigation
Hallucinated extractions on poor-quality scans. Mitigation: confidence scoring per field, mandatory citation back to source PDF coordinates, human-in-the-loop on any low-confidence field, and a regression suite of historical edge cases.
Use case 03
Underwriting copilots
30-50% ↑ underwriter capacity
Underwriters spend most of their time gathering and reconciling risk signals: loss runs, MVRs, prior carrier histories, public records, third-party data. A copilot retrieves and structures those signals against the policy precedent corpus, surfaces similar-risk historical bindings with their outcomes, and drafts a first-pass recommendation memo. Final decision stays with the underwriter — the copilot just kills the 4 hours of data-gathering.
Failure mode + mitigation
Anchoring bias: underwriters rubber-stamp the copilot's recommendation. Mitigation: show the copilot's confidence and dissenting prior cases, require underwriters to write a short justification, and run periodic blind A/B comparisons of copilot-assisted vs. unassisted decisions.
Use case 04
Customer-service deflection
25-40% auto-resolution on policy & billing queries
Most policyholder calls are repetitive: 'what's my deductible,' 'when's my payment due,' 'can you email my ID card.' A retrieval agent with read access to policy admin, billing, and document systems handles these in 30 seconds without escalation. Critically: it knows what it doesn't know (anything claim-status, coverage-interpretation, or refund-related) and warm-transfers to a human with full context preserved.
Failure mode + mitigation
Hallucinated payment confirmations. Mitigation: anything that says 'your payment is processed' must come from a confirmed billing-system response, never the LLM. Test set explicitly contains payment-confirmation traps.
Use case 05
Fraud-signal scoring
20-35% ↓ false positives on SIU referrals
Traditional fraud scoring combines structured signals (claim frequency, prior carrier flags, loss ratios). Modern fraud detection adds the narrative dimension: language patterns in recorded statements, contradictions across documents, and known-fraud-ring graph features. An agent layers narrative scoring on top of your existing model, explains its reasoning in adjuster-readable form, and links every signal to its evidence source.
Failure mode + mitigation
Bias risk. Mitigation: bias auditing across protected classes, no decision authority without human SIU review, full audit trail of inputs and reasoning, separate model evaluation against held-out fairness test set.
Use case 06
Subrogation identification & recovery
15-25% ↑ subrogation recoveries
Subro opportunities get missed because adjusters working at scale don't have time to read every police report or look for liable third parties. A subro agent reads each closed claim packet, flags potential third-party liability, drafts the initial demand letter, and tracks the recovery workflow. Most carriers we work with see 15-25% recovery uplift in the first 6 months because more files get worked, not because the legal logic is better.
Failure mode + mitigation
Frivolous subro letters to non-liable parties damaging carrier reputation. Mitigation: minimum-confidence threshold for auto-drafting, mandatory adjuster review before any letter goes out, monthly QA sampling on letters sent.
Use case 07
Producer/agent enablement copilots
30-50% ↑ quote velocity
Agents and brokers spend hours reconciling carrier appetite, navigating raters, and answering basic policyholder questions. An enablement copilot embedded in the agent portal answers product/appetite questions, walks through quoting, and drafts coverage explanations in the agent's voice. Particularly useful for carriers selling through independent agents who carry 8-12 markets and can't keep every guideline straight.
Failure mode + mitigation
Outdated appetite data. Mitigation: nightly sync from underwriting guidelines, dated source citations on every answer, and flagging of any retrieval older than 30 days.
Compliance posture
What state DOIs actually care about.
Insurance AI rollouts get blocked at the compliance gate more often than at the engineering gate. Three things that need to be in place before a carrier signs off:
- Audit logs that survive market-conduct exams. Every AI decision logged with input, retrieval sources, model version, prompt hash, and the human reviewer's response. Retained per your state's record-retention rule (typically 5–7 years).
- HIPAA-equivalent PHI handling on bodily-injury claims. Auto, GL, and workers' comp claims include medical records. BAAs with model providers, PHI redaction at boundary, and no PHI in vector indexes shared across customers.
- Disparate-impact testing. Fraud models, pricing copilots, and any AI that influences claim outcomes need bias audits across protected classes with documented mitigation when disparities surface. NAIC's AI bulletin (Aug 2024) is the baseline.
Build vs buy
When off-the-shelf works and when it doesn't.
Insurance-specific AI platforms (Tractable, Shift Technology, Clara Analytics, EvolutionIQ) cover narrow vertical slices well. For commodity workflows — photo-based auto damage estimation, basic FNOL chatbots, off-the-shelf fraud scoring — buying is faster and the vendor takes regulatory risk.
You should build custom when: (1) the workflow touches your proprietary policy admin or claims system in ways a vendor won't integrate, (2) you have product lines a vertical vendor doesn't cover (specialty, surety, complex commercial), (3) the competitive advantage is the workflow itself rather than generic AI capability, or (4) your data is the differentiator and you don't want it pooled into a multi-tenant vendor model. Most carriers we work with run a hybrid: buy where commodity, build where differentiated. See our build vs buy comparison for the framework.
Where to start
Discovery sprint structure for insurance.
A 2-week paid discovery sprint with us for an insurance carrier covers: (a) workflow shadowing in claims/underwriting/service to find the highest-friction repeat tasks, (b) data audit (what's in PAS, claims system, document repo, recorded calls — and what's accessible), (c) regulatory walkthrough with your compliance team to define non-negotiables, (d) a ranked backlog of 4-6 AI use cases with rough payback estimates, and (e) a fixed-price proposal for the top 1-2.
See how to build an AI agent for the engineering pattern, and cost of building an AI agent for budget templates. Typical insurance first-build lands between $80K and $180K depending on system access and regulatory exposure.