Chain-of-thought (CoT) prompting forces the AI to show its work. On a CGL trigger analysis where the wrong answer produces a coverage dispute, on an Anti-Concurrent-Cause property analysis where the wrong wind-vs-flood allocation produces a $400K denial reversal, on an L&H accelerated-UW decision where the wrong rate class produces a discrimination claim, on a cat-XOL cession recommendation where the wrong layer carved produces a treaty-renewal disaster, on a Schedule P development review where the wrong method weighting produces an inadequate IBNR - chain-of-thought is what turns a black-box answer into a defensible memo that the credentialed reviewer can sign and the DOI examiner can audit. This lesson shows five worked CoT prompts across the L3 insurance surfaces (coverage analysis, ACC property reasoning, L&H accelerated UW, cat-XOL treaty cession, Schedule P development pattern review). Real prompts in code blocks. Real outputs with named numbers and named case-law citations. Real failure modes the chain catches before the AI ships an unsignable conclusion. The headline rule: a credentialed reviewer signing an AI-assisted decision must see the reasoning chain, not just the conclusion, because the conclusion alone is unsignable and the audit trail without reasoning is undefendable in DOI exam, bad-faith litigation, or treaty-renewal due diligence.

CoT Prompt 1 - CGL Trigger Analysis

Scenario: contractor sued for property damage from work performed 2019-2021; multiple CGL policies in effect during that window across three carriers; the carrier asks AI to analyze trigger of coverage. The naive prompt - "Is there coverage?" - produces a one-paragraph yes/no that misses the multi-policy allocation question entirely.

CoT prompt structure.

You are a coverage analyst reviewing a property-damage claim against a contractor. Apply the standard CGL trigger analysis. Show your reasoning step by step:
1. Identify the operative trigger theory in the controlling jurisdiction (manifestation, continuous trigger, injury-in-fact, exposure).
2. Identify each policy in effect during the operative period (named carriers, policy numbers, effective dates, limits, deductibles).
3. Apply the controlling jurisdiction's allocation methodology (pro-rata by time on risk, pro-rata by exposure, all-sums, joint-and-several).
4. Identify each policy's deductible, limit, and any aggregate considerations.
5. Identify potential exclusions and exceptions per policy (CGL Coverage A exclusions; products-completed-operations vs. operations; pollution exclusion; AI-related exclusions if any).
6. Reach a coverage conclusion with allocation by carrier; identify referral or escalation triggers.
Facts: [insert detailed claim facts, named insured, policy schedule, jurisdiction, claim period, damages, plaintiff's theory of liability].
Show your reasoning. Cite specific case law where applicable.

What the CoT chain catches. The naive prompt's one-paragraph answer might say "yes, coverage triggers under the 2020 policy." The CoT chain produces: the trigger theory in the controlling jurisdiction (continuous trigger in California under Aerojet-General Corp. v. Transamerica Insurance Co., 17 Cal. 4th 38 (1997); manifestation trigger in some other jurisdictions); the policies in effect 2019-2021 across three carriers (Carrier A 2019-2020 at $5M limit; Carrier B 2020-2021 at $3M limit shared with auto; Carrier C 2021-2022 at $5M limit); pro-rata-by-time allocation under the jurisdiction's continuous-trigger framework; per-policy deductibles ($25K each occurrence on A and C; $50K on B); specific exclusions (products-completed-operations excluded on Carrier A's policy for the 2020 portion; pollution exclusion applies on all three; no AI-specific exclusion on any); a coverage conclusion with an allocation table showing how the $1.8M property-damage claim distributes across the three carriers. The credentialed reviewer (the coverage analyst with AIC or coverage counsel with JD) can audit each step; the analyst can challenge the trigger-theory choice; the carrier's coverage counsel can validate the case-law citations via Westlaw lookup.

CoT Prompt 2 - Anti-Concurrent-Cause Property Analysis

Scenario: residential property damaged by Hurricane Ian (2022) in Florida; the loss includes wind damage to the roof and walls plus flood damage to the ground floor plus sewer-backup damage in the basement. The Anti-Concurrent-Cause (ACC) language in the policy says "we do not cover loss caused by any of the following whether or not any other cause or event contributes concurrently or in any sequence to the loss." The question: what does the policy cover after applying ACC, and what does it not cover?

CoT prompt.

Analyze coverage under ACC language for a hurricane loss with mixed-cause damage. Step through:
1. Identify the ACC language in the policy verbatim with form-edition reference (ISO HO 00 03 ACC language varies by edition).
2. Identify each covered peril in the policy.
3. Identify each excluded peril.
4. For each loss element, identify the operative cause(s) (wind only, flood only, sewer backup only, multiple causes contributing).
5. Apply ACC to each loss element: if any excluded peril contributed concurrently or in sequence, the element is excluded; if only covered perils contributed, it is covered.
6. Identify jurisdiction-specific case law on ACC enforceability (Florida has generally enforced ACC; California has been more skeptical).
7. Reach a per-element coverage conclusion with dollar allocation if available.
8. Identify escalation triggers (gray area on causation, large-loss threshold).
Facts: Hurricane Ian, Florida residential property at [address], [insert detailed loss elements with damage amounts]. Policy: ISO HO 00 03 form with standard ACC language. Show your reasoning.

What CoT surfaces. Per-element analysis: wind damage to the roof (covered - only wind cause; $95K); wind damage to walls and windows (covered - only wind cause; $48K); interior wind damage to upper floors (covered - wind-driven rain through compromised roof and windows; $42K); ground-floor flood damage (excluded - flood is excluded peril; ACC bars even if wind contributed; $145K); sewer backup damage in basement (excluded - sewer backup is excluded peril without endorsement; ACC bars even if hurricane contributed; $75K). Jurisdiction analysis: Florida courts have generally enforced ACC language (cite Wallach v. Rosenberg, 527 So. 2d 1386 (Fla. 3d DCA 1988); Florida Insurance Guaranty Association v. Smith, 161 So. 3d 600 (Fla. 5th DCA 2015)). Final allocation: covered $185K (roof plus walls plus interior wind damage); excluded $220K (flood plus sewer backup). Escalation triggers: the $48K wall damage requires confirmation that storm surge did not contribute (storm surge is treated as flood under Florida case law); the file is large enough to warrant claims-supervisor review under the $200K-aggregate threshold.

Without CoT, the AI's one-line answer might say "partial coverage" without the per-element breakdown that the adjuster needs to actually settle the claim, write the partial-denial letter, or defend the partial denial against a Florida CRN under §624.155.

CoT Prompt 3 - L&H Accelerated-UW Decision

Scenario: life insurance application for a 42-year-old female, $500K term policy, accelerated-underwriting pathway eligible per the carrier's accelerated-UW criteria (no medical exam if certain conditions are met; MIB hits and Rx history checked via the standard MIB and Milliman IntelliScript flows). The AI scores the application against accelerated-UW criteria. The decision: accelerate or send to full underwriting; or partial-accelerate with limited additional tests.

CoT prompt.

Apply accelerated-underwriting decision logic for a life insurance application. Step through:
1. Identify the carrier's accelerated-UW criteria (age band typically 18-50 or 25-60; face amount cap typically $500K-$1M; BMI band; prior-condition exclusions; prescription history flags; MIB hit handling; ECDIS / Rx hit handling).
2. Evaluate the applicant against each criterion.
3. Identify any flags (e.g., BMI 31 vs. cap 30; statin prescription suggesting cardiac risk management; MVR with one moving violation; prior declined application 8 years ago surfaced via MIB).
4. Determine if any flag warrants exiting the accelerated pathway, partial-accelerating with limited tests, or proceeding fully accelerated.
5. Apply Colorado Reg 10-1-1 Phase 3 (L&H 2026), NY DFS Circular Letter 2024-7, and state fair-discrimination rules to the decision.
6. Reach a decision: accelerate, partial-accelerate, or full UW.
7. Document the rationale and any consumer-disclosure language required (the disclosure framework's L&H overlay).
8. Identify escalation triggers (MIB hit interpretation, Rx ambiguity, BMI threshold close to cap).
Facts: [insert applicant detail including age, sex, height, weight, BMI, MIB hits, Rx history, MVR, prior applications]. Carrier criteria: [insert criteria]. Show your reasoning.

The chain catches the discrimination risk. The naive prompt might output "accelerate based on overall risk score" without engaging with the specific flag. The CoT chain produces step-by-step evaluation: applicant age 42 (within accelerated band 25-50), face $500K (within face cap $1M), BMI 31 (above cap 30 - flag for review), statin prescription (managed-cardiac flag - not disqualifying alone), no MIB hits beyond one application 8 years ago that resulted in standard issue, no MVR violations in 5 years. Decision: partial-accelerate - limited tests required given the BMI flag (likely a quick paramedical exam to confirm vitals; possibly an HbA1c to rule out undiagnosed diabetes); the statin alone is insufficient to exit the pathway because cardiac risk is being managed. Consumer disclosure: must advise the applicant of partial-acceleration and the basis (BMI threshold); under the disclosure framework, the applicant has the right to request human review of the AI's decision. Bias-test reference: BMI cap calibration documented against the carrier's home-state DOI guidance and the Colorado Reg 10-1-1 Phase 3 L&H framework rolling out in 2026.

CoT Prompt 4 - Cat-XOL Cession Recommendation

Scenario: 2026 hurricane season approaches; the cat-XOL treaty $50M xs $100M expires 12/31; renewal negotiation is underway with three reinsurers (Munich Re, Hannover Re, Swiss Re); the chief actuary asks the AI for a cession-strategy recommendation given the cat-model output, the competing reinsurer terms, the 2026 market conditions, and the carrier's RBC headroom.

CoT prompt.

Recommend cat-XOL treaty cession strategy for the 2026-2027 renewal. Step through:
1. Identify the carrier's exposure profile (TIV by territory, COPE attributes, current cat-aggregate consumption).
2. Identify the cat-model PML curve (RMS Touchstone primary, Verisk AIR comparison, KCC reconciliation; 1-in-50, 100, 250, 500, 1000 return periods).
3. Identify current treaty terms vs. expiring; identify proposed renewal terms from three competing reinsurers with rate-on-line, reinstatement, hours clauses, AI representations.
4. Per layer, evaluate: gross-net effective coverage at each PML point; reinstatement terms; aggregate cap; exclusions; co-insurance; UNL definition.
5. Identify second-event reinstatement implications and 2026 hard-market reinstatement constraints.
6. Identify any single-risk exposure exceeding the treaty per-risk limit ($25M); recommend facultative cession on over-line.
7. Compare reinsurer quotes against the PML curve; identify the best-layer-by-reinsurer placement.
8. Recommend retention / cession / retrocession strategy across all layers.
9. Document residual capital exposure at 1-in-250 and 1-in-500 against the carrier's RBC tolerance.
Facts: [insert TIV, PML curve, quotes, AI representations, market conditions]. Show your reasoning.

What CoT produces. Layered analysis: $0-100M retention (carrier balance sheet absorbs); $100-150M cat-XOL primary layer (Munich Re quote favorable on reinstatement at 22.8% ROL with 2x reinstatement; Swiss Re quote less favorable on AI representations but lower ROL at 21.5%; Hannover Re positioned between the two); $150-250M cat-XOL secondary layer (Hannover Re quote favorable on rate at 14.2% ROL with 1x reinstatement reflecting hard-market constraint); above $250M unhedged (the board accepts this exposure level given current RBC headroom 230% and ORSA scenario testing). Three buildings exceed the $25M single-risk treaty limit (Miami hotel $38M, Houston manufacturing $42M, Phoenix data center $54M) → facultative cession needed for $59M aggregate over-line (per the facultative-cession lesson). Residual exposure: 1-in-250 PML $1.24B gross → $260M net retention after primary and secondary cat-XOL; 1-in-500 PML $1.62B gross → $385M net retention; the board acknowledges that at 1-in-500 the RBC ratio drops to 195% which still clears the 175% threshold. AI representation analysis: Munich Re representations align with the carrier's documented AI governance posture; Swiss Re's tighter representations require general counsel review before signing.

Without CoT, the AI's answer might be "place full layer with Munich Re at 22.8% ROL" without the layer-by-layer reinsurer best-fit, the facultative-cession carve-out, the residual-exposure transparency, or the AI-representation analysis that the chief actuary and CRO need to sign the cession decision memo.

CoT Prompt 5 - Schedule P Development-Pattern Review

Scenario: the Appointed Actuary is reviewing the AY 2023 commercial-auto BI Schedule P development pattern at year-end 2026; the pattern is running 8-12% above the prior accident-year band; the AI is asked to surface the drivers of the unfavorable development.

CoT prompt.

Review the AY 2023 commercial-auto BI Schedule P development pattern. Step through:
1. Compare AY 2023 link ratios (12-to-24, 24-to-36, 36-to-48 months) to the prior accident-year band of comparable ratios.
2. Identify magnitude and direction of deviation; flag link ratios outside the 90th percentile of prior-year range.
3. Identify potential driver categories: claim-handling change (settlement-velocity shift), jurisdiction mix shift (more cases in social-inflation states), severity inflation (medical CPI, attorney fees), social inflation (jury awards, plaintiff-friendly verdicts), claim-cohort composition (large-loss tail).
4. For each driver category, evaluate evidence in the data (jurisdiction-segmented link ratios; ALAE-to-loss trends; case-reserve adequacy signals).
5. Identify specific jurisdictions and their AY 2023 development patterns (California, Illinois, New York, Georgia, Pennsylvania, Texas - the social-inflation jurisdictions for commercial auto).
6. Identify ALAE-to-loss trend by jurisdiction.
7. Identify case-reserve adequacy signals (post-evaluation case-reserve increases as severity emerges).
8. Recommend method weighting for IBNR estimation (chain-ladder paid, chain-ladder incurred, Bornhuetter-Ferguson, ELR method).
9. Recommend disclosure language for the Schedule P narrative.
10. Identify escalation triggers (Appointed Actuary RMAD analysis if reserves at risk of material adverse development).
Facts: [insert triangle data, ALAE, case reserves, jurisdiction segments]. Show your reasoning.

What CoT produces. Link-ratio comparison: AY 2023 12-to-24 ratio 1.42 vs. prior band 1.26-1.32 (above the 90th percentile of prior years); 24-to-36 ratio 1.18 vs. prior band 1.10-1.14; 36-to-48 ratio 1.09 vs. prior band 1.04-1.06. Driver analysis: social-inflation jurisdictions concentrated in AY 2023 mix; ALAE-to-loss in California, Illinois, New York, Georgia, Pennsylvania, and Texas running 4-7 points above non-litigation jurisdictions; case-reserve adequacy showing systematic increases at re-evaluation post-mediation. Method recommendation: weight chain-ladder paid 40%, Bornhuetter-Ferguson 30%, chain-ladder incurred 30% on this accident year; expected-loss-ratio method weight 0% (the ELR method assumes prior-year patterns hold; AY 2023 evidence suggests they do not). Schedule P narrative: documents AY 2023 unfavorable development of $14M; attributes to social inflation in litigation jurisdictions; references the method weighting; references the Appointed Actuary's RMAD (range of reasonable actuarial development) analysis under ASOP No. 36 and ASOP No. 43. The CoT chain gives the Appointed Actuary the audit trail to sign the AY 2023 Schedule P entry, the Statement of Actuarial Opinion, and the year-end reserve opinion.

When to Use CoT vs. When Not

CoT is overkill for simple lookups - form-edition lookup, sublimit-against-treaty constraint check, single-record data extraction. CoT is essential for: multi-step coverage analyses (CGL trigger with multiple policies); allocation decisions (ACC property, multi-claimant BI); accelerated-UW decisions with multiple flags; reinsurance cession strategy across multiple layers; reserve-method selection on edge cases (unusual development patterns); Schedule P narratives; SAO documentation; ORSA narrative drafting.

The 2026 best-practice: CoT prompts for all credentialed-reviewer signed artifacts; direct prompts for verification-layer-handled deterministic checks (form-edition validation, named-insured cross-check, sublimit comparison). The credentialed reviewer must be able to audit the chain; the verification layer handles the deterministic checks at the same speed and accuracy as a CoT prompt could but without the reasoning overhead.

The CoT plus persona plus RAG triple. The strongest L3 prompting pattern combines CoT (the reasoning structure) plus persona (the role's voice and decision frame, per the persona-engineering lesson) plus RAG (retrieval from the carrier's actual documents, per the RAG lesson). The result: role-specialized structured reasoning grounded in the carrier's actual claim-handling manual, UW guide, treaty wording, and ASOP posture. The credentialed reviewer signs; the DOI examiner audits; the treaty reinsurer reads the AI-representation disclosure favorably; the AM Best analyst reviews the 2026 readiness survey response with documented evidence.

Failure Modes CoT Catches vs. Misses

What CoT catches. Multi-step reasoning errors (the AI applies the wrong trigger theory and the reviewer can see the mistake at the trigger-identification step); allocation mistakes (the AI computes per-element coverage incorrectly and the reviewer can see which element); citation gaps (the AI offers a conclusion without a case citation and the reviewer flags the gap); jurisdictional mismatches (the AI applies California case law to a Texas claim and the reviewer catches it).

What CoT misses. Fundamental factual errors (the AI mis-reads a policy number or a date; the verification layer catches this at the deterministic-check stage); hallucinated citations (the AI invents a case citation that does not exist; the catch-hallucinations lesson's discipline catches this via Westlaw cross-check); rubber-stamp review by the reviewer (the reviewer skims the chain without engaging; the consumer-disclosure-and-professional-judgment lesson's rubber-stamp discipline addresses this).

CoT is one layer of the defensible-output stack. Verification layer plus credentialed-reviewer engagement plus RAG plus persona engineering plus drift monitoring round out the full discipline.

Key Takeaways

• Chain-of-thought forces AI to show reasoning step by step. Five worked prompts: CGL trigger, ACC property, L&H accelerated UW, cat-XOL cession, Schedule P development. Each turns a black-box answer into a defensible memo with named numbers, named case-law citations, and named escalation triggers.
• CGL trigger analysis CoT surfaces trigger theory, multi-policy allocation, per-policy exclusions, and case-law citations. Naive prompt's one-paragraph answer misses multi-policy allocation that drives the actual settlement; Aerojet-General Corp. v. Transamerica continuous-trigger reference in California cited.
• Anti-Concurrent-Cause CoT produces per-element analysis. Wind damage covered $185K (roof, walls, interior wind); flood plus sewer backup excluded $220K. Florida case-law on ACC enforceability cited (Wallach v. Rosenberg, FIGA v. Smith). Adjuster has per-element breakdown to actually settle and to defend any partial denial against a Florida CRN.
• L&H accelerated-UW CoT catches the discrimination risk. Step-by-step evaluation against carrier criteria; BMI 31 vs. cap 30 flagged; statin prescription noted (cardiac risk managed); partial-acceleration decision with consumer disclosure; Colorado Reg 10-1-1 Phase 3 L&H reference for the 2026 rollout.
• Cat-XOL cession CoT layered-strategy analysis. $100M retention plus $100-150M Munich Re primary at 22.8% ROL plus $150-250M Hannover Re secondary at 14.2% ROL plus above $250M unhedged with board acknowledgment. Facultative cession flagged for three buildings exceeding $25M ($59M aggregate over-line). Residual exposure documented at 1-in-250 ($260M net) and 1-in-500 ($385M net) against carrier RBC tolerance.
• Schedule P development review CoT produces link-ratio comparison, driver analysis, jurisdiction-specific patterns, method weighting, narrative language. AY 2023 12-to-24 ratio 1.42 vs. prior band 1.26-1.32; social-inflation jurisdictions identified; chain-ladder paid 40% / BF 30% / chain-ladder incurred 30% / ELR 0%; ASOP No. 36 and No. 43 references. Audit trail for Appointed Actuary signature on Schedule P, SAO, and reserve opinion.
• CoT for credentialed-reviewer artifacts; direct prompts for verification-layer-handled deterministic checks. 2026 best-practice. Verification layer handles deterministic at same speed; CoT handles judgment-laden artifacts where reasoning chain is the defensible documentation.
• Without CoT, the credentialed reviewer is signing a black-box conclusion. With CoT, the reviewer audits each step, challenges the chain, validates the citations, and signs a reasoned memo that supports DOI exam, bad-faith defense, treaty-renewal due diligence, and AM Best 2026 readiness survey response.
• CoT plus persona plus RAG together is the strongest L3 prompting pattern. Persona provides the role's voice and decision frame; CoT provides the reasoning structure; RAG provides grounding in the carrier's actual documents. The triple supports the credentialed reviewer's signature across the L3 work surface.