The claims-side mirror of Lesson 1 is the FNOL-to-reserve pipeline, and it is not symmetric. A submission-to-quote workflow runs across 20 steps with five swim lanes and an elapsed-time floor measured in hours. An FNOL-to-reserve workflow runs across roughly 22 steps with six swim lanes, an elapsed-time floor measured in hours-to-days for fast-track and weeks-to-months for complex, and a reserving discipline that compounds for years on long-tail claims. The platforms are different - Hi Marley for conversational intake, Five Sigma for auto-summary coverage check, Roots Automation for cognitive intake workers, Tractable for photo damage AI, EagleView for aerial roof reports, Snapsheet for virtual inspection, Shift Technology for fraud network analysis, ISO ClaimSearch for cross-carrier claim history, CCC for auto estimating, Carfax for vehicle history - and the regulatory exposure is sharper because claims handling carries the bad-faith risk that underwriting does not. This lesson is the FNOL-to-reserve map. Twenty-two steps. Six lanes. Five §4 reason-code checkpoints attached to claims-side decisions (coverage verification, triage, reserving recommendation, fraud routing, denial/payment). Three regulatory guardrails tied to NAIC Unfair Claims Settlement Practices Act, the three-state bad-faith framework (Texas Insurance Code §541, Florida §624.155, California Cumis/Brandt), and the medical-records HIPAA boundary for L&H/WC/BI claims.

The Twenty-Two-Step Claims Pipeline

The FNOL-to-reserve pipeline at a 2026 P&C or L&H carrier runs as follows. Each step has a named owner, named SLA, named AI platform where applicable, and a placement in the regulatory-exposure architecture.

Step 1 - FNOL intake. AI-driven: Hi Marley SMS intake or a call captured by a conversational AI receptionist. Insured texts "I had an accident at the intersection of Peachtree and 10th in Atlanta, the other driver ran a red light, my airbag deployed, I'm at Emory Hospital." Hi Marley extracts: claim type (auto BI/PD), location, time, parties, injury indicator, severity indicator. SLA: under 90 seconds to structured intake. Owner: claims operations.

Step 2 - Claim creation in the claims-management system. AI-driven: structured intake data populates Guidewire ClaimCenter, Duck Creek Claims, Sapiens ClaimsPro, or Snapsheet. Claim number assigns. Loss-date timestamps. Notice-to-carrier clock starts. SLA: under 2 minutes from intake completion. Owner: claims operations.

Step 3 - Coverage check. AI-driven: Five Sigma auto-summary or Roots Automation cognitive intake worker reads the policy declarations, endorsements, and exclusions; produces a one-page coverage analysis identifying available limits, applicable deductibles, named-insured status, exclusions triggered, and reservation-of-rights candidates. SLA: under 4 minutes. Owner: claims operations + adjuster.

Step 4 - Coverage verification by adjuster. Human-supervised AI: adjuster reads the Five Sigma summary, opens the actual policy in the PAS, verifies the AI's analysis, and signs off on coverage status. Override authority on AI's coverage call carries §4 reason chain. SLA: under 15 minutes. Owner: adjuster.

Step 5 - Triage routing - fast-track vs. complex. AI-driven: the workflow engine (Snapsheet, Roots, or in-platform triage in Guidewire ClaimCenter) scores the claim against the carrier's triage rules. Fast-track candidates: clear coverage, no injury, single-party loss under $7,500 (auto) or $15,000 (property), no exclusions triggered, no SIU red flags. Complex candidates: injury, multi-party, large-loss, coverage gray-area, reservation-of-rights, prior-claim density. SLA: under 4 minutes. Owner: claims operations.

Step 6 - Adjuster assignment. AI-driven: workload-balancing assigns to the right desk. Fast-track to a fast-track unit with high-volume tooling. Complex to a senior adjuster with the specialty (auto BI, property large-loss, WC permanent partial, L&H disability, BI demand package). SLA: under 4 minutes. Owner: claims operations.

Step 7 - First-touch contact SLA enforcement. Human-only with AI assistance: adjuster contacts insured within the regulatory first-touch window (24 hours for most lines, sometimes 48 for non-injury, 12 for catastrophe under state DOI bulletins). AI drafts the first-touch script and the recorded-statement outline. Adjuster makes the call. SLA: 24 hours from intake (state-specific). Owner: adjuster.

Step 8 - Recorded statement and transcription. AI-driven with adjuster sign-off: recorded statement captured on the call; Hi Marley or in-platform speech-to-text transcribes; Roots Automation or Five Sigma summarizes into structured fields (claimant statement of facts, injury description, treatment history, employment status, prior-claim history). Adjuster reviews and approves transcription. SLA: under 24 hours from call. Owner: adjuster.

Step 9 - Investigation plan and assignment. Human-only with AI assistance: adjuster drafts investigation plan. AI surfaces patterns from comparable claims: investigative steps, document requests, third-party assignments, expected timeline. SLA: same business day. Owner: adjuster.

Step 10 - Photo damage AI on auto/property. AI-driven: Tractable (auto and property), CCC (auto), or Snapsheet (auto + light property) ingests claimant-submitted or adjuster-collected photos. Tractable produces a damage-area heatmap, estimate of repair vs. total-loss probability, and parts/labor cost prediction. CCC produces an estimate against the OEM repair guide with line-item parts and labor hours. SLA: under 30 minutes. Owner: adjuster + estimator.

Step 11 - Aerial roof report on property claims. AI-driven: EagleView retrieves pre-loss and post-loss aerial imagery, computes roof slope, area, condition, and damage indicators. Output integrates with Xactimate or Symbility for property estimating. SLA: under 4 hours from order. Owner: adjuster + estimator.

Step 12 - Virtual inspection. AI-driven with adjuster guidance: Snapsheet ClaimXperience guides claimant through a self-inspection video; AI extracts angles, damage indicators, and VIN/serial numbers. Adjuster reviews the captured video. SLA: 24-48 hours from invite. Owner: adjuster.

Step 13 - Reserve recommendation. AI-driven with adjuster sign-off: reserving engine (Snapsheet, Five Sigma, or in-platform in ClaimCenter) recommends initial indemnity and ALAE reserves based on claim characteristics, comparable claim outcomes, and severity-prediction model. SLA: under 24 hours from coverage verification. Owner: adjuster + supervisor.

Step 14 - Reserve approval. Human-supervised AI: adjuster reviews reserve recommendation against carrier's reserve-authority schedule; supervisor approves if above adjuster authority. §4 reason chain captures the approval rationale. SLA: under 8 hours from recommendation. Owner: adjuster + supervisor.

Step 15 - Fraud screen. AI-driven: Shift Technology scores the claim against the carrier's fraud model + cross-carrier patterns. ISO ClaimSearch matches against prior claim activity across the industry. NICB referral check on auto. Network analysis surfaces clinic-attorney-claimant clusters. SLA: under 24 hours from intake. Owner: SIU.

Step 16 - SIU referral if scored. Human-only with AI assistance: if Shift score above SIU threshold or ISO ClaimSearch flags repeat patterns, AI drafts the SIU referral memo with NAIC Unfair Claims §4 reason codes. SIU investigates. SLA: 1-3 business days for triage. Owner: SIU investigator.

Step 17 - Medical-records review (BI/WC/L&H). AI-driven in HIPAA-eligible environment: Azure OpenAI with BAA, AWS Bedrock with BAA, or on-prem deployment ingests the 600-page medical-records PDF, produces chronological treatment timeline, ICD-10/CPT code analysis, AME/IME vs. treating-physician separation, AMA Guides 6th Edition impairment-rating drivers (WC), FCE results (L&H disability), demand-response framework (BI). SLA: 1-3 business days. Owner: claims + medical reviewer.

Step 18 - Subrogation triage. AI-driven: AI identifies subrogation potential from FNOL facts and police report. Carfax surfaces vehicle ownership history. Liability analysis flags responsible third party. SLA: under 5 business days. Owner: subrogation unit.

Step 19 - Settlement negotiation. Human-only with AI assistance: AI drafts settlement proposal, demand-letter response, structured-settlement analysis. Adjuster negotiates. SLA: varies. Owner: adjuster.

Step 20 - Reserve adequacy review. AI-driven: monthly portfolio review of reserves against actual outcomes; drift detection on severity, frequency, ALAE; champion-challenger model comparison. SLA: monthly. Owner: claims actuary.

Step 21 - Payment and closure (or denial). Human-only with AI assistance: AI drafts payment documentation, release, or denial letter with §4 reason chain. Adjuster signs. SLA: per state UCSPA. Owner: adjuster.

Step 22 - Post-closure analytics. AI-driven: closed claim feeds back into severity model, ALAE model, fraud model, triage model for retraining. SLA: continuous. Owner: claims data science.

The Fast-Track vs. Complex Sub-Pipelines

Steps 1-22 describe the maximum pipeline. In practice, the triage decision at Step 5 routes to one of two sub-pipelines with very different shapes.

Fast-track sub-pipeline. Auto PD under $7,500 with no injury and clear coverage: Steps 1-3 (intake + creation + coverage) run in under 7 minutes. Step 5 triages fast-track. Step 6 assigns to fast-track unit. Step 10 Tractable photo AI estimates damage in 30 minutes. Step 13 reserve recommendation auto-approves under adjuster authority. Step 21 payment fires within 48-72 hours of FNOL. Steps 7, 8, 9, 17, 19 typically skip. Total elapsed: 2-3 business days from FNOL to payment. Pre-AI baseline 7-14 business days; compression 70-80%.

Complex sub-pipeline. Auto BI with injury, multi-party, ambiguous liability: all 22 steps run. Step 7 first-touch within 24 hours. Step 8 recorded statement with adjuster sign-off. Step 9 investigation plan. Step 10 photo + Step 11 aerial (if property involved) + Step 12 virtual inspection. Step 15 fraud screen with Step 16 SIU referral if scored. Step 17 medical-records review with HIPAA discipline. Step 18 subrogation triage. Step 19 negotiation cycles. Step 13 reserve adjusts as new information lands. Step 21 settlement or denial after months or years. Pre-AI baseline 12-36 months on long-tail; AI-enabled timeline depends on litigation but reserve adequacy and severity prediction improve materially.

The Five §4 Reason-Code Checkpoints on the Claims Side

The submission-to-quote pipeline carried five §4 checkpoints. The FNOL-to-reserve pipeline carries five different ones, keyed to claims-side bad-faith exposure.

Checkpoint 1 - Coverage verification (Step 4). The AI's coverage analysis (Step 3) and the adjuster's verification (Step 4) together produce the coverage determination. If the adjuster overrides the AI (e.g., AI says "covered," adjuster says "reservation-of-rights candidate"), the override carries reason chain. Coverage analysis is the highest bad-faith exposure point in the pipeline; reservation-of-rights letters that diverge from the AI's reasoning need defensible documentation.

Checkpoint 2 - Triage routing (Step 5). Why fast-track vs. complex? Reason chain references the triage rules (claim type, coverage clarity, severity indicators, SIU flags). If a complex claim gets routed fast-track and loss develops adversely, the misrouting becomes a finding in a market-conduct exam.

Checkpoint 3 - Reserve recommendation (Step 13) and approval (Step 14). Reserve adequacy is a §4.4 documentation requirement and an annual-statement integrity requirement. Reason chain includes severity model output, comparable-claim references, ALAE benchmarks. Supervisor approval reason chain references reserve-authority schedule.

Checkpoint 4 - Fraud screen and SIU referral (Steps 15-16). NAIC Model Bulletin §4 applied to fraud: which model signals fired, which network connections surfaced in Shift, which ISO ClaimSearch matches landed. SIU referral memo cites §4 reason codes explicitly. Adverse-determination implications on the claimant flow under FCRA when a fraud denial converts to a refusal-to-pay.

Checkpoint 5 - Payment or denial decision (Step 21). The final coverage decision and the AI's role in supporting it. Denial letters with §4 reason chains survive bad-faith litigation; denial letters without survive less well. Texas Insurance Code §541, Florida §624.155, and California's Cumis/Brandt framework all examine the documentation of the denial decision.

The Three Regulatory Guardrails

Guardrail 1 - NAIC Unfair Claims Settlement Practices Act first-touch and acknowledgment SLAs. Most states adopt UCSPA with state-specific timelines: acknowledgment within 15 days (some 10), substantive response within 30 (some 15), denial-with-rationale within prescribed windows. The pipeline encodes these at Steps 2 (acknowledgment timestamp), 7 (first-touch SLA), and 21 (denial timeline). The AI workflow engine reminds, but the human owns the SLA compliance.

Guardrail 2 - Three-state bad-faith framework on settlement conduct. Texas Insurance Code §541 (unfair settlement practices, knowing-violation treble damages), Florida §624.155 (civil remedy notice, 60-day cure window), California Cumis/Brandt (independent counsel obligations when conflict exists, fees recoverable). The pipeline's negotiation step (19) and denial step (21) carry the most exposure. AI-drafted settlement proposals and denial letters need human sign-off, §4 reason chains, and conflict-screen documentation.

Guardrail 3 - HIPAA boundary on medical-records review (Step 17). Medical records on L&H disability, WC permanent-impairment, and BI demand-package review carry PHI. The AI must operate in a HIPAA-eligible environment with a BAA: Azure OpenAI with BAA, AWS Bedrock with BAA, or on-prem deployment. Consumer-grade tools (ChatGPT, Claude consumer, Gemini consumer) are categorically prohibited. The pipeline encodes the BAA-chain at Step 17 and the minimum-necessary rationale in the §4 reason chain. Privilege and work-product overlay attaches when defense counsel is involved; the AI's output to defense counsel may be work-product, the AI's output to the carrier's claims file is not. Lesson 10 in this chapter walks the workflow in detail.

The Six Swim Lanes of Claims Ownership

The pipeline runs across six swim lanes: claims operations, adjuster (with specialty sub-lanes for auto BI, property large-loss, WC, L&H, BI demand-package), SIU, subrogation, claims actuary, and claims data science. The handoffs are different from underwriting because claims tail extends over months and years.

Claims operations lane. Steps 1, 2, 5, 6. Owns intake-to-assignment. KPI: intake-to-coverage-verified elapsed time, intake-to-first-touch elapsed time, triage accuracy (downstream reroute rate).

Adjuster lane (with specialty sub-lanes). Steps 4, 7, 8, 9, 13, 14, 19, 21. Owns the claim file. KPI: cycle time, leakage rate, customer satisfaction (CSAT), bad-faith complaint frequency, severity prediction accuracy.

SIU lane. Steps 15, 16. Owns fraud investigation. KPI: referral-to-confirmation rate, recovery dollars, false-positive rate.

Subrogation lane. Step 18. Owns recoveries. KPI: identified-to-recovered ratio, recovery cycle time.

Claims actuary lane. Step 20. Owns reserve adequacy. KPI: reserve development triangles, IBNR accuracy, ALAE projection accuracy.

Claims data science lane. Steps 10, 11, 12, 17, 22. Owns the AI infrastructure for Tractable/CCC/EagleView/Snapsheet integration, medical-records LLM pipeline, severity model, fraud model, triage model. KPI: model accuracy, drift, ROI per model.

Why the Claims Map Is Harder Than the Underwriting Map

The submission-to-quote pipeline ends at a bound policy. The FNOL-to-reserve pipeline doesn't end; reserves develop over months and years until ultimate. The map has to handle three layers of temporal complexity the underwriting map does not.

First, the long tail. A liability claim opened in 2026 may close in 2031. The pipeline's Step 20 reserve adequacy review runs monthly for the life of the claim. Severity models retrain on the tail data; drift detection has to distinguish model drift from book-of-business drift from inflation drift.

Second, the litigation overlay. Once a claim is in litigation, the pipeline's AI involvement has to respect privilege and work-product. AI outputs to defense counsel may be work-product; AI outputs to the claims file are not. Document discovery includes the file but not the work-product. The carrier's AI governance memo must address this.

Third, the regulatory complexity. Coverage litigation, bad-faith litigation, market-conduct exams, DOI complaints, FCRA adverse-action implications, MHPAEA non-quantitative treatment limit signals on L&H mental-health claims, ERISA preemption on certain L&H matters, HIPAA on every L&H/WC/BI medical record. The claims pipeline carries more regulatory surface than the underwriting pipeline by a factor of three to five.

These layers make the map more important, not less. The carrier that operates a 22-step claims pipeline without naming the steps is the carrier that produces the 18.5% loss-ratio variance AM Best identifies as the AI-enabled-vs.-laggard gap, except going the wrong direction.

Key Takeaways

• The FNOL-to-reserve pipeline at a 2026 carrier is approximately 22 named steps running from Hi Marley SMS intake to post-closure analytics. Six swim lanes own the steps: claims operations, adjuster (with specialty sub-lanes), SIU, subrogation, claims actuary, claims data science.
• The pipeline branches at Step 5 triage into fast-track and complex sub-pipelines. Fast-track: Steps 1-3 + 5-6 + 10 + 13 + 21, 2-3 days FNOL-to-payment. Complex: all 22 steps, months-to-years on long-tail claims with regulatory and litigation overlay.
• Five §4 reason-code checkpoints attach to claims-side decisions: coverage verification (Step 4), triage routing (Step 5), reserve recommendation and approval (Steps 13-14), fraud screen and SIU referral (Steps 15-16), and payment or denial decision (Step 21). The denial documentation is the highest bad-faith exposure point.
• Three regulatory guardrails: NAIC UCSPA SLAs (acknowledgment, first-touch, denial timelines), three-state bad-faith framework (Texas §541, Florida §624.155, California Cumis/Brandt), and HIPAA boundary on medical-records review with BAA-chain discipline (Azure OpenAI BAA, AWS Bedrock BAA, on-prem).
• Named platforms across the claims stack: Hi Marley for conversational intake, Five Sigma and Roots Automation for coverage check and intake worker automation, Guidewire ClaimCenter / Duck Creek Claims / Sapiens ClaimsPro / Snapsheet for claims management, Tractable for photo damage AI, CCC for auto estimating, EagleView for aerial roof reports, Snapsheet ClaimXperience for virtual inspection, Shift Technology for fraud, ISO ClaimSearch for cross-carrier history, Carfax for vehicle history.
• The fast-track sub-pipeline compresses FNOL-to-payment from 7-14 days to 2-3 days at Tractable-enabled carriers. The complex sub-pipeline's compression lives in reserve adequacy and severity prediction rather than cycle time on disputed claims.
• The claims map is harder than the underwriting map because of three temporal layers: the long tail (years of reserve development), the litigation overlay (privilege and work-product), and the regulatory complexity (UCSPA + bad-faith + FCRA + MHPAEA + ERISA + HIPAA depending on line). Carriers that skip the map produce the wrong-direction 18.5% loss-ratio variance.
• Coverage verification (Step 4) is the single highest bad-faith exposure point in the pipeline. The AI's coverage call plus the adjuster's verification plus the override reasoning (when applicable) all carry §4 reason chains. Reservation-of-rights letters that diverge from the AI's reasoning need defensible documentation that survives Cumis/Brandt scrutiny in California and §541 scrutiny in Texas.