The Atlanta auto-and-GL claims adjuster's Tuesday diary at 9:00 a.m. has 15 open files. The first ten minutes of the morning have already touched seven AI tools: Tractable's mobile photo estimate on the total-loss Highlander, CCC Intelligent Solutions' litigation prediction on the slip-and-fall GL with $42K ALAE on a $30K reserve, EagleView's aerial roof report on three April-storm hail claims (the architectural-vs-three-tab miscall on one is $4,200 of RCV difference), Shift Technology's fraud network signal on a third-party medical clinic appearing on five plaintiff-attorney-represented soft-tissue claims in 60 days, Hi Marley's SMS thread (47 messages) with the water-damage insured who hasn't responded in 11 days, Five Sigma's auto-coverage summary citing the wrong policy edition (HO 00 03 when the dec page shows HO 00 05), and the carrier's enterprise LLM drafting the three-days-overdue Reservation-of-Rights letter on the CGL faulty-workmanship-exclusion question. CCC's April 2026 report put total-loss frequency at a record 23.1%. Bain estimates 30-50% claims leakage reduction is achievable. Snapsheet publishes ~20% LAE reduction. Tractable's Admiral Seguros case put digital completion at 70-75%. Five Sigma's Starr Insurance 2025 deployment and Sutherland partnership reported 40% faster claim resolution, 35% claims-handling cost reduction, 60% faster general-queue email response, 70% accuracy improvement. Hi Marley ranked #244 on Deloitte's Tech Fast 500 in 2025. Shift Claims launched as agentic-AI claims orchestration; Covéa deployed it end-to-end in 2026 for fraud/risk/claims. This lesson maps every named claims AI platform - by surface, by category, by governance hook - with the 15-file Atlanta diary as the narrative anchor. By the end, you can walk into any claims room and name the AI on every desk.

Tractable - The Auto and Property Photo CV Anchor

Tractable is the 2026 reference for computer-vision-driven auto and property claims estimating. The Admiral Seguros case study put digital completion at 70-75% - the share of damaged-vehicle estimates closing on Tractable's AI-generated estimate without an in-person appraisal. The mobile-self-service workflow has the insured upload photos from a phone prompt; Tractable's CV returns ACV with a confidence band and per-zone damage attribution.

On the Atlanta auto file: Tractable returned $19,400 ACV on the 2019 Toyota Highlander; the insured produced three Atlanta-area dealer listings averaging $23,800; the dispute escalates the decision to comparable-vehicle review and salvage-vs-retain analysis. The override discipline - documented comparable-set review, local-market range check, escalation above threshold - is the L2 verification anchor. The L4 metric is "AI ACV accepted with documented insured concurrence," not raw acceptance rate.

Category and governance. Tractable's photo CV is predictive. NAIC §4 demands reason chain on the ACV figure separate from the score; Texas Insurance Code §541, Florida §624.155, and California Cumis/Brandt frame bad-faith exposure when AI estimates close without engaging legitimate disputes; the L4 algorithm-inventory entry names the model card, fairness assessment (geographic and vehicle-class disparate-impact), drift monitoring (AUC on ACV vs. settled), and override discipline.

CCC Intelligent Solutions - Severity, Litigation Prediction, and the 35,000-Shop Network

CCC Intelligent Solutions' 2026 footprint covers 35,000+ repair facilities and 350+ insurance companies. CCC's April 2026 report put total-loss frequency at a record 23.1% - meaningful because the speed of total-loss decisioning becomes a hard P&L lever. CCC's severity prediction, litigation prediction, and estimating workflows integrate with the Mitchell network and the repair-shop ecosystem.

On the Atlanta GL slip-and-fall file: CCC's litigation prediction rated the claim "moderate-high severity" with the ALAE trajectory supporting a $125K indemnity reserve and $50K mediation reserve. The model card includes KS in the 30-50 range and AUC 0.80-0.90 on validated cohorts. The supervisor's reserve recommendation memo cites the CCC prediction as a signal and enumerates the human-reviewable facts (the claim, the defense-counsel input, the ALAE trend, the jurisdictional comparables). NAIC §4 reason-chain discipline applies.

Category and governance. CCC's models are predictive; the workflow layer compounding them is increasingly agentic. The L4 algorithm-inventory entry names the deployment surface (severity, litigation, estimating), the model card per module, and the override discipline. The L3 verification step on a CCC severity flag is the supervisor's reserve memo enumerating facts independent of the score.

Shift Technology - Fraud, SIU, and the Agentic Claims Orchestration

Shift Technology's fraud and risk detection is the 2026 reference. Shift Claims - the agentic capability launched as end-to-end orchestration - combines predictive scoring with generative referral drafting and deterministic workflow routing. Covéa (UK) deployed Shift Claims in 2026 for end-to-end fraud/risk/claims integration. The underlying predictive layer blends GBMs with graph features: the third-party medical clinic appearing on five plaintiff-attorney-represented soft-tissue claims in 60 days is a graph signal; geographic clustering is another; ISO ClaimSearch prior-loss match is a third.

On the Atlanta auto file: Shift's score is 0.87 on the cluster signal. The SIU referral memo enumerates the human-reviewable facts - the clinic, the prior-loss match, the geographic cluster within a single mile, the shared attorney across five files. The Shift score is referenced as a signal that triggered the review; the facts justify the referral. NAIC §4 reason-chain discipline applies; the algorithm-inventory entry for Shift Claims names the variables, the bias-test exhibit, the SIU-manager review checkpoint.

Category and governance. Shift's fraud scoring is predictive; Shift Claims orchestration is agentic. NAIC §4, Colorado Reg 10-1-1, NY DFS Circular Letter 2024-7 proxy test, and Colorado SB 21-169 bias testing all apply. The 82% Hispanic-surname cluster failure from the previous lesson is the structural failure mode; the bias-test exhibit and SIU manager review are the structural mitigations.

Snapsheet, EagleView, and ClaimXperience - Virtual Inspection

Snapsheet's virtual-inspection workflow reports approximately 20% LAE reduction in mature deployments. EagleView's aerial roof reports cover roof age, material, area, and slope from satellite and drone imagery; the architectural-vs-three-tab miscall on the Atlanta hail file's $4,200 RCV difference is the canonical predictive override moment. ClaimXperience adds video-driven virtual inspection layered with the customer-self-service workflow.

Together they replace field-adjuster trips on 50-70% of property claims and 50-60% of auto claims; LAE per claim drops; cycle time compresses. The override discipline on edge cases - non-standard roof materials, mixed-material roofs, weather-obscured damage, non-OEM aftermarket modifications - is the L2 verification anchor. The L4 algorithm-inventory entries name each tool, the override threshold, and the field-adjuster escalation pattern.

Category and governance. All three platforms are predictive (computer vision on photos and aerial imagery). NAIC §4 reason-chain discipline applies on every dispute or override; the L3 escalation workflow surfaces edge cases to senior adjusters.

Hi Marley - The Conversational SMS Layer

Hi Marley is the 2026 reference for claims customer SMS. The platform produces LLM-tuned conversational SMS to insureds, supervised in a human queue, with transcripts preserved as structured records. The carrier's claim-handling tone shapes the fine-tune; the supervised queue catches off-tone outputs. Hi Marley ranked #244 on Deloitte's Tech Fast 500 in 2025; carrier deployments span the top-25 personal-lines book.

On the Atlanta water-damage file: the 47-message SMS thread is preserved in Hi Marley; the 11-day non-response from the insured triggers the cycle-time alert; the adjuster's outreach is drafted in Hi Marley with the carrier's tone applied. The L2 verification step is the adjuster's read of the drafted message before send; the L4 algorithm-inventory entry names the system prompt version, the temperature setting, and the fine-tune corpus discipline (sanitized of NPI per GLBA Safeguards).

Category and governance. Hi Marley is generative. State Unfair Claims Settlement Practices Act (UCSPA) regimes capture SMS interactions as records; NAIC §4 documentation applies on any adverse-claim-handling impact; GLBA Safeguards governs the NPI handling in the fine-tune corpus and the production message stream.

Five Sigma - The AI-Native Claims Core

Five Sigma is the 2026 AI-native claims core platform. Starr Insurance selected Five Sigma in 2025; the Five Sigma + Sutherland partnership announced 2025 supports claims modernization at scale. Customer reports of 40% faster claim resolution, 35% claims-handling cost reduction, 60% faster general-queue email response, and 70% accuracy improvement on Five Sigma deployments anchor the carrier-side ROI case.

On the Atlanta water-damage file: Five Sigma's auto-coverage summary cites ISO HO 00 03 when the dec page shows HO 00 05 - the wrong-edition failure from the previous lesson. The adjuster overrides with a documented file note; the L3 RAG architecture would have loaded the dec page and endorsement schedule into context, preventing the failure structurally. Five Sigma's draft FNOL summary on the water-damage intake captures the loss facts but misses the ordinance-or-law endorsement that bears on the rebuild estimate; the adjuster catches the gap via verification.

Category and governance. Five Sigma's draft summaries and coverage positions are generative; the workflow layer is increasingly agentic. NAIC §4 reason-chain discipline applies on every adverse coverage position; Colorado Reg 10-1-1 algorithm-inventory entry names Five Sigma with the system prompt archive, the override discipline, and the prompt-log retention pattern.

Roots Automation, Sprout.ai, and Clearcover - Intelligent Claims Automation

Roots Automation provides intelligent automation across the claims lifecycle - document classification, exception routing, task automation. Sprout.ai provides claim summarization and adjuster-facing generative assistance. Clearcover's claims-AI integration drives the customer-facing experience on a digital-first personal-lines book. Each compounds with the broader stack; carrier choice depends on existing platform investments and the carrier's digital-first vs. traditional posture.

Category and governance. Roots is predictive plus deterministic; Sprout is generative; Clearcover is full-stack across categories. Algorithm-inventory entries and NAIC §4 documentation apply uniformly.

ISO ClaimSearch - The Deterministic Reference Layer

ISO ClaimSearch is the industry-wide claim-history reference database. On the Atlanta SIU referral, the prior-loss match on the named insured for a similar 2023 injury is the deterministic signal that joins the human-reviewable fact list. The match is rule-bound (named insured, SSN, address, prior loss type); not predictive; not generative. The L2 verification step is the adjuster's cross-check on the match identifiers; the L3 escalation routes confirmed matches to SIU when the cluster supports it.

Category and governance. ISO ClaimSearch is deterministic. NAIC §4 reason-chain naming includes ISO ClaimSearch matches as documented facts; FCRA-adjacent disclosure may apply where consumer rights touch.

How the 15-File Atlanta Diary Traces the Platforms

Walk back to 9:00 a.m. Tuesday Atlanta. The 15 files name the platforms.

File 1 (Total-loss Highlander): Tractable returns $19,400 ACV; insured disputes with $23,800 comparable; salvage decision pending; subrogation triage against third-party driver's policy limits.

File 2 (Slip-and-fall GL): CCC's litigation prediction rates moderate-high severity; $42K ALAE on $30K reserve; defense counsel asks $125K indemnity + $50K mediation; supervisor's EOD memo cites CCC as signal and enumerates facts.

Files 3-5 (Three April hail claims): EagleView aerial reports on each; one's architectural-vs-three-tab miscall produces $4,200 RCV override; Snapsheet virtual inspection on customer-self-service workflow.

File 6 (Overdue ROR on CGL faulty workmanship): carrier enterprise LLM drafts the ROR with RAG over ISO CG 00 01 04 13 and the venue case-law corpus; adjuster verifies form edition, exclusion language, controlling case; file note logged.

File 7 (MHPAEA NQTL escalation): AI utilization-review tool denied 14 of 22 IOP requests vs. 4 of 22 medical/surgical; L&H claims supervisor flagged; 60-day quantitative parity test triggered; corrective-action memo drafted.

File 8 (Hi Marley SMS on water-damage): 11-day non-response from insured; cycle-time alert; outreach drafted; coverage summary from Five Sigma cites HO 00 03 when dec page shows HO 00 05 - adjuster overrides with documented file note.

Files 9-15 (Mix of routine and complex auto/property/GL with Tractable, CCC, EagleView, Shift, Hi Marley, Five Sigma, and the enterprise LLM touching each).

Across the 15 files, Shift's network signal on the soft-tissue cluster spans multiple claims; the SIU referral memo enumerates the third-party clinic, the prior-loss match (ISO ClaimSearch), the geographic clustering, the shared attorney. The supervisor's monthly loss-ratio review is due EOD; the AI-touched artifacts in each file feed the supervisor's narrative.

The L4 Algorithm Inventory for the Claims Operation

The L4 governance program documents every claims AI platform in the algorithm inventory: Tractable (model card, bias-test exhibit by vehicle class and geography, override discipline), CCC (severity and litigation modules separately documented, KS/AUC thresholds), EagleView (roof classification model with confidence routing), Snapsheet/ClaimXperience (virtual-inspection workflow and override patterns), Shift Technology + Shift Claims (fraud and orchestration, bias-test per cluster, SIU-manager checkpoint), Hi Marley (system prompt archive, fine-tune corpus sanitization, supervised queue protocol), Five Sigma (system prompt archive, RAG architecture, override discipline), Roots Automation (workflow rules + exception routing), Sprout.ai (system prompt archive), ISO ClaimSearch (deterministic reference, match-validation discipline). Each entry names the accountable executive (VP of claims, VP of SIU, chief claims officer, chief medical officer for L&H), the model card or system prompt, the bias-test exhibit, the drift-monitoring runbook, the override discipline, the FCRA / state UCSPA / MHPAEA chain attachment, and the incident-response runbook.

Key Takeaways

• The Atlanta 9:00 a.m. claims diary touches at least seven named AI products in the first ten minutes: Tractable, CCC, EagleView, Shift Technology, Hi Marley, Five Sigma, and the carrier's enterprise LLM. Plus ISO ClaimSearch as the deterministic reference layer and Snapsheet/ClaimXperience for virtual inspection. The lesson maps every platform by surface, category, and governance hook.
• Tractable's Admiral Seguros case put digital completion at 70-75%; CCC covers 35,000+ repair facilities and 350+ insurance companies with severity and litigation prediction (KS 30-50, AUC 0.80-0.90); CCC's April 2026 total-loss frequency hit a record 23.1%. Each predictive output requires an override discipline on edge cases (the architectural-vs-three-tab roof miscall is the canonical example); the L4 metric is documented insured concurrence, not raw acceptance.
• Shift Technology launched Shift Claims as agentic-AI claims orchestration combining predictive fraud scoring (GBM + graph features, AUC 0.80-0.90, KS 30-50), generative SIU referral drafting, and deterministic workflow routing. Covéa (UK) deployed end-to-end in 2026 for fraud/risk/claims. The SIU referral memo enumerates human-reviewable facts (clinic, prior-loss match, geographic cluster, shared attorney) separately from the Shift score per NAIC §4.
• Snapsheet publishes ~20% LAE reduction; EagleView covers aerial roof age/material/area/slope; ClaimXperience adds video virtual inspection. Together they replace field-adjuster trips on 50-70% of property and 50-60% of auto claims with documented override discipline on edge cases.
• Hi Marley's LLM-tuned conversational SMS - supervised in a human queue with preserved transcripts - covers the customer-facing claims interaction surface. Hi Marley ranked #244 on Deloitte Tech Fast 500 in 2025. State UCSPA regimes capture SMS as records; GLBA Safeguards governs the fine-tune corpus and the production stream.
• Five Sigma is the AI-native claims core platform. Starr Insurance deployed in 2025; the Sutherland partnership supports claims modernization at scale; published outcomes of 40% faster resolution, 35% cost reduction, 60% faster email response, 70% accuracy improvement. The wrong-edition failure (HO 00 03 vs. HO 00 05) and missing ordinance-or-law endorsement are the L3 RAG architecture's structural fix.
• Roots Automation, Sprout.ai, and Clearcover compound the stack - intelligent automation, claim summarization, full-stack digital-first. ISO ClaimSearch is the deterministic reference layer that joins predictive signals with audit-defensible facts.
• The L4 algorithm inventory documents every claims platform with named accountable executive, model card or system prompt, bias-test exhibit, drift-monitoring runbook, override discipline, regulatory chain attachment (FCRA, state UCSPA, MHPAEA), and incident-response runbook. The L5 board narrative ties combined-ratio movement, ALAE reduction, severity drift, cycle-time compression, and complaint-ratio drop to the stack's verification maturity.