In a vendor demo at a top-25 carrier in March 2026, the chief underwriter watched a Cytora salesperson say "agentic AI" twelve times in eight minutes. The chief actuary, sitting one chair over, was running a GLM in Akur8 and trying to figure out why his SHAP plot showed roof age as the third-most-important feature when his own intuition said it should be second. Down the hall, the SVP of claims was reading a Shift Technology brief about a fraud network detected through GBM with an AUC of 0.87 and a KS of 41 - and on the next page, a Hi Marley case study about an LLM tuned to the carrier's tone via fine-tuning and grounded in retrieval-augmented generation against the claim-handling manual. Five rooms in the same building, five different uses of the word "AI," forty different technical terms in play, and no one on the executive committee can fluently translate them all. This lesson is the glossary that closes the gap. Forty terms - predictive, generative, and agentic - anchored to named 2026 insurance platforms (Cytora Autopilot, Federato RiskOps, Akur8 Discover, Earnix, Shift Claims, Tractable, CCC, Hi Marley, Five Sigma, Coalition Control 2.0, Send Flow, Convr Risk 360, Indico, Hyperscience Hypercell) and to named regulatory artifacts (NAIC AIA, Colorado ECDIS, FCRA adverse-action, MHPAEA NQTL, SERFF, Schedule P, ACORD 130, ISO CG 00 01, COPE, SOV, BI worksheet, ROR, EUO, BOR, AOR). Read once, you can read a DOI bulletin, a vendor RFP response, a reinsurance treaty AI clause, and a SERFF rate filing without flinching.

The Three Categories, Named by Platform

Before the glossary, the platform map by category - because the term "AI" in 2026 insurance is meaningful only when you can point at which species is doing the work.

Predictive AI platforms. Akur8 (transparent GLM/GBM pricing platform, post the January 2026 Matrisk acquisition that added SERFF filings intelligence and the RSM regulatory partnership), Earnix (dynamic pricing and real-time AI rating engine), Cytora's appetite scoring layer, Federato RiskOps' triage and portfolio analytics, Convr's Risk 360 enrichment scoring, Shift Technology's fraud detection (and the predictive layer of Shift Claims), Tractable's auto and property computer vision, CCC Intelligent Solutions' severity and litigation prediction, EagleView's aerial imagery and roof-age classification, Coalition's external attack-surface scoring via Control 2.0, Munich Re's risk-assessment platform, Swiss Re Magnum, RGA AURA NEXT, SCOR Velogica, LexisNexis Risk Solutions for L&H, ICEYE for parametric flood data, Guidewire Predict.

Generative AI platforms. Hi Marley (LLM-tuned claims SMS), Five Sigma's auto-coverage summaries and FNOL drafts, Send (renewal stewardship narrative drafting through Send Flow), Outmarket (wholesale broker narrative and cover email generation), Sprout.ai's claim summarization, the carrier's enterprise LLM behind Microsoft Copilot for M365, Azure OpenAI with data residency, AWS Bedrock with Anthropic Claude, Google Vertex with no-training contractual terms. Hyperscience Hypercell's IDP with Claude on Bedrock is hybrid - predictive extraction with generative reconciliation. Indico is similar.

Agentic AI platforms. Cytora Autopilot (the post-Applied Systems acquisition agentic capability, named a 2026 AI Excellence Award winner with reports of 4-6x submission throughput per underwriter), Federato's agentic UW workbench, Shift Claims (Shift Technology's agentic-AI claims orchestration capability; Covéa deployed it 2026 for end-to-end fraud/risk/claims), Earnix's agentic decisioning across UW-pricing-rating-personalization, and the next-wave Five Sigma claims-agent functionality. Agentic compounds the categories - predictive scoring plus generative drafting plus deterministic workflow orchestration - into a multi-step task executor with human-oversight checkpoints.

Terms 1-10 - The Model Mechanics

1. GLM (Generalized Linear Model). The credentialed-actuary's workhorse. A statistical model with a linear predictor and a link function (log for Poisson frequency, log for gamma severity, logit for binary classification). Akur8's transparent-GLM platform is the 2026 reference. Coefficients are interpretable; bias-testing exhibits are readable in SERFF; the model card lives in the algorithm inventory under Colorado Reg 10-1-1.

2. GBM (Gradient Boosting Machine). A tree-ensemble model - XGBoost, LightGBM, CatBoost - that achieves higher accuracy than a GLM on most insurance datasets at the cost of interpretability. Akur8 supports GBM alongside GLM with SHAP-based explainability. CCC's litigation-prediction models use GBM variants. Shift Technology's fraud network detection blends GBM with graph features.

3. SHAP (SHapley Additive exPlanations). The 2026 standard for predictive-model explainability. Computes the contribution of each feature to each prediction. SHAP values for the Dallas commercial property GLM might show territory contributing +12% to expected loss cost, roof age contributing +8%, sprinkler contributing -15%. NAIC AI Systems Evaluation Tool Exhibit C high-risk-system documentation expects SHAP or equivalent. NY DFS Circular Letter 2024-7's proxy-test analysis uses SHAP at the variable level.

4. Embedding. A learned vector representation of a token, phrase, document, or image. Convr's Risk 360 embeds COPE narratives into a vector space so similar narratives cluster. RAG architectures retrieve via embedding similarity. Hyperscience Hypercell uses embeddings for ACORD-form layout classification.

5. RAG (Retrieval-Augmented Generation). The architecture pattern that grounds a generative model in a curated knowledge base. At prompt time, the system retrieves relevant chunks (ISO form excerpts, claim-handling-manual sections, prior SERFF filings, treaty wording paragraphs) and includes them in the context window. The LLM drafts grounded in the retrieved content. The L3 RAG-and-persona chapter builds this out for the carrier's appetite guide, claim-handling manual, market access list, and ASOP No. 23/38/41/56 library.

6. Agent. An AI system that can take multi-step action across tools, with reasoning between steps. Cytora Autopilot orchestrates intake → IDP → enrichment → appetite scoring → triage routing → UW workbench update → quote generation; Shift Claims orchestrates FNOL → coverage check → fraud score → SIU referral draft → human review checkpoint. Agentic systems compound predictive, generative, and deterministic surfaces; each step has its own governance hook.

7. Fine-tune. Continuing training of a base model on domain-specific data to adapt behavior. Hi Marley fine-tunes against carrier-specific claim-handling tone and vocabulary. Fine-tuning carries data-leakage risk if the training corpus contained NPI or PHI; carrier governance under GLBA Safeguards and HIPAA requires the fine-tune corpus to be sanitized.

8. Drift. The change in a model's behavior or environment over time. Three categories matter on the insurance desk: data drift (incoming-feature distribution shifts), concept drift (target-variable distribution shifts), and model drift (the deployed model's calibration moves). PSI, KS, AUC are the typical detection metrics. The L3 drift-monitoring runbook combines them with business-impact drift (loss-ratio movement, hit-ratio movement, severity trend).

9. Proxy variable. A feature that correlates strongly with a protected class without naming it. ZIP code as a stand-in for race; credit-based insurance scoring as a stand-in for protected class; vehicle-make as a stand-in for income or geography. NY DFS Circular Letter 2024-7's proxy test demands written analysis of any variable that may proxy a protected class. Colorado SB 21-169 quantitative bias-testing operationalizes the test.

10. Disparate impact. A protected class experiences materially worse outcomes from the model even without explicit use of the class. The disparate-impact ratio (outcome rate for the disadvantaged group / outcome rate for the advantaged group) is the standard metric; thresholds below 0.8 typically trigger remediation. Colorado Reg 10-1-1 algorithm inventory entries name the disparate-impact analysis and the mitigation.

Terms 11-20 - Governance and Adverse Action

11. Adverse action. Any decision that materially harms an applicant or insured - decline, non-renew, rate increase, claim denial, SIU referral, accelerated-UW knockout. NAIC Model Bulletin §4 demands documented, human-reviewable reasons. FCRA §615 requires a pre-notice and an adverse-action letter when consumer-report data was used.

12. Model card. A structured document that records a model's intended use, training data, performance metrics, fairness assessment, limitations, accountable owner, and version history. Required for Colorado Reg 10-1-1 algorithm inventory; expected for NAIC AI Systems Evaluation Tool Exhibit C; standard practice in carrier model-risk-management programs.

13. System prompt. The carrier-tuned instructions placed before user input that shape the generative model's role, constraints, tone, and format. Send Flow's renewal-narrative system prompt is product-team-tuned across the carrier's seven target markets. The L2 prompt-engineering chapter builds carrier-specific system prompts for UW, claims, producer, and actuarial roles.

14. Context window. The maximum tokens an LLM attends to at once. 32K to 1M+ tokens in 2026 commercial deployments. Endorsement schedules outside the window are invisible; the L3 RAG architecture exists to handle artifacts that exceed the window.

15. Hallucination. Fluent, confident, factually wrong generative output. The four insurance-desk patterns: fabricated ISO/AAIS form numbers (CG 00 03), fabricated case citations, fabricated bulletin numbers (NY DFS 2024-9 instead of 2024-7), fabricated cost basis or sublimits.

16. Prompt injection. Hidden instructions in user content (PDF metadata, document text layers) that hijack a downstream LLM's behavior. IDP input sanitization at Hyperscience, Indico, and Cytora's intake layer is the structural mitigation.

17. Jailbreak. A crafted user prompt that bypasses the system prompt's constraints. Mitigation is layered: hardened system prompts, output classifiers, audit-log surveillance, user-permission scoping.

18. Fairness metric. A quantitative measure of disparate treatment or impact across protected classes. Demographic parity, equalized odds, calibration within groups, and the disparate-impact ratio. The Colorado Reg 10-1-1 bias-test exhibit selects and documents the relevant metric per use case.

19. Calibration. A model's predicted probability matches the empirical frequency. A predictive frequency model that says 5% but produces 8% actual losses on the cohort is mis-calibrated. Calibration plots and the Brier score are the standard tools. Drift in calibration is the canary on a deployed pricing model.

20. PSI (Population Stability Index). A drift metric that measures how much an incoming feature distribution has shifted from training. PSI > 0.25 is the typical alert threshold; PSI > 0.10 warrants investigation. The L3 drift-monitoring runbook tracks PSI per feature per quarter.

Terms 21-30 - Performance Metrics and Actuarial Anchors

21. KS (Kolmogorov-Smirnov statistic). Measures separation between two distributions - typically positive vs. negative classes in a binary classifier. Higher KS = better separation. CCC's litigation-prediction model and Shift's fraud-network detector cite KS in their model cards.

22. AUC (Area Under the ROC Curve). Binary-classifier performance, ranging 0.5 (random) to 1.0 (perfect). Shift's fraud models commonly cite AUC 0.80-0.90 on validated cohorts. AUC alone is insufficient - calibration, lift, and business impact accompany it.

23. AIA (Algorithmic Impact Assessment). A structured assessment of a model's risk, used in some governance frameworks and required for high-risk systems under the NAIC AI Systems Evaluation Tool Exhibit C. Sister concept to a model card but broader: includes harm scenarios, mitigation, and accountability assignments.

24. ECDIS (External Consumer Data and Information Sources). Colorado Reg 10-1-1's term for the third-party data sources a carrier uses in life-insurance underwriting - credit, prescription history, MIB code, MVR, public records, behavioral data. ECDIS inventory is required; bias-testing per source is required; consumer disclosure of adverse routing is required.

25. NQTL (Non-Quantitative Treatment Limitation). A Mental Health Parity and Addiction Equity Act (MHPAEA) term for limits on behavioral-health benefits that are not numeric (utilization review, prior authorization, medical-necessity criteria). When an AI utilization-review tool denies behavioral-health claims at a rate disparate from medical/surgical benchmarks (the Atlanta scenario flagged 14 of 22 IOP requests vs. 4 of 22 medical/surgical), an NQTL analysis is triggered.

26. ALAE (Allocated Loss Adjustment Expense). Defense counsel, expert witnesses, court fees, mediation costs - expenses tied to a specific claim. The Atlanta GL slip-and-fall file had $42K ALAE on a $30K reserve, prompting the defense panel's $125K indemnity reserve recommendation plus $50K mediation reserve. AI-assisted ALAE projection is an L2 reserving topic.

27. IBNR (Incurred But Not Reported). Losses that have occurred but not yet been reported to the carrier. Schedule P loss development triangles drive IBNR estimation. AI-assisted commentary flags anomalous development factors; the credentialed actuary owns the estimate under ASOP No. 56.

28. ACV (Actual Cash Value). Replacement cost minus depreciation. Tractable's photo damage estimate produced $19,400 ACV on the 2019 Toyota Highlander in the Atlanta scenario; the insured disputed the comparable-vehicle set. AI-assisted ACV is a predictive output; the policy form (HO 00 03 vs. HO 00 05) determines whether ACV or RCV applies.

29. RCV (Replacement Cost Value). The cost to replace the damaged property with new property of like kind and quality. EagleView's roof reports drive RCV on hail and wind claims. The $4,200 RCV difference between architectural and three-tab shingles on the Atlanta hail file is the canonical predictive-model override moment.

30. COPE (Construction, Occupancy, Protection, Exposure). The four pillars of commercial property risk. "Construction: mostly masonry, some frame; Occupancy: mixed retail/residential; Protection: Class 4 PPC, sprinklered except buildings 12 and 14; Exposure: no significant" - the Dallas COPE narrative parsed by Convr's NLP layer into structured features for Akur8's GLM.

Terms 31-40 - Artifacts and Distribution

31. SOV (Schedule of Values). The location-by-location spreadsheet on a multi-location commercial property submission. 47 buildings in the Dallas scenario, $182M TIV, 18 rows with missing roof age. The IDP layer (Hyperscience, Indico) extracts SOV fields at 99%+ accuracy; the underwriter handles below-threshold confidence rows.

32. BI (Business Interruption). Lost net income plus continuing expenses while a damaged business is being restored. The Dallas Business-Income worksheet had a crossed-out gross-earnings line with a higher number written in pen - the classic broker-edit pattern that an LLM-drafted summary will gloss over without prompting the underwriter to ask why.

33. ROR (Reservation of Rights). The carrier's notice to the insured that coverage is being investigated and the carrier reserves the right to deny in full or part. The Atlanta CGL file had a three-day-overdue ROR on the faulty-workmanship-exclusion question. LLM-drafted ROR letters require grounding (paste the actual form, name the venue case) and human verification.

34. EUO (Examination Under Oath). A formal sworn statement taken from an insured during a claim investigation, typically on suspect first-party theft, fraud, or material-misrepresentation claims. EUO outlines drafted with LLM assistance follow the L2 carrier playbook.

35. SERFF (System for Electronic Rate and Form Filing). The NAIC's electronic filing system. SERFF rate-filing memoranda for predictive pricing models reference model cards, bias-testing exhibits, and proxy-test analyses. Akur8 Discover's NLP search across SERFF filings is the 2026 reference for comparative-filing intelligence; the January 2026 Matrisk acquisition consolidated this capability.

36. BCAR (Best's Capital Adequacy Ratio). AM Best's capital-adequacy metric, used in the Best's Credit Rating process. AI capability under the April 2026 Best's Special Report is a survey-and-readiness assessment input - not a standalone rating product. Carriers calibrate their AM Best narrative to readiness categories (data, governance, cyber, legacy integration).

37. ORSA (Own Risk and Solvency Assessment). A carrier's documented assessment of its own risk profile and capital adequacy under various stress scenarios. AI-driven cat modeling, AI-influenced pricing, and AI-touched claims handling each appear in the ORSA narrative. FCAS Exam 9 (Financial Risk and Rate of Return) overlaps with AI-assisted capital modeling.

38. MGA (Managing General Agent). An intermediary that underwrites and administers business on behalf of a carrier under delegated authority. MGA bordereau reporting, fronting agreements, and the VIPR platform for delegated-authority compliance are the 2026 reference points. Vertafore's 2026 MGA Outlook documents the operational and credit exposure for fronting carriers.

39. BOR (Broker of Record letter). The instrument by which an insured transfers their account from one broker to another. State-specific timing rules (incumbent's 5-day or 10-day cooling-off), e-signature requirements, and notice substance all bind. LLM-drafted BORs require venue-state verification.

40. AOR (Agent of Record). The agent designated to service a policy; conceptually adjacent to BOR but typically used for agent-side rather than broker-side relationships. AOR cycles, BOR cycles, and POR (Producer of Record) notes are the producer's daily distribution-paperwork surface.

How the Glossary Lands on the Three Tuesdays

Walk back to the three desks. The terms snap into place.

Dallas, 8:14 a.m. The Cytora triage score (predictive) is read against the appetite guide; the Federato RiskOps view shows the Tier-1 wind aggregate (predictive plus deterministic treaty constraints); the Akur8 GLM produces a rate indication with SHAP attributions (predictive, GLM term 1, SHAP term 3); the COPE narrative (term 30) is parsed by Convr; the SOV (term 31) has 47 rows; the BI worksheet (term 32) shows the crossed-out gross-earnings; the LLM drafts the quote-with-restriction memo (generative, with system prompt and temperature, terms 13-14). The chief underwriter signs the memo with documented reason codes (term 12 model card + term 18 fairness metric + term 23 AIA).

Atlanta, 9:00 a.m. The 15-file diary surfaces Tractable's ACV (term 28) on the Highlander; CCC's litigation prediction (GBM term 2, KS term 21, AUC term 22) on the slip-and-fall GL with $42K ALAE (term 26) on a $30K reserve; EagleView's RCV (term 29) on the three hail claims with the architectural-vs-three-tab override; the Reservation-of-Rights (term 33) overdue on the CGL faulty-workmanship question; the Mental Health Parity NQTL flag (term 25) on 14 of 22 IOP denials vs. 4 of 22 medical/surgical; the Hi Marley SMS (generative) on the water-damage claim; the Shift fraud signal (predictive plus graph features) with the SIU referral that has to enumerate human-reviewable facts. ISO ClaimSearch's prior-loss match (term 11 adverse-action documentation). The adjuster's file notes name each AI by category.

Boston, 4:00 p.m. The Send-drafted renewal narrative (generative, RAG term 5 over carrier appetite guides); the Coalition cyber risk score (predictive) with the affirmative AI endorsement; the four-carrier comparison with sublimit, deductible, endorsement, exclusion differences; the BOR letter (term 39) draft; the AOR cycle (term 40) on the certificates; the chatbot disclaimer for the agency website (generative governance, NAIC §4 documented). The CCO's supervisory procedure attaches at every AI-generated client communication.

The actuary in the back office is running the Akur8 GLM (terms 1, 3), comparing to a GBM (term 2), running a SHAP analysis (term 3) for the SERFF filing (term 35), pulling comparable filings via Akur8 Discover post-Matrisk, drafting the rate-filing memo with the LLM (generative), documenting the NY DFS Circular Letter 2024-7 proxy test (term 9), running the Colorado Reg 10-1-1 disparate-impact analysis (term 10), and signing the actuarial certification under ASOP No. 41. The Schedule P (term 27 IBNR) triangle is in the next workbook; the ORSA narrative (term 37) is due to the CRO in two weeks; the BCAR (term 36) story for AM Best is in the appendix.

Key Takeaways

• The 40-term glossary is the working vocabulary of 2026 insurance AI. Every DOI bulletin, vendor RFP, treaty AI clause, SERFF filing memo, and NAIC AI Systems Evaluation Tool exhibit uses subsets of it. Fluency is the foundation of the L1 corridor.
• Predictive AI platforms - Akur8, Earnix, Cytora's appetite layer, Federato RiskOps, Convr Risk 360, Shift Technology, Tractable, CCC, EagleView, Coalition Control 2.0, Munich Re, Swiss Re Magnum, RGA AURA NEXT, SCOR Velogica, Guidewire Predict - output a number, score, or class. Governance attaches at the model card, bias test, drift monitoring, and algorithm inventory.
• Generative AI platforms - Hi Marley, Five Sigma's FNOL drafts, Send Flow, Outmarket, Sprout.ai, the carrier's enterprise LLM (Copilot for M365, Azure OpenAI, Bedrock with Claude, Vertex with no-training terms) - output text. Governance attaches at the system prompt, temperature, RAG architecture, citation-fact-check suite, and prompt-log archive.
• Agentic AI platforms - Cytora Autopilot (4-6x submission throughput per underwriter, 2026 AI Excellence Award winner), Federato's agentic UW workbench, Shift Claims (Covéa 2026 deployment), Earnix agentic decisioning - orchestrate multi-step tasks across predictive, generative, and deterministic surfaces with human-oversight checkpoints. Governance attaches at every step.
• The actuarial vocabulary - GLM, GBM, SHAP, calibration, PSI, KS, AUC, ALAE, IBNR, ACV, RCV, COPE, SOV, BI, BCAR, ORSA - is non-negotiable for reading SERFF filings, Schedule P narratives, and reinsurance treaty negotiations. The L2 actuarial chapters build operational fluency; L1 secures the read-without-flinching threshold.
• The governance vocabulary - proxy variable, disparate impact, adverse action, model card, system prompt, context window, hallucination, prompt injection, jailbreak, fairness metric, AIA, ECDIS, NQTL - operationalizes NAIC Model Bulletin §4, Colorado Reg 10-1-1, NY DFS Circular Letter 2024-7, MHPAEA, and FCRA §615. The L4 governance chapters draft the Exhibits A-D response packet against this vocabulary.
• The artifact vocabulary - ROR, EUO, SERFF, ACORD 125/126/127/130/140, ISO CG 00 01, HO 00 03, BOR, AOR, POR, MGA bordereau, fronting agreement - is the surface every AI tool operates on. The L2-L5 corridor walks each artifact through AI-assisted production with verification chains attached.
• The three Tuesdays - Dallas UW 8:14 a.m., Atlanta claims 9:00 a.m., Boston producer 4:00 p.m. - are the worked examples that anchor every term. Read once, you can name the AI on your own desk and the governance regime that binds it. That is the L1 outcome.