The pricing actuary's lifecycle for an AI-built personal-auto or commercial-property GLM in 2026 is no longer "build the model and hand it to filing." It is a ten-stage discipline that runs from data prep through Colorado Reg 10-1-1 quantitative fairness testing, NAIC Model Bulletin §4 documentation, a SERFF-defensible model card, monitoring KPIs, and retirement criteria written before deployment. Akur8 (transparent GLM with automated penalization, $220M Series C in early 2025, deployed at over 200 insurers including The Hartford, MS Amlin, AXA Direct), Earnix (dynamic pricing optimizer, deployed at Allianz, Aviva, Travelers), and Guidewire Garage AI (the experimentation tier inside InsuranceSuite Cloud, Predict tier overlay) are the three tools the pricing actuary most often runs against in 2026. This lesson walks the lifecycle on a worked example - a personal-auto BI severity GLM rebuild for a $1.2B premium book filed in CO, CA, NY, NJ - and shows what the model card, the bias-test exhibit, the monitoring plan, and the retirement memo actually contain. The headline rule: a model that cannot survive the Colorado Reg 10-1-1 quantitative testing exhibit in front of a Division of Insurance examiner does not get into production, no matter how clean the validation AUC looks.

The Lifecycle in Ten Stages

The lifecycle is data prep, feature engineering, candidate-model build (GLM, GBM, ensemble), out-of-time validation, fairness and proxy testing, model card and ASOP No. 41 communication, SERFF filing prep, parallel-run shadow period, production deployment with monitoring, and retirement-criteria pre-commitment. Each stage has named artifacts and named owners. The credentialed actuary signs ASOP No. 41 communication and ASOP No. 56 modeling attestation. The MRM (model risk management) function owns the registry entry and the drift monitoring. The Chief Actuary signs the SERFF filing and owns the production model. The naive sequence - build, validate, file - leaves the carrier exposed when a Colorado examiner asks for the unadjusted disparate-impact ratio across protected-class proxies and the answer "we didn't run that test" produces a market-conduct finding.

The personal-auto BI severity rebuild example. $1.2B written premium, 1.4M policies in force, BI severity GLM last refit in 2023, ASOP No. 23 data-quality review surfaces three years of new claims data, a new telematics partner (Cambridge Mobile Telematics from the LexisNexis acquisition), and the loss ratio drift inside Tier-2 zip clusters in Florida and Texas demands a refit. Project sponsor: VP of Pricing. Credentialed actuary: FCAS in pricing. MRM lead: ARM-Risk-trained MRM analyst. Filing manager: rate-filing analyst with SERFF credentials. Total project budget: $480K (Akur8 SaaS at $180K/year, internal labor, external bias-testing consultant). Timeline: 14 weeks from data freeze to first state filing.

Stages 1-2 - Data Prep and Feature Engineering

Data freeze on 12/31/2025. Three-year experience window 2023-2025; BI severity capped at $250K per claim (basic-limits cap for personal-auto GLM); 642,000 closed BI claims; 1.4M policies in force on the freeze date; exposure measured in earned car-years. Data sources: policy admin (Guidewire PolicyCenter), claims (Guidewire ClaimCenter), telematics events (Cambridge Mobile Telematics + Arity), credit-based insurance score (LexisNexis Risk Classifier), MVR (LexisNexis ChoicePoint), CLUE prior loss history, garaging-address geocode (PreciselyID), territory codes (proprietary).

The data-quality review under ASOP No. 23. Materiality threshold: any feature with greater than 0.5% missingness on the modeling sample requires documented imputation or exclusion. Telematics has 38% coverage at the bind point (new business with optional telematics participation); the actuary's choice is two-model architecture - a base GLM scored on full population, and an uplift layer trained only on telematics-participating policies. Credit score has 6% missingness (no-hit and thin-file insureds); imputation is the conservative neutral percentile. The data-quality memo lists every imputation rule, every exclusion threshold, every cap; it is the artifact ASOP No. 23 requires and the artifact the SERFF examiner reads first.

Feature engineering with Akur8. Akur8's Engine module produces candidate features automatically: penalty-tuned univariate transforms (binning of driver age, vehicle age, prior incidents), interactions (driver age × prior accidents, vehicle symbol × annual mileage), and time-decay weights for older accident history. The actuary's job is not to write the binning code; the actuary's job is to constrain feature selection to features that survive the fairness test. Three features that score well on AUC and lift get excluded before they reach the candidate model: garaging-zip clustered at the census-tract level with material correlation to majority-minority census tracts; education attainment (where allowed by jurisdiction but with proxy-test failure in NY DFS framework); occupation grouping with proxy correlation to gender. Akur8 logs each exclusion with reason code; the log is appended to the model card.

Stages 3-4 - Candidate-Model Build and Validation

Three candidate models. Champion: Tweedie GLM with log link, severity-only target, Akur8 penalization for variable selection. Challenger A: GBM (LightGBM) with same feature set, monotonic constraints applied on driver age (older drivers cannot increase severity prediction beyond a documented bound), prior accidents (more accidents must increase predicted severity), credit score (lower credit cannot reduce predicted severity). Challenger B: Earnix ensemble with GLM base and GBM residual layer. The candidate models are scored on out-of-time validation (2025 claims held out from training), with the production decision driven by Gini, calibration on each decile, segment-level lift in the top three risk segments, and material-error metrics on the bottom decile.

The validation results. Champion GLM produces Gini of 0.412 on out-of-time validation; challenger GBM produces Gini of 0.448; Earnix ensemble produces 0.451. Pure AUC argues for the ensemble. The model-selection memo argues for the champion GLM on four grounds: SHAP explainability for SERFF examiner walkthrough is materially simpler on a transparent GLM; ASOP No. 56 peer review burden is lower; the bias-test artifacts (disparate-impact ratio, geographic-proxy test, surname-geocoded proxy) compute cleanly on a GLM and require approximation methods on the ensemble; and the 0.04 Gini gap on the ensemble produces 1.3 loss-ratio points of estimated lift, which the carrier is unwilling to spend its filing-defense complexity on. The decision is documented. The challenger GBM is retained as the monitoring benchmark; if it materially out-predicts the champion in the live monitoring window, the retirement criterion triggers.

Stage 5 - Fairness and Proxy Testing

The Colorado Reg 10-1-1 (effective for personal-auto pricing models filed after 2024) requires quantitative testing of the model against protected classes for unfair discrimination. Colorado does not provide protected-class membership data, so the carrier must use a defensible proxy methodology. The 2026 industry-standard proxy is Bayesian Improved Surname Geocoding (BISG) augmented with the carrier's own validation; the alternative is the BIFSG (Bayesian Improved First-name Surname Geocoding) which adds first-name signal. NY DFS Insurance Circular Letter No. 1 (2019) on the use of external consumer data and Letter No. 7 (2025 update) on AI in underwriting and pricing impose similar proxy-testing expectations on New York filings.

The disparate-impact ratio computation. Run the production model scoring on the in-force book. Stratify by BISG-imputed race/ethnicity (probabilistic - every policy receives a vector of imputed probabilities across White, Black, Hispanic, Asian, Other). Compute the average predicted severity within each group, weighted by the imputed probability. The disparate-impact ratio is the ratio of average predicted severity for the lowest-predicted protected group to the highest-predicted reference group (the four-fifths rule from EEOC discrimination doctrine, imported into the insurance fairness framework). A ratio below 0.80 is presumptively problematic and requires either feature-set adjustment, model-form constraint, or a documented business-necessity defense. The auto BI rebuild produces DI ratios of 0.91 (Black vs. White reference), 0.94 (Hispanic vs. White), 0.97 (Asian vs. White). Above the 0.80 threshold; the model proceeds to filing.

The geographic-proxy test. Independent of BISG, the actuary tests whether geographic features (garaging zip, territory code, census-tract clusters) correlate materially with majority-minority census tract concentration. The test: compute correlation between feature value and the census-tract minority-share variable from ACS data; flag any feature with absolute correlation greater than 0.30. Three of the original territory features failed; they were dropped at the feature-engineering stage and the test is rerun on the final model to confirm. NY DFS proxy test applies the same methodology with a 0.25 threshold; the NY filing uses the tighter threshold.

SHAP, PDP, ALE. The explainability artifacts attached to the model card. SHAP (Shapley) values produce per-policy contribution decomposition - the SERFF examiner can ask "for this representative policyholder, what is each feature contributing to predicted severity?" and the actuary returns a quantitative decomposition. PDP (partial dependence plot) shows the marginal effect of each feature averaged across the book. ALE (accumulated local effects) corrects for feature-correlation bias that contaminates PDPs on a GLM with correlated inputs (driver age × vehicle age). All three are computed and attached to the SERFF supporting documentation; the Colorado examiner explicitly asked for ALE on the last filing cycle.

Stages 6-7 - Model Card and SERFF Prep

The model card is the single authoritative document on the model. Length: 35-55 pages. Sections: (1) model purpose and scope; (2) training and validation data, including ASOP No. 23 data-quality memo by reference; (3) feature list with definitions; (4) feature-engineering decisions and exclusions with reason codes; (5) candidate-model comparison and model-selection memo; (6) validation results with Gini, calibration plots, decile lift, segment lift; (7) fairness exhibit with DI ratios, geographic-proxy test, BISG methodology, sensitivity analysis; (8) SHAP, PDP, ALE artifacts; (9) monitoring plan with thresholds; (10) retirement criteria; (11) ASOP No. 41 communication; (12) ASOP No. 56 attestation by credentialed actuary; (13) ASOP No. 38 catastrophe-model integration if relevant; (14) signoffs.

The SERFF filing package. Rate filing memorandum (10-15 pages, prose), rate-change indication exhibit (the actuarial indications by territory and class with overall loss-ratio target reconciliation), bias-testing exhibit (the DI ratios and geographic-proxy results), model card (by reference, attached as supporting documentation), actuarial certification (one page, signed FCAS or ACAS), response-to-objection template (pre-drafted answers to the eight most common DOI objections - "explain the credit factor," "explain the credibility methodology on the territory factor," "demonstrate Colorado Reg 10-1-1 compliance," "explain how telematics participation is offered without coercion"). Akur8 Discover/Matrisk provides competitor-filing intelligence - the actuary can see the indicated rate-change pattern across the named-peer set for the same state-line of business and reposition the filing memo accordingly.

Stages 8-9 - Shadow Deployment and Monitoring

Before the production model takes underwriting authority, the model runs in shadow mode for 90 days. Every new business quote and every renewal scores against both the legacy model (production decision) and the candidate model (shadow). Differences over a defined threshold (greater than 12% predicted-loss-cost delta or greater than 4% rate-change at the policy level) are flagged for actuarial review. The shadow window also produces the operational data the carrier needs to confirm that pre-prod model inferences match training-time distributions - the PSI (population stability index) on each feature should be below 0.10 in the shadow window; PSI above 0.25 means the deployment population already differs materially from training population and the model requires retraining before going live.

The monitoring plan. Six dashboards: PSI on each feature with red threshold 0.25; KS statistic on the predicted-vs-actual claim-cost distribution with red threshold 0.20; AUC drift quarterly with red threshold 0.04 below validation; calibration drift in the top three deciles (predicted-actual ratio outside 0.90-1.10 triggers); segment-level loss ratio in territory and class clusters; business-impact metrics (overall loss ratio, hit ratio, retention). Each dashboard has a named owner, a refresh cadence (daily for PSI on telematics features, monthly for calibration), and a documented escalation path on red.

Stage 10 - Retirement Criteria, Written Before Deployment

The retirement memo is pre-committed at deployment. The model retires (planned full refit) when one of three conditions triggers: planned three-year refit cycle (this model retires on 12/31/2028 absent earlier trigger); a major regulatory change (e.g., Colorado expands Reg 10-1-1 to require additional protected classes, a new state adopts a binding AI law); or a monitoring trigger (PSI above 0.25 on two consecutive months on a material feature, AUC drift greater than 0.04 sustained, calibration outside band sustained, or the GBM challenger out-predicts the champion by greater than 2 loss-ratio points sustained). Documenting retirement criteria before production deployment is what separates a defensible model from an indefensible one in an MRM audit - Colorado examiners ask explicitly "when does this model retire and what triggers an earlier retirement?" The answer is in the memo, signed at deployment.

Akur8, Earnix, Guidewire Garage AI - The Tool Comparison

Akur8's strength is transparent GLM with automated penalization and a model-build workflow that the SERFF examiner can walk through feature-by-feature. The bias-testing module is mature; the Discover/Matrisk competitor-filing module is differentiated. Pricing in 2026 starts at $180K-$420K annual SaaS depending on premium volume; deployed at 200+ insurers globally. Earnix's strength is dynamic pricing optimization - price elasticity modeling, willingness-to-pay estimation, multi-objective optimization across loss ratio and retention. The actuary's role with Earnix is heavier on optimization-constraint specification (regulatory caps, fairness floors, retention guardrails). Pricing in 2026 starts higher, $350K-$900K annual, deployed at the Tier-1 personal-auto carriers (Allianz, Aviva, Travelers, Munich Re). Guidewire Garage AI is the in-platform experimentation tier inside InsuranceSuite Cloud - the carrier's actuarial sandbox for testing candidate models against historical PolicyCenter and ClaimCenter data. Strength: tight integration with the production stack, no second SOC 2 review. Weakness: the explainability and bias-testing tooling is shallower than Akur8's. The 2026 procurement question is rarely "Akur8 OR Earnix" - most Tier-1 carriers run Akur8 for transparent GLM build and Earnix for optimization downstream of the GLM, with Guidewire Garage as the integration layer.

Key Takeaways

• The lifecycle is ten stages, not three. Data prep, feature engineering, candidate-model build, validation, fairness testing, model card, SERFF prep, shadow deployment, production monitoring, retirement criteria. The naive build-validate-file sequence loses to a Colorado examiner asking for the unadjusted DI ratio.
• Akur8 transparent GLM beats GBM ensemble at filing time even when the ensemble's Gini is higher. The 0.04 Gini gap (0.448 vs. 0.412) is 1.3 loss-ratio points; the filing-defense complexity of explaining the ensemble to SERFF examiners is not worth it. Retain the GBM as monitoring benchmark; document the trade-off.
• Colorado Reg 10-1-1 requires quantitative fairness testing on any personal-auto pricing model. BISG-imputed disparate-impact ratios against four-fifths threshold (0.80 EEOC rule); auto BI rebuild produced 0.91, 0.94, 0.97 - above the threshold and proceeds. Below 0.80 requires feature adjustment, model-form constraint, or documented business-necessity defense.
• The geographic-proxy test runs independent of BISG. Feature-vs-census-tract-minority-share correlation above 0.30 flags (Colorado); above 0.25 flags (NY DFS). Three territory features dropped at feature engineering on this rebuild; retest on final model confirms.
• SHAP, PDP, ALE are non-negotiable on the model card. Colorado examiners explicitly asked for ALE on the last filing cycle. PDP alone is contaminated by feature correlation on a GLM with driver-age × vehicle-age inputs; ALE corrects it.
• The 90-day shadow window catches deployment-time PSI drift before production. Feature PSI greater than 0.10 triggers actuarial review; greater than 0.25 means retraining is required before go-live. The shadow window is not optional and the budget is $30K-$90K of analyst time.
• The monitoring plan is six dashboards with named owners, refresh cadence, and escalation paths. PSI, KS, AUC, calibration, segment-level loss ratio, business impact. Each has a red threshold pre-committed. The MRM analyst owns the dashboards; the chief actuary owns the escalation.
• Retirement criteria are written at deployment, not at retirement. Three-year planned refit, regulatory-change trigger, monitoring trigger (PSI on two consecutive months, AUC drift sustained, calibration outside band, GBM challenger out-predicts by greater than 2 LR points). Examiners ask "when does this model retire" and the answer must be in the model card.
• Akur8 + Earnix + Guidewire Garage is the most common 2026 Tier-1 stack. Akur8 for transparent GLM build and bias-testing; Earnix for dynamic-pricing optimization on top of the GLM; Guidewire Garage as the integration sandbox into PolicyCenter / ClaimCenter. Total run cost on a $1.2B book: $530K-$1.3M annual SaaS spend plus actuarial labor.