The rate-filing memorandum is the document the DOI rate reviewer reads in detail, the actuarial certification is the signed attestation that creates personal professional responsibility under ASOP 41, and the bias-testing exhibit is the section that the consumer-protection division, the attorney general's office, and any future class-action plaintiff will examine first. The three artifacts move together - a strong filing memo without a bias exhibit invites disparate-impact challenge, a clean bias exhibit without filing-memo methodology is rate-foundation-thin, and a certification on an under-documented memo is professional-conduct exposure. The 2026 SERFF-ready package addresses NAIC Model Bulletin §4 governance, NY DFS Circular Letter 2024-7 proxy testing, Colorado Reg 10-1-1 algorithm inventory, and the state-specific filing requirements layered on each line. AI compresses the structural drafting, integrates the model card and SHAP/PDP exhibits from the Akur8-style workflow, and produces the bias-testing exhibit with disparate-impact ratios per variable and alternative-variable analysis where applicable. The pricing actuary signs the memo. The chief actuary signs the certification. This lesson is the workflow for producing the three artifacts to a standard that survives a hostile state DOI examination and a class-action discovery request three years later.

The Three Artifacts and Their Audiences

The rate-filing memorandum is read by the DOI rate reviewer - typically a credentialed actuary or actuarial analyst at the state insurance department who reviews the methodology, the variable selection, the rate-change calculation, and the supporting analysis. The reviewer's questions, follow-ups, and ultimate approval or disapproval flow from the memo. The actuarial certification is read by the DOI rate reviewer for compliance and by the state actuary, state attorney general, and class-action plaintiff for professional-responsibility documentation. The bias-testing exhibit is read by the rate reviewer for methodological adequacy and by the consumer-protection division, the AG's office, and litigation plaintiffs for disparate-impact documentation. Each audience reads differently; each section serves multiple audiences.

The 2026 best-practice structure has the memo as the methodological centerpiece (15-25 pages), the certification as a signed statement (1-2 pages) referring to the memo, and the bias-testing exhibit as a dedicated section (8-15 pages) with the disparate-impact ratios, proxy-testing methodology, alternative-variable analyses, and business-necessity documentation. The three artifacts cross-reference each other; the filing package is internally consistent and externally defensible.

The Rate-Filing Memorandum as the Methodological Centerpiece

The memorandum walks the DOI rate reviewer through ten sections: (1) Introduction and Purpose with explicit ASOP citations (23, 38, 41, 56); (2) Data Sources and Time Period with data-quality assertions and lineage; (3) Methodology with GLM family, link function, regularization details, and software platform (Akur8, Earnix, SAS Viya); (4) Variable Selection with per-variable LASSO results, actuary override decisions, and documented rationale; (5) Proxy Testing and Bias Analysis referencing the dedicated bias-testing exhibit; (6) Model Performance with holdout deviance, Gini, lift charts, and GBM comparator gap analysis per ASOP 38; (7) Rate-Change Calculation with methodology for translating model output to filed rates; (8) Comparable-Carrier Context with Akur8 Discover or Matrisk findings supporting filing posture; (9) Limitations and Operating Constraints with model lifecycle, monitoring discipline, and recalibration cadence; (10) Certification placeholder referencing the separate signed certification.

The actuary's voice is critical. The memo is signed by the appointed actuary and reflects the actuary's professional judgment; AI assists with structural drafting and exhibit integration, but the methodology narrative, the variable-selection rationale, and the bias-testing interpretation are the actuary's work product in the actuary's voice. SERFF reviewers can tell when a memo is AI-paste vs. actuary-authored - paste-language has tells (over-formal, generic phrasing, missing carrier-specific context); actuary-authored language reflects the carrier's data, the carrier's prior filings, and the actuary's specific judgment about this particular submission.

The Actuarial Certification and the ASOP 41 Discipline

The certification is the chief actuary's signed attestation. The standard ASOP 41 (Actuarial Communications) form has six elements: (1) identification of the actuary issuing the certification with credentials (FCAS, MAAA, or equivalent); (2) statement of the actuary's qualifications to issue the certification per the Code of Professional Conduct and the Qualification Standards; (3) statement of the work performed, including reliance on internal data, vendor data, and model platforms (Akur8, Earnix, SAS Viya); (4) statement of the methodologies applied and the ASOPs followed (23, 38, 41, 56, plus any state-specific or line-specific ASOPs); (5) statement of the conclusions reached and the basis for those conclusions; (6) signature, credential designation, date, and contact information.

The certification is short (typically 1-2 pages) but carries personal professional responsibility. The certifying actuary must have personal knowledge of the work - not delegated, not paste-acknowledged. The workflow's discipline supports the certification: the chief actuary reviews the memorandum, the model card, the bias-testing exhibit, and the supporting analysis before signing; the chain-of-custody of AI assistance is preserved but the certification reflects the chief actuary's personal review. AI assistance does not change the professional standard; it accelerates the work that the chief actuary personally evaluates.

The Bias-Testing Exhibit and the Disparate-Impact Ratios

The bias-testing exhibit is the section the consumer-protection division and the AG's office read first. The exhibit has six sections: (1) Methodology - the proxy-test methodology applied, BISG inference details, statistical-significance thresholds; (2) Per-Variable Disparate-Impact Ratios - for each rating variable, the ratio computed and the comparison to the state-specific threshold; (3) Actions Taken - for variables with material disparate impact, the action (drop, constrain, document business necessity); (4) Alternative-Variable Analyses - for retained proxy variables, the alternative-variable analysis required by NY DFS Circular Letter 2024-7 showing the variable's predictive power is not duplicable through proxy-free alternatives; (5) Business-Necessity Documentation - for retained proxy variables, the documented business necessity per state-specific framework; (6) Ongoing Monitoring - the discipline for monitoring disparate impact in production, the drift-detection triggers, and the response protocol if monitoring surfaces concerns.

For the worked-example personal-auto frequency model with credit-based insurance score retained at 1.40x cap: the exhibit shows the BISG methodology details, the 14% disparate-impact ratio (Black vs. White inferred), the constraint applied (1.40x cap), the alternative-variable analysis showing 3.2% Gini loss if dropped, the business-necessity documentation, and the monitoring discipline. The exhibit is internally consistent with the rate-filing memorandum, the model card, and the algorithm-inventory entry. Cross-references are explicit; the rate reviewer can trace any claim to its supporting documentation.

NY DFS Circular Letter 2024-7 Proxy Test in Writing

NY DFS Circular Letter 2024-7 (2024) introduced specific guidance on insurer use of external consumer data and artificial intelligence in underwriting and rating. The proxy-test framework imposes the alternative-variable analysis requirement: where a variable produces disparate impact, the insurer must demonstrate that proxy-free alternative variables produce materially worse predictive performance. The Circular Letter further specifies (a) the documentation expected in the SERFF filing, (b) the ongoing monitoring discipline, and (c) the response expected when monitoring surfaces drift.

For New York-filing personal-auto rate changes, the bias-testing exhibit must include a section specifically addressing the Circular Letter 2024-7 proxy test. Language template: "Per NY DFS Circular Letter 2024-7, the carrier has conducted alternative-variable analysis on retained variables that produce disparate impact above [threshold]. For credit-based insurance score: alternative-variable analysis tested telematics-derived behavioral predictors, vehicle-telematics-derived utilization predictors, and deeper claims-history features. The proxy-free alternative model produced [X]% Gini loss vs. the model with credit-based insurance score; the predictive performance loss meets the materiality threshold under the Circular Letter's framework. Business necessity is documented in [appendix]. Ongoing monitoring per the Circular Letter's framework: quarterly disparate-impact ratio testing, threshold-triggered escalation, annual recalibration review." The language is the actuary's; AI assists with structure but the substance is the actuary's professional judgment applied to the carrier's data.

The State-Specific Filing Overlay

Each state has its own filing requirements layered on the NAIC baseline. Colorado Reg 10-1-1 requires the algorithm-inventory entry attached to the filing. NY DFS Circular Letter 2024-7 requires the proxy-test framework. California's Insurance Commissioner has issued guidance restricting certain variables (credit-based insurance score prohibited in personal auto). Connecticut, Washington, Maryland, and others have specific guidance on disparate-impact testing. Texas, Florida, and others have state-specific rate-filing-form requirements. The AI workflow loads the state-specific overlay matrix and surfaces the requirements per filing.

The state-overlay matrix is a living document maintained by the actuarial team or the filing manager. Quarterly review picks up new bulletins, legislative changes, and DOI-published guidance; annual review aligns the matrix with the carrier's filing cadence. Multistate carriers maintain a unified matrix; single-state carriers maintain a state-specific matrix. The matrix discipline ensures every filing references the controlling state-specific requirements and the package addresses them in the appropriate sections. Carriers without matrix discipline file inconsistent packages across states; carriers with the discipline file with state-specific calibration.

The SERFF-Ready Package Assembly

The SERFF-ready package contains six items in a structured organization: (1) Cover Transmittal Letter - short letter addressed to the state DOI rate reviewer, summarizing the filing purpose and effective date; (2) Rate-Filing Memorandum - the methodological centerpiece; (3) Rate-Change Exhibit - the proposed rates by territory, class, and tier vs. current; (4) Bias-Testing Exhibit - the dedicated section on proxy testing and disparate-impact analysis; (5) Actuarial Certification - the signed attestation; (6) Supporting Appendices - model card, SHAP/PDP plots, alternative-variable analysis details, comparable-carrier intelligence summary, algorithm-inventory entry per state requirements. Each document references the others; the package is internally consistent and externally defensible.

AI accelerates the assembly: pulls the model output from the Akur8-style workflow, generates the memorandum structure, populates the bias-testing exhibit from the proxy-test results, drafts the cover transmittal, and references the algorithm-inventory entry. The pricing actuary reviews and customizes; the chief actuary reviews and certifies. The SERFF submission timing typically targets 90-120 days before the requested effective date to allow for examination and response cycles; carriers using AI-assisted workflows compress the internal preparation cycle (model build through package assembly) from 14-18 weeks to 4-8 weeks, allowing the carrier to file more responsively to loss-cost trends.

What the Pricing Actuary and Chief Actuary Each Own

The pricing actuary owns the model build, the variable-selection decisions, the bias-test results documentation, the model card, the memorandum drafting (with AI assistance), the alternative-variable analyses, and the algorithm-inventory entry updates. The chief actuary owns the actuarial certification under ASOP 41, the personal-knowledge review of the entire package, the professional-conduct assessment, and the relationship with the DOI rate reviewer during examination. The two roles do not overlap; both are non-delegable under the Qualification Standards. AI compresses the documentation and structural drafting; the professional judgment remains with the actuaries.

The rate-filing manager owns the SERFF submission mechanics - uploading the package, tracking the submission, monitoring examiner-response timing, escalating delays. The chief underwriting officer owns the implementation when the rate is approved - coordinating with PolicyCenter or Duck Creek to build the new rates, updating underwriting guidelines, training the underwriters. The general counsel or compliance officer owns the regulatory-defense posture if the filing is challenged or if the rate-implementation produces consumer complaints. AI accelerates each role's work; the role accountabilities remain distinct.

Key Takeaways

• The three rate-filing artifacts: rate-filing memorandum (methodological centerpiece, 15-25 pages, read by DOI rate reviewer), actuarial certification (signed attestation, 1-2 pages, under ASOP 41 personal responsibility), bias-testing exhibit (dedicated section, 8-15 pages, read by consumer-protection division and AG's office). The three artifacts cross-reference and must be internally consistent.
• The memorandum has ten sections: Introduction with ASOP citations, Data Sources, Methodology, Variable Selection, Proxy Testing and Bias Analysis, Model Performance, Rate-Change Calculation, Comparable-Carrier Context, Limitations, Certification placeholder. Actuary's voice is critical; SERFF reviewers can detect AI-paste vs. actuary-authored language.
• The actuarial certification under ASOP 41 has six elements: actuary identification with credentials, qualifications statement, work performed, methodologies and ASOPs followed, conclusions, signature with date. Personal professional responsibility; not delegable to AI; the certifying actuary must have personal knowledge.
• The bias-testing exhibit has six sections: Methodology (BISG inference, statistical thresholds), Per-Variable Disparate-Impact Ratios, Actions Taken (drop/constrain/document), Alternative-Variable Analyses per NY DFS 2024-7, Business-Necessity Documentation, Ongoing Monitoring. Read first by consumer-protection division and AG's office.
• NY DFS Circular Letter 2024-7 imposes the proxy-test framework: alternative-variable analysis showing proxy-free alternatives produce materially worse predictive performance, ongoing monitoring discipline, response protocol on drift. SERFF filing must include specific language addressing the Circular Letter's framework.
• State-specific filing overlay layers on NAIC baseline: Colorado Reg 10-1-1 algorithm inventory, NY DFS 2024-7 proxy test, California Commissioner guidance on prohibited variables, Connecticut/Washington/Maryland disparate-impact testing, Texas/Florida state-specific filing forms. Matrix discipline ensures every filing references controlling state-specific requirements.
• SERFF-ready package contains six items: cover transmittal, memorandum, rate-change exhibit, bias-testing exhibit, actuarial certification, supporting appendices. Internal preparation cycle compresses from 14-18 weeks (pre-AI) to 4-8 weeks (AI-assisted); SERFF submission timed 90-120 days before effective date.
• Pricing actuary owns model build, variable selection, bias-test documentation, memo drafting; chief actuary owns the certification, personal-knowledge review, professional-conduct assessment, DOI examination relationship. Roles non-delegable under Qualification Standards.
• NAIC Model Bulletin §4 chain-of-custody on every AI-touched artifact: model and version, prompt template, input documents, reviewer sign-off. Filing memo draft, bias-testing exhibit, comparable-carrier intelligence integration each enter the audit trail.