The SERFF (System for Electronic Rate and Form Filing) submission is the last gate between a model and the production rating engine. NAIC's SERFF processed approximately 220,000 filings in 2025 across the 50 state DOIs; the median time-to-decision on a personal-auto rate filing was 47 days in 2025 (range 22-180 by state); the average filing has 2.4 objections during review; the eight DOIs that examine pricing-model fairness most aggressively in 2026 are Colorado, New York, California, Connecticut, Washington, Massachusetts, New Jersey, and Illinois. This lesson walks the AI-assisted rate-filing workflow on the personal-auto BI severity GLM rebuild from L3 Lesson 15: drafting the rate filing memorandum, the actuarial certification per ASOPs 23/41/56, the bias-testing exhibit per Colorado Reg 10-1-1 and NY DFS Circular Letter 2024-7, the rate-change indication exhibit, the response-to-objection template; using Akur8 Discover and Matrisk (acquired by Akur8 in January 2026) for competitor-filing intelligence and DOI objection-history pattern matching; running the SERFF submission across a multi-state filing (CA, CO, NY) with state-specific objection patterns and approval timelines. The headline rule: a filing memo that anticipates the eight most common DOI objections gets approved 30-60 days faster than a filing that responds to them ad hoc, and the difference is roughly $4-12M of premium-collection delay on a $1.2B book.

The Multi-State Filing Context

The personal-auto BI severity GLM rebuild filed simultaneously into Colorado (Reg 10-1-1 quantitative fairness with the October 15, 2025 expansion and the July 1, 2026 first compliance report), California (Prop 103 prior-approval state with the most intensive actuarial review and a California-specific ban on credit-based insurance scoring), and New York (DFS Circular Letter 2024-7 effective July 11, 2024 plus the legacy Circular Letter No. 1 plus Regulation 64). Each state has a different filing pattern: Colorado requires the bias-testing exhibit in the filing form with the AISET Exhibit A bias-testing prompts addressed; California requires Class Plan exhibits with the rating-factor table and the Prop 103 reasonable-classification justification plus the no-credit-score compliance; New York requires the DFS proxy-test results per Circular Letter 2024-7 §III, the source-data review per Letter No. 1, the model card by reference with the commit hash, and the senior-management governance report. The same GLM, three different filing memos with significant content overlap and significant state-specific divergence.

The filing-team org chart. Filing manager (SERFF-credentialed, owns the submission and the multi-state coordination). Pricing FCAS (drafts the memorandum, signs the actuarial certification under ASOP 41). MRM analyst (provides the bias-testing exhibit and the model card with the version control). Chief actuary FCAS, MAAA (signs the filing actuarial certification, owns the production model, signs the ASOP 56 modeling attestation through the external peer reviewer engagement). General counsel (reviews the filing prior to submission, signs off on consumer-disclosure language if applicable, reviews the response-to-objection template for legal exposure). State-relations lead (liaison with each DOI's market-conduct examiner, maintains the relationship register and the prior-filing history per state).

AI Drafting the Rate Filing Memorandum

The rate filing memorandum is the 10-15 page prose document that sits at the top of the filing. Sections: (1) Purpose and Scope; (2) Indicated Rate Change Summary; (3) Methodology Overview; (4) Data and Experience Period; (5) Trend, Development, Catastrophe Load; (6) Indications by Coverage and Territory; (7) Rating Plan Changes; (8) Bias-Testing Summary; (9) Model-Governance Statement; (10) Effective Date and Implementation. AI-drafting workflow: the pricing FCAS feeds Akur8's filing-generation module the candidate model from Akur8 Deploy, the rate-change indication exhibit, and a set of constraints (state-specific language, prior filing references, DOI-objection history from Matrisk). The module produces a 12-page draft in roughly 4 minutes. The FCAS then spends 6-10 hours customizing - adding the company-specific narrative, reconciling against the company's prior filing language, and inserting state-specific disclosures.

Where AI saves time and where it does not. AI saves time on structure (consistent section ordering, regulatory language patterns, ASOP-citation placement), boilerplate (data sources, methodology overview, trend rationale, the standard sections on data quality and model governance), and cross-state consistency (the same model, three filings with state-specific overlays managed by the AI rather than by manual cross-walk). AI does not save time on the company-specific narrative (why this book differs from competitors, how the company's experience supports the indication, why specific decisions were made about coverage, territory, or class definitions). The FCAS must own the narrative. The 30-50% time savings on the memo (from 18-24 hours manual to 10-14 hours AI-assisted) is real and audited at 2026 Tier-1 carriers; it does not eliminate the FCAS's role; it lets the FCAS spend more time on the narrative and less on the structural drafting.

The Akur8 ecosystem in 2026. Akur8 Rate Repo stores the candidate models with version control; Akur8 Deploy puts the approved model into production with monitoring; Akur8 Discover provides competitor-filing intelligence; Matrisk (acquired January 2026) catalogs the DOI objection history across all 50 states. The RSM partnership and AAIS partnership announced in 2026 expand Akur8's reach into the mid-market and the small-carrier segment. The Branch Insurance case study (published 2026) documented the rate-filing cycle compression from 6-8 weeks to 3-4 weeks with the full Akur8 stack.

The Actuarial Certification

The actuarial certification is the one-page document signed by the chief actuary (or designated FCAS/ACAS) stating that the filed rates are not excessive, inadequate, or unfairly discriminatory, and that the rates comply with the state's rating laws. ASOP No. 41 governs the communication; ASOP No. 12 governs the risk classification; ASOP No. 56 governs the modeling. The certification cites the model card by reference (with the commit hash), cites the bias-testing exhibit by reference, and states the credentials of the signing actuary (FCAS, MAAA, with the year of fellowship). Filing rejection on certification defects is common: missing reference to ASOP No. 12 (California), missing credentials statement (New York), missing date of certification (multiple states), missing reference to the rating laws of the filing state.

AI-drafting the certification. The certification is highly templated. Akur8's filing module produces the draft with the correct ASOP references, the correct credential blocks, the correct state-specific rating-law citations (California Insurance Code §1861, Colorado Revised Statute §10-4, New York Insurance Law §2304-2310). The chief actuary reviews and signs. Total time: 30 minutes to review, 5 minutes to sign. Pre-AI: 2-4 hours of FCAS time drafting and checking citations against the current ASOP publication and the current state insurance code.

The peer-review dimension. ASOP 56 modeling attestation typically requires an independent peer reviewer for material model changes. The peer reviewer (an external FCAS or MAAA not employed by the carrier) signs a separate attestation. The peer-review engagement is structured at model release; the verification log and the model card become the peer reviewer's primary working papers. The peer-review attestation attaches to the filing as a supporting document.

The Bias-Testing Exhibit

The bias-testing exhibit is a separate document attached to the filing. Length: 12-20 pages. Sections: methodology (BISG - Bayesian Improved Surname Geocoding, geographic proxy, sensitivity), in-force book disparate-impact ratios, sensitivity analysis, geographic proxy results, feature-level SHAP contribution to disparate impact, model-form constraints, explainability artifacts (SHAP, PDP, ALE - Accumulated Local Effects). Colorado Reg 10-1-1 requires the exhibit; NY DFS Circular Letter 2024-7 requires equivalent disclosure with the proxy-test methodology specified; California has signaled in 2026 it will require similar documentation under SB 1058 (passed 2025). The exhibit is not boilerplate - the numbers are specific to the model and the in-force book at the filing snapshot.

The exhibit's purpose at filing. Demonstrate that the model has been tested for unfair discrimination using accepted methodology and that the results pass the regulatory threshold. The exhibit does not eliminate examiner questions; it preempts the threshold question (has the carrier tested?) and shifts the conversation to specific feature-level decisions (why was credit-score retained at this magnitude given the SHAP contribution to disparate impact?). The eight most-aggressive DOIs (CO, NY, CA, CT, WA, MA, NJ, IL) all expect this exhibit by 2026; filings without it draw an objection requesting it within 14 days of submission and add 14-30 days to the approval cycle.

The NAIC AISET Exhibits A/B/C/D context. The NAIC AI Systems Evaluation Tool (AISET) was piloted in early 2026 across 12 states (California, Colorado, Connecticut, Florida, Iowa, Louisiana, Maryland, Pennsylvania, Rhode Island, Vermont, Virginia, Wisconsin). The re-exposure period in September-October 2026 collected industry comments; adoption is expected at the NAIC Fall National Meeting in November 2026. Exhibit A addresses bias-testing methodology prompts (BISG, geographic proxy, sensitivity protocols); Exhibit B addresses model card and explainability artifacts; Exhibit C addresses governance documentation and the algorithm inventory; Exhibit D addresses consumer disclosure and adverse-action notice. The carriers that align the bias-testing exhibit to the AISET prompts in 2026 filings position for the 2027 nationwide adoption with minimal rework.

The Rate-Change Indication Exhibit

The rate-change indication exhibit is the quantitative rationale for the proposed rate change. Sections: experience window, exposure base, premium volume, claim counts, indicated overall rate change, indicated by territory (22 territories with credibility weights and per-territory caps), indicated by class (vehicle symbol × driver classification), indicated by coverage (BI, PD, Medical, UM, Collision, Comprehensive each separately), loss-ratio target reconciliation (current 78.4% → target 65.0%), trend factors (severity +4.2%, frequency -1.6%), development factors (loss-development triangle progression), large-loss treatment ($250K basic-limits cap), catastrophe load.

State-specific overlays on the indication. California Prop 103 requires the indication to be filed with class plans and rating-factor tables aligned to the state's classification system and the Prop 103 reasonable-classification justification. Colorado Reg 10-1-1 requires the indication by territory to demonstrate territory-credibility methodology and per-territory rate-change caps. New York requires the indication to show consistency with prior approved indications and any deviations to be justified per DFS Circular Letter 2024-7 §III.4. The same overall +5.8% indication produces three different exhibit structures depending on state; the FCAS spends time on the state-specific structural adjustments while the AI produces the cross-state consistent base.

Akur8 Discover and Matrisk - Competitor-Filing Intelligence

Akur8 Discover/Matrisk ingests every SERFF rate filing across all 50 states and produces competitor-filing intelligence by line of business, state, and named peer set. The pricing actuary queries: "personal-auto BI in Colorado from competitors X, Y, Z over the last 24 months." Discover returns: filings, effective dates, indicated rate-change percentages, approved rate-change percentages, DOI objection patterns, time-to-approval, and the response-to-objection language used by each competitor. The actuary can position the company's +5.8% indication against the peer median (+5.4%) and the peer range (+3.2% to +7.9%), citing the data in the filing memorandum's competitive-context section. DOI examiners read the same data; the actuary's positioning prevents the "your rate change is materially above market" objection.

Matrisk and the objection-history database. Matrisk catalogs every objection issued by each DOI on each filing. The actuary can search: "Colorado DOI personal-auto objections on credit-score factor 2024-2026." Matrisk returns the actual objection language, the carrier's response language, and the resolution (approved, modified, withdrawn). The response-to-objection template draws on this database - pre-drafted answers to the eight most common objection patterns. Carrier-specific tailoring is still required, but the template gets 70% of the answer right at zero incremental drafting time.

The Matrisk acquisition context. Akur8 acquired Matrisk in January 2026, integrating the objection-history database with the Akur8 Discover competitor-intelligence feed. The integrated platform produces: peer-positioning analysis on rate magnitude, peer-positioning on filing structure, DOI objection-pattern matching by state and topic, and the response-to-objection template seeded with the carrier's prior response language adjusted for current model. The integration compresses the response-to-objection cycle further - the carrier's filing manager moves from response-pattern lookup to response customization in a single workflow.

The Response-to-Objection Template

SERFF objections arrive 7-30 days after filing submission. Eight patterns dominate the personal-auto pricing objection space in 2026: (1) "explain and justify the credit-based insurance score factor"; (2) "explain credibility methodology on territory"; (3) "demonstrate Colorado Reg 10-1-1 / NY DFS Circular Letter 2024-7 compliance"; (4) "explain telematics participation and any pricing differential"; (5) "justify indicated rate change against competitor benchmarks"; (6) "explain candidate-model selection (GLM vs. ensemble) and SHAP transparency"; (7) "document data-quality controls and ASOP No. 23 compliance"; (8) "state model retirement criteria and review cadence per ASOP 56 model governance." The template pre-drafts each response; the FCAS customizes per-filing with carrier-specific numbers.

The objection-response workflow. A DOI objection received via SERFF. The filing manager logs into Akur8/Matrisk and queries the objection pattern. The pre-drafted response is retrieved. The FCAS customizes with carrier-specific numbers (e.g., "for this filing the credit-score factor magnitude at the top decile is 1.42 vs. bottom decile 0.71; the SHAP contribution to disparate impact under BISG is 8 basis points on the in-force book DI ratio"). General counsel reviews. The chief actuary signs if the response is material. The response is submitted via SERFF. Cycle time: 4-12 hours per objection vs. 12-30 hours unassisted. Three objections on a personal-auto filing now resolve in 1-2 weeks instead of 3-6 weeks; the cumulative cycle compression on a multi-state filing with multiple objections is 30-60 days on the time-to-approval.

The SERFF Submission Mechanics

SERFF submission via the SERFF portal (login.serff.com). State-by-state filing forms vary; California uses the CDI Prior Approval forms; Colorado uses the DORA Insurance Division forms; New York uses the DFS rate filing forms. The filing package per state: company filing information, contact information, filing description, rate-filing memorandum, actuarial certification, rate-change indication exhibit, bias-testing exhibit, model card (by reference with the URL and commit hash, or attached as a PDF), supporting actuarial documentation, prior-filing references, effective date request, expected impact on policyholders. The SERFF tracking ID is assigned at submission; objections arrive via SERFF; the filing manager monitors the SERFF dashboard daily and the multi-state coordinator runs a weekly status review.

Multi-state timing. The three filings (California, Colorado, New York) submitted on the same day. California Prop 103 prior-approval typically 75-180 days. Colorado Reg 10-1-1 review typically 30-90 days. New York DFS prior-approval typically 45-120 days. The first effective date achievable depends on the slowest state; the carrier sometimes splits the filings to take effective dates state-by-state as approvals arrive. The rate-implementation team must be ready for staggered approvals and the system must handle multi-effective-date rate plans across the same model family.

State-Specific Objection Patterns and 2026 Trends

Colorado DOI under Reg 10-1-1 (expanded October 15, 2025; first compliance report July 1, 2026): aggressive on bias-testing exhibit completeness, SHAP/PDP/ALE artifacts, BISG methodology validation, sensitivity analysis depth, and the algorithm inventory entry per Reg 10-1-1's algorithm-inventory requirement. California CDI under Prop 103: aggressive on classification plans, rating-factor tables, prior-approval rationale, no-credit-score compliance, broad consumer-impact analysis. New York DFS under Circular Letter 2024-7 (effective July 11, 2024) plus Letter No. 1: aggressive on external-consumer-data sourcing, proxy-test results with the proxy-correlation threshold (0.25 is the trigger), model card completeness, telematics consent, and the senior-management governance reporting requirement that Circular Letter 2024-7 introduced. Connecticut MC-25-8, Washington (which has signaled active examiner training), Massachusetts (which has pursued auto-pricing fairness aggressively for two decades), New Jersey, and Illinois are catching up in 2026 with AI-specific scrutiny and emerging state-specific bulletins.

The 2026 cross-state harmonization question. NAIC's AI Working Group continues to push for harmonization of AI-pricing standards across states. The current state is fragmentation - Colorado's quantitative disparate-impact threshold (0.80 four-fifths, with 0.85 expectation in practice), NY DFS's proxy correlation threshold (0.25), California's classification-plan justification, Connecticut's MC-25-8 incoming AI rule, Nevada Bulletin 24-006. A carrier filing in all 50 states maintains a 50-state filing matrix tracking which state requires which exhibits and which language. The matrix is itself an AI-assisted artifact in 2026; the carrier's filing manager updates it as state regulations evolve and as the AISET Exhibits A/B/C/D move toward nationwide adoption at the NAIC Fall National Meeting in November 2026.

The treaty-renewal dimension. The rate-filing approval timeline affects the treaty renewal because the next 1/1 or 4/1 treaty placement is informed by the carrier's projected loss-ratio trajectory. A carrier that compresses the rate-filing cycle from 6-8 weeks to 3-4 weeks (Branch Insurance case study) signals operational discipline to the treaty broker and to the reinsurance panel; the signal supports better treaty terms at the next placement.

Key Takeaways

• Multi-state filing the same GLM, three different memos. Colorado (Reg 10-1-1 fairness exhibit with the October 15, 2025 expansion), California (Prop 103 class plans + no-credit-score + reasonable-classification justification), New York (DFS Circular Letter 2024-7 plus Letter No. 1 proxy-test). The actuary cannot file one memo into three states; each requires state-specific overlay.
• AI drafts the memorandum in 4 minutes; the FCAS customizes in 6-10 hours. 30-50% time savings on the memo (18-24 hours manual to 10-14 hours AI-assisted). AI saves time on structure, boilerplate, cross-state consistency. AI does not save time on company-specific narrative - the FCAS still owns it. The Akur8 ecosystem (Rate Repo, Deploy, Discover, Matrisk) integrates the workflow with the Matrisk acquisition in January 2026.
• The actuarial certification is one page, highly templated. AI drafts in seconds with correct ASOP references (12, 41, 56), credential blocks, state-specific rating-law citations (California Insurance Code §1861, Colorado Revised Statute §10-4, New York Insurance Law §2304-2310). Chief actuary signs. 30 minutes review + 5 minutes sign vs. 2-4 hours pre-AI. ASOP 56 peer-review attestation attaches as supporting document.
• The bias-testing exhibit is a separate 12-20 page document. Required by Colorado Reg 10-1-1; equivalent disclosure required by NY DFS Circular Letter 2024-7; California signaled under SB 1058 (passed 2025). All eight aggressive-fairness DOIs (CO, NY, CA, CT, WA, MA, NJ, IL) expect this exhibit by 2026. NAIC AISET Exhibits A/B/C/D piloted early 2026 with November 2026 expected adoption - carriers that align in 2026 filings position for 2027 nationwide adoption.
• Akur8 Discover/Matrisk gives competitor-filing intelligence. Query "personal-auto BI in Colorado, peers X/Y/Z, last 24 months"; returns filings, indicated rate-change, approved rate-change, DOI objection patterns, time-to-approval. Position the +5.8% indication against peer median +5.4% and peer range +3.2-+7.9%. Matrisk acquired January 2026 integrates objection-history database with Discover competitor-intelligence feed.
• Eight common DOI objection patterns dominate 2026 personal-auto filings. Credit factor; territory credibility; Reg 10-1-1 / Circular Letter 2024-7 compliance; telematics participation; competitor benchmarks; model selection; data-quality controls; retirement criteria. The pre-drafted response template gets 70% of each answer right at zero incremental drafting time; FCAS customizes the remaining 30%.
• SERFF objection-response cycle compresses from 12-30 hours to 4-12 hours with AI assistance. Three objections on a personal-auto filing now resolve in 1-2 weeks vs. 3-6 weeks. The cumulative time-to-approval drops 30-60 days; on a $1.2B book that is $4-12M of premium-collection delay avoided. Branch Insurance case study documented 6-8 weeks to 3-4 weeks compression with the full Akur8 stack.
• The same GLM produces three filings on the same submission day, three approval timelines. California Prop 103 75-180 days; Colorado 30-90 days; New York DFS 45-120 days. The rate-implementation team is ready for staggered approvals; the carrier sometimes splits effective dates state-by-state. The treaty broker reads the rate-filing cycle compression as operational discipline signaling at the next 1/1 or 4/1 placement.
• The 50-state filing matrix is an AI-assisted artifact. Tracks which state requires which exhibits, which language, which thresholds (Colorado 0.80 four-fifths with 0.85 expectation; NY DFS 0.25 proxy correlation; California reasonable-classification justification; Connecticut MC-25-8; Nevada Bulletin 24-006). NAIC AI Working Group pushes harmonization; current state is fragmentation; the carrier's filing manager updates the matrix as state regulations evolve toward the AISET nationwide adoption.