Key Takeaways
Why a rating engine is a filing rather than code, what quoting owes that binding does not, and the testing any model in underwriting requires.
- Your rate logic goes on file with each state through SERFF. Changing a factor is a filing event, not a config change.
- SERFF Filing Access publishes rate and form filings, so your filed rates are readable by competitors and theirs by you.
- NAIC’s own surveys put AI or machine learning use at 88% of 193 auto insurers and 92% of 93 health insurers.
- New York requires discrimination testing before an AI system goes into production, and says an insurer may not rely on a vendor’s claim alone.
- Colorado’s Regulation 10-1-1 requires an officer-signed attestation by 1 December, filed annually through SERFF.
Your rating engine is a filing before it is code
The single thing that makes insurance software different from every other vertical is that your pricing logic is a document lodged with regulators.
Quick Verdict: Before you scope an insurance build, find out whether your rating logic is already on file and in which states. If it is, your release process is a regulatory process and your sprint cadence has to match a filing cadence. Treat it like any other regulated web application build, where the compliance artifact is a deliverable rather than a document someone writes afterward. The closest comparison on this site is a HIPAA-compliant application, where the same principle applies: the rules shape the architecture, not the other way round.
Rates and forms are submitted to states through SERFF, the electronic filing system operated by the National Association of Insurance Commissioners. NAIC describes it as the most cost-effective and efficient way to submit rates and forms filings to the states and jurisdictions, and says it accelerates the pace of market-entry for new and renewing products.
Read that from an engineering seat. Market entry has a pace, and the pace is set by a filing queue.
So the thing your developers call a rating table is, from the regulator’s side, an exhibit. The factors, the relativities, the territory definitions and the rules that combine them all exist in a submitted document that your code is supposed to implement exactly.
| What your team calls it | What the regulator sees |
| Rating table | A filed exhibit |
| Config value | A rate factor requiring a filing |
| Feature flag on pricing | An unfiled rate in use |
| Hotfix to a rating bug | Potentially a deviation from the filing |
| Deploy | Nothing, unless the filing changed |
The fourth row is the one that causes real trouble. A rounding bug in production is a defect to your team and a mismatch between filed and charged rates to an examiner.
Do this: Ask for the current filed rate exhibits for every state you operate in, and diff them against what your code actually computes.
Assume one approval regime per jurisdiction, not one deploy
Launching nationally is not one launch. SERFF’s own state selector lists 52 entries, including the Interstate Insurance Product Regulation Compact, and NAIC notes that other states may be in production without having provided links.
That is the real shape of a national rollout: dozens of separate reviewers, each with their own standards and timelines, looking at your filing.
States differ in how much they review before you may use a rate. Some require approval before use, some allow use once filed, and some allow use with a filing to follow. Those three regimes produce completely different release plans for the same code change, and your roadmap has to carry the slowest one.
The practical consequence is that state becomes a first-class dimension in your architecture, not a column on a table. Rate versions are effective-dated per state. Two customers quoted on the same day in different states are priced from different filings. A rollback in one state is not a rollback in another.
Teams that discover this late end up rebuilding the pricing layer, which is the single most expensive thing in an insurance system to rebuild. Multi-state versioning looks a lot like multi-tenant architecture problems, and the same early decisions decide whether it stays manageable.
Do this: Write down, per launch state, which filing regime applies and what the review window has actually been. Use real past filing dates, not published targets.
Your competitors can read your filed rates, and you can read theirs
This is the fact that surprises every software team and almost no actuary.
NAIC’s SERFF Filing Access provides consumers and stakeholders with convenient online access to publicly available rate and form filings. Your filed rating plan is not a trade secret in the way your codebase is. Much of it is a public record.
Two consequences follow, and they point in opposite directions.
The uncomfortable one: a differentiating pricing approach is visible once filed. If your whole advantage is a clever factor, expect it to be read.
The useful one: you can read the competition before you build. Filed rates in your target states tell you what the market actually charges, which territories carry which relativities, and which carriers have recently refiled. That is competitive intelligence most startups pay consultants for, sitting in a public system.
My working model: build your advantage into things that are not filed. Speed of quote, quality of the application flow, the data you enrich with, how fast you bind, how well you handle the awkward cases. Those are not exhibits.
Do this: Pull three competitor filings in your largest target state before you finalize your own rating approach.
Build quoting and binding as separate systems with separate obligations
A quote and a bound policy are not two states of one record. They are different regulated acts, and merging them in your data model creates problems that surface at audit rather than in testing.
A quote is an indication produced from filed rates and whatever information the applicant supplied. It can expire, it can be refreshed, and it may be produced before full verification. A bind creates coverage. An issued policy carries filed forms, and those forms are their own filings alongside the rates.
| Stage | What is filed | What you must retain |
| Quote | The rate exhibits used | Inputs, rate version, timestamp, output |
| Bind | Rates plus eligibility rules | What changed between quote and bind, and why |
| Issue | Policy forms and endorsements | The exact form versions delivered |
The retention column is where software teams underbuild. If a regulator asks why a specific applicant was quoted a specific number on a specific day, the answer has to be reconstructable: which filing version, which inputs, which rules fired, what the output was.
That means immutable quote records with the rate version stamped on them. Not a recomputation. A recomputation using today’s filing cannot answer a question about last March.
This is ordinary event-sourcing discipline, and it is the same reasoning behind the audit trails in a policy or claims administration system. Build it in version one, because retrofitting an audit trail onto a mutable record is a rewrite.
Do this: Stamp every quote with the rate filing version used, and make quote records append-only.
Any model touching underwriting owes documented testing before production
If you put a model or an external data source anywhere near pricing or eligibility, you have taken on a testing obligation, and in some states it is explicit.
New York’s Department of Financial Services issued Insurance Circular Letter No. 7 on 11 July 2024, covering the use of AI systems in insurance underwriting. Its language is unusually direct for a regulator.
On timing, it states that unfair or unlawful discrimination testing and analysis should be administered prior to putting AI systems into production. Before, not after.
On explainability, it expects an insurer to be able to explain, at all times, how its AI system operates, and to articulate a logical relationship between external consumer data and other model variables.
On method, it encourages insurers to use multiple statistical metrics, naming the Adverse Impact Ratio and Denials Odds Ratios among them. It also warns that external consumer data and information sources may reflect systemic biases and that their use can reinforce and exacerbate inequality.
This is not a niche concern. NAIC’s Big Data and Artificial Intelligence Working Group surveyed insurers by line on whether they use, plan to use, or plan to explore AI and machine learning models.
| Line | Share of respondents | Respondents | Report issued |
| Health | 92% | 93 insurers | May 2025 |
| Private passenger auto | 88% | 193 insurers | December 2022 |
| Home | 70% | 194 insurers | August 2023 |
| Life | 58% | 161 companies | December 2023 |
Most of the market is already here. The testing obligation is not a future problem for your roadmap.
Do this: For every model input, record where the data came from and whether it is an external consumer data source. That list is the scope of your testing obligation.
You cannot push the testing onto your vendor
This is the clause that catches teams who bought a scoring product and assumed the compliance came with it.
New York’s circular letter says plainly that the responsibility to comply with anti-discrimination laws remains with the insurer at all times, that insurers retain responsibility for understanding any tools they use, and that an insurer may not rely solely on a third party’s claim of non-discrimination.
So a vendor certificate is not a defense. You need your own testing, your own documentation, and contractual terms that give you what you need to produce both. The circular letter expects insurers to develop written standards for third-party oversight and, where appropriate and available, to include terms in their contracts with vendors.
Regulators are building tooling for this. NAIC formed a Third-Party Data and Models Working Group in 2024 to develop a framework helping regulators evaluate third-party AI data and models used by insurers, and its AI Systems Evaluation Tool was being piloted by 12 participating states as of March 2026, with adoption expected at the 2026 Fall National Meeting.
Read that as a schedule. The evaluation tool exists to ask you questions, and the questions will be about documentation you either have or do not.
Practically, this reshapes vendor selection. The question stops being which score performs best and becomes which vendor will hand you the variable list, the methodology and the audit rights. A vendor that will not is a vendor whose output you cannot defend, which is adjacent to the evidence discipline behind a SOC 2 program. Build your model inventory the way you would any AI-assisted product, with provenance recorded at the point of use.
Do this: Ask every data or model vendor in writing for the variable list, the methodology summary and audit rights. Treat a refusal as disqualifying.
Put the filing calendar in the same place as the release calendar
Insurance has dated obligations that have nothing to do with your sprint board, and they are the ones that get missed.
Colorado is the clearest example because its requirement reaches insurers who think it does not apply to them. Under SB21-169, the Colorado Division of Insurance adopted amended Regulation 10-1-1, Governance and Risk Management Framework Requirements, covering life insurers, private passenger automobile insurers and health benefit plan insurers, effective 15 October 2025.
Here is the part that catches people. An insurer that does not use external consumer data and information sources, or algorithms or predictive models that use them, still has to say so. The Division requires that attestation to be filed no later than 1 December, signed by an officer of the insurer, submitted annually through SERFF as an Annual Report or Annual Certification, with a separate filing for each insurer.
So the same system you file rates through carries an annual compliance filing, on a fixed December date, which needs an officer’s signature rather than an engineer’s. That date comes round every year and does not move for your release schedule.
The broader point is that insurance work has a calendar you inherit: rate filing windows, form refilings when a statute changes, annual certifications, data calls. None of it appears in a feature backlog unless someone puts it there.
Do this: Create one shared calendar holding filing deadlines and release dates together, and give it an owner who is not the engineering lead.
What software cannot fix here
Some of this is genuinely outside an engineering team’s reach, and saying so early saves a quarter.
You cannot shorten a state’s review. If a jurisdiction takes weeks to approve a rate change, that is your timeline regardless of how good your deployment pipeline is.
You cannot engineer around an unfiled rate. There is no architecture in which charging a rate you have not filed becomes acceptable, and a feature flag does not change that.
You cannot buy your way out of the testing obligation, because New York has said the responsibility stays with the insurer at all times.
I also cannot tell you from here which regime applies to your product in your states, or whether your specific model triggers a testing requirement. That is work for a regulatory attorney and your appointed actuary. Nothing here is legal, actuarial or compliance advice. What I can tell you is which questions change your build, and they are the seven above.
One honest limit on this piece itself. The NAIC model bulletin on AI systems was adopted in December 2023, and the NAIC’s own state-adoption tracker that I could retrieve is dated 30 April 2024, listing eleven states that had adopted it by then. The real number today is higher and I could not find a current published count, so this piece deliberately does not state one. Check your own states rather than trusting any figure in an article, including this one.
Where we fit: if you have your filing regimes identified and need the quoting, versioning and audit layer built properly, that is a contained engagement and the first step is Get your build plan, which is complimentary, comes back in two business days and includes the scope, the stack, the cut list with reasons, a fixed price and a launch date. No sales call. We are the wrong call if you have no filings and no appointed actuary yet, because your blocker is regulatory rather than technical and no developer shortens it. Get your build plan and we will tell you which of those two you are. You can see how we scope and sequence work and read what previous builds actually produced first.
Do this: Confirm your filing regime per state before commissioning a single screen of the rating engine.
Frequently Asked Questions
Why does changing a rate factor take so long in insurance software?
Because the factor is part of a filing lodged with each state through SERFF, not a configuration value. Depending on the state, you may need approval before using a new rate, or you may file and use it immediately. Your release timeline is set by the slowest reviewing state, not by your deployment pipeline.
Are filed insurance rates public?
Largely yes. NAIC’s SERFF Filing Access provides consumers and stakeholders with online access to publicly available rate and form filings. That means competitors can read your filed rating plan and you can read theirs. Build differentiation into things that are not filed, such as quote speed, application flow and binding experience.
Do I need bias testing if I only use traditional rating variables?
Possibly not, but you still may owe a filing saying so. Colorado’s Regulation 10-1-1 requires insurers that do not use external consumer data and information sources, or models relying on them, to file an officer-signed attestation by 1 December each year through SERFF. Confirm your own state’s position with counsel.
Can my scoring vendor handle compliance for me?
No. New York’s Insurance Circular Letter No. 7 states that the responsibility to comply with anti-discrimination laws remains with the insurer at all times, and that an insurer may not rely solely on a third party’s claim of non-discrimination. You need your own testing, documentation and contractual audit rights from the vendor.
How should quotes be stored in an insurance system?
Append-only, with the rate filing version stamped on every quote alongside the inputs and the output. If a regulator asks why a specific applicant received a specific number on a specific date, you must reconstruct it from the filing in force then. Recomputing with today’s rates cannot answer that question.
Resources
- Artificial Intelligence, National Association of Insurance Commissioners. The December 2023 adoption of the model bulletin, the Big Data and Artificial Intelligence Working Group survey findings by line of business, the Third-Party Data and Models Working Group formed in 2024, and the AI Systems Evaluation Tool piloted by 12 states as of March 2026.
- System for Electronic Rates and Forms Filing, National Association of Insurance Commissioners. What SERFF is, the 52 jurisdiction entries in its state selector, and SERFF Filing Access for publicly available rate and form filings.
- Insurance Circular Letter No. 7: Use of Artificial Intelligence Systems in Insurance Underwriting and Pricing, New York State Department of Financial Services, 11 July 2024. Pre-production discrimination testing, the explainability expectation, the Adverse Impact Ratio and Denials Odds Ratios metrics, and the limits of relying on a vendor’s assurance.
- SB21-169: Protecting Consumers from Unfair Discrimination in Insurance Practices, Colorado Division of Insurance. Amended Regulation 10-1-1 effective 15 October 2025, the lines covered, and the officer-signed attestation due by 1 December annually through SERFF.
- Implementation of NAIC Model Bulletin: Use of Artificial Intelligence Systems by Insurers, National Association of Insurance Commissioners, status as of 30 April 2024. The eleven states listed as having adopted the bulletin by that date, and the four states listed under insurance-specific regulation.
M. Sabir Junaid
Founder • CEO
M. Sabir Junaid is founder and CEO of Velcod, the senior product studio. He leads product and delivery across MVPs, SaaS platforms, and mobile apps — system architecture, no-code and low-code engineering, and AI integration. An official Bubble partner, Sabir has shipped products across fintech, marketplaces, edtech, and creator platforms. He works with founders raising pre-seed to Series A who need a product built to launch and scale in weeks, not a six-month agency build.
Share
- Previous
- Next
Ready to build?
Your idea, live in 3 weeks.
Book a free strategy call and get a fixed-price roadmap for your MVP, SaaS, or app.