AI Underwriting Automation in US Insurance: 2026 Rules
Published 3 November 2026 · 10 min read
What US insurers must prove before automating underwriting with AI: NAIC bulletin, Colorado Reg 10-1-1, NY DFS Circular 7, plus a build checklist.
The short answer
In 2026, a US insurer can automate large parts of underwriting with AI, but it has to be able to prove three things on demand: that a documented AI systems program governs the model, that the model has been tested for unfair discrimination against protected classes, and that a declined applicant can be told which specific data drove the decision. There is no single federal AI insurance law. What you actually face is a stack of state rules led by the NAIC model bulletin, Colorado's Regulation 10-1-1, and New York's Circular Letter No. 7 — plus the ordinary privacy and data security obligations covered on our US privacy compliance page and in our United States practice.
Stratgik is a technology firm, not a law firm. Nothing here is legal advice. What follows is the engineering and governance view: what regulators now ask to see, and what you have to build so that the answer exists.
What "AI underwriting automation" means in practice
Most insurers we talk to are not replacing underwriters. They are automating four distinct things, and each carries a different level of regulatory risk:
- Intake and data extraction. Pulling fields out of applications, medical records, loss runs, ACORD forms and broker emails. Low risk — it is OCR and classification, and a human still decides.
- Triage and straight-through processing. Routing clean, low-face-amount cases to auto-accept and flagging the rest. Medium risk, because the routing rule itself can adversely affect a protected class even when no human declines anyone.
- Risk scoring and pricing. Models that set a class, a rating factor or a decline. This is the regulated core, and it is where every rule below bites.
- Post-bind monitoring and fraud flags. Medium to high, depending on whether a flag can trigger a rescission or a claim denial.
Scope your first automation to intake and triage, keep the decision boundary with a human, and move the decision itself only once governance and testing are genuinely in place. Teams that invert that order end up rebuilding.
The rulebooks that apply in 2026
| Rule | Who it binds | Status as of November 2026 | What it demands |
|---|---|---|---|
| NAIC Model Bulletin on the Use of AI Systems by Insurers | Insurers in adopting states | Adopted by the NAIC in December 2023; a May 2025 Holland & Knight review counted 24 states that had adopted it, with four more issuing related regulation or guidance | A documented AIS program, governance, risk management and internal controls, third-party vendor oversight, testing and audit, records produced on regulator request |
| Colorado SB21-169 and Amended Regulation 10-1-1 | Life, private passenger auto and health benefit plan insurers in Colorado | Effective October 15, 2025 | Governance and risk management framework for external consumer data and information sources (ECDIS), algorithms and predictive models; testing for unfair discrimination; annual filing. Insurers that do not use such tools file an officer-signed attestation, due December 1 each year |
| NY DFS Insurance Circular Letter No. 7 (2024) | Insurers authorized in New York, Article 43 corporations, HMOs, fraternal benefit societies, NYSIF | Issued July 11, 2024 | Board-level oversight, annually reviewed written policies, AIS inventory, pre-production and ongoing discrimination testing, specific adverse-decision reasons, written notice within 15 days, vendor audit rights |
| Colorado SB 26-189 (replacing the 2024 Colorado AI Act) | Deployers of covered automated decision-making technology, including in insurance | Adopted May 2026, effective January 1, 2027 | Pre-use notice; after an adverse outcome, disclosure of the right to access and correct personal data and to obtain human review and reconsideration |
| NAIC Insurance Data Security Model Law (#668) | Licensees in adopting states | NAIC's adoption map shows status as of October 8, 2026 | Information security program, risk assessment, third-party service provider due diligence, breach investigation and notification |
What the NAIC bulletin actually asks you to build
The bulletin is not a technical standard. It is a demand for evidence. The NAIC expects an insurer to adopt and maintain a written AI systems ("AIS") program proportionate to the consumer harm the system could cause, covering governance, risk management and internal controls, third-party vendor oversight, and testing and auditing designed to detect error, bias and unfair discrimination. Regulators are told to expect to examine that program and to ask for documentation of how a model was developed, used and overseen.
Two engineering consequences follow. First, a model inventory stops being optional: you need a register that names every AI system touching a regulated process, its owner, its training and input data, its last test date and its risk tier. Second, nothing can live only in a notebook. Model cards, test results, approval records and change logs have to sit in a system that can produce a defensible audit trail years later.
The NAIC is also building the examination instrument. Its Big Data and Artificial Intelligence Working Group has developed an AI Systems Evaluation Tool for use in market conduct and financial examinations; as of March 2026, 12 participating states were piloting it, and adoption is expected at the 2026 Fall National Meeting. If you are planning a 2027 underwriting automation launch, assume a structured regulator questionnaire exists by the time you go live.
Colorado: the state that put testing in writing
Colorado remains the most concrete US regime. Under SB21-169, insurers must test big data systems — including external consumer data sources, algorithms and predictive models — for unfair discrimination against protected classes, correct any harm found, and demonstrate that testing to the Division of Insurance. Amended Regulation 10-1-1, effective October 15, 2025, applies this to life, private passenger auto and health benefit plan insurers.
The detail that catches teams out is the negative filing. An insurer that uses none of these tools still has to file an annual attestation signed by an officer, due December 1 each year. "We don't use AI" is a position you have to assert formally — which means you need to know, accurately and company-wide, whether any vendor tool in your stack quietly qualifies. That inventory question is usually what starts the real project.
Separately, Colorado's broader AI statute moved. SB 26-189 replaced the 2024 Colorado AI Act in May 2026 with a lighter, disclosure-focused framework effective January 1, 2027. It drops the original risk-management-policy and impact-assessment mandates but keeps consumer notice and, after an adverse outcome, the right to request data, correct inaccuracies and obtain human review. Enforcement sits with the Attorney General, with a 60-day cure period that does not apply to knowing or repeated violations and penalties up to $20,000 per violation. Note also that the FTC issued a proposed policy statement on July 1, 2026 raising potential preemption of state AI laws under Section 5 of the FTC Act — so treat the state-law map as unsettled and design for the strictest regime you operate in.
New York: the most prescriptive on notices and vendors
If you only read one document before building, read Circular Letter No. 7. It is the clearest statement of what "fair" testing means in operational terms, and it sets a three-step assessment: check for disproportionate adverse effects on protected classes, establish a legitimate and fair rationale for any difference, and search for less discriminatory alternatives — repeating that search at least annually. DFS expects multiple statistical metrics plus a qualitative explanation of how the system works, and an assessment of whether external data acts as a proxy for a protected class.
Two provisions have direct product implications. Applicants declined through an AI-based process should receive written notice within 15 days, with the reasons identifying the sources of information relied on and a route to check that data's accuracy. And an insurer cannot hide behind a vendor's proprietary algorithm to avoid giving specific reasons — DFS states plainly that insurers retain responsibility for understanding the tools, data and AI systems they use in underwriting and pricing, and should secure audit rights and regulator-cooperation clauses in vendor contracts.
In other words, reason codes are not a phase-two feature. A scoring model you cannot explain at the individual-decision level is not deployable in New York, however accurate it is.
Build or buy?
Most insurers end up with a hybrid: a bought document-ingestion and data-enrichment layer, and an owned decision layer. The reason is the explainability obligation above. You can outsource extraction; you cannot outsource the duty to explain a decline. Our build vs buy calculator runs the cost side, but on the compliance side ask any vendor these questions and insist on written answers:
- Can you produce per-decision reason codes traceable to specific input fields, not just global feature importances?
- What protected-class proxy testing have you run, with which metrics, on which populations, and will you share the results?
- Will the contract grant us audit rights and commit you to cooperate with state regulator inquiries?
- Where is model and applicant data stored and processed, and which subprocessors touch it?
- What is your process when we report that an input data point about a consumer is wrong?
- How are model versions pinned, and will you notify us before a material change to the model?
A vendor that cannot answer the first and third questions is a compliance liability whatever its accuracy benchmarks. For the owned decision layer, our custom software and managed technology teams build the inventory, logging and reason-code plumbing these rules assume exists.
Pre-launch checklist
- Maintain a model inventory covering every AI system in a regulated process, with owner, inputs, risk tier and last test date.
- Write the AIS program document and have it formally approved; schedule the annual review.
- Record board or senior-management oversight of AI outcomes in minutes, not just in a policy PDF.
- Run pre-production discrimination testing with multiple metrics; document the less-discriminatory-alternatives search and diarize it annually.
- Test external data sources for proxy effects on protected classes before they enter a rating or decline path.
- Build per-decision reason codes and an adverse-action notice that names data sources and the correction route; meet New York's 15-day window.
- Log every automated decision immutably: model version, inputs, score, outcome, and any human override.
- Put audit rights, regulator-cooperation and change-notification clauses into every AI and data vendor contract.
- Diarize the filings: Colorado's December 1 attestation, and any state-specific AI reporting where you are licensed.
- Align the information security program and third-party due diligence with your state's version of Model #668.
- Define and rehearse the rollback: what happens operationally the day a test shows disparate impact in production.
Cost and timeline, realistically
For a mid-sized carrier or MGA, intake and triage automation is typically a 10–16 week build and pays back fastest, because the volume is high and the regulatory surface is small. Moving the decision itself is a different project: expect the governance, testing and logging work to take as long as the model work, and to continue as a run cost rather than a one-off. The sequencing that works is inventory first, then intake automation, then a shadow-mode decision model scored against human underwriters for a full cycle, then limited straight-through processing on one narrow, well-tested segment. Each step produces the documentation the next step's regulator will ask for.
Frequently asked questions
Is there a federal US law governing AI in insurance underwriting?
No single statute. Insurance is state-regulated, so the operative rules are state adoptions of the NAIC model bulletin plus state-specific regulation such as Colorado's Regulation 10-1-1 and New York's Circular Letter No. 7. Federal exposure comes from general law — unfair and deceptive practices, fair lending and civil rights statutes where applicable — and the FTC's July 2026 proposed policy statement signals that the federal-versus-state boundary is actively contested.
Does the NAIC model bulletin apply to us automatically?
Only in states that have adopted it. A May 2025 Holland & Knight review counted 24 adopting states, with four more issuing related regulation or guidance. Because adoption keeps moving, check the current status in every state where you are licensed rather than relying on a count — and if you write in multiple states, building to the strictest is usually cheaper than maintaining variants.
We only use a vendor's model. Are we still responsible?
Yes. Both the NAIC bulletin and the New York circular place third-party oversight squarely on the insurer, and DFS states that insurers retain responsibility for understanding the tools and data they use and cannot rely on a vendor's proprietary algorithm to avoid giving specific adverse-decision reasons.
What does Colorado require if we do not use AI at all?
Under Amended Regulation 10-1-1, insurers that do not use external consumer data sources, algorithms or predictive models still file an annual attestation signed by an officer, due December 1 each year. Confirm the current filing mechanics with the Division of Insurance, and confirm internally that no vendor tool in your stack falls within scope.
Can we automate declines entirely, with no human in the loop?
Technically yes in many lines, but the obligations do not relax: you still owe specific reasons tied to identified data sources, a route for the consumer to check that data, and — once Colorado's SB 26-189 takes effect on January 1, 2027 — disclosure of a right to human review and reconsideration after an adverse outcome. Most carriers keep a human review path because it is cheaper than engineering around the notice and reconsideration duties.
Stratgik builds and runs the underwriting automation layer for insurers and MGAs: document ingestion, decision services, model inventories, reason-code and audit logging, and the vendor oversight evidence your examiners will ask for. We are a technology firm, not a law firm, and we work alongside your counsel rather than replacing them. See our United States practice and US privacy compliance work, or book a free 30-minute session and we will walk your current stack against the checklist above and tell you plainly where the gaps are.
Keep reading
More from Stratgik

17 Nov 2026
Fractional CTO in the US: Cost, Contracts and 1099 Rules
What a fractional CTO costs in the US in 2026, how to structure the contract, 1099 classification after the DOL's propos...

16 Nov 2026
How to Pay a Fractional CTO: Retainer, Day Rate or Equity
Retainer, day rate or equity for a fractional CTO? Compare structures, see 2026 advisor-equity data, and the India, UK a...

15 Nov 2026
AI Accounts Payable Automation in the US: Rules and Cost 2027
What AI accounts payable automation costs in the US in 2027, plus the Nacha, IRS 1099 e-filing and California ADMT rules...

