Skip to main content
    Back to Blog
    24 min readLast reviewed:
    Insurance
    NAIC
    AI Compliance
    Underwriting
    EU AI Act

    AI in Insurance 2026: NAIC Model Bulletin, Colorado Reg 10-1-1, and ChatGPT for Underwriting

    T

    Team Sonomos · Sonomos

    The Sonomos research team writes about AI privacy, data protection, and how to use generative AI safely at work.

    Short answer: Insurers can use ChatGPT, Claude, and Gemini in 2026, but the regulatory ground has moved enough that the question is no longer "can we?" and increasingly "how do you show you did it correctly?" The NAIC Model Bulletin on the Use of Artificial Intelligence Systems by Insurers (adopted December 4, 2023) had been adopted by 29 jurisdictions as of the NAIC's November 3, 2026 status map; Colorado Regulation 10-1-1 requires a governance and risk-management framework for life, private passenger auto, and health benefit plan insurers that use external consumer data, algorithms, and predictive models; NY DFS Insurance Circular Letter No. 7 (2024) sets expectations for AI in underwriting and pricing, including proxy and quantitative testing; HIPAA still governs health plans; GLBA Title V still governs customers' nonpublic personal information, enforced for insurers by state insurance regulators; and the EU AI Act classifies AI used for risk assessment and pricing of natural persons in life and health insurance as high-risk under Annex III. Letting applicant or insured personal data into a consumer-tier AI tool, with no governance record and no technical control to keep the data out, runs against nearly every one of these regimes at once. This guide walks through what each authority actually says and the controls that map to it.

    What the NAIC Model Bulletin actually requires

    The NAIC Model Bulletin: Use of Artificial Intelligence Systems by Insurers, adopted by the NAIC's Executive Committee and Plenary on December 4, 2023, is not law on its own. It is a template that each state insurance department can issue, modify, or pass over. It does not create new legal standards either: it reminds insurers that decisions "made or supported" by AI systems must comply with existing law, naming the state versions of the Unfair Trade Practices Act, the Unfair Claims Settlement Practices Act, the Corporate Governance Annual Disclosure Act, the Property and Casualty Model Rating Law, and the Market Conduct Surveillance Model Law.

    Adoption keeps moving. The NAIC's implementation map, status as of November 3, 2026, lists 29 jurisdictions that have adopted the bulletin: 27 states (Alaska, Arkansas, Connecticut, Delaware, Hawaii, Illinois, Iowa, Kentucky, Maryland, Massachusetts, Michigan, Mississippi, Nebraska, Nevada, New Hampshire, New Jersey, North Carolina, North Dakota, Oklahoma, Pennsylvania, Rhode Island, South Carolina, Vermont, Virginia, Washington, West Virginia, and Wisconsin) plus the District of Columbia and Puerto Rico. The most recent entries are Mississippi (July 22, 2026), South Carolina (September 24, 2026), and North Dakota (October 5, 2026). The same map lists four states with their own insurance-specific AI regulation or guidance instead: California (Bulletin 2022-5), Colorado (3 CCR 702-10), New York (Circular Letter No. 7), and Texas (Bulletin B-0036-20). The NAIC notes that the map does not determine whether a state's action contains every element of the model.

    The bulletin points to the AI Principles the NAIC adopted in 2020 as guidance, which it summarizes as:

    1. Fairness and ethical use of AI.
    2. Accountability.
    3. Compliance with state laws and regulations.
    4. Transparency.
    5. A safe, secure, fair, and robust system.

    The operative part asks each insurer to "develop, implement, and maintain a written program (an 'AIS Program')" for AI systems that make or support decisions related to regulated insurance practices, scaled to the insurer's own assessment of the risk to consumers. The program is expected to cover:

    • Governance, with responsibility vested in senior management "accountable to the board or an appropriate committee of the board," and a documented accountability structure (committees, decision rights, escalation, training).
    • Risk management and internal controls: data practices (lineage, quality, bias analysis and minimization), an inventory and description of predictive models, validation and testing, model drift, protection of non-public information, and record retention.
    • Third-party AI systems and data: due diligence on vendors and, "where appropriate and available," contract terms giving the insurer audit rights and requiring the vendor to cooperate with regulatory inquiries.
    • Consumer notice: processes "providing notice to impacted consumers that AI Systems are in use," with information appropriate to the stage of the insurance life cycle.

    Section 4 of the bulletin lists what a regulator may ask for in an investigation or market conduct action: the written AIS Program, model inventories, data and bias documentation, validation and drift testing, vendor due diligence, and vendor contracts. The bulletin also says its goal "is not to prescribe specific practices or to prescribe specific documentation requirements," and that insurers may show compliance by other means. It names the NIST AI Risk Management Framework as one framework an AIS Program may rely on (see our NIST AI RMF governance guide).

    For an insurer using ChatGPT or Claude in a workflow that touches regulated practices (underwriting, claims, marketing, fraud detection), that list is effectively the evidence file an examiner can request.

    Colorado Regulation 10-1-1: the most detailed state rule

    Colorado's rules rest on SB21-169, signed July 6, 2021 and codified at C.R.S. § 10-3-1104.9. The statute restricts insurers' use of external consumer data, algorithms, and predictive models that unfairly discriminate based on race, color, national or ethnic origin, religion, sex, sexual orientation, disability, gender identity, or gender expression, and directs the Commissioner to adopt rules for specific types of insurance after a stakeholder process.

    The implementing rule is Regulation 10-1-1 in 3 CCR 702-10 (Unfair Discrimination). It first took effect for life insurers on November 14, 2023. An amended version adopted August 19, 2025 took effect October 15, 2025 and, under its new title ("Governance and Risk Management Framework Requirements for Life Insurers', Private Passenger Automobile Insurers', and Health Benefit Plan Insurers' Use of External Consumer Data and Information Sources, Algorithms, and Predictive Models"), applies to insurers offering individually issued life insurance, private passenger auto insurance, and health benefit plans.

    What the amended regulation text requires:

    • A scoped definition of ECDIS. "External consumer data and information sources" include credit scores, social media habits, locations, purchasing habits, home ownership, educational attainment, licensures, civil judgments, court records, consumer-generated Internet of Things data (telematics for auto), biometric data, and risk scores derived from them. For health benefit plan insurers, ECDIS excludes an individual's medical records.
    • A risk-based governance framework designed to determine whether ECDIS and the models that use it "potentially result in unfair discrimination with respect to race," and to remediate it if detected through quantitative testing requirements established by the Division.
    • Board oversight and senior-management accountability: the framework must be overseen by the board or a board committee, with a documented cross-functional governance group.
    • An up-to-date inventory with version control of every ECDIS, algorithm, and predictive model, plus documented testing, drift monitoring, vendor selection, a complaints process that gives consumers "information necessary to take meaningful action" after an adverse decision, and an annual review of the framework.
    • Vendor accountability: an insurer using third-party vendors "remains responsible" for meeting the requirements.
    • For health benefit plans, a provider acting for the insurer must be ultimately responsible for prior or concurrent authorization decisions that ECDIS-based models inform.
    • Reporting: life insurers filed a progress report by June 1, 2024 and a compliance report by December 1, 2024 and annually after; auto and health insurers filed a progress report by December 1, 2025 and owe a compliance report by July 1, 2026 and annually after. Compliance reports must be signed by an officer attesting to compliance, or come with a corrective action plan.

    Colorado's quantitative-testing rule exists so far only as a draft. In September 2023 the Division released a draft testing regulation for life insurance underwriting that would estimate applicants' race and ethnicity with BIFSG and compare decline rates and premiums across groups, with a 5% difference triggering further testing. As of this review, the Code of Colorado Regulations lists the October 15, 2025 amendment of 3 CCR 702-10 as its current version, and we could not find a final quantitative-testing rule in effect. Check the Division's current rules before relying on any particular test design.

    For LLM use in Colorado, the practical reading is that an LLM-based tool that uses ECDIS, or feeds a model that does, in a covered line belongs in the framework and on the inventory. Colorado's broader AI statute is a separate regime: the provisions of SB 24-205, covered in our Colorado AI Act guide, were repealed and reenacted in May 2026 by SB26-189, an automated decision-making technology law that lists insurance among its consequential-decision areas.

    New York DFS Insurance Circular Letter No. 7 (2024)

    New York's Department of Financial Services issued Insurance Circular Letter No. 7 on July 11, 2024, setting out its expectations for insurers' use of external consumer data and information sources (ECDIS) and AI systems in underwriting and pricing. It applies to insurers authorized in New York, Article 43 corporations, HMOs, fraternal benefit societies, and the New York State Insurance Fund. It does not address claims or other stages of the product life cycle.

    Key expectations from Circular Letter No. 7:

    • Justification before use: an insurer "should not use ECDIS or AIS for underwriting or pricing purposes unless" it can establish that the data source or model "does not use and is not based in any way on any class protected pursuant to Insurance Law Article 26," and not where the use would result in unfair discrimination. External data should also be supported by generally accepted actuarial practice.
    • Proxy assessment: insurers should evaluate whether external data correlates with protected-class status and, if it does, whether its use is justified by a legitimate business need.
    • Quantitative assessment: testing for disproportionate adverse effects, a search for less discriminatory alternatives, and suggested metrics such as adverse impact ratios and denial odds ratios, both before deployment and regularly afterward.
    • Governance: board oversight, senior-management responsibility, written policies, and internal audit.
    • Vendor accountability: "The responsibility to comply with anti-discrimination laws remains with the insurer at all times," and insurers may not rely solely on a vendor's claim of non-discrimination.
    • Transparency: disclosure that AI systems and external data are used, and adverse underwriting notices with specific reasons, citing New York Insurance Law §§ 3425, 3426, and 4224. Insurers may not rely on the proprietary nature of a vendor's algorithm to justify a vague reason.

    The proxy expectation is the one most often underestimated. An LLM that never receives race or national origin can still produce outputs that track them, because free-text inputs carry proxies such as name, address, occupation, and vocabulary. DFS expects the testing to reach those proxies.

    The federal layer: GLBA, HIPAA, FCRA

    State insurance regulation is the foreground; federal law sets a floor underneath it.

    EU AI Act: life and health insurance pricing is high-risk under Annex III

    For insurers established in the EU, and for providers and deployers elsewhere whose AI output is used in the Union, the AI Act's Annex III, point 5(c) classifies as high-risk "AI systems intended to be used for risk assessment and pricing in relation to natural persons in the case of life and health insurance." That covers life and health underwriting and pricing. It does not, by itself, make claims handling or property and casualty pricing high-risk.

    Who carries which obligation matters:

    • Providers (whoever develops the system and places it on the market or puts it into service) must ensure the system meets the Section 2 requirements: risk management, data governance, technical documentation, logging, transparency to deployers, human oversight, and accuracy, robustness, and cybersecurity (Articles 9 to 15). An insurer that builds its own pricing tool on top of an LLM can end up as the provider.
    • Deployers, such as an insurer using a vendor's system, have Article 26 obligations: use the system according to its instructions, assign competent human oversight, make sure input data they control is relevant, monitor operation, keep logs for at least six months, and tell people that a high-risk system is used in decisions about them.
    • Fundamental rights impact assessment: Article 27 specifically requires deployers of Annex III point 5(b) and (c) systems, which includes life and health insurance pricing, to carry one out before first use.

    Timing changed in 2026. According to the Future of Privacy Forum's AI Act timeline, the AI Omnibus (Regulation (EU) 2026/1744), published in the Official Journal on July 24, 2026, postponed application of the Annex III high-risk requirements from August 2, 2026 to December 2, 2027. Our EU AI Act checklist for deployers covers the deployer side in more depth.

    What goes wrong in real insurance workflows

    Four patterns show where the rules above bite:

    1. A personal ChatGPT account in an underwriter's workflow. An underwriter pastes an application's free-text fields into a personal account to "summarize and flag concerns." The applicant's NPI is now with a vendor under terms the insurer never signed, with no confidentiality contract of the kind GLBA's service-provider exception assumes, and outside the AIS Program's inventory and vendor diligence. This is a textbook case of shadow AI.

    2. A consumer Claude plan in health claims triage. A health plan's claims handler asks Claude to summarize a long set of medical records from a Free or Pro account. Those plans run under Anthropic's consumer terms, not the commercial terms its BAA sits on, so PHI has been disclosed to a vendor without a business associate agreement.

    3. AI-drafted adverse underwriting notices. An insurer lets staff use an LLM to draft adverse underwriting decision letters. The drafts give generic reasons, or their wording varies with applicant demographics. In New York, Circular Letter No. 7 expects specific reasons and testing for disproportionate adverse effects; in Colorado, a covered insurer's framework is supposed to test for and remediate unfair discrimination. Nobody tested the drafting step.

    4. Vendor LLMs nested inside a vendor platform. A claims-management vendor adds a "smart summary" feature powered by an LLM provider's API. The insurer has a contract with the vendor; the vendor has one with the LLM provider. Whether the GLBA confidentiality terms or the HIPAA business associate chain reach the end of that chain is often unclear, and the NAIC bulletin, Colorado Regulation 10-1-1, and Circular Letter No. 7 all leave responsibility with the insurer.

    Control layers that map to the rules

    A 2026 insurance AI deployment that can stand up to these rules is a stack:

    Governance layer (NAIC bulletin)

    • A written AIS Program, owned by senior management accountable to the board or a board committee.
    • An AI inventory that lists every LLM-influenced workflow and assigns a risk tier.
    • Vendor due diligence files for OpenAI, Anthropic, Google, Microsoft, and any platform that subprocesses to them: security attestations (SOC 2, ISO 27001), DPAs, BAAs where applicable, and audit or cooperation terms where available.
    • Pre-deployment and ongoing testing, with quantitative bias analysis where state rules expect it.
    • Consumer notice processes, and a complaint path that gives consumers enough information to act on an adverse decision.

    Contract layer

    • A business tier for any LLM that touches NPI, PHI, or applicant data: for example ChatGPT Enterprise or the OpenAI API, Claude for Work or the Anthropic API, Gemini in Google Workspace, or Microsoft 365 Copilot. Anthropic says it does not train on commercial product inputs by default; OpenAI says API data is not used for training unless the customer opts in; Microsoft says Copilot prompts and responses are not used to train foundation models.
    • DPA / BAA signed and retained, with the BAA's covered products matched to the products actually in use.
    • Sub-processor lists monitored.
    • Data residency chosen deliberately for European-applicant data where the vendor offers it.
    • Audit and regulator-cooperation terms, which the NAIC bulletin lists as contract terms to seek "where appropriate and available."

    Tier-and-settings layer

    • No personal-account use by underwriters, claims handlers, or customer-service staff. That takes policy, training, and an enforcement mechanism, not just a memo.
    • Training disabled on every sanctioned tier where the vendor offers the setting.
    • Retention configured as low as the workload allows. OpenAI's Zero Data Retention for the API, for example, is subject to its prior approval.
    • Memory features off unless individually evaluated.

    Data-minimization layer (where the breach prevention happens)

    This is where Locke sits. Pasting an applicant's free-text health-history disclosure into a sanctioned ChatGPT Enterprise account may be contractually permitted but is operationally undesirable: the more PHI and NPI that crosses the wire, the more data is in scope for breach analysis if a future incident occurs. Data minimization reduces that. Locke 1.0 performs local-first redaction on-device, detecting names, account numbers, dates of birth, medical record numbers, ICD-10 codes, NPI numbers, drug names, diagnosis terms, and similar entities and replacing them with one-way placeholders before the prompt leaves the device. Reversible tokenization is planned shortly after 1.0.

    For insurance specifically, this layer is the difference between "we have a BAA" and "the BAA covered less data to begin with."

    Testing layer (Colorado, New York, EU AI Act)

    • Pre-deployment bias testing with a documented methodology, sample, and protected-class breakdown.
    • Proxy analysis, which Circular Letter No. 7 expects, examining whether the LLM's outputs correlate with protected classes through non-explicit inputs.
    • Production drift monitoring with thresholds and an incident-response process when thresholds are crossed. Both the NAIC bulletin and Colorado Regulation 10-1-1 name model drift.
    • An officer-signed annual compliance report, which Colorado requires of covered insurers that use ECDIS.

    Frequently asked questions

    Does the NAIC Model Bulletin apply to claims handling, not just underwriting?

    Yes. The bulletin expects the AIS Program to address AI systems "across the insurance life cycle, including areas such as product development and design, marketing, use, underwriting, rating and pricing, case management, claim administration and payment, and fraud detection." An LLM used to summarize claim documents or draft adjuster correspondence falls within it in a state that has adopted the bulletin. The controls can be lighter than for underwriting, because the bulletin scales them to the degree of potential harm to consumers, but the use still belongs on the inventory. New York's Circular Letter No. 7, by contrast, addresses only underwriting and pricing.

    Can we use ChatGPT for general-purpose tasks if no insurance data is in the prompt?

    The NAIC bulletin is aimed at AI systems that make or support decisions related to regulated insurance practices. Using ChatGPT to help an HR team draft a job posting or summarize a vendor's product brief is not that. The line moves the moment a prompt touches applicant or insured data, claims content, or rating factors, and GLBA and HIPAA attach to customer data whatever the task.

    What does Colorado Reg 10-1-1 mean for a national life insurer that does not write business in Colorado?

    Regulation 10-1-1 applies to insurers authorized to do business in Colorado that offer individually issued life insurance, private passenger auto insurance, or health benefit plans. A life insurer with no Colorado business is outside it. Its framework (board oversight, an inventory with version control, documented testing, officer-signed annual reports) overlaps heavily with what the NAIC bulletin and New York's Circular Letter No. 7 ask for, so a program built to the Colorado standard covers much of the ground those regimes cover.

    How do we satisfy disparate-impact testing for an LLM used in underwriting?

    None of the regulators' documents reviewed here prescribes a test specific to LLMs. Circular Letter No. 7 describes a three-step quantitative assessment: measure disparities in outcomes across protected classes, assess whether any disparity has a legitimate explanation, and search for less discriminatory alternatives, using metrics such as adverse impact ratios and denial odds ratios, before deployment and regularly after. Colorado's 2023 draft testing rule for life insurers would estimate race and ethnicity with BIFSG and compare decline rates and premiums. For an LLM, that usually means identifying the decision it influences, building samples that vary proxy attributes (name, geography, occupation, vocabulary), running the LLM on each, and analyzing the outputs for distributional differences. The protected classes come from the applicable state law, such as the list in Colorado's § 10-3-1104.9.

    Is an AI-generated denial letter a "decision" or just a draft?

    The regulators look at function rather than labels. The NAIC bulletin scales controls partly by "the extent to which humans are involved in the final decision-making process," so a nominal human sign-off lowers risk only if the reviewer can and does override the AI. Circular Letter No. 7 requires adverse underwriting notices to give specific reasons and bars hiding behind a vendor's proprietary algorithm. Colorado's Regulation 10-1-1 requires that a provider acting for a health benefit plan insurer be ultimately responsible for authorization decisions that ECDIS-based models inform.

    Do we need a separate program for LLMs versus traditional ML?

    The NAIC bulletin does not require one. Its definitions cover predictive models and generative AI alike, and it says validation and testing "should be reflective of the underlying components of the AI System, whether based on Predictive Models or Generative AI." One AIS Program with risk-tiered procedures can cover both. LLMs do bring specific risks, such as prompt injection, fabricated identifiers, and behavior changes driven by upstream model updates, that are worth naming explicitly in the program.

    The bottom line

    The riskiest assumption an insurer can make in 2026 is that AI is a productivity tool the law hasn't caught up with. It has, through the NAIC Model Bulletin in 29 jurisdictions, through Colorado Regulation 10-1-1 and New York's Circular Letter No. 7, through HIPAA, GLBA, and FCRA, and through the EU AI Act's high-risk classification of life and health insurance pricing.

    The insurers that handle this best aren't the ones who buy the most sophisticated AI tools. They're the ones who treat AI like every other regulated process: governed, inventoried, tested, contracted, and kept away from applicant and insured data unless that data has been minimized first. A technical control between the underwriter's keyboard and the prompt doesn't replace the governance file, but it shrinks how much personal data the governance file has to account for.

    Sources and primary references

    Reviewed November 5, 2026. Regulations are cited to their official text where it could be retrieved. Vendor statements are cited to the vendor's own documentation as it read on that date; vendors revise these pages, so confirm before relying on them.

    Sonomos publishes this blog and makes Locke, which detects and masks sensitive information on your device before it reaches an AI model. How Locke works

    This article is general information about privacy and AI, not legal advice, and reading it does not create a solicitor–client or attorney–client relationship. Regulatory requirements turn on your specific facts and jurisdiction and change over time. Consult qualified counsel before relying on anything here, and check the publication and last-reviewed dates above. We may revise posts, but we do not undertake to keep any post current as the law changes.

    Reference pages for this post