AI in insurance that underwriting, claims and your regulators can all trust
AI in insurance now sits under explicit regulatory expectations. We design the platforms, governance programs and testing evidence that let carriers use AI across claims, underwriting and service while meeting state requirements on fairness and oversight.
What is different about AI in insurance
AI in insurance is regulated at the level of outcomes. State insurance regulators care less about which model you chose and more about whether its use leads to decisions that are inaccurate, arbitrary or unfairly discriminatory. That lens covers underwriting, pricing, claims and marketing, and it applies to models you buy from vendors as much as models you build.
Insurers are also document businesses. Applications, loss runs, medical records, adjuster notes, policy forms and endorsements are exactly the unstructured material generative AI handles well. The opportunity is large, but so is the exposure, because the same documents carry health information, personally identifiable data and facts that decide whether a claim is paid.
The carriers moving fastest treat governance as a design input. They keep a written AI program, inventory every AI system including third-party ones, and build testing for bias and accuracy into the platform so that a market conduct exam finds evidence rather than intentions.
Where AI earns its place
Claims intake and first notice of loss
Extract facts from FNOL calls, emails, photos and forms, populate the claim file and flag coverage questions and potential fraud indicators for the adjuster. The model accelerates triage; coverage and payment decisions stay with licensed staff and documented rules.
Adjuster and examiner copilots
Summarize long claim files, medical records and correspondence, draft reservation of rights letters and status updates, and answer questions about the policy language that actually applies. Every answer cites the endorsement or document it came from.
Commercial underwriting submission triage
Read broker submissions, loss runs and schedules, extract exposures into structured data, check appetite and prioritize the queue. Underwriters spend time on risks they can write, and the extraction accuracy is measured field by field against reviewed samples.
Policy form and product knowledge
A retrieval system over filed forms, endorsements, rate manuals and underwriting guidelines, versioned by state and effective date. Service, underwriting and claims staff get answers tied to the form edition in force, not whatever document the search found first.
Subrogation and recovery identification
Scan closed and open claim files for recovery opportunities that manual review misses, with a written rationale for each referral. Recovery teams review the referral, and precision is tracked so the model does not flood them with weak leads.
Agent, broker and policyholder service
Answer billing, coverage and endorsement questions across chat and contact center, with handoff to licensed staff when a question requires advice. Guardrails prevent the assistant from binding coverage, interpreting policy beyond approved answers or making claims promises.
The regulatory constraints that shape insurance AI
Insurance AI regulation is state by state and still moving, so we design for the common core and document state-specific choices. These are the constraints we plan around from day one.
- NAIC Model Bulletin on the Use of AI Systems by Insurers
- Adopted by the NAIC in December 2023 and since adopted in some form by a large number of states, the bulletin expects a written AI systems program, governance and risk management controls, oversight of third-party AI, and documentation regulators can review during examinations.
- Unfair discrimination and unfair trade practices laws
- Existing state laws already prohibit unfairly discriminatory rating, underwriting and claims practices, regardless of whether a human or a model made the decision. We build testing for proxy discrimination and disparate outcomes into model development and ongoing monitoring.
- Colorado SB21-169 and state-specific rules
- Colorado's law on external consumer data, algorithms and predictive models requires governance and testing to show those tools do not unfairly discriminate, with implementing regulations first applied to life insurers and other lines under consideration. New York DFS has also issued guidance on AI and external data in underwriting and pricing.
- Health and personal information
- Claim files and applications can contain protected health information subject to HIPAA where the insurer is a covered entity, plus state insurance privacy laws and consumer privacy statutes. Data flows, retention and model provider contracts must reflect the most sensitive data a use case touches.
- Claims handling and adverse decisions
- Unfair claims settlement practices laws require prompt, fair investigation and explanation. An AI system that influences denials or reductions must be explainable to the claimant and the regulator, and the licensed professional must remain accountable for the decision.
- Third-party models and data vendors
- Regulators expect insurers to answer for vendor models and data they use. We design vendor due diligence, contractual audit and testing rights, and platform controls so vendor AI is inventoried, monitored and replaceable like anything built in-house.
How we architect AI for insurers
We start with the AI systems program the NAIC bulletin describes and make it operational: an inventory that the platform populates automatically, risk tiers that decide which controls apply, and a clear line between systems that assist staff and systems that influence consumer outcomes. The second group carries the heavier testing and documentation load.
The platform itself routes every model call through a gateway that tags the line of business, state and use case. That tagging matters in insurance because obligations differ by state and line. When a regulator asks how AI is used in private passenger auto claims in a given state, the answer comes from logs, not from a survey of business units.
Document intelligence is built for claim and underwriting files: extraction pipelines with field-level confidence scores, human review queues for low-confidence fields, and retrieval that respects form editions and effective dates. Protected health information is segregated, and models that see it run under agreements that prohibit training and retention.
Fairness testing is designed in, not bolted on. For pricing, underwriting and claims models we define protected-class testing approaches suited to the data available, often using accepted inference methods where demographic data is not collected, and we set monitoring thresholds that trigger review when outcomes drift.
AI in insurance: common questions
What does the NAIC Model Bulletin require insurers to do about AI?
The NAIC Model Bulletin, adopted in December 2023, expects insurers to maintain a written program for the responsible use of AI systems that make or support decisions affecting consumers. That program covers governance, risk management and internal controls, oversight of third-party AI and data, and documentation. Regulators can request that documentation during investigations and market conduct examinations in states that adopted it.
How do insurers test AI models for unfair discrimination?
Insurers test for unfair discrimination by comparing model outcomes across protected classes and checking whether inputs act as proxies for them. Because many insurers do not collect race or ethnicity, testing often uses accepted inference methods such as surname and geography-based estimation. Results, thresholds and remediation decisions should be documented and repeated when models or data change.
Can generative AI be used in claims handling?
Yes, generative AI is well suited to claims work such as summarizing files, extracting facts from documents and drafting correspondence. The safeguards are that licensed adjusters make coverage and payment decisions, outputs cite their sources, and any AI influence on a denial or reduction is documented and explainable under unfair claims settlement practices rules.
Are insurers responsible for AI models they buy from vendors?
Yes. State regulators expect insurers to answer for third-party AI systems and data they use in regulated decisions, just as for models built internally. That means due diligence before purchase, contract terms that allow audit and testing, inclusion in your AI inventory, and ongoing monitoring of vendor model performance and outcomes.
Where should an insurance carrier start with AI?
Most carriers should start with internal productivity in claims and underwriting, such as file summarization and submission extraction, because the value is immediate and the consumer-outcome risk is lower. At the same time, stand up the AI systems program and platform controls the NAIC bulletin expects, so pricing and decisioning use cases later inherit governance that already works.
Other industries
- Financial Services
- Healthcare and Life Sciences
- Public Sector
- Manufacturing
- Legal and Professional Services
- Technology and SaaS
In the capital region? We are based in Rockville, Maryland, and run architecture workshops in person. See enterprise AI consulting in Washington DC, Maryland and Virginia.
Bring us the architecture question you are stuck on.
Kelsey, our AI assistant, can answer questions and book you a 30-minute briefing with a senior architect in about two minutes, by voice or by typing.
- Vendor-neutral advice. We resell nothing.
- Senior architects on the call, not a sales team.
- Mutual NDA available before we talk specifics.
Book with Kelsey
She finds a time that works, books it on our calendar and sends the invite. No forms.
By talking or chatting with Kelsey, you agree we may text and email you about your request. Message and data rates may apply. Reply STOP to opt out. Terms and Privacy.
Prefer a person? Call (301) 301-5765.