AI for Insurance Companies: Underwriting, Claims, Subrogation and Servicing, and the Rules Carriers Work Under

How is the insurance industry using AI?

Insurers have used predictive models for decades, in rating plans, fraud scores and reserving. What changed in the last few years is that language models can now read and write the documents that carry most of a carrier’s work: ACORD forms and broker emails, loss runs, adjuster notes, medical reports, policy wordings, endorsements and regulator correspondence.

The NAIC’s line-of-business surveys give a sense of scale among US carriers. Of the companies that responded, 88% of private passenger auto insurers, 70% of homeowners insurers, 58% of life insurers and 92% of health insurers said they were using, planning to use or planning to explore AI or machine learning. The surveys also found that roughly half of the marketing models in use came from third-party vendors, while auto and home carriers mostly built pricing and underwriting models themselves.

This guide follows the value chain from submission to subrogation and then covers the rules, the tools and a sequencing plan. Regulatory sources were read on the regulators’ own sites on 25 October 2026. The insurance AI use case hub lists the individual workflows in more detail.

Submission intake and AI in insurance underwriting

Commercial underwriters spend a large share of their week reading. A mid-market property submission might include a broker cover email, an ACORD application, a statement of values with hundreds of locations, five years of loss runs and engineering reports. Much of the work is getting that material into a usable shape before any judgement is applied.

Useful tasks for a model:

• Extraction. Pull named insured, occupancy, construction, protection, values, requested limits and deductibles into the carrier’s intake format, with the page each value came from.
• Appetite check. Compare the extracted risk with the written underwriting guidelines and state which rules it meets, which it breaks and which need more information.
• Loss run summary. Total losses by year and cause, call out large losses and open claims, and note where loss runs and the application disagree.
• Referral note. Draft the note an underwriter sends to a referral authority, in the house format, citing the submission documents.

The underwriter keeps acceptance, terms and price. That line matters for regulation as well as quality. In personal lines and life, models that use external consumer data to support underwriting or rating are the core subject of Colorado’s regulation and New York’s circular letter, both covered below. The underwriting submission and risk review guide goes deeper on how an agent grades a submission against appetite.

Claims: FNOL triage, file summaries and correspondence

Claims is where most carriers see the largest volume of documents per employee, and where a weak control turns into a regulatory finding fastest.

First notice of loss

At intake, a model can classify the loss type, identify the policy and the coverage parts that might respond, list the documents still needed and suggest a handling route, such as fast track, desk adjuster or field. Combined with existing severity scores, it helps assign the claim to the right person on day one.

File summaries during the claim

Bodily injury and complex property files grow quickly: medical records, repair estimates, expert reports, recorded statement transcripts and months of diary notes. A model that produces a dated chronology, a summary of treatment or damage, open issues and a draft reserve rationale saves a handler from rereading the whole file at each review. Every point should link to the source page.

Letters and status updates

Acknowledgement letters, requests for information and status updates follow templates and state-specific timing rules. Drafting them from the file, for the adjuster to approve, is low-risk and popular. Coverage position letters and denials need far more care and should always be written or edited by the adjuster who owns the decision.

The Unfair Claims Settlement Practices Act standards apply regardless of the tool used, which is the point the NAIC bulletin makes in its first section. The claims processing guide and the claims processing network playbook show how intake, document reading, rules and approval routing fit together with an audit trail.

Subrogation and fraud signal summaries

Subrogation. Recovery opportunities are often visible in the file but missed: a police report naming another driver, a contractor’s invoice for the work that failed, a product mentioned in a fire investigator’s notes. A model that reads each closing or reserved file for third-party involvement, and drafts a short recovery memo with the evidence and the limitation dates to check, gives recovery specialists a ranked worklist instead of a sample.

Fraud signals. Most carriers already run scoring models and rules in their special investigations unit. Language models add value on the narrative side: summarising why a claim scored high, comparing statements across interviews, noting inconsistencies between the loss description and the photos or estimates, and drafting the referral summary. The decision to refer, and anything said to the claimant, stays with the investigator. California’s 2022 bulletin on racial bias named the flagging of claims from particular ZIP codes for SIU referral as one of the practices under review, which is a reminder to test referral patterns, not just model accuracy.

Policy servicing and broker communications

Servicing teams answer the same questions many times a day. Does this policy cover a home office? What changes if the insured adds a vehicle? When does the cancellation notice period start? Retrieval over approved policy forms, endorsements, rating manuals and service procedures lets a model draft an answer with the clause quoted, for a service representative to send or adjust.

Broker communication is a similar pattern at higher stakes. Brokers email underwriters for quote status, clarifications and mid-term changes. A model can classify the inbox, draft replies from the file and flag messages that change the risk, such as a new location or a claim, for the underwriter. Anything that binds, changes terms or states a coverage position should be approved by an authorised person before it goes out.

From 2 August 2026, the EU AI Act’s Article 50 transparency rules apply to chatbots that interact with people, so a carrier serving EU policyholders through a chat assistant must tell users they are dealing with an AI system. In the US, the NAIC bulletin asks carriers to have processes for telling consumers that AI systems are in use, at a level suited to the stage of the insurance life cycle.

Actuarial, finance and regulatory reporting

Actuarial and finance teams already work with code and statistical tools, so language models help mainly with the text around the numbers. Examples include drafting the narrative of a reserve review from the actuary’s exhibits, explaining large movements in a quarterly report, summarising rate filing objections from a department of insurance and drafting the response for the actuary to finish, and preparing first drafts of sections of the ORSA summary report or the Corporate Governance Annual Disclosure.

Keep calculation in the actuarial tools. A model can explain what a triangle shows, but the triangle itself should come from the reserving system, and every figure in a regulatory document should trace back to it. The EU AI Act guide for banks and insurers notes that AI used to calculate insurers’ capital requirements under Union law is not treated as high-risk, which keeps most actuarial modelling outside Annex III.

Generative AI in the insurance industry: where it helps and where it stays out

A useful rule of thumb divides insurance AI into two families.

Predictive models score, rank and price. Generalised linear models, gradient-boosted trees and similar methods sit behind rating plans, fraud scores, propensity models and reserving. They are validated statistically, filed with regulators where rates are concerned and tested for unfair discrimination.

Generative models read and write. They extract data from documents, summarise files, answer questions from policy wordings and draft correspondence and reports. They should not set prices or make coverage decisions, and they are poor at arithmetic unless they call a calculation tool.

The NAIC bulletin covers both. It defines generative AI explicitly and says validation, testing and auditing should reflect whether a system rests on predictive models or generative AI. In practice, that means accuracy testing for extraction tasks, review sampling for drafted letters and clear records of which model version produced which output.

The regulatory frame in brief

US insurance regulation of AI runs through the states, mostly by applying existing unfair trade, unfair discrimination and claims practices laws to new tools.

• NAIC model bulletin. Adopted on 4 December 2023, it expects every carrier to keep a written AI systems program covering governance, risk management, internal controls and third-party systems, and lists the documents a department may request in an examination. The NAIC adoption map, with status as of 8 October 2026, shows 29 jurisdictions that have adopted it and four with their own insurance-specific rules or guidance.
• Colorado. SB21-169 bars insurers from using external consumer data, algorithms or predictive models in ways that unfairly discriminate on protected characteristics. Its governance and risk management regulation, 3 CCR 702-10, has applied to life insurers since November 2023, and amendments effective 15 October 2025 extended it to private passenger auto and health benefit plan insurers.
• New York. Insurance Circular Letter No. 7, issued on 11 July 2024, covers AI systems and external consumer data in underwriting and pricing. It expects carriers to show their systems do not unfairly discriminate, using quantitative testing, to keep board and senior management oversight, to stay responsible for vendor tools and to give specific reasons when a decision is adverse.
• California. Bulletin 2022-5 reminded insurers that marketing, rating, underwriting, fraud investigation and claims must be free of bias, and said the department may review the algorithms and models behind them.
• EU AI Act. For carriers with European business, AI used for risk assessment and pricing of individuals in life and health insurance is high-risk under Annex III point 5(c). Under the Digital Omnibus, Regulation (EU) 2026/1744, those obligations apply from 2 December 2027.

This is a summary for planning. The bulletin’s requirements, adoption by state and the NAIC’s 2026 follow-on work deserve a closer read before a carrier writes its program.

Choosing AI tools for insurance companies

The market splits into four groups, and most carriers end up with more than one.

AI inside core systems. Policy administration, claims and billing vendors increasingly ship assistants and document processing within their products. These suit work that lives in one system. Ask which model provider processes your data, where, under what retention terms and whether you can export prompts, outputs and sources for your own records and for an examiner.

General enterprise assistants. Business plans of the large assistants handle drafting and summarising well under commercial terms that exclude training on customer data. They are a reasonable home for internal, non-consumer material. Claim files with medical records or a policyholder’s financial details need a closer look at contracts, data location and logs.

Insurtech point solutions. Specialist vendors cover FNOL intake, submission extraction, subrogation detection or fraud. Under the NAIC bulletin and New York’s circular, you remain responsible for their outcomes, so contracts should include audit rights and cooperation with regulatory inquiries.

Platforms you run yourself. Carriers that want claims, underwriting and servicing data to stay on their own infrastructure, or that want one governed layer across many systems, deploy a platform in their own data centre or private cloud. Our procurement guide for AI platforms sets out the questions to ask in an RFP and a proof of concept.

Where to start with AI for insurance companies

A workable AI strategy for insurance companies sequences projects by consumer impact and data sensitivity, which keeps the first ones useful and defensible. The table is a starting point for a carrier’s own AIS program inventory, not a substitute for it.

Two habits pay off across all of them. Keep an inventory from day one, listing each AI system, its purpose, owner, data sources, model version and the human decision it supports. And log every answer with the sources used, because the NAIC bulletin’s examination section asks for exactly that kind of evidence.

How VDF AI fits an insurance carrier

VDF AI is a platform a carrier installs in its own data centre, in a private cloud or in an air-gapped network. When it is set to local models only, claim files, submissions, policy wordings, retrieval indexes and logs never leave that environment. VDF AI holds no regulatory approval or certification of its own, so the deployment is governed by the carrier’s AIS program, vendor management and state examinations like any other system.

For underwriting, each line or team can have its own knowledge collection, and permission-aware retrieval only returns passages from documents the user may open, which keeps commercial submissions, medical underwriting evidence and claim files apart. Governed agents handle steps such as submission extraction, FNOL classification, file summaries and subrogation memos. Each role gets only the tools it needs, and a human approval step sits in front of anything that reaches a policyholder, broker or regulator. The audit trail records who asked, which model answered, which sources it used and which tools it called, which maps directly to the inventory and documentation requests in the NAIC bulletin.

The VDF AI Router registers open-weight models served through Ollama, vLLM or other OpenAI-compatible endpoints, and can admit approved cloud models for low-sensitivity tasks under policy, with budgets and per-call logging. Staff sign in through Microsoft Entra ID single sign-on on-premises, while Okta, Keycloak and other SAML or OIDC providers connect through an SSO-aware reverse proxy, and role-based access control is included on every plan. A carrier can run its first claims or underwriting pilot on a three-month platform trial before committing to a seat bundle.

The insurance solution page and the insurance agent hub describe the claims and underwriting agents in more detail. To find out whether data access, governance or skills will slow a first project, take the free AI readiness assessment. This guide is general information, not legal advice on any state’s requirements.