AI in insurance

Segments covered: property and casualty (P&C) carriers, with examples from personal auto, homeowners, and small commercial lines; notes on how life and annuity carriers and agencies and brokerages differ. Health, reinsurance, and title are outside this guide's lens. Linked workflows: order intake as the pattern for staged claim intake, support replies, knowledge retrieval. Linked roles: support leaders, operations leaders, finance leaders.

An insurer sells a promise and keeps it later, sometimes years later, when a loss occurs. Between the sale and the payment sit underwriting, pricing, policy service, claims handling, and fraud review, and nearly every step produces a decision a state regulator can examine. The work is document-heavy enough to invite automation, and the decisions it feeds are closely supervised.

This guide combines three kinds of material. Sector figures come from National Association of Insurance Commissioners (NAIC) financial snapshots, the Census Bureau, and the Bureau of Labor Statistics. Regulatory statements go no further than the regulators' own documents. Recommendations about sequencing, ownership, and controls are proposed practice: editorial reasoning, not findings. The carrier in the sequencing section and the $500 million carrier in the economics section are invented. No insurer was interviewed and no deployment was observed for this chapter.

How the business makes money

A P&C insurer collects premiums in advance, pays claims later, and invests the money it holds in between. The combined ratio adds claims, claim-handling cost, and acquisition and administration cost as shares of premium; below 100% means an underwriting profit before investment income.

The NAIC's year-end 2025 snapshot reports, for the US P&C industry, adjusted to eliminate affiliated amounts:

Source: NAIC, Industry Snapshots for the Period Ended December 31, 2025. The population is insurers filing annual statements with the NAIC; figures may change with later filings or amendments.

By our own arithmetic, not as NAIC-published ratios: loss adjustment expense, the cost of investigating, adjusting, and defending claims, was about 9.0% of earned premium in 2025 ($86.0 billion ÷ $958.7 billion), and losses about 57.5%; together they make the 66.5% loss ratio. The expense ratio is measured against written premium. Investment income ($88.9 billion) exceeded the underwriting gain ($68.7 billion), so the timing of claim payments matters as well as their size.

The ratios move with weather and the pricing cycle. The NAIC's 2024 full-year report attributes that year's underwriting profit, the first in four years, to rate increases and a turnaround in personal lines, and cites NOAA's count of 27 US events with at least $1 billion in insured losses in 2024. NAIC, 2024 Full Year Results

How other segments differ. Life and annuity carriers earn from long-duration contracts and investment spread; the same snapshot reports 2025 life and accident-and-health net earned premium of $796.1 billion, benefits of $362.9 billion, commissions and general expenses of $168.1 billion, and net investment income of $258.9 billion. Underwriting can draw on medical records, and service over decades of in-force business weighs more than claims. Agencies and brokerages earn commissions without bearing underwriting risk; their economics turn on account-manager capacity for submissions, renewals, and certificates. This guide does not develop either segment.

For scale, BLS estimates August 2026 employment at 1,574,800 for insurance carriers (NAICS 5241) and 1,358,100 for agencies, brokerages, and other insurance-related activities (NAICS 5242), seasonally adjusted and preliminary. BLS CES5552410001, CES5552420001 Those counts say nothing about how work in the sector is changing.

Where AI could touch the value chain

No Field Guide chapter is written for insurance. The table names the chapter whose pattern transfers; the notes after it say what differs. "Not yet covered" means the guide has no chapter for it.

Claim intake. The order-intake chapter separates extraction, validation, and authority, stages records for a person to approve, and prevents duplicates on retry. Those transfer directly. What differs: the reporter may be an agent, claimant, body shop, or attorney; one loss can arrive through several channels; and notices can contain health information. Creating a draft claim record is staging; deciding coverage, setting a reserve, or authorizing payment is not.

Photos. The computer vision chapter shows why base rates and capture conditions decide whether image classification helps. A defensible first use is checking that usable photos arrived; estimating a repair amount feeds the settlement decision and belongs with the uncovered rows.

Service. The support chapter drafts replies from approved evidence for a person to send. What differs: a coverage answer can be read as the insurer's representation, so route coverage questions to licensed staff.

Wordings. The knowledge retrieval chapter enforces permissions before search and answers from cited sources. For insurers the hard part is versioning: the form edition and endorsements in force on the date of loss govern, and a system that returns the current form for a two-year-old loss is wrong in a way that looks right.

Calls. The speech systems chapter covers accuracy, speaker attribution, and consent; state recording-consent rules are a question for counsel.

Why the uncovered decisions differ. Underwriting, pricing, settlement, and fraud scoring are where a model's output becomes a decision adverse to a consumer: a declined application, a higher premium, a reduced or denied claim, an investigation. The regulatory documents below address those decisions directly, through unfair discrimination standards, claims settlement standards, adverse-decision notices, and the expectation that outcomes can be explained. They also act on losses, the largest part of the combined ratio. Their omission is a coverage gap, not a judgment that AI has no role there.

Constraints that shape every project

The governance expectation. The NAIC Model Bulletin on the Use of Artificial Intelligence Systems by Insurers, adopted December 4, 2023, reminds insurers that "decisions or actions impacting consumers that are made or supported by" AI systems "must comply with all applicable insurance laws and regulations," including unfair trade practice and unfair discrimination laws. It expects a written "AIS Program," with senior management accountable to the board, controls proportionate to the potential harm to consumers, and coverage of underwriting, pricing, marketing, "claim administration and payment, and fraud detection." It covers vendor-built systems, suggests vendor contract terms for audit rights and cooperation with regulators, and says an insurer under investigation "can expect to be asked about its development, deployment, and use of AI Systems" whether or not it has a written program. NAIC Model Bulletin

A model bulletin takes effect only where a state issues it. The NAIC's implementation map, status as of October 8, 2026, lists 29 jurisdictions as adopting it (27 states, the District of Columbia, and Puerto Rico), the latest being South Carolina on September 24 and North Dakota on October 5, 2026, and separately lists insurance-specific AI rules or guidance in California, Colorado, New York, and Texas; it does not determine whether each adoption contains every element of the model. NAIC implementation map Whether an intake tool "supports" a regulated decision is a question for counsel; this guide proposes listing every AI system in the program, including those that only stage data, and letting risk tiering set the controls.

New York. Insurance Circular Letter No. 7, issued July 11, 2024, says insurers should not use external consumer data or AI systems in underwriting or pricing "if such use would result in unfair discrimination." It sets out a three-step assessment ending in a documented search for a less discriminatory alternative. It keeps responsibility for vendor tools with the insurer and describes written notice within 15 days when an applicant cannot be underwritten using those data or systems. It "is not intended to address phases of the insurance product lifecycle other than underwriting and pricing." NYDFS Circular Letter No. 7 (2024)

Colorado. SB21-169, signed July 6, 2021, bars insurers from unfairly discriminating on race, color, national or ethnic origin, religion, sex, sexual orientation, disability, gender identity, or gender expression in any insurance practice, including through external consumer data. SB21-169 Its regulation, 3 CCR 702-10, first applied to life insurers (effective November 14, 2023); as amended effective October 15, 2025 it also covers private passenger auto and health benefit plan insurers that use external consumer data, or models using it, in any insurance practice. It requires a board-overseen framework to detect and remediate unfair discrimination with respect to race through the Division's quantitative testing, with auto insurers' annual compliance reports due from July 1, 2026. 3 CCR 702-10

Federal uncertainty. On December 16, 2025 the NAIC said a federal executive order on state AI laws "introduces legal uncertainty" for oversight of rate setting, underwriting, and claims processing. NAIC statement How the order bears on a state's insurance rules is a question for counsel.

Claims handling standards. The bulletin names state unfair claims settlement practices acts among the laws AI-supported decisions must meet. When an AI-staged notice counts as received, and what happens if review backs up past a state deadline, are questions for counsel and claims compliance.

Claim-file data. Injury claims carry medical information; property claims carry photos of homes and third parties; recorded statements carry voices. Which privacy and retention rules apply is a question for advisers; this guide states none. Proposed practice: inventory claim content by type before a pilot and read vendor retention and training terms against it.

Form versions. If the insurer cannot say which forms were attached to a policy on a past date, retrieval and service drafting will produce confident answers from the wrong text. Repairing that linkage is ordinary data work, and it comes first.

Catastrophe surges. First-notice volume after a hurricane, wildfire, or hailstorm can rise far above normal for weeks. A system tested in a quiet month has not been tested. Size the exception queue at surge volume, as the operations leaders guide shows, and plan the manual fallback.

The sector's economics lens

For your finance lead

The decisive question is which part of the combined ratio an effect touches: losses, loss adjustment expense, or underwriting expense. They behave differently and should not be added without a counterfactual for each.

The illustration uses an invented carrier with $500 million of net earned premium and the 2025 industry ratios above, assuming written equals earned premium and splitting the 66.5% loss ratio into 57.5% losses and 9.0% loss adjustment expense by our arithmetic. These are aggregates applied to a fictional firm; a real carrier's ratios differ by line, state, and year.

Loss adjustment expense. An intake project that removes $500,000 a year of paid cost lowers loss adjustment expense by 1.11% ($500,000 ÷ $45.0M) and the combined ratio by 0.1 point ($500,000 ÷ $500M), equal to 1.41% of the underwriting gain ($500,000 ÷ $35.5M). It exists only if a paid cost stops: fewer outside adjusters in a surge, less overtime, an unfilled vacancy. Loss adjustment expense also includes costs intake does not touch, such as defense of litigated claims, so the addressable share is a firm-level number. The economics chapter explains the capacity-to-cash distinction.

Losses. A 1% change in losses is $2.875 million, or 0.575 point, almost six times the intake example. Vendor pitches call it reducing claim leakage, but what claims cost the insurer is what claimants receive, set through settlement decisions that state law governs and this guide does not cover. A lower loss ratio from underpaying valid claims is not a gain; it is the outcome the regulatory documents above exist to prevent. Treat any projected loss-ratio effect as a hypothesis for actuarial and compliance review, never as an efficiency benefit.

Underwriting expense. Service drafting, submission intake, and retrieval act on the 25.8%, part of which is agent and broker commission that internal speed does not change.

Double counting. Adjuster time saved fixing intake errors is already inside the intake saving. A cycle-time gain claimed as better retention is an unmeasured revenue hypothesis. Faster payment slightly reduces invested funds. Results gross and net of reinsurance differ; say which.

What is measured about adoption and outcomes

For your analyst

Adoption, from the Census Bureau. The Business Trends and Outlook Survey (BTOS) asks employer businesses every two weeks: "In the last two weeks, did this business use Artificial Intelligence (AI) in any of its business functions?" For the collection period September 21 to October 4, 2026 (reference period September 7 to 20; published October 8, 2026), the estimate for insurance carriers and related activities (NAICS 524) was 37.4% yes, standard error 2.59 points; the three prior cycles, in date order, were 35.7%, 32.8%, and 37.5%. In the same cycle 46.5% (standard error 2.15) expected to use AI in the next six months, against 23.8% current use for all businesses. BTOS subsector data, national data The broader finance and insurance sector (NAICS 52), which combines banks, securities firms, and insurers, was 40.1%. BTOS sector data

Read these narrowly:

• NAICS 524 includes carriers, agencies, brokerages, and claims adjusting firms. Estimates are shares of businesses, not of employment or premium, so a small agency and a national carrier count equally. BTOS methodology, version 5
• A 90% confidence interval is about 4.3 points either side (1.645 × 2.59). The rise from 30.4% four cycles earlier to 37.4% now is not clearly distinguishable from sampling noise, and the subsector and NAICS 52 estimates overlap.
• Multi-unit companies in several sectors or states are reported separately, so large groups may be outside the estimate.
• The question was reworded in November 2025, so earlier estimates are not comparable. Census Bureau, May 2026

The BTOS AI supplement (reference period November 2025 to January 2026) reports only at the NAICS 52 level, banking and insurance together. Among all such businesses, 16.2% reported AI use in customer service in the last six months and 11.8% in legal and compliance. Among AI users, 54.9% used it to supplement an employee's task and 13.7% to perform a task previously done by an employee; 95.0% reported no change in total employment from AI. BTOS AI supplement, 2026, reference period

Adoption, from regulators. In December 16, 2025 testimony to New York legislative committees, the NAIC reported that its surveys found adoption or planned adoption of AI or machine learning by approximately 88% of auto insurers, 70% of homeowners insurers, and 58% of life insurers surveyed, citing survey reports from December 2022, August 2023, and December 2023, and listed claims uses including help determining settlement amounts, damage assessment, and fraud detection. NAIC testimony The figures combine use, planned use, and exploration, and the auto survey collected responses in 2021. NAIC auto survey page They measure reported use, not outcomes.

Outcomes. We found no attributable, independently documented measurement of AI effects on claims cycle time, loss adjustment expense, service handling time, or loss ratios, and this guide reports none. Vendor case studies exist in this market; none is cited here.

Who is involved

The governance owner sits across all of these: claims and service projects belong in one program with one inventory and one vendor diligence process. The vendor diligence playbook gives a procedure; add the bulletin's suggested audit and cooperation terms.

Expertise specific to this industry. A general AI builder does not bring insurance governance knowledge: how a system fits the AIS Program and its risk tiering where a state has issued the NAIC bulletin, the vendor audit and cooperation terms the bulletin suggests, unfair claims settlement practices standards for claims intake, and, for any New York underwriting or pricing use, the Circular Letter No. 7 assessment ending in a search for a less discriminatory alternative. This guide places that expertise in-house, with the AI governance owner and the claims leader, and sends legal questions to counsel as adviser. See Staffing an AI pilot and Choosing and working with an AI partner.

Sequencing: a first move and what follows

The following is a fictional, proposed sequence. Halverd Mutual is an invented regional P&C carrier writing personal auto and homeowners in four states, three of which have adopted the NAIC bulletin, plus small commercial business through about 300 independent agencies.

First: inventory and repair. Halverd lists every AI tool in use, including features inside its claims and contact-center software, and adds them to its AIS Program with owners. It then finds that for older homeowners policies its system cannot say which forms were attached on a past date. Repairing that linkage is the first investment, and needs no model.

Second: staged intake for agent-reported claims. About a third of Halverd's notices arrive from agencies by email with attachments, and staff re-key them. Halverd pilots extraction that stages a draft claim record for an intake specialist to approve, following the order-intake pattern. The system cannot confirm coverage, set a reserve, or assign payment, and it matches against open claims before anything is created, because the same loss also arrives by phone. It runs in a quarter with no declared catastrophe, measured per notice, field, and duplicate prevented; the cash test is whether the next storm needs fewer outside adjusters for intake than a comparable past event.

Third: claims call transcription. With counsel's answer on consent in each state, Halverd transcribes recorded claims calls and drafts file notes the adjuster edits, measured by speaker and condition as the speech systems chapter describes.

Later: retrieval over wordings. Once form linkage is reliable, Halverd tests retrieval that returns the governing form text for a policy and date, with citations. It locates text; the adjuster decides coverage. Pricing models, settlement estimates, and fraud scores stay with their existing owners and governance throughout.

Counterexample: when to wait. Suppose Halverd had found that 80% of notices already arrive as structured data from its own web and app forms, that its claims system cannot hold a staged record without creating a live claim, and that it has no AI inventory. Then extraction would address little volume, any error would create a real claim, and a regulator's question about AI use would have no documented answer. The right call would be the inventory, form linkage, and better notice forms, revisiting AI intake later.

Questions for operators, investors, and vendors

• Which AI systems are in use today, including features inside purchased software, and is each in the AIS Program with an owner and risk tier?
• Does the system make or support a decision about a consumer? If it only stages data, what stops its output becoming a decision without review?
• Can the policy system identify the forms and endorsements in force on a past date of loss?
• What share of first notices arrives unstructured, by channel, and how does that change in a catastrophe?
• Does any projected benefit come from paying less on claims? Who in actuarial and compliance reviewed it?
• What claim data, including medical information and photos of third parties, would the vendor receive, retain, or train on, and do the contracts give audit rights and regulatory cooperation?
• For a vendor claim, what were the insurer population, line, baseline, and period, and was it independently verified?

When to wait

Defer if there is no inventory or governance program for the AI already in use, if form history cannot be tied to policies, if the claims system cannot hold a staged record apart from a live claim, or if notices and service requests already arrive structured. Defer anything touching underwriting, pricing, settlement, or fraud scoring until the testing, documentation, and adverse-decision processes the applicable state rules describe are in place, and take the question of what applies to counsel. Better notice forms, form-library cleanup, and a written AI inventory may remove more risk and work with fewer uncertainties. The Field Guide treats those as legitimate outcomes, not failures to adopt AI.

Evidence still needed and related reading

Attributable operator accounts are missing for how carriers run these processes: unstructured notice share by channel, intake error and duplicate rates, how form history is stored, and how loss adjustment expense divides among intake, adjusting, and defense. No independently documented outcome measurement of AI in insurance operations has been identified. The NAIC survey reports behind its testimony were not read directly, and the California and Texas guidance on the NAIC map has not been reviewed.

Continue with purchase-order intake for the staged-write pattern, document extraction, support replies, retrieval, operations leaders, and the economics chapter.

Hiring in New York: open roles at NYC AI companies building for finance and insurance, read daily from the companies' own job boards. These companies sell AI into the industry; they show which skills vendors are hiring for, not what firms in the industry are hiring.