Insurance AI and Data Trends 2026: What Carriers Are Deploying

TL;DR: Insurance AI trends in 2026 are defined by a widening gap between adoption and production. Conning's survey found full AI adoption among insurers rose from 8% to 34% in a single year and large language model adoption from 18% to 63%, yet roughly 22% of carriers report reaching full production despite more than 90% testing AI during 2025. Regulation arrived in parallel: 25 states had adopted the NAIC Model Bulletin on insurers' use of AI as of July 2026, and the NAIC is piloting a multistate AI Systems Evaluation Tool with twelve states.

Insurance is further into AI adoption than most industries and no further into returns. Conning's survey data shows full AI adoption among insurers climbing from 8% to 34% in one year, with large language model adoption tripling from 18% to 63%, and roughly 90% of carriers somewhere on the pilot-to-production path.

The production numbers tell a different story. Around 22% of carriers report AI in full production despite more than 90% testing it during 2025, which is a familiar shape: the technology works in evaluation and stalls at the boundary where it has to touch a policy administration system, a claims file and a filed rate at the same time.

What follows is eight trends shaping carrier AI in 2026, each anchored to named survey or regulatory data, written for the people who own the combined ratio rather than the people who own the model.

Key takeaways

• Adoption quadrupled in a year. Conning's survey found full AI adoption among insurers rose from 8% to 34%, and large language model adoption from 18% to 63%.

• Production lags badly. Roughly 22% of carriers report AI in full production, against more than 90% testing AI during 2025.

• Underwriting leads the use cases. GlobalData's July 2026 survey found 34.4% of respondents expect underwriting and risk profiling to benefit most from AI, with claims management second at 18.3%.

• Regulation is now examinable. As of July 2026, 25 states had adopted the NAIC Model Bulletin on insurers' use of AI, requiring a written AI program and a named accountable person or committee.

• Examiners are getting a tool. The NAIC's Big Data and Artificial Intelligence (H) Working Group is piloting a multistate AI Systems Evaluation Tool with twelve participating states.

• Health carriers are furthest along. The NAIC's first major health insurance AI survey found 84% of health insurers use AI or machine learning in some capacity.

• Data readiness is the gate. Gartner's survey of 248 data-management leaders found 63% either lack AI-ready data practices or are unsure whether they have them.

• Agentic projects carry cancellation risk. Gartner predicts more than 40% of agentic AI projects will be canceled by the end of 2027 on cost, unclear value and inadequate risk controls.

Which insurance AI trends in 2026 actually matter to a carrier?

Carrier AI adoption has effectively saturated, and the useful metric for 2026 is the share of carriers running AI in production rather than the share experimenting. Conning's survey found full AI adoption rising from 8% to 34% year over year, with about 90% of insurers somewhere between pilot and production.

The gap between those figures and the roughly 22% reporting full production is where the industry's real work sits. A carrier with a model in evaluation has spent money and changed nothing. A carrier with a model in production has changed a workflow, which means it has also resolved data access, governance and human review.

This mirrors the cross-industry pattern. McKinsey's State of AI research across about 2,000 respondents found more than 80% of organizations report no tangible enterprise-level EBIT impact from generative AI. Insurance is not behind that curve. It is on it, with more pilots than most.

For an executive, the practical consequence is to stop measuring AI programme health by pilot count. Count workflows in production with a named owner and a tracked metric. Everything else is evaluation.

Where do carriers expect AI to create the most value?

Underwriting and risk profiling is the use case carriers expect to benefit most from AI in 2026. GlobalData's July 2026 survey found 34.4% of respondents naming it, nearly double the 18.3% who named claims management.

That ordering is rational. Underwriting is where selection happens, and a small improvement in selection compounds through the book in a way that a claims efficiency gain does not. It is also the workflow most constrained by submission volume: capacity limits how many submissions get quoted at all.

The constraint is document-shaped rather than model-shaped. A submission arrives as a broker email, an ACORD form that is partly blank and a loss run that is a scan. Turning that into a scored risk requires reading all three and joining them to prior claims history, which is an integration problem before it is an underwriting problem.

Carriers that have moved this into production generally did so by narrowing scope hard: one line, one class, a defined appetite, and a human underwriter approving every bind. You can see that shape in the insurance submission and triage workflow on this site.

Claims recovery is the most underworked opportunity on the book

Claims recovery is the trend carriers talk about least and quantify worst, because the money at stake is invisible in normal reporting. A claim that was adjusted correctly, paid correctly and closed does not appear in any exception queue, even when the file contains evidence that a third party was liable.

The structural reason is that nothing re-reads a claim after it is reserved unless the number moves. Police reports, adjuster notes and repair estimates are scanned into the file at intake and never opened again, and subrogation rights expire on a statutory clock that no system is watching.

This is a document problem that AI is unusually well suited to, and it is why claims sits second in GlobalData's July 2026 use-case ranking at 18.3% despite the underwriting focus. The work is reading every document in every open file against a fixed set of questions, which is exactly what a person cannot do at volume and a model can.

The governance requirement is strict, though. Any recovery demand has to cite the document and the line, because the counterparty's carrier will contest it. That makes provenance a functional requirement rather than a compliance nicety. The claims and subrogation workflow walks through what that evidence chain looks like.

Regulation moved from guidance to examination in 2026

AI regulation for insurers stopped being advisory in 2026. As of July 2026, 25 states had formally adopted the NAIC Model Bulletin on the use of artificial intelligence systems by insurers, with additional states moving through approval, and four more issuing their own AI-specific insurance regulation.

The bulletin's substance is governance rather than technology. Insurers are expected to maintain a written program for the responsible use of AI systems that make or support decisions in regulated insurance practices. That program must address risk management and internal audit, and name a responsible person or committee with documented authority up to senior leadership.

Enforcement is arriving with it. The NAIC's Big Data and Artificial Intelligence (H) Working Group is piloting a multistate AI Systems Evaluation Tool with twelve participating states, structured as an examiner questionnaire covering AI governance, risk management and model oversight.

For a carrier, the operational read is that AI decisions are now examinable artifacts. A model whose inputs cannot be reconstructed is a finding waiting to happen, which pushes provenance and access control from the nice-to-have column into the build requirements.

Health carriers are furthest ahead and most scrutinized

Health insurance is the segment with the deepest AI penetration and the sharpest regulatory attention. The NAIC's first major health insurance AI survey found 84% of health insurers use AI or machine learning in some capacity.

That level of adoption in a segment where AI touches coverage determinations explains why examination tooling arrived when it did. The regulatory concern is not whether models are used, it is whether an adverse determination can be explained and whether the process is auditable after the fact.

The lesson generalizes to property and casualty carriers who have not yet been examined. The controls that health carriers are being asked to demonstrate, documented governance, a named accountable owner, traceable inputs and human review on adverse outcomes, are the same controls the model bulletin describes for everyone else.

Building those controls after deployment is materially more expensive than building them into the workflow, which is the strongest practical argument for treating governance as a design input rather than a compliance follow-up.

How bad is the policy file integrity problem?

Policy file integrity is the quiet trend of 2026, and it is a data problem rather than an AI one. Limits live in a policy administration system, endorsements live in a document store and effective dates live in the filings, and on a meaningful share of most books those three do not agree.

The mechanism is mundane. A signed endorsement is executed, stored and never indexed back to the policy record, so the administration system continues to report the old limit. Nothing is wrong with either system individually. There is simply no process reconciling them.

MuleSoft's 2026 Connectivity Benchmark gives the scale of the underlying condition: organizations run an average of 897 applications, of which only 27% are connected to one another. A carrier is not an exception to that statistic.

The exposure is asymmetric. An unpriced limit carried on a policy is a claim the reserve does not anticipate, and a file that cannot be reconstructed is a market conduct finding. The policy file integrity workflow shows how that reconciliation is done across systems of record.

Agentic pilots are running ahead of the guardrails

Agentic AI is the direction of travel for carriers and the highest-risk category in the portfolio. Gartner predicts more than 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls.

Agents intensify the integration problem rather than changing its nature. An agent that cannot reach the claims system cannot act in it, so the connector work that a drafting tool could defer arrives immediately and at larger scale. The permission question arrives with it, because an agent reading across systems inherits every access decision the organization has postponed.

Gartner's data-management survey of 248 leaders found 63% either lack AI-ready data practices or are unsure whether they have them. A carrier in that position that commissions an agentic programme is buying an infrastructure project it has not scoped.

The carriers doing this successfully are fencing autonomy tightly: agents that read broadly, act narrowly, and hand every outbound artifact to a licensed human for approval. That is slower than the marketing suggests and it is the version that passes an examination.

Should carriers build AI or buy it?

The sourcing question is close to settled by evidence. MIT Media Lab's Project NANDA, reviewing more than 300 enterprise deployments, found AI systems from specialized external vendors succeeded in roughly 67% of cases against about 33% for internally built systems.

For carriers this cuts against a strong internal instinct. Insurance companies have large technology organizations and long histories of building, and the pattern the data describes is that building the plumbing, connectors, entity resolution, provenance and permission inheritance, becomes permanent work that competes with the rate and product roadmap.

The same research found general-purpose assistants reached production in only about 5% of evaluations despite high user satisfaction, which is the other half of the answer. A tool that cannot reach the policy administration system will not underwrite, however good it feels to use.

S&P Global Market Intelligence, surveying more than 1,000 respondents across North America and Europe, found 42% of businesses scrapped most AI initiatives in 2025 against 17% the prior year, and that the average organization abandoned 46% of proof-of-concepts before production. The sourcing decision is one of the few levers that measurably changes those odds. A fuller treatment sits in custom AI versus plug and play.

Where carrier AI budgets go, and what each buys

The three routes a carrier can take differ less in model quality than in where the integration work lands and whether the output survives an examination.

• General-purpose assistant for staff: What it costs Low per seat, Time to value Days, Fails when The question spans policy, claims and billing, or an examiner asks for the basis of a decision

• Build the layer in-house: What it costs Senior engineering, permanently, Time to value Quarters to years, Fails when Connector and entity maintenance becomes a standing team competing with the rate and product roadmap

• Purpose-built workflow on a governed layer: What it costs Implementation plus subscription, Time to value Weeks to months per workflow, Fails when The carrier will not name a unit of value or an accountable owner, so nothing is redesigned

None is universally right. The first suits generic knowledge work, the third suits workflows that cross systems and carry regulatory exposure, and the second mainly suits carriers whose technology is itself the product.

What separates the carriers in production from the ones still piloting

The carriers running AI in production did one thing first that the piloting majority skipped: they picked a single unit of value and built the data layer for it before building anything on top.

In our own work with carriers, the failure pattern we watch most often runs in the opposite order. A carrier commissions a submission triage pilot against a curated set of clean ACORD forms from one broker. It performs, approval follows, and scaling begins. Then the real submission flow arrives. Three brokers with different formats. Loss runs as image scans. A prior claim filed under a former entity name that appears in ISO but not on the runs, and a schedule in a spreadsheet nobody standardized.

At that point three costs land at once. Entity resolution becomes infrastructure rather than a lookup. Provenance becomes mandatory, because an underwriter will not bind on a figure they cannot trace. And permissions stop being theoretical, because the system now reads across policy, claims and billing.

Every carrier we have seen avoid this did the same thing: they scoped the pilot against the ugliest data they had rather than the cleanest. That single choice converts a pleasant demonstration into a real test, and it costs the price of a pilot rather than the price of a programme. It is the cheapest risk reduction available in an AI budget and it is rarely taken, because the incentive during evaluation is to make the pilot succeed.

How OutcomeCatalyst fits

OutcomeCatalyst is a governed intelligence layer that connects the systems a carrier already runs into one structured, permissioned context that underwriters, claims teams and AI agents can reason over, without replacing the policy administration or claims platform.

For an insurance operation, the sequence we work in is deliberate. Start from the unit the business counts, a submission, a policy or a claim. Resolve that unit across the systems holding it, including the document store nobody indexed. Attach provenance so every figure keeps a link to the page it came from, which is what makes an underwriting decision defensible and a recovery demand collectible. Inherit the access rules already in force rather than inventing new ones. Only then build the workflow, with a licensed human approving anything that leaves the building.

That order is what makes the output examinable under the NAIC Model Bulletin rather than merely accurate. Worked examples run through submission intake and triage, policy file integrity and claims and subrogation, and the broader vertical view sits on the insurance industry page.

Common questions about insurance AI in 2026

How many insurers are using AI in 2026?

Adoption is close to universal in evaluation and much narrower in production. Conning's survey found full AI adoption among insurers rising from 8% to 34% year over year with large language model adoption going from 18% to 63%, and roughly 90% of carriers somewhere between pilot and production. Around 22% report AI in full production despite more than 90% testing it during 2025.

What is the NAIC Model Bulletin on artificial intelligence?

It is model regulatory guidance setting expectations for how insurers govern AI systems used in regulated insurance practices. It calls for a written AI program covering governance, risk management and internal audit, and a designated responsible person or committee with documented authority to senior leadership. As of July 2026, 25 states had adopted it.

Will insurance AI use be examined by regulators?

Yes, and the tooling is being built now. The NAIC's Big Data and Artificial Intelligence (H) Working Group is piloting a multistate AI Systems Evaluation Tool with twelve participating states, structured as an examiner questionnaire on AI governance, risk management and model oversight. The practical implication is that AI-supported decisions need reconstructable inputs.

Which insurance workflow should a carrier automate first?

Usually the one with a queue going unworked for lack of capacity rather than the one with the slowest task. Submissions that are in appetite but never quoted, and claims with a recoverable third party that nobody re-read, both create revenue rather than trimming cost, which is a stronger business case than a time saving distributed across a department.

Why do carrier AI pilots fail to reach production?

Most fail at the boundary where the model has to touch multiple systems of record at once. Gartner's survey of 248 data-management leaders found 63% either lack AI-ready data practices or are unsure whether they have them, and MuleSoft reports the average organization connects only 27% of its 897 applications. The pilot tests the model against clean data. Production tests the integration.

Should carriers build their own AI or buy it?

The evidence favors buying the fit. MIT's Project NANDA found specialized external vendor systems succeeded in roughly 67% of deployments against about 33% for internal builds, across more than 300 deployments reviewed. Building means owning connectors, entity resolution and governance permanently, which competes with the rate and product roadmap.

Does AI in underwriting create regulatory risk?

It creates documentation obligations rather than an automatic problem. The NAIC Model Bulletin is concerned with governance, accountability and the ability to explain decisions affecting consumers. An underwriting workflow where every input traces to a source document and a licensed underwriter approves the bind is considerably easier to examine than an opaque score.

Sources

• Conning, insurer AI adoption survey, 2025 (full AI adoption among insurers rose from 8% to 34% year over year; large language model adoption from 18% to 63%; approximately 55% reporting early or full deployment and 90% somewhere between pilot and production): conning.com

• National Association of Insurance Commissioners, Implementation of NAIC Model Bulletin: Use of Artificial Intelligence Systems by Insurers, adoption map current to July 2026 (25 states formally adopted; written AI program, designated responsible person or committee, governance and internal audit expectations): naic.org

• National Association of Insurance Commissioners, Big Data and Artificial Intelligence (H) Working Group, 2026 (multistate AI Systems Evaluation Tool piloted with twelve participating states; examiner questionnaire covering AI governance, risk management and model oversight): naic.org

• National Association of Insurance Commissioners, health insurance AI/ML survey (84% of health insurers use AI or machine learning in some capacity): naic.org

• GlobalData, insurer AI use-case survey, July 2026 (34.4% of respondents expect underwriting and risk profiling to benefit most from AI; 18.3% name claims management): globaldata.com

• Gartner, press release, 25 June 2025 (more than 40% of agentic AI projects predicted to be canceled by end of 2027 due to escalating costs, unclear business value and inadequate risk controls): gartner.com

• Gartner, data management survey, 248 data-management leaders, 2025 (63% either lack AI-ready data practices or are unsure whether they have them): gartner.com

• MIT Media Lab, Project NANDA, The GenAI Divide: State of AI in Business 2025, 300+ enterprise deployments, 52 case studies, 153 leadership interviews, August 2025 (specialized external vendor systems succeeded in ~67% of deployments against ~33% for internal builds; general-purpose tools reached production in ~5% of evaluations): mlq.ai

• McKinsey & Company, The State of AI: Global Survey, approximately 2,000 respondents, 2025 (more than 80% report no tangible enterprise-level EBIT impact from gen AI): mckinsey.com

• MuleSoft (Salesforce), 2026 Connectivity Benchmark Report (average of 897 applications per organization, only 27% connected to one another): mulesoft.com

• S&P Global Market Intelligence, AI adoption survey, more than 1,000 respondents across North America and Europe, 2025 (42% of businesses scrapped most AI initiatives, up from 17%; 46% of proof-of-concepts abandoned before production): ciodive.com

OutcomeCatalyst connects the systems you already run into a governed intelligence layer your team and your agents can reason over. Demos on this site use fictional data. To see this on your own book, start a conversation.