Insurers test AI gains, but data and governance slow scale-up

The industry has moved past the “should we try AI?” phase

The sharper question now is whether an insurer’s operating stack can actually support AI without breaking audit trails, slowing workflows, or exposing new risk. That is the real maturity check running through current insurance AI adoption: pilots are easy to approve, but production-grade deployment is far less forgiving.

The pattern is clear across the market. AI is already showing up in IT and developer productivity, administrative operations, underwriting, and claims. Those are the places where insurers can see fast returns, because the tasks are repeatable and the output is easy to measure. But the farther AI moves from a contained experiment into policy, claims, and service workflows, the more it runs into the same old problems: inconsistent data, fragmented processes, and systems that were never designed for machine-driven decision support.

Where insurers are getting real value first

The early wins are not glamorous, but they are practical. In IT and development, AI can speed coding, summarization, and internal support work. In administrative operations, it can help route work, extract data, and reduce manual handling. Underwriting and claims are also seeing adoption because they are packed with documents, repetitive review steps, and triage decisions that can be narrowed with the right model and the right guardrails.

That matters because it shows how insurers are thinking about value. They are not chasing AI as a shiny front-end feature. They are using it where it can shave time from highly structured tasks and where the cost of a mistake is still manageable. The problem is that these wins do not automatically translate into enterprise-scale readiness. A good pilot in one workflow does not mean the carrier can safely wire AI into the rest of the business.

The bottleneck is no longer model availability

The industry has plenty of models to choose from. The harder problem is whether the carrier’s own data, approvals, and workflow design are mature enough to support AI reliably. That is why data quality keeps coming up first. If policy data is incomplete, claims files are inconsistent, or reference data lives in disconnected systems, the model may still produce an answer, but it will not be one the business can trust.

AM Best’s survey of about 150 rated carriers and managing general agents in November 2025 makes that tension hard to ignore. About 60% of respondents expected AI to significantly transform their business model within one to three years. Even so, only about one in five said their AI implementation was already at an advanced stage. The middle of the market is not standing still, but it is also not ready to scale blindly.

The same survey shows how carriers are thinking about themselves. Fifty-three percent called themselves cautious pacesetters rather than first movers. That is a useful label because it captures the mood accurately. Insurers do not want to miss the upside, but they also know that one bad deployment in underwriting, claims, or servicing can create operational noise that is much harder to unwind than a bad dashboard demo.

What is slowing the jump from pilot to production

The top blockers are not mysterious. AM Best found that data readiness was the biggest issue, cited by 45% of respondents. Security and privacy followed at 43%, with legacy system integration close behind at 41%. Those are not side issues. They are the core operating constraints that determine whether AI can move beyond a narrow test and into the daily rhythm of a carrier.

Data readiness is especially unforgiving in insurance because so much of the work depends on context. A claims file is not just a pile of documents. It is a sequence of events, decisions, exceptions, and handoffs. If that record is messy, the model may misread the case, or worse, it may look accurate while missing the operational nuance that a human adjuster would catch.

Legacy integration is the other hard stop. Even a strong model cannot fix a workflow that has to jump across old policy admin systems, separate claims platforms, and manual approval steps. If AI creates another layer on top of that mess, the result is hidden operational risk, not efficiency.

Regulators are forcing the conversation about governance

The regulatory backdrop is getting more serious, and that is pushing insurers toward governance disciplines they cannot skip. The National Association of Insurance Commissioners says AI use in insurance has been driven by large amounts of data, faster and cheaper computing, cloud technology, and tools such as large language models. That combination is why adoption is spreading, but it is also why supervision is tightening.

In 2025 and 2026, the NAIC Big Data and Artificial Intelligence Working Group has been working on an AI Systems Evaluation Tool for regulators. At the same time, states are formalizing expectations. As of March 2025, 24 states had adopted the NAIC Model Bulletin on the Use of AI Systems by insurers, with Wisconsin among the later adopters. That bulletin expects written AI programs, governance, risk management, internal audit functions, vendor management, consumer notice, and documentation across underwriting, pricing, policy servicing, claim management, and fraud detection.

That list tells you everything about where the bar is moving. Insurers are no longer being judged only on whether AI works. They are being judged on whether they can explain it, control it, audit it, and keep it aligned with consumer obligations and business rules.

The market is excited, but production remains thin

This caution is not unique to AM Best’s respondents. Conning’s 2025 survey found nearly 90% of insurance executives viewed AI as a top strategic initiative, but only 22% had AI solutions in production. Another 25% were testing discrete use cases, while 45% were still in exploration. Roots Automation’s 2025 survey landed in the same neighborhood, with 22% of insurers reporting production AI and 45% still exploring applications.

Those numbers are the best proof that the industry is split between ambition and execution. Executives clearly see AI as strategically important. The adoption curve, though, is still shallow once you look past the conference slides and proof-of-concept demos. Production deployment requires cleaner inputs, better controls, and workflow fit that most carriers are still building.

What software buyers should demand next

For P&C software buyers, this is where the conversation changes. Feature announcements are not enough anymore. A platform that can demo a chatbot or summarize a claim note is not automatically ready for the real work of insurance operations. The next buying criteria need to be more demanding:

• Can the system connect policy, claims, and administrative data without creating another manual handoff?

• Can it manage approvals and exceptions in a way that fits existing underwriting and claims processes?

• Can it preserve auditability, so the carrier can see what the model did and why?

• Can it support vendor oversight, documentation, and internal controls that line up with evolving regulatory expectations?

Those questions are now more important than the headline promise of automation. The carriers that scale AI successfully will be the ones that treat it as an operating discipline, not a novelty. The next phase of adoption will not be decided by enthusiasm. It will be decided by whether the data is clean enough, the workflows are tight enough, and the governance is strong enough to let AI run in the real business without becoming a liability.