The hardest part of insurance AI is no longer proving that a model can work in a controlled setting. It is deciding who owns it once the pilot ends.
That distinction is becoming urgent as insurers move from experimentation to production. BCG reported in 2025 that only 7% of insurance companies had successfully brought AI systems to scale, even though the sector is unusually well suited to AI because of its data reserves and analytic workforce. A separate 2025 insurance AI adoption survey found 45% of respondents still exploring AI, 25% testing discrete use cases and only 22% running live solutions in production.
For Rob Galbraith, CEO of Forestview Insights and a former insurance innovation leader, the gap between those numbers is not simply a vendor problem. It is a product-management problem inside the carrier.
“There’s a lot of work that goes around turning a technology into a product,” Galbraith said.
Galbraith draws a line between a technology that is impressive and a product that can actually be sold, serviced, updated and supported in an insurance organization. The difference is not cosmetic. A product has an account owner, billing support, implementation support, a process for updates, a path for patches and a defined way to handle issues when performance changes.
That maturity matters most when the technology is genuinely new. Insurers with very large innovation budgets can afford to explore frontier tools before they are ready for daily operations. Most carriers cannot. They need AI that has moved beyond experimentation into something stable enough for regular business use.
“You need to have an account representative if there’s any issues,” Galbraith said. “Somebody you work with, with billing, if you have any issues with the product implementation, how are you going to get updates, how are you going to get patches?”
The same test applies to models built internally. A carrier may have strong data scientists, access to internal data and the subject-matter expertise needed to build a useful model. That does not mean the organization has built the operating structure to keep it useful.
Galbraith has seen the failure mode clearly. A company funds a project to build a new AI model. Data scientists work with business experts to develop it. The pilot succeeds well enough to launch. Then the project closes.
“Once it implements, those resources are gone,” he said. “Those data scientists move on to a different project. The subject matter experts on the business side, they move on to their regular day jobs. And then if you need to upgrade or you need some monitoring or whatever, you’re kind of stuck.”
That is the quiet risk in many AI business cases. The launch budget covers development, proof of concept and implementation. It may not cover monitoring, retraining, documentation, user support, performance review, issue escalation or model retirement. Yet AI systems do not remain static. Data changes, workflows change, user behavior changes and regulatory expectations change.
NIST’s AI Risk Management Framework, released in 2023 and under revision in 2026, frames AI risk management as work that extends across the design, development, use and evaluation of AI systems. For insurers, that lifecycle view matters because the accountability for an AI-assisted decision does not end when the model goes live.
The practical question for insurance executives is not just whether an AI pilot has an owner during implementation. It is whether the business unit is prepared to own the model after implementation. Galbraith said one real test is whether the business unit is willing to pay the ongoing cost once the innovation or project budget disappears.
That ownership should include more than budget. It should include the operational discipline to know when the model is still performing, when outputs require review, when a data-quality issue has emerged and when business rules have changed enough that the model needs adjustment. IT and data science matter, but the business cannot disappear after launch because the business understands the workflow, the decision and the customer impact.
This is where the Transformation lesson is clearest. AI capability is becoming easier to buy or build. AI operating discipline is not. The carriers that scale AI will not be the ones that run the most pilots. They will be the ones that turn successful pilots into maintained products: monitored, funded, supported and still connected to the business problem they were built to solve.
