Celent: 48% of Insurers Run GenAI in Production, Pushing Past the Early-Majority Threshold

Celent's third annual GenAI-oneers survey puts 48% of global insurers with at least one generative AI use case in production, up from roughly 8% in 2023, 28% in 2024 and 44% at mid-2025. Celent projects late-majority status during 2026.

The number behind the headline is the qualifying bar. "In production" means one live use case past pilot, which is why the same industry reads as 48% adopted and 7% at scale.

Key Takeaways

• 48% in production counts any insurer with a single GenAI use case past proof of concept, so one claims summarization tool in one regional office qualifies alongside enterprise orchestration.
• 7% at scalable success is Sedgwick's measure of AI running across lines and claim types with demonstrable improvement, and the distance between the two figures is the entire deployment-depth problem.
• 68% of organizations have moved 30% or fewer of their GenAI experiments into full production, per Deloitte, which places most of the 48% cohort alongside a larger pool of stalled pilots.
• 22% plan agentic AI by year-end 2026, against roughly 14% today and a Celent projection of 70% by 2028.
• 24% of insurance executives told Grant Thornton they were confident of passing an independent AI governance review within 90 days.

Four Numbers Measuring Four Different Things

The surveys are not in conflict. They set different bars, and reading them as one figure produces the wrong competitive conclusion.

Survey / Source	Metric	Finding
Celent (Q1 2026)	At least one GenAI use case in production	48% of global insurers
Sedgwick (March 2026)	Scalable AI success across enterprise	7% of insurers
Deloitte (Q3 2025)	Moved 30%+ of experiments to production	Only 32% of organizations
AM Best / Insurance Journal (May 2026)	Self-described advanced AI stage	20% of ~150 carriers/MGAs
Grant Thornton (2026)	Could pass independent AI governance review in 90 days	Only 24% of insurance leaders
EY (2026)	Early or full GenAI adoption	55% of insurers

Celent counts breadth: one workflow live. Sedgwick counts depth: enterprise scale with measurable improvement in speed, accuracy and cost. Deloitte counts conversion rate, and finds most deploying organizations have converted under 30% of their experiments. AM Best's survey of about 150 rated carriers and MGAs adds self-assessment: 53% describe themselves as cautious pacesetters rather than first movers, and only 20% put their implementation at an advanced stage.

Read together they describe one distribution rather than four disagreements. A large share of carriers have initial production experience, a smaller group has real deployment depth, and a much smaller group operates at enterprise scale with governance that would survive examination.

Where the live use cases sit is consistent across sources. Evident's Q4 2025 tracker puts claims management at 37% of GenAI and agentic deployments, underwriting and pricing at 21% each, and direct policyholder interaction at 36% against a 7% historical average.

Why the Gap Decides the Expense Assumption

The distinction matters most where it reaches a filing, because an expense ratio assumption built on the 48% figure is being built on the wrong denominator.

For carriers in the scalable tier the improvement is documented and specific. Chubb reports roughly 1.5 points of automation savings against an 84% combined ratio, and Morgan Stanley's projection of $9.3 billion of P&C operating income by 2030 rests on 200 basis points of expense ratio compression. Those are the figures a rate filing can support with evidence.

For the rest of the 48% the same improvement is prospective rather than realized, because a single production workflow in one business unit does not move a companywide expense ratio. Modelling one expense trajectory across both populations either overstates savings for the shallow deployers or understates them for the few operating at depth, and the surveys give no way to tell which carrier is which from outside.

Loss adjustment expense is where the effect should surface first and unevenly. With 37% of GenAI deployments sitting in claims management, allocated loss adjustment expense is the category most directly exposed, and personal lines is where claim volume and standardization make integration easiest. That is a testable proposition on a two to four quarter horizon: FNOL acceleration and AI triage should show up as compressed development in early reporting periods for the deepest deployers, and not at all for the shallow ones.

Governance Is the Constraint, Not Capability

If 48% are in production and 7% are at scale, the interesting question is what holds the rest in place, and the evidence points at governance infrastructure rather than model capability.

Grant Thornton's 2026 survey of 100 insurance executives found 44% citing governance or compliance challenges as contributors to AI project failure, and only 24% confident of passing an independent AI governance review within 90 days. AM Best's top implementation barriers point the same way: data readiness at 45%, security and privacy at 43%, legacy system integration at 41%. Only 13% felt very confident measuring AI return on investment at all.

Regulatory reach is expanding into exactly that space. Twenty-three states plus Washington, D.C. have adopted the 2023 NAIC AI Model Bulletin in some form, the Spring 2026 meeting advanced a multi-state pilot using an evaluation tool that assesses AI systems, data sources, governance practices and high-risk use cases, and a proposed vendor registry would give regulators visibility into third-party tools. AM Best found 68% of carriers source AI from third parties while 18% consider third-party model risk a challenge, so the registry would be reporting on dependencies most carriers have not scoped themselves.

That is the constraint the 22% agentic figure runs into. Autonomous multi-step workflows raise the governance requirement at the same time they raise the capability, and the survey population moving toward them is the one that has not yet demonstrated it can document what it already runs.