Insurer AI Adoption Hits 82% But Only 7% Reach Full Scale: Diagnosing the Maturity Gap

Sedgwick's March 2026 report, "Future-ready property claims," puts adoption of claims AI at between 58% and 82% of insurers. It puts mature capabilities at 12% and scalable success at 7%.

The 82% figure has travelled as evidence that AI is transforming claims. The arithmetic underneath it says something narrower: three quarters of the carriers using AI somewhere cannot make it work across an operation, because the gains are locked inside individual steps.

Key Takeaways

• 58% to 82% adoption, 12% mature, 7% scalable is the same population measured three ways, and the drop from 12% to 7% says even carriers with reliable individual deployments struggle to repeat them across lines of business.
• FNOL intake fell from 10 days to 36 hours, roughly 85% faster, which is the clearest quantified gain and stops at the first handoff if downstream steps still run manually.
• Only 24% of insurance leaders are very confident they could pass an independent AI governance review within 90 days, against 62% who rate themselves as scaling across multiple functions.
• 68% say AI controls exist but the evidence is fragmented across teams and tools, which is the governance mirror of the vendor fragmentation blocking the technology.
• Over 90% of carriers tested AI in 2025 and 22% reached full production, while spend is projected to grow more than 25% in 2026.

Three Tiers of the Same Population

The report's value is that it measures the same carriers against three different bars, and the distance between them is the finding.

Broad adoption at 58% to 82% counts any carrier using AI in at least one claims function, which usually means document data extraction, a chatbot, or basic triage routing: point solutions, often vendor-supplied, handling one step of a multi-step workflow. Mature capabilities at 12% requires AI running reliably across multiple claims functions with consistent data flow between tools, so outputs from one step feed the next without manual re-entry. Scalable success at 7% means enterprise scale across lines of business and claim types with measurable improvement in speed, accuracy and cost.

Where the gains are real they are substantial, and they are step-local.

Metric	Before AI	After AI	Improvement
FNOL intake processing	10 days	36 hours	~85% faster
Low-severity claims cycle time	Baseline	AI-assisted	80% faster
Photo-based claim handling	Manual review	AI-powered analysis	Up to 54% efficiency gain
Documentation productivity	Baseline	AI-assisted	50% productivity gain
Adjuster time on low-value tasks	~30% of workday	Targeted for automation	Capacity reallocation

Sedgwick projects that 80% to 85% of simple claims could eventually reach straight-through processing, against claims handlers currently spending roughly 30% of their time on low-value administrative work. David Guaragna, Sedgwick's managing director of property operations, puts the constraint plainly: "Strategy isn't optional; it's the new competitive advantage."

Why the Gains Stop at the Handoff

The reason 82% does not become 7% at the operation level is architectural, and it shows up as data that does not survive being passed along.

A typical 2026 claims stack runs one vendor for FNOL intake and triage, another for photo-based damage estimation, a third for document extraction on coverage verification, a fourth for fraud models and a fifth for settlement recommendations. Each can perform well inside its scope. When a damage estimate from one tool feeds a settlement engine from another, the two may use different field definitions, damage categories or severity scales, so the downstream model receives inputs it was not trained on and produces output requiring human correction. That correction is the efficiency gain going back out.

Line of business compounds it. A carrier writing personal auto, homeowners, commercial property, general liability and workers' compensation is running five claim types with different data requirements, regulatory frames and settlement patterns, and a model trained on personal auto photo damage does not generalize to commercial roof assessment. Scaling means separate models per line, each with its own training data, validation and governance. Legacy claims systems then supply the last constraint: without API connectivity, AI tools sit layered on top of existing platforms rather than embedded in workflow, adding a layer instead of removing one.

The reserving consequence follows directly from where the gains sit. FNOL intake accelerating from 10 days to 36 hours should show up as faster initial reserve posting and, if the models are calibrated, more accurate first estimates, which would compress early development volatility. That benefit is real only to the extent the record arriving at the reserve is reliable, and it is produced by the tool whose outputs the next system may not read cleanly. A carrier can book the cycle-time improvement and get no development-pattern improvement at all, and the two are measured in different places by different people.

The same split governs loss adjustment expense. The 80% faster low-severity processing and 50% documentation productivity gain are real for carriers at the scalable tier, where they should reach allocated LAE within a few quarters. For the 93% below it, a prospective LAE reduction in a rate indication is an assumption about integration work that has not happened rather than about technology that has.

The Governance Gap Repeats the Pattern

The complication is that the organizational readiness to fix the fragmentation is measured to be in the same state as the fragmentation itself.

Grant Thornton's 2026 AI Impact Survey covers 950 business leaders including 100 insurance respondents. 62% rate their AI maturity as scaling across multiple functions, 13 points above the cross-industry average and far above the 12% mature rate Sedgwick measures operationally. Only 24% are very confident they could pass an independent AI governance review within 90 days, meaning 76% cannot demonstrate governance on demand. 68% say controls exist but the evidence is fragmented across teams and tools, only 7% believe their workforce is fully ready, and 56% name regulatory or compliance uncertainty as a top barrier to scaling.

The self-assessment gap between 62% scaling and 24% review-ready is the governance version of the 82% against 7%. Both measure the distance between having something and being able to operate it.

The regulatory direction closes on the same seam. By late 2025, 23 states and Washington, D.C. had adopted the NAIC's 2023 AI Model Bulletin in some form, evaluation tool pilots are running with insurers selected by market share and anticipated AI reliance, and the proposed third-party vendor registry would give regulators visibility into the tools carriers depend on without relieving anyone of vendor diligence. A carrier running five vendors across its claims workflow inherits five sets of model documentation, five validation frameworks and five sets of performance metrics to monitor, and each of those vendors' governance practices becomes part of its own compliance position.

Spending is moving the other way from all of it. More than 90% of carriers tested AI in 2025 and 22% reached full production, while spend is projected to grow more than 25% in 2026, with more than 35% of insurers expected to deploy AI agents across at least three core functions. Adding tools to a fragmented architecture produces more fragmentation, and the 7% is the count of carriers for whom that is not what the next purchase will do.