Insurers race to close AI execution gap

The bottleneck is not that insurers lack pricing intelligence; it is that model speed now exceeds organizational execution speed. As Analysis reported, a pricing model can be built and validated in hours while the resulting price can still take months to reach customers, because data access, validation, deployment and governance sit between the model and the market, and even after actuarial approval insurers still must configure business rules, integrate systems, run testing, complete compliance checks and coordinate reversibility before a recommendation becomes an operational price.

That execution drag cuts actuarial throughput because faster scenario generation does not remove the need for cross-functional approvals or human oversight. Earnix’s 2026 Industry Trends Report surveyed 400 executives and found only 30% said their organisations can quickly access the information they need to make business decisions, while 46% believe their technology provides the required speed; it also found 92% conduct formal AI reviews, but fewer than one in three are fully confident those reviews keep pace with regulatory change, and 56% favor keeping human intervention in place for at least the next three years.

The new insurance tooling is trying to move AI from recommendation engines into the operating layer where decisions actually get carried out. Appian and Synechron’s Open Underwriting Stack is explicitly pitched as an “operational AI layer” that modernizes underwriting without ripping out legacy systems, while connecting “legacy cores, modern SaaS tools, and real-time risk data” and keeping data “open, auditable, and underwriter-controlled”; that same direction showed up when, among recent developments, Appian’s April 2026 platform upgrades expanded AI-assisted development and deeper data fabric integrations, including with Snowflake Cortex AI.

What makes that shift notable is that the model is no longer just smarter analytics but software designed to plug into existing stacks, data pipelines, and applications. Meta’s Muse Spark 1.3 released “two pricing models,” including a cheaper option “if you will let us switch off the Sierra data retention policy and let us train on your data,” with pricing described as “one to$4 dollars per million tokens,” and Meta said this could “open up a whole new opportunity to create apps”; Lemonade shows the insurance end state, having been “built from the ground up over the course of ten years using a single, unified AI first computer system,” including Tesla-linked pricing that allows a 50% discount for every mile travelled using Tesla’s full self-driving software.

The scale of the execution gap looks industry-wide rather than anecdotal. An Omdia poll sent to 22,000 partners worldwide found that 43% of MSPs are evaluating an alternative to user-based pricing, another 17% know they need to make a change but have not decided what kind, and a full 40% are not even thinking about modifying the per-user, per-month rates they have charged for years—evidence that only a minority are positioned to translate pricing insight into timely market action.

Vendor commentary in insurance points to the same broad pattern of slow, traditional execution. Earnix says group benefits renewals are still typically built by combining manual rating with experience rating, while actuaries, underwriters, pricing teams and sales staff often work across separate systems, spreadsheets and manual workflows; that setup encourages broad whole-case rate actions instead of faster, more granular changes, and supports the view that the bottleneck is common across portfolios rather than confined to a few laggards.