For much of the generative AI boom, the insurance industry appeared to be moving cautiously.
Insurers tested copilots, automated documents and experimented with models that could help employees search data or summarize claims. But the most consequential decisions in insurance, including how risk is priced, which policies are underwritten and why a claim is accepted or rejected, largely remained beyond AI’s reach.
That is beginning to change.
A new generation of insurance technology companies is moving AI away from peripheral productivity tools and into the operational systems where financial decisions are made. The emerging contest is no longer simply about who can give insurers access to the most powerful model. It is about who can build an industry-specific layer capable of using those models safely, continuously and at scale.
Earnix entered that race in June with the launch of AIOS, an insurance-focused AI orchestration system designed to bring agents, models, data, workflows and human oversight into a single governed environment. The company introduced the platform to more than 200 insurance executives at an industry event in Boston later that month.
The launch comes as established insurance software providers are pursuing similar ground. Duck Creek unveiled an insurance-native agentic AI platform in April, while Guidewire is positioning its own platform around predictive and agentic AI operating on insurance-specific data and infrastructure.
The competition reflects a wider shift in enterprise AI. General-purpose models may provide the underlying intelligence, but regulated industries increasingly require systems that understand their data, processes, terminology and legal obligations.
For insurance, those requirements are particularly demanding.
Pricing and underwriting decisions cannot simply be generated by a model and accepted at face value. Insurers must be able to demonstrate how decisions were reached, apply internal policies consistently and respond to regulators or customers when an outcome is challenged. An AI system operating in that environment must therefore do more than produce a plausible answer. It must operate within defined authorities, preserve an audit trail and allow human intervention.
Erel Margalit, chairman of Earnix and founder and chairman of venture capital firm Jerusalem Venture Partners, argues that this is precisely why insurance is becoming an important proving ground for vertical AI.
“Because regulation is the crucible where generic AI melts,” Margalit said. “A black box can’t write a rate filing or defend a declined policy to a regulator. Insurance demands explainability, auditability and deep domain context just to get in the door, so if you can make AI trustworthy and governed here, you’ve cracked the hardest version of the problem.”
Earnix has spent more than two decades developing software used by insurers to support pricing, rating and underwriting decisions. AIOS is an attempt to extend that position into agentic AI, allowing specialized agents and applications to operate over existing insurance data and decisioning infrastructure.
The system is structured around three components. Applications and workflows provide the interface through which insurance teams use the technology. An Agent Hub supplies specialised AI agents. Underneath them, a governance layer known as the Mesh coordinates agents, data and computing resources while enforcing identity, permissions and auditability.
The ambition is to move AI from advising employees to participating directly in live decision-making, while keeping those decisions within a controlled operating framework.
That distinction matters. Many insurers have already invested heavily in digitisation, moving policies, claims and customer records into modern core platforms. But digitising information is not the same as using it to make continuously adaptive decisions.
“The industry already went through one digitization wave,” Margalit said. “Guidewire and Duck Creek moved the core systems of record into the modern era. But those are systems of record, not systems of intelligence: they store the policy, they don’t outthink the risk.”
That characterization inevitably understates the AI ambitions of both companies. Guidewire and Duck Creek are themselves investing heavily in intelligence and automation, and both now describe their platforms as foundations for agentic AI. The result is likely to be a competitive struggle over which layer of the insurance technology stack ultimately governs AI-driven decisions.
The stakes extend beyond operational efficiency.
Insurance companies are being asked not only to adopt AI, but also to understand and insure the new risks it creates. A recent industry report warned that much of insurers’ exposure to AI agents may already sit silently inside conventional policies that were never written with autonomous systems in mind. At the same time, regulators are raising concerns about accountability, concentration among a small number of model providers and the movement of AI into customer-facing financial decisions.
Insurers will therefore need better systems for analyzing fast-changing risks while proving that their own use of AI remains controlled. That combination makes insurance an unusually demanding market, but also a potentially valuable one for vertical AI companies.
Margalit believes that the economic value of AI will increasingly migrate away from the foundation models themselves and toward the systems that apply them inside specific industries.
“Foundation models are the engine, but engines commoditize,” he said. “There’ll be a handful of giant ones and they’ll get cheaper every year. The lasting value is in the chassis built around them: the governed, domain deep layer that turns raw horsepower into real business outcomes inside a specific industry.”
That argument is increasingly influencing the wider software market. The first phase of generative AI rewarded companies that built models and broadly applicable interfaces. The next phase may favor businesses with proprietary industry data, established customer relationships and the ability to embed AI inside complex operational processes.
Insurance provides an early test of that thesis because the difference between an AI demonstration and a production system is especially stark. A chatbot can be impressive despite occasional errors. A pricing or underwriting platform cannot.
The companies that succeed will need to connect AI with live data and existing systems, but also determine what agents are allowed to do, when a human must intervene and how every consequential decision can be reconstructed later.
“The winners will rewire how they operate, not bolt AI onto legacy plumbing,” Margalit said. “The laggards will still be running shiny pilots that never make it to production or touch the P&L.”
That is the real race now unfolding across insurance technology. It is not simply a race to add AI features. It is a race to determine which platforms insurers will trust to make decisions when money, regulation and customer outcomes are all on the line.
