Short answer. Around 60% of insurers are still stuck at the AI pilot or proof-of-concept stage, and the block is rarely the model. Pilots stall on governance, data readiness, and integration with core policy and claims systems. To reach production, build the pilot on the same engineering, security, and model-governance standards as the live system, write measurable success criteria before you pick a model, and assign a named owner for the model after go-live. Ableneo takes roughly 4 of 5 projects to production.
A pilot proves a model can work on a clean sample. Production proves it works every day, on live data, inside the systems underwriters and claims handlers already use. The gap between the two is where most insurance AI dies, and the reasons are operational, not scientific.
The pilots that cross the gap share a pattern. They are built as the first version of the production system, not as a throwaway demo. They connect to governed data from day one. They carry a business owner who is accountable for the result, not only a data science team that is accountable for the model.
The difference is concrete:
Ableneo shipped 34 production AI projects in 2025, and roughly 4 of 5 reached production. In one production deployment, document validation ran 14 times faster, from 7 hours to 30 minutes, at over 90% accuracy. Insurance runs on documents: claims files, medical reports, policy schedules. That is where a production-grade system earns back its cost.
An insurer moving a pilot to production does not start from zero on governance. Under Solvency II, insurers already run model risk management for actuarial and capital models: validation, documentation, and sign-off. The task is to extend that discipline to AI, not invent it. DORA, in force since January 2025, adds operational resilience and third-party risk obligations that apply the moment an AI model or its infrastructure comes from a vendor.
The regulatory clock is a reason to industrialize now, not to wait. AI used for risk assessment and pricing in life and health insurance is listed in Annex III of the EU AI Act as high-risk, with obligations for these stand-alone systems applying from 2 December 2027. A pilot that skips logging, human oversight, and technical documentation has to be rebuilt to meet those obligations. Building them in from the pilot stage reaches production and compliance in one motion.
Around 60% of insurers are stuck at the pilot or proof-of-concept stage; the block is governance and integration, not the model.
Because the pilot optimizes for a demo, not for daily operation. It runs on exported data, skips integration with the core policy and claims platforms, and has no owner once the data science team moves on. Industry surveys place around 60% of insurers at the exploration or proof-of-concept stage.
The failure is commercial and organizational, not technical. There is no funded business case beside the model, no change plan for the underwriters and claims handlers who must use it, and no clean path to live data. The model works in the demo and dies on contact with the operation.
Six things a pilot can skip and production cannot: governed live data instead of a static export, integration with the core policy and claims systems, monitoring with drift detection, human oversight with a low-confidence fallback, an audit trail with technical documentation, and a named owner accountable for the result.
A pilot can demo well without any of these. Each one is work, and that work is the actual project. Treating it as scope from the first day is what separates a system that ships from a demo that impresses and then stops.
AI used for risk assessment and pricing in life and health insurance is classified as high-risk under Annex III of the EU AI Act. Obligations for these stand-alone high-risk systems apply from 2 December 2027. In practice that makes risk management, data governance, event logging, human oversight, and technical documentation legal requirements rather than good habits.
This is an advantage for the insurer that plans ahead. The controls the Act requires are the same controls that make a pilot production-ready. Build them once, at the pilot stage, and the system clears both bars at the same time instead of being reworked under a deadline.
A named business owner, accountable for the outcome, supported by a model risk management process. Insurers already run this for actuarial and capital models under Solvency II. Extend it to AI: version the model, log every decision, set a revalidation schedule, and define who signs off on a model change before it goes live.
When the model or its infrastructure comes from a vendor, DORA requires the insurer to manage that third-party risk and prove operational resilience. Ownership does not transfer to the vendor. The insurer stays accountable for the decision the model makes, which is why a named internal owner is not optional.
Pick one use case with a measurable business result, a clean data path, and a willing business owner. Write the success criteria and the production requirements before selecting a model. Build the pilot on production engineering standards so it does not need rebuilding when it works.
Small entry, real proof, then expand to the next use case. That sequence costs less than a large program that stalls at the demo, and it produces a reference the rest of the organization can trust.
Ableneo builds AI that reaches production in regulated industries across Central Europe. Across 34 projects shipped in 2025, roughly 4 of 5 reached production, and the average client runs 1.7 projects with us, a signal that the first system works well enough to earn the second. The method holds across banking, insurance, and telecom: define the business result first, build the pilot to production standards, and assign a named owner before go-live. For the general version of this problem, see how to get a stalled AI pilot into production.
Key takeaways
Planning AI in a regulated business? Ableneo takes systems from classification to governed production.