Short answer. Start with a use case the EU AI Act does not classify as high-risk, such as fraud detection or claims triage. The Act lists AI for risk assessment and pricing in life and health insurance as high-risk (Annex III, point 5(c)), while AI used to detect financial fraud is excluded from that category. That single distinction sets the sequence: a fraud or claims use case carries a lighter compliance load and can reach production faster, while high-risk pricing systems face a 2 December 2027 deadline and a governance program that typically runs 9 to 14 months.
For most insurers the question is not whether to use AI but which use case to put into production first. The EU AI Act answers part of that question for you. It sorts insurance AI into risk tiers, and the tier decides how much governance, documentation, and oversight you must build before the system goes live. Starting with the wrong use case means months of compliance work before you have a single result to show the board.
Capgemini’s 2026 World Property and Casualty Insurance Report found that 60% of property and casualty insurers are still in exploration or proof of concept on AI, and only about 10% reach advanced maturity. The report also noted that the typical insurer spends 72% of its AI budget on technology and infrastructure and 28% on change management. The pattern behind the stall is consistent: insurers pick ambitious, high-risk use cases, hit the governance wall, and never ship.
A sharper sequence looks like this:
Insurance is one of the most explicitly regulated AI sectors in Europe. The EU AI Act names insurance risk assessment and pricing in Annex III, the same list that covers credit scoring and biometric identification. On top of the Act, EIOPA published its Opinion on Artificial Intelligence Governance and Risk Management on 6 August 2025, which sets governance expectations for insurance AI that is not high-risk, including data governance, fairness, human oversight, and explainability.
The two layers do not conflict, they stack. A high-risk pricing model must meet the full Annex III obligations: risk management, data governance, technical documentation, logging, human oversight, and a conformity assessment. A fraud model that falls outside the high-risk tier still has to meet the EIOPA governance baseline. Knowing which layer applies to which use case is the difference between a six-week sprint and a multi-quarter program, and it is why sequencing is a regulatory decision, not only a maturity one.
Start with a use case outside the EU AI Act high-risk tier, such as fraud detection or claims triage, so a first system reaches production in weeks.
Annex III, point 5(c) names AI systems used for risk assessment and pricing of natural persons in life and health insurance as high-risk. That is the clear case. AI used to detect financial fraud is explicitly excluded from the high-risk category. Many non-life lines, such as motor and property, are not named in that provision, so their classification depends on how the system is used and must be read against the Annex III text directly rather than assumed.
The practical takeaway: do not treat all insurance AI as high-risk, and do not assume none of it is. Build an inventory of every AI system in pricing, underwriting, claims, and fraud, then classify each one against the Annex III wording. Vendor marketing and even some legal summaries disagree on the edge cases, so the inventory has to cite the Regulation, not a secondary source.
Fraud detection and claims triage are the fastest path to a production AI system that a board can see working. They sit outside the Annex III high-risk tier, so they avoid the conformity assessment, the registration, and the heaviest documentation load. They also produce a number quickly: fewer fraudulent payouts, faster claims handling, lower manual review cost. A first use case that ships in weeks and shows a measurable result builds the internal confidence that the harder use cases need.
This is the opposite of the common failure mode, where an insurer starts with AI-driven pricing because it looks most valuable, then spends two quarters on governance before anything reaches customers. The lower-friction entry point earns the right to tackle underwriting and pricing later, with a governance model that already works.
A use case outside the high-risk tier is not a use case without rules. EIOPA’s 2025 Opinion asks insurers to apply a risk-based, proportionate approach across the AI lifecycle for all AI that affects customers, including fraud and claims systems. That means documented data governance, testing for fairness and bias, a human in the loop for decisions that affect a policyholder, and enough explainability that a claims handler can understand why the model flagged a case.
Building this baseline on a lower-risk use case first is deliberate. The same controls, documentation habits, and oversight roles scale up to the high-risk systems later. An insurer that runs a well-governed fraud model has already built most of the operating model that an Annex III pricing system will demand.
The high-risk obligations for standalone Annex III systems, including insurance risk assessment and pricing, apply from 2 December 2027. That sounds distant, but a full compliance program for a high-risk AI system typically runs 9 to 14 months end to end, covering inventory, classification, risk management, data governance, documentation, oversight design, and conformity assessment. Counting back from the deadline, the window to start a high-risk underwriting or pricing engagement opens in late 2026.
Separately, insurers running customer-facing chatbots or generative AI tools already face the EU AI Act Article 50 transparency obligation, live since 2 August 2026, with a machine-readable content-marking requirement from 2 December 2026. That is a shorter, nearer deadline than the Annex III one and should not be conflated with it.
A workable first year has three moves. First, ship a non-high-risk use case, fraud or claims triage, to production in the first quarter and measure one hard number. Second, run an AI inventory and Annex III classification in parallel, so you know exactly which systems carry the high-risk load. Third, use the second half of the year to start the governance program for the first high-risk pricing or underwriting system, with the 2 December 2027 deadline as the fixed point to plan back from.
The sequence keeps a visible result in front of the board early, spreads the compliance work across months instead of stacking it before launch, and makes sure the hardest use case starts with enough runway. It turns a regulatory constraint into a delivery plan.
Ableneo works with regulated financial services and insurance clients across Central Europe, where roughly 4 of every 5 AI projects reach production. That production rate is the point: in a sector where 60% of insurers are stuck in proof of concept, the discipline that gets a system past the governance wall and into live operation is the differentiator. Our work starts with the regulatory classification and the business result, then builds the data, oversight, and documentation that let the system ship and stay compliant. See Ableneo’s AI transformation FAQ for the full set of questions regulated buyers ask before they start.
Key takeaways
Planning AI in a regulated business? Ableneo takes systems from classification to governed production.