Short answer. Most claims AI is not high-risk under the EU AI Act. Annex III point 5(c) names only risk assessment and pricing of natural persons in life and health insurance, so document reading, triage, and fraud flagging in claims sit outside the high-risk tier. Start where the work is document-heavy, high-volume, and low-consequence, run the model in shadow mode against real claims, and keep a named human accountable for every decision. Ableneo ships roughly 4 of 5 AI projects into production using this sequence.
Claims is the most document-heavy process an insurer runs. A single motor claim can carry a police report, repair estimates, photographs, and medical notes. AI earns its place first where it reads and structures those documents, not where it decides the payout.
Ableneo’s strongest production proof for this pattern comes from outside insurance and maps directly onto it. At T-Mobile Czech Republic, an Ableneo document-validation system cut consent checking from 7 hours to 30 minutes, 14 times faster, at over 90% accuracy, and the client approved the result for public reference. Claims intake is the same shape of problem: high document volume, repetitive judgment, a clear accuracy target.
The early use cases that fit this profile:
Across its 2025 portfolio, Ableneo shipped 34 production AI projects for 20 clients, and 94% of them used large language models. The pattern that reaches production starts narrow and measurable, then widens.
Insurance is a regulated industry, so the question is never only whether the model works. It is also who is accountable and under which rule. The EU AI Act sets the risk tier. Annex III point 5(c) classifies AI used for risk assessment and pricing of natural persons in life and health insurance as high-risk, and those obligations apply from 2 December 2027 under the amended timeline. Claims handling, property and casualty pricing, and fraud detection are not named in point 5(c), so most claims AI falls outside the high-risk tier.
Outside the high-risk tier is not outside regulation. EIOPA’s Opinion on AI governance and risk management, published on 6 August 2025, tells national supervisors that existing insurance law already requires sound data governance, record-keeping, fairness, explainability, and human oversight, in proportion to the risk of each system. AI literacy obligations under Article 4 of the EU AI Act have applied to every provider and deployer since 2 February 2025, whatever the risk tier. For insurers inside DORA, the IT that runs a claims model also sits under operational-resilience rules. The practical effect is that governance, not the model, is what lets a claims pilot go live.
Most claims AI is not high-risk: Annex III point 5(c) covers only risk assessment and pricing in life and health insurance.
The dividing line is the decision the system drives, not the department it sits in. Annex III point 5(c) is narrow: risk assessment and pricing of natural persons in life and health insurance. A model that prices a life policy is high-risk. A model that reads a motor claim, extracts the repair cost, and routes the file is not, because it does not assess risk or set a price for a person. Fraud detection is excluded from the point as well, though a fraud model still needs fairness and oversight controls under existing law.
Two checks keep the classification honest. First, confirm the use case is not caught by another Annex III category, for example a model that also scores creditworthiness. Second, if a pricing-adjacent model genuinely poses no significant risk to a person’s health, safety, or rights, Article 6(3) can take it out of the high-risk tier, but that call must be documented, not assumed. When in doubt, read Annex III against the actual decision the model makes.
Start where three conditions meet: the work is document-heavy, the volume is high, and the consequence of a single error is low. First-notice-of-loss intake and document validation fit all three. The model reads and structures, a handler decides, and a wrong extraction is caught before it reaches a customer. This is the same profile as the T-Mobile Czech Republic document-validation case, which is why work of this shape reaches production fast.
Leave the payout decision and life or health pricing for later, after the governance is proven on the low-consequence work. Scaling automation to the consequence of error is the rule that separates claims programs that ship from ones that stall. A model that flags is easier to govern than a model that decides, and flagging delivers most of the time saving.
Pilots stall because a pilot tests whether a model can produce an answer, while production tests whether the insurer can rely on it. Pilots often run on clean or synthetic data, so real claims, with missing pages and inconsistent formats, expose the gap on day one. Industry reporting puts insurer AI projects reaching production in the low double digits, with one analysis at 11%.
Clearing the gap takes four moves. Design the pilot on real production data from the start. Run the model in shadow mode next to the current process and compare outputs before trusting it. Put a governed layer between the model and the live workflow so decisions are logged and reversible. Define who owns the decision and who can override it before go-live, not after. Ableneo treats these four as the entry criteria for production, which is why roughly 4 of 5 of its projects get there.
The EIOPA Opinion sets the floor: proportionate data governance, record-keeping, fairness testing, explainability, and human oversight, sized to the risk of each system. For a document-reading model that means a logged audit trail of every extraction, a measured accuracy threshold, and a defined point where a handler takes over. For a fraud-triage model it means bias testing across customer groups and a reason code a person can read.
Two obligations apply whatever the tier. Article 4 of the EU AI Act requires that the people operating the system have enough AI literacy to use it safely. Insurers inside DORA must also cover the model’s supporting IT under their operational-resilience and third-party risk controls. Governance defined once, at this level, carries over when a higher-risk use case arrives later.
Accountability means one named role owns the outcome, not the model. In a production claims system that role sees the model’s output and the reason behind it, can change it, and signs the decision that reaches the customer. The model proposes, the person disposes. Explainability exists to make that override possible: if a handler cannot see why the system flagged a claim, they will override every time, and the automation delivers nothing.
Design the oversight into the workflow rather than bolting it on. Set the accuracy threshold below which a case always goes to a person. Log every automated and human action against the claim. Review the overrides monthly to find where the model is weak. This is what turns a claims model from a demonstration into a system an insurer can run and defend.
Ableneo builds AI into regulated operations as a system, not a single model: the data flow, the governance, the human oversight, and the audit trail that let a claims team run it and a regulator read it. The production track record sits behind this, roughly 4 of 5 projects reaching production across 34 deployments in 2025, with named work in financial services and insurance including UNIQA, and the T-Mobile Czech Republic document-validation case as the closest public proof of the claims-intake pattern. For the step before this one, see where an insurer should start with AI under the EU AI Act.
Key takeaways
Planning AI in a regulated business? Ableneo takes systems from classification to governed production.