Is an Insurer's AI Fraud Detection High-Risk Under the EU AI Act?

6 min readAbleneo AI transformation team

Short answer. In most cases, no. A standalone AI system whose main purpose is detecting financial fraud is kept off the EU AI Act’s high-risk list by the explicit exception in Annex III, point 5(b). The classification changes to high-risk the moment the same model is built into a life or health insurance pricing or risk-assessment system, because point 5(c) carries no fraud exception. Annex III obligations apply from 2 December 2027, so the way an insurer separates fraud detection from pricing today sets its compliance load for years.

1What This Means in Practice

The EU AI Act sorts systems by risk, and only a short list in Annex III counts as high-risk for financial services. Two entries matter for insurers. Point 5(b) covers AI used to evaluate the creditworthiness of natural persons or to set their credit score, and it ends with one carve-out: systems used for the purpose of detecting financial fraud are excluded. Point 5(c) covers AI used for risk assessment and pricing of natural persons in life and health insurance, and it has no such carve-out. That gap is the whole decision.

The classification follows the system’s main intended purpose and how it is wired into the business. Three cases show the line:

The safe architecture is separation. Ableneo runs AI in production for insurers across the same regulated EU markets this rule governs, with property underwriting live in 3 of 6 markets for one insurance group. That production work is where the high-risk line gets drawn in practice. Keep the fraud model as its own system, with its own purpose, its own logs, and a person in the loop, and it stays on the right side of point 5(b).

2Why This Matters for Regulated Industries

For a CEE insurer, the stakes are concrete. A high-risk classification under Annex III triggers the full set of duties in Articles 8 to 15: a risk management system, data governance, technical documentation, event logging, human oversight, and a conformity assessment before the system goes live. Deployers of high-risk insurance systems also owe a fundamental rights impact assessment under Article 27. Those obligations apply from 2 December 2027, the date set for Annex III high-risk systems after the Digital Omnibus.

Getting the classification wrong is costly in both directions. Treat a plain fraud detector as high-risk and you load a low-risk system with documentation and assessment work it never needed, then watch a deployment you wanted in production slow to the pace of a conformity file. Treat a pricing engine with a fraud feature as low-risk and you miss duties a supervisor will expect. The work is to read each system against the exact wording of point 5, not against the word “fraud” on a slide.

A standalone AI fraud detector is excluded from high-risk by Annex III point 5(b); a fraud capability inside a life or health pricing system is high-risk under point 5(c), which has no fraud exception.

3What Separates a Fraud Model That Stays Low-Risk From One That Becomes High-Risk?

The test is the system’s main intended purpose and how it is connected. Point 5(b) excludes AI used for the purpose of detecting financial fraud, so a model built and documented as a fraud detector stays off the list. The exception narrows when that model drives a decision the Act does regulate. If a fraud score feeds a credit or pricing decision, or sits inside a life or health insurance pricing system, it travels with that system’s classification. For insurance this is sharp: a fraud capability folded into a point 5(c) pricing engine is high-risk, while the same capability run as a separate claims-fraud system is not. The architecture decision and the legal classification are the same decision.

4What Still Applies if Your Fraud AI Is Not High-Risk?

Low-risk does not mean unregulated. A fraud model that makes or heavily shapes an automated decision about a person falls under Article 22 of the GDPR, which gives that person a right to human intervention and an explanation. For a financial entity, DORA has applied since January 2025 and treats the fraud system as ICT that needs resilience testing, incident reporting, and third-party oversight. EIOPA’s Opinion on AI governance and risk management, published on 6 August 2025, tells insurance supervisors to apply the rules insurers already live under, on data governance, record-keeping, fairness, and human oversight, to AI systems whether or not they are high-risk. A fraud model outside Annex III still answers to all three.

5How Does the Fraud Exception Differ Between Banking and Insurance?

The wording is not symmetrical, and the difference matters. Banking’s credit-scoring entry, point 5(b), names the fraud exception directly, so a bank can keep a transaction-fraud model off the high-risk list on the face of the text. Insurance’s entry, point 5(c), names no exception, so an insurer cannot lean on an Annex III fraud carve-out the way a bank can for credit. Recital 58 adds a broader principle: AI provided for by Union law to detect fraud in the offering of financial services, and to calculate the capital requirements of credit institutions and insurance undertakings, should not be treated as high-risk. The qualifier “provided for by Union law” limits how far that principle reaches, so an insurer should record the legal basis of its fraud system rather than assume the recital covers it.

6What Should an Insurer Do Before the December 2027 Deadline?

Start with an inventory. List every AI system that touches fraud, underwriting, pricing, and claims, and classify each against Annex III point 5 on its real purpose, not its product name. Where a fraud capability sits inside a life or health pricing engine, decide whether to split it into a standalone system, which is the cleanest route to keeping it low-risk. Document the main intended purpose of each model, the legal basis wherever you rely on Recital 58, and the human oversight around any automated decision. An insurer that does this in 2026 has until 2 December 2027 to build, instead of a scramble once the deadline is in sight.

7The Ableneo Perspective

Ableneo has delivered 49 AI projects, with 35+ running in production for 26 clients across finance, energy, and telco, so where the high-risk line falls is a question it answers inside live systems rather than on paper. The firm passed an ECB vendor security assessment and runs insurance AI in regulated EU markets, which is the setting where the gap between a standalone fraud model and a pricing system stops being theoretical. For insurers weighing where to draw the line before a system reaches production, the Ableneo AI transformation FAQ collects the governance decisions that come up first.

Key takeaways

Sources

Planning AI in a regulated business? Ableneo takes systems from classification to governed production.

Talk to Ableneo