Where Should an Insurer Start With AI: Claims, Fraud, or Underwriting?

7 min readAbleneo AI transformation team

Short answer. Start with claims-fraud triage. It returns measurable savings inside one quarter, it sits outside the EU AI Act’s high-risk category, and it builds the data and governance muscle you need before you touch underwriting. Across the market, 60% of property and casualty insurers are still stuck in exploration or proof of concept (Capgemini, 2026). The ones that ship pick a narrow, provable first use case and expand from there.

1What This Means in Practice

“Where do we start with AI” is a sequencing decision, not a technology decision. The insurers that reach production treat the first use case as a way to earn the right to the second one. The pattern that works is simple: start where the value is measurable, the scope is bounded, and a person still makes the final call.

Claims-fraud triage fits all three. A model scores incoming claims by fraud probability and orders the queue so the Special Investigation Unit works the highest-risk cases first. The investigators still decide. The insurer keeps a baseline (recovery rate and investigation hours before the model) and measures the lift after. That number lands on a board slide in one quarter, which is what unlocks the budget for the next step.

The market data shows why discipline matters more than ambition. Capgemini’s 2026 report found that 42% of insurers track no AI metrics at all, 55% report unclear return on investment, and the typical insurer spends 72% of its AI budget on technology and infrastructure, leaving only 28% for the change management and training that decide whether people actually use the system. A pilot with no baseline and no owner is the default way an insurer joins the 60% stuck in proof of concept.

A workable first-year sequence for most insurers looks like this:

2Why This Matters for Regulated Industries

The order you pick is also a regulatory choice. The EU AI Act classifies AI used for risk assessment and pricing in life and health insurance as high-risk (Annex III, point 5c). The same Act explicitly excludes AI used to detect fraud in financial services from the high-risk category. So the two most common starting points, fraud and pricing, sit on opposite sides of the regulatory line. Starting with fraud gives you a high-value, measurable use case with a lighter compliance load, which is the right place to learn before the stakes rise.

That matters for timing. High-risk obligations for insurance pricing and risk assessment phase in to the Annex III deadline of 2 December 2027, and EIOPA’s Opinion on AI governance and risk management (6 August 2025) already sets expectations for data governance, fairness, human oversight, and explainability even for AI that is not classified as high-risk. An insurer that spends 2026 and 2027 shipping non-high-risk use cases arrives at underwriting with a working governance model, an audit trail, and a team that has done this before. DORA, in force since January 2025, adds the operational-resilience requirements on top. None of this is a reason to wait. It is a reason to start where the compliance load is lowest and build up.

Start with claims-fraud triage: measurable in one quarter, bounded scope, human in the loop.

3Why Do Insurers Usually Start With Claims Fraud, Not Underwriting?

Three reasons. The data already exists: claims systems hold years of labeled outcomes, so a model has something to learn from on day one. The value is measurable in money and hours, which answers the board’s first question. And the use case keeps a human in the loop by design, because the model orders the work rather than deciding it, so the governance burden stays proportionate.

Underwriting and pricing invert all three. The decision affects a person’s access to cover, the data is more sensitive, and the model’s output is the decision rather than a prompt for one. That is why fraud triage is the training ground and pricing is the destination.

4What Makes Underwriting and Pricing Harder to Start With?

Risk assessment and pricing in life and health insurance is high-risk under the EU AI Act, so it carries the full obligation set: a risk management system, data governance, logging, technical documentation, human oversight, and transparency to the people affected. On top of the Act, insurers answer to Solvency II and the Insurance Distribution Directive, and actuarial control over the model is non-negotiable. Fairness and non-discrimination testing is not optional here, because a pricing model that produces a disparate outcome is both a regulatory and a reputational failure.

The practical consequence is that the value arrives later. An underwriting-automation pilot needs its governance scaffold built before it can go live, while a fraud-triage pilot can show a number first and formalize governance as it scales. Insurers that start with pricing often stall, not because the model fails, but because the compliance work was not ready when the model was.

5How Do You Choose the First Use Case? A Simple Test

Score each candidate on three questions and start with the highest total. First, can you measure the value in one quarter against a baseline you already have? Second, can you build it on data you already hold, without a new collection program? Third, does it sit outside the EU AI Act’s high-risk category, so the governance load is proportionate to a first pilot? A use case that scores well on all three, like claims-fraud triage or claims document processing, earns the budget and the trust for the harder ones. A use case that scores well only on value, like automated pricing, is the second or third project, not the first.

6How Do You Know a Pilot Is Ready for Production?

A pilot is ready when four things are true. The measured lift beats the baseline by a margin the business accepts. Monitoring is live, so model drift and data-quality breaks are caught before they reach a customer. Human oversight and a rollback path are defined and tested, not assumed. And one named person owns the system in production. An insurer that cannot answer all four is running an experiment, not a pilot headed for production, and experiments are how the 60% got stuck.

7The Ableneo Perspective

In Ableneo’s portfolio, 4 of 5 AI projects reach production, and 94% use large language models in the delivered system. That production rate is the point: it is the difference between a vendor who ships a demo and a partner who ships a system an insurer can run, audit, and defend to a regulator. Ableneo works inside financial services and insurance in Central Europe, with relationships that include UNIQA, so the regulatory world of Solvency II, the EU AI Act, and DORA is the environment the team builds in, not a constraint it discovers late. See Ableneo’s AI transformation FAQ for the decision-by-decision detail behind each step.

Key takeaways

Sources

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

Talk to Ableneo