Short answer. Measure AI ROI in insurance against the metrics that already move your combined ratio: loss adjustment expense, claims leakage recovery, underwriting cycle time, and loss ratio. Set a baseline on each one before the project starts, then track the delta after the model reaches production. A pilot that never leaves the lab returns nothing, so ROI starts the day the system runs on live policies. Roughly 4 of 5 of the AI projects Ableneo ships reach production, which is where the return is.
Insurance ROI is unusual because the money sits in ratios, not in a single line item. A property and casualty insurer runs on the combined ratio: claims plus expenses divided by premium. Anything below 100% is an underwriting profit. AI earns its return by moving the parts of that ratio you can measure, not by the accuracy score of a model in a lab.
Three places carry most of the measurable return:
Each of these has a figure you can read before and after launch. That is the ROI. A vendor estimate of what might happen is not.
An insurer cannot book the return and ignore the rulebook. Under the EU AI Act, AI used for risk assessment and pricing in life and health insurance is classified high-risk under Annex III, which brings documentation, human oversight, and data governance obligations that carry real cost. EIOPA’s August 2025 Opinion on AI governance confirms that Solvency II and the Insurance Distribution Directive already apply to AI, so the governance work is not optional overhead you can defer to a later release.
That changes the ROI math. The cost of compliance, model documentation, monitoring, and human review belongs inside the business case, not outside it. An honest insurer ROI model counts the governance cost against the ratio improvement. A model that leaves it out looks better on a slide and fails at the first supervisory review.
Measure AI ROI in insurance against ratios you already report: loss ratio, LAE ratio, combined ratio, and recovery rate.
Start with the four that supervisors, actuaries, and the board all recognize: loss ratio, loss adjustment expense (LAE) ratio, combined ratio, and claims recovery rate. Loss ratio and LAE ratio isolate where AI touches the claim. Combined ratio shows whether the improvement survived once expenses and acquisition costs are included. Recovery rate captures subrogation and salvage, the leakage AI is best placed to catch.
Add two operational metrics that lead the financial ones: underwriting cycle time and cost per claim. The operational metrics move first, the ratios follow a quarter or two later. Track both, or you will declare failure before the financial result has had time to appear.
You cannot prove a return against a number you never measured. Before the model goes live, pull twelve months of history on each target metric and segment it by product line, channel, and claim type. Leakage in particular hides inside averages, so a formal audit of a sample of closed claims tells you how much of your LAE is missed subrogation, over-reserving, or coverage error.
That audit becomes the baseline. After the system runs for a quarter, you compare like for like against it. Without the baseline, every number after launch is an assertion, and a board that has seen AI promises before will treat it as one.
Claims return is faster and easier to read. A claims model touches a high-volume, repetitive process, so cost per claim and leakage recovery show up within a quarter. Underwriting return is slower and shows up in the loss ratio over the life of the book, because a better-priced policy only proves itself as claims develop over the following years.
That difference should shape the sequence. Most insurers get a cleaner first ROI story from claims and document processing, then reinvest it into the longer underwriting case. Measuring an underwriting model on this quarter’s numbers understates it, and measuring a claims model over three years hides the win that could have funded the next project.
For life and health pricing and risk assessment, the AI Act treats the system as high-risk, so the business case has to fund a conformity path: technical documentation, logging, human oversight, and post-market monitoring. These are recurring costs, not a one-time setup fee.
The return still clears the bar for most claims and document use cases, which sit outside the high-risk category, and that is one reason insurers start there. For high-risk pricing, the honest ROI line puts the governance cost in the denominator and still expects a positive result, because an unauditable model that a supervisor can halt has a return of zero.
The clock starts at production, not at the pilot. A claims or document use case that reaches live policies typically shows an operational delta inside one quarter and a ratio delta inside two to three quarters. The risk to that timeline is not the model. It is the last mile: integration with the policy and claims systems, sign-off from risk and compliance, and adoption by the handlers who have to use it every day.
Projects that stall at the last mile return nothing, however strong the pilot looked. This is why the production rate matters more than the demo, and why an insurer evaluating a partner should ask how many of that partner’s projects actually reached live policies.
Ableneo measures an AI project by whether it reaches production and moves a number the business already tracks, not by the quality of the pilot. Across the 2025 portfolio, roughly 4 of 5 projects reached production and 94% use LLMs, and the firm’s insurance work in the CEE market, including UNIQA, runs in the same Solvency II and EU AI Act environment that shapes every insurer’s ROI case. The return lives in the ratio, and the ratio only moves once the system is live. See Ableneo’s AI transformation services for how that production-first approach is structured.
Key takeaways
Planning AI in a regulated business? Ableneo takes systems from classification to governed production.