Generative AI adoption is accelerating—65% of insurers are already using it1 —but most initiatives remain stuck in proof of concept. The gap is rarely about model capability. It is about execution: selecting the right process, redesigning it with domain experts, scaling on the right platform, and driving adoption with governance and change management.
That was the practical message from Milliman’s webinar, “Artificial intelligence for insurance—Real-world success stories,” led by José Silveiro, Milliman’s consulting lead in Spain and Portugal, and Alexandre Boumezoued, Milliman principal based in France. As Boumezoued noted, only a minority of AI pilots translate into measurable business outcomes, creating a generative AI divide between success stories and stalled projects.
The webinar grounded the discussion in concrete examples. In one claims operation handling large volumes of PDFs—ID cards, signed letters, and other unstructured documents—Milliman combined specialist vision models with large language models to structure data reliably. Using a specified model design choice for tasks like signature detection and document isolation, then applying a different model only to the relevant extracted content, our experts were able to improve accuracy while controlling compute cost. The result: Information extraction was correct in eight cases out of 10, and the shift from manual data collection to targeted validation cut operational processing time by a factor of three.
Fraud management was another high-impact arena, especially as manipulation of photos and evidence becomes easier. The proposed approach embeds AI into the claims workflow: checking consistency between policy inception data and claim details, analyzing communications for anomalies, using vision models to assess whether damage matches the reported event, and comparing repair estimates to market benchmarks. In the example of a household water-damage claim, a comparison of the claim cost against a market benchmark identified an overestimate of more than 25%, triggering escalation rather than delaying every honest customer.
Silveiro summarized the objective clearly: “Fraud detection is one of the most impactful areas for the integration of AI in the insurance sector. The goal is not only to get greater accuracy and operational efficiency, but also to ensure that the detection of suspicious patterns does not create friction or delays for honest customers.”
The third case study focused on actuarial and financial reporting. For a bancassurance life entity facing increasing reporting complexity and tighter timelines, an agent-based system automated analysis and generated an executive-ready PowerPoint deck. What previously took five days was produced in about 15 minutes with a short validation step, while expanding coverage to more funds and surfacing drivers that had not been observed before.
Across all three use cases, the common thread was governance by design: human-in-the-loop oversight, explainability, access controls, monitoring, and auditability. Boumezoued’s point on pragmatism is a useful test for any executive team: “Working with our current clients, we deliver tangible use cases within a three-month time frame, enabling them to accurately evaluate the ROI before committing to broader-scale integrations.”
1 European Insurance and Occupational Pensions Authority. (2025, February 1). Generative AI market survey: Outlook, use cases and risk management. Retrieved 11 May 2026 from https://www.eiopa.europa.eu/document/download/bec886e2-dea0-4bbe-9624-d5f23f85700a_en?filename=EIOPA-BoS-25-679-GenAI-Report.pdf.