Generative AI and insurance: lessons from the field

Introduction

The enthusiasm surrounding generative AI in the insurance industry is palpable. Yet, the reality on the ground reveals that its application to risk analysis is far more complex and counterintuitive than it appears. This article uncovers the hard-learned truths from real-world projects, aimed at those who want to understand what’s really happening behind the technology beyond the marketing talk.

1. The Blank Page Challenge: When Data Exists but the Blueprint Doesn’t

The first major difficulty lies in the nature of insurance data itself. It’s scattered across multiple systems, requiring extensive work on compilation and versioning. It’s also deeply heterogeneous — not only in format (text, tables, images) but also in business logic, covering domains with opposite risk profiles, such as energy plant insurance versus medical liability.

Faced with dense and highly technical documentation, the greatest challenge isn’t just collecting data. It’s building a data model from scratch to extract insights whose nature and structure are not yet known. It’s a true battle with the blank page — structuring the unknown to uncover value.

The real challenge is not extracting what you already know you’re looking for, but discovering what you didn’t even know you needed to find.

2. The Model Paradox: Precision Before Prestige

Contrary to popular belief, the insurance industry is far more pragmatic than prestigious when it comes to AI models. Training a proprietary model is a complex endeavor, less for technical reasons than for strategic ones: few actors are willing to share the large volumes of proprietary data required for such a project. This data heterogeneity largely explains why, although tempting, training custom models is rarely the preferred path.

Field experience shows that combining existing foundation models with advanced pre- and post-processing techniques — such as retrieval augmentation — can achieve highly acceptable results (often around a 70/30 ratio). Here lies the paradox: although the industry demands extreme precision, it often opts for proven, efficient solutions rather than costly, bespoke models.

3. Compliance: The Gatekeeper That Can Stop the Game Before It Starts

In the insurance world, compliance is not just a box to tick — it’s a critical factor that can halt an initiative before it even begins. Regulations such as GDPR and the upcoming AI Act are central concerns from the design phase onward. Security is equally essential, requiring not only penetration testing (Pentests) for nearly all projects but also rigorous audit trails (traceability, logs, assessments) to guarantee total transparency.

Compliance is such a critical issue that it can kill a project before it even gets off the ground.

This absolute demand for security and regulatory compliance is one of the strongest specificities of the insurance sector. It sharply distinguishes it from other industries exploring generative AI under less restrictive conditions and enforces a constant culture of rigor.

4. Beyond the Chatbot: The Hidden Complexity of User Adoption

AI adoption among teams is more nuanced than it seems. While conversational agents are an excellent icebreaker for introducing non-technical teams to AI, they are only the tip of the iceberg. The real challenge — and often the most surprising — lies in collecting structured, relevant user feedback, which is essential for measuring ROI and improving tools.

This complexity pushes projects to evolve beyond simple chatbots. Increasingly, the focus is on deep UI/UX work to integrate AI seamlessly and usefully into business processes. This technological shift also introduces new roles, such as the 'knowledge base maintainer,' responsible for ensuring the accuracy and relevance of the knowledge on which AI relies.

Conclusion

Ultimately, deploying generative AI for risk analysis is a demanding journey — navigating fragmented data, extreme accuracy requirements, and an unforgiving regulatory landscape. Yet, despite these challenges, the observed return on investment remains high, fully justifying the effort. The question is no longer whether generative AI will transform insurance, but who will master its complexity well enough to lead the way.