EIOPA survey on GenAI and insurance: Where are we now?

On 2 February 2026, EIOPA published its EU-wide survey on the adoption of Generative AI (GenAI) in the insurance sector.

The survey, based on responses from 347 insurance companies in 25 countries, shows that GenAI is already widely used: 65% of insurers have implemented the technology and a further 23% plan to adopt it in the next three years. But most applications are still in the proof-of-concept phase, reflecting a cautious and gradual approach.

 

The main use cases

Sixty-four percent of the use cases that emerged from the survey relate to internal productivity functions, operations (document analysis, coding assistants, data extraction), and decision-making support. Customer-facing applications (36%) such as chatbots, voicebots, and marketing content are less mature.

It’s significant that most use cases aren’t linked to a specific line of business. Many are cross-functional and can be implemented in both life and non-life insurance, particularly in back-office operations and initiatives aimed at improving overall process efficiency.

In terms of the degree of autonomy of the systems, those designed to support human assessments and decisions are common. More autonomous systems and AI agents are on the rise.

 

Risks, obstacles and governance

Among the main risks highlighted by survey participants are hallucinations (inaccurate, complacent, misleading outputs), specific IT vulnerabilities, potential customer data breaches, and problems with explainability of how a particular output was obtained.

Added to these is the challenge of regulatory compliance: several companies report that navigating such a complex regulatory landscape is a significant obstacle, including industry regulations, the AI Act, privacy, and intellectual property.

The human factor is another critical issue. Many companies have highlighted a significant shortage of internal skills, both in finding suitable profiles and in training existing staff. This gap makes it difficult to build in-house teams capable of effectively developing, managing, and governing GenAI systems.

As for governance, the survey shows that 49% of companies now have a dedicated AI policy. But far fewer companies have gone further and adopted comprehensive AI governance.

 

Make or buy? How companies procure AI

The survey revealed a prevailing “make or buy” strategy. Companies generally choose the AI system based on the use case. For general, non-strategic needs, such as increasing internal productivity, insurers tend to buy standard “off-the-shelf” solutions. For core processes where they’re looking for a competitive advantage, companies prefer to develop their own solutions. But this doesn’t mean having to create Gen AI models from scratch, but rather developing customized in-house applications based on existing third-party or open-source models.

By doing this, companies have greater control over the final application and data, while leveraging the power of market-leading technologies.

 

GenAI customization: Between RAG and fine-tuning

On the data front, companies mainly adopt two strategies: RAG and fine-tuning. The survey shows that 38% of respondents use RAG techniques. These allow the model to be enriched with proprietary or contextual knowledge without changing its weights, maintaining control over sources and reducing the risk of hallucinations, with a good balance between cost and performance. In contrast, 21% said they use fine-tuning, a more expensive and complex option that involves retraining the model on internal data to achieve deeper customization. Finally, 27% use neither RAG nor fine-tuning, relying on pre-trained models as they are.

 

Agentic AI

The survey also focuses on agentic AI: of the 957 use cases identified, only 84 fall into this category and are mainly in the early stages of development. Insurers expect an impact mainly in customer-facing applications (49 cases), thanks to more autonomous chatbots and voicebots, while 35 cases relate to back-office activities.

Some use cases are already in production, such as chatbots that provide information on claims or systems that automatically summarize customer calls. In the medium term (3-5 years), wider adoption is expected, especially in core areas such as claims, underwriting, and fraud detection, albeit with significant challenges related to the explainability, traceability, and reliability of autonomous systems.

 

How to interpret this data and key points to consider

The figures show an acceleration in AI adoption in the insurance sector. At the same time, large-scale adoption is being held back by a number of factors. The insurance sector should consider very carefully:

• the complex and evolving regulatory framework, including the AI Act, which requires a shift from sporadic initiatives to an orderly compliance program: mapping use cases, classifying them by risk level, monitoring regulatory developments, and adopting a horizontal and multidisciplinary approach;
• third-party risk with regard to procuring AI from third-party suppliers, to be managed through targeted contract review and new sets of clauses for AI, as part of a procurement process that truly assesses cyber risk, concentration, and business continuity; and
• governance, which must go beyond a mere “AI policy” and become a cross-functional process in the relevant business functions, with clear rules, defined roles and responsibilities, and concrete controls over how AI is used.

For insurance companies, the message is simple, but not one that can be implemented overnight: they need to move quickly, but methodically.

Our Algorithm to Advantage hub is a dedicated space for our latest insights, analysis and thought leadership on AI. It's designed to keep you informed with regularly updated content intended to give you actionable insights you can turn into your competitive edge. Visit the hub.