One of the biggest barriers to effective AI adoption in insurance is data quality. While poor or unstructured data has long been seen as an obstacle – especially for predictive analytics – today’s reality is more complex. Insurers face multiple challenges: legacy systems that produce paper documents or operate on scanned images instead of structured data, fragmented databases, and a lack of integrated workflow tools for end-to-end process management. Historically, key data could even go unrecorded – because processing was done manually and business logic or parameters of each case did not leave digital trace in data. It could remain in the employee’s head, in a paper form or just as a vague text in a descriptive field (which is not structured data).

Yet the landscape is shifting.  AI now can transform unstructured, externally received documents into structured data, removing reliance on manual entry and unlocking automation. But data alone isn’t enough. True transformation requires redesigning the entire data architecture – supported by modern data platforms, digital workflows, human-in-the-loop (HITL) interfaces, and automated business logic. Critically, this doesn’t mean centralising all data. Approaches like data mesh enable decentralised, step-by-step integration – vital for processes like claims, where automation hinges on well-structured policy data. Business rules, workflow engines, and smarter system design must converge to drive this evolution. Without these foundational changes, AI cannot effectively enable intelligent, large-scale decision-making.

Leading organisations are tackling this step-by-step, often beginning with specific product lines or regions where data is more manageable. The result is a gradual, focused improvement in the quality and availability of data – a necessary step towards making AI work at scale. In fact, the Sollers AI Report identifies document and email data extraction as the top automation priority, with 69% of insurers using AI in this area – highlighting just how foundational structured data has become. With increasing structured data volumes, the systematic data governance and proper data tools are becoming increasingly important.

The growing popularity of generative AI and LLMs is accelerating interest in large-scale deployments. Yet speed without proper oversight can be risky. AI systems implemented without clear governance, ethical controls, and robust data management may do more harm than good.

To strike the right balance, many insurers are adopting “hub-and-spoke” operating models – combining central governance with decentralised experimentation and adoption. This structure allows business units to innovate while ensuring consistency in standards, quality, and regulatory compliance. This aligns with current trends described in the Sollers AI Report, as on average 54% of insurers now use or are planning a centralised governance model for their AI automation units, in contrary to on average 20% for the decentralized models.