Almost every insurer today has an AI strategy. Many have multiple pilots running simultaneously across claims handling, fraud detection, underwriting support, and customer service. Boards are asking where AI can reduce costs. Vendors are promising transformation at extraordinary speed. The question is no longer whether to adopt AI in insurance. The question is why so many insurance AI programmes, despite genuine investment and genuine commitment, are still not delivering at scale.
The answer, in almost every case I have observed across enterprise deployments, is not the model. It is the data the model is asked to work with.
The insurance industry has some of the most comprehensive data in any organisation. There are decades’ worth of information stored in insurance companies about policies, claims and fraud. The problem is where that data sits. Insurers have different claim handling systems, underwriting systems, and policy management systems that cannot communicate with one another since they were never built to interact. According to an FSI forum AI implementation survey in 2025, data quality was identified by 55% of financial and insurance organisations as the number one challenge that prevented the implementation of AI technology. A separate study found that 83% of financial institutions lack real-time access to transaction data and analytics due to fragmented systems.
When AI is built on top of that fragmentation, the results are predictable. The model sees only a fragment of the picture and produces incomplete or misleading outputs. The scale of the data preparation challenge in insurance AI is rarely discussed honestly. Before a model is trained, before a single workflow is automated, the foundational work of cleaning, connecting, and governing data consumes the majority of implementation effort. In insurance specifically, that proportion is consistently higher than in most other industries, because the data is older, more fragmented, and spread across more legacy systems than almost any comparable sector. The organisations that treat this as a preliminary task to get through quickly are the ones whose AI programmes stall. The ones that treat it as the primary engineering challenge are the ones that build AI systems capable of performing reliably when they reach production.
{{/usCountry}}When AI is built on top of that fragmentation, the results are predictable. The model sees only a fragment of the picture and produces incomplete or misleading outputs. The scale of the data preparation challenge in insurance AI is rarely discussed honestly. Before a model is trained, before a single workflow is automated, the foundational work of cleaning, connecting, and governing data consumes the majority of implementation effort. In insurance specifically, that proportion is consistently higher than in most other industries, because the data is older, more fragmented, and spread across more legacy systems than almost any comparable sector. The organisations that treat this as a preliminary task to get through quickly are the ones whose AI programmes stall. The ones that treat it as the primary engineering challenge are the ones that build AI systems capable of performing reliably when they reach production.
{{/usCountry}}The insurance organisations that are generating real, measurable value from AI have understood this distinction. They are not competing on model sophistication. They are competing on data quality, data connectedness, and the governance frameworks that make AI outputs trustworthy enough to act on in a regulated environment.
Consider what this looks like in practice across the functions where insurance AI investment is most concentrated.
In underwriting, AI is only as accurate as the data it assesses risk against. The quality of an underwriting AI system is a direct reflection of the quality of the data it was built on. Structured, complete, and well-governed historical data produces models that assess risk accurately and consistently. The inverse is equally true, and the consequences in a regulated, high-stakes environment like insurance are not abstract. The output is underwriting decisions that look automated but carry the errors of the data they were built on. In stark contrast to that, insurers that have done the legwork of ingesting data, validating at entry, and tracing lineage throughout their underwriting platforms are experiencing the true power of AI as promised in the form of faster decisions, better risk selection, and reduced loss ratios.
In claims, the same principle applies at higher speed and higher stakes. A skilled claims adjuster brings the full picture to every decision: The policy record, the customer's history, known fraud indicators, and the current transaction details, all considered together. AI that cannot access all of those simultaneously is not replicating that judgement. It is automating a fraction of it, and leaving the most consequential part of the decision exactly where it was before. It handles the easy part of the decision-making process while creating a new bottleneck.
In fraud detection, the competitive advantage is almost entirely a data advantage. Fraud patterns are identified by finding correlations across large volumes of historical data. The richer, cleaner, and more connected that data is, the earlier and more accurately fraud signals can be detected. Insurers with fragmented data environments detect fraud after the fact. Insurers with integrated data environments detect it at the point of claim submission, or before.
The implication for insurance leadership is straightforward, though the execution is not. Investing in AI without first investing in data architecture is not a shortcut. It is a guarantee of limited returns. Data governance is not the unglamorous back-end of an AI programme. It is the foundation that every other part of the programme rests on. Data lineage, access controls, quality monitoring, validation at ingestion, and integration across policy, claims, and underwriting systems are not implementation details. They are the decisions that determine whether AI produces outcomes a CFO can defend, a regulator can audit, and an underwriter can trust. The insurers that treat this work as strategic are the ones whose AI programmes scale. The ones that treat it as overhead are the ones still running pilots two years later.
The future competitive advantage in insurance will not go to the companies with the highest investments in Artificial Intelligence. It will belong to the ones that treated data infrastructure as a strategic investment before they asked AI to perform.
(The views expressed are personal)
This article is authored by Abhishek Rungta, founder and CEO, Indus Net Technologies Ltd.