What Saudi insurers want from AI is claims, not chatbots

Ask an insurance technology vendor what its AI does and the demonstration will usually start with a customer asking a question. Ask an insurer where its costs are and the answer is somewhere else entirely. The economics of an insurance company are dominated by claims. Premium comes in, claims go out, and the difference between those two numbers, adjusted for the cost of running the business, is the result. A customer service improvement changes the expense ratio at the margin. A change in how claims are assessed, priced or detected as fraudulent changes the loss ratio, and the loss ratio is the business. That is why the useful AI in insurance is unglamorous. A claim arrives as documents: a form, photographs, a police report, an invoice, a medical record. Somebody reads them, checks them against a policy, decides whether the claim is covered and for how much, and looks for signs that it is not what it appears to be. Every one of those steps is document understanding and pattern recognition, which is precisely what current systems are good at, and none of it faces a customer. The Saudi regulatory environment has been opening space for exactly this. The Insurance Authority, established as an independent regulator in 2024 with responsibility for licensing, solvency, market conduct and consumer protection, expanded its regulatory sandbox capacity in June 2025 to allow additional insurtech companies to test innovative products including parametric insurance, peer-to-peer coverage models and AI-driven underwriting. Parametric insurance is the item on that list worth pausing on, because it changes the claims problem rather than automating it. A parametric policy pays a fixed amount when a defined, measurable trigger occurs, a rainfall level, a temperature threshold, a flight delay, rather than when a loss is assessed. There is no claims adjustment, because there is nothing to adjust. For risks with reliable external data, it removes the most expensive part of the process entirely. At LEAP 2026 the Insurance Authority and HUMAIN signed a cooperation agreement to explore AI compute, products and secure data access for the sector. The verb matters and TechScoop has reported it as written. Exploring compute and data access is a long way from an insurer running a model in production, and nothing published indicates a deployment. The data access element is the part that would matter most if it materialises. Insurance AI is limited by data availability far more than by model capability. An insurer sees its own policies and its own claims. It does not see the market, which means it cannot easily detect a claimant presenting the same loss to three insurers, and it has limited history for pricing risks it has not written before. Sector-level data infrastructure, with appropriate controls, is what would change what any individual insurer's models can do. It is also the hardest thing to build, because it requires competitors to contribute data. Fraud detection is where the near-term return is clearest, and it connects to work already happening in Saudi financial services. The compliance and fraud platforms built for banks, including Mozn's FOCAL, are solving a structurally similar problem: identify anomalous patterns in transaction and identity data, with acceptable precision, in a regulated environment that requires the decision to be explainable afterwards. Insurance fraud is a different domain with the same shape. The caution to carry into any insurtech conversation is about pricing. AI-driven underwriting produces more granular risk assessment, which is another way of saying it charges different customers different prices based on inferred characteristics. That is what underwriting has always done, and doing it with a model rather than a table raises questions about which characteristics are being used, whether they are permissible, and whether the customer can be told why they were priced as they were. Market conduct sits within the Insurance Authority's remit, and a regulator that supervises conduct will eventually ask an insurer to explain a price. Widely cited claims about AI cutting insurance claim costs by fixed percentages circulate in vendor and market research material without a verifiable methodology behind them, and are not used here. The questions worth putting to any insurtech vendor at Money20/20 are the same ones that apply anywhere: which insurer is running this in production, on what line of business, and what changed in the loss ratio. Everything before that answer is a demonstration.