The story is simple: what insurers could not price three years ago—a $20 billion AI campus—is now routine. Private capital is flooding into AI infrastructure at speeds that have left traditional underwriting flat-footed. For UK mid-market insurers, this is not a trend to watch. It is a forced migration. The FCA Consumer Duty (PS22/9) still requires firms to demonstrate they understand the risks they are pricing. You cannot honestly claim that when you are issuing bespoke policies on assets that barely existed in historical datasets. Your brokers are asking for quotes on facilities you cannot model. Your actuaries are asked for premiums based on assumptions, not data. This is the insurance sector's version of Y2K—except the deadline is now.

What this story reveals is that the insurance industry is being caught between two worlds. One world—the old one—relied on decades of claims data, industry tables, and published loss frequencies. The other world—this one—demands that you price risk on infrastructure types, failure modes, and supply-chain dependencies that have no actuarial history. Private credit markets are not waiting for you to catch up. They are moving capital into AI infrastructure at a pace that makes traditional insurance cycles look glacial. The result is chaos disguised as opportunity. Every major insurer is spinning up specialist teams. Every Lloyd's syndicate is writing bespoke contracts. But most of them are still using the same tools—spreadsheets, expert judgment, and committee sign-off—to manage risks at a scale and speed that demand something fundamentally different: real-time data integration, predictive modeling on emerging asset classes, and the ability to update your underwriting framework weekly, not annually.

This is where the conversation about 'AI for insurance' usually goes wrong. Vendors will tell you that generative AI systems—or retrieval-augmented generation (RAG) tools like Harvey or Legora—can help underwriters process policy documents faster. They can. But that solves the wrong problem. Processing documents faster does not solve the problem of not having the data you need to price the risk. What matters now is the ability to ingest, synthesize, and model fragmentary data about infrastructure that is being built in real time. You need systems that can pull regulatory filings, supply-chain intelligence, operational metrics, and market signals into a live risk model—and update your underwriting assumptions when the picture changes. That is not a document-reading problem. That is a decision-making infrastructure problem. Trovix Watch was built for exactly this: firms that need to monitor regulatory and market change in real time and translate it into actionable updates to their frameworks. Most generative AI tools in insurance today are still treating underwriting as a clerical task. The real challenge is making it a data-driven one.

What should a mid-market insurer or broker actually do right now? First, stop pretending that bespoke pricing is sustainable. It is a short-term revenue play that masks the fact that you do not have a scalable underwriting model for this asset class. Second, map what data you actually need—not what you wish you had—to make informed decisions about AI infrastructure risk. That includes technical specs, supply-chain concentration, geographic exposure, and operational history. Third, build or buy systems that can ingest that data continuously and flag when your pricing assumptions need to change. Fourth, engage with the FCA early. Your Consumer Duty obligations require you to understand the risks you are selling. 'We are learning as we go' is not a compliance posture. You need to show that you have a framework for managing model risk under conditions of extreme uncertainty. This is not about having perfect data. It is about having a defensible process for updating your understanding when the data changes.

Source: CNBC