AIG's Eric Andersen told Bloomberg this month what UK insurers already know but won't admit: AI data center projects are maxing out their capacity and forcing P&C underwriters to make coverage decisions on assets they barely understand. A single hyperscaler facility now demands layered coverage spanning project finance, environmental liability, business interruption, cyber, and operational risk—often across multiple jurisdictions and novel failure modes that have no historical data. The FCA's Consumer Duty (PS22/9) places clear accountability on firms to understand and price risk accurately. Yet most mid-market insurers are still extracting risk data manually from unstructured documents: PDF feasibility studies, architectural drawings, regulatory filings from three different countries, equipment specifications buried in appendices. This is not a knowledge problem that generic large language models can solve. It is a data extraction and structured risk assessment problem that requires purpose-built systems.

This story reveals something deeper: the insurance industry is facing a capacity cliff that mirrors the one that hit Lloyd's in the 1990s, except this time the tail risk is being concentrated in AI infrastructure rather than spread across shipping and natural catastrophe. Every major insurer—AIG, Munich Re, Swiss Re, Allianz—is simultaneously trying to underwrite the same cohort of hyperscale projects. Capacity is finite. Pricing discipline is breaking down. And the firms getting rich fastest are the ones who can process risk information 10 times faster than their competitors. The ones who cannot will retreat, leaving market gaps and pushing mid-market UK insurers into risky corners of the market they do not understand. This is not new. What is new is the speed. A data center build moves from conception to full operational load in 18 months. Traditional underwriting cannot scale to that cadence without fundamentally changing how risk data is processed.

Trovix's position on this is simple: generic AI assistants and chatbots built on LLMs are not the answer. Tools like ChatGPT, Microsoft Copilot, and even specialized legal AI like Harvey or Legora were designed to answer questions and generate prose—not to ingest messy, multi-format technical documentation and extract the precise, structured risk signals that an underwriter needs. They hallucinate. They miss nuance. They create compliance exposure for firms subject to FCA and PRA oversight (PRA SS1/23 requires firms to understand their operational AI systems before deployment). The answer is document intelligence paired with domain-specific risk frameworks. Trovix Sift does exactly this: it extracts and structures technical data from the unstructured noise—site plans, grid connection agreements, cooling specifications, disaster recovery protocols—and presents risk factors in the language underwriters already speak. It does not replace judgment. It radically compresses the time between 'I need to understand this asset' and 'I can price this risk.' That is the real competitive edge in a capacity crisis.

If you underwrite or manage risk in a mid-market UK insurance firm, do three things immediately. First, audit your current underwriting intake for AI data center projects—how much time is actually spent on document extraction versus risk analysis? Second, map the specific data fields and risk indicators your team needs to make a coverage decision, then test whether your current AI tools can reliably extract those fields from the documents you actually receive. (Most cannot.) Third, commission a proof of concept with a system designed for structured data extraction, not conversational AI. The firms that survive this capacity squeeze will be the ones who move from 'let me ask the AI a question' to 'let me extract the risk data I need.' It is not sexy. It is entirely necessary.

Source: Bloomberg News