AIG's warning that AI infrastructure buildout is pushing P&C providers to their limits has landed like a live wire in the UK insurance market. The story is straightforward: data center construction and operation create novel, complex risks—from power supply interdependencies to cooling system failures to cyber exposure spanning development and operations phases. No insurer's historical claims data covers this. No standard model fits it. And the FCA's Consumer Duty (PS22/9) now demands that insurers understand and price these risks accurately or face enforcement action for treating customers unfairly. For mid-market UK insurers already stretched by the transition to remote underwriting and the Lloyd's Blueprint Two modernisation drive, this is not a problem of capacity alone. It is a problem of knowledge. And that matters intensely to the regulated firms buying insurance from them.

This story is part of a larger pattern: the insurance sector has been sold a lie about AI. The narrative has been that ChatGPT, Claude, Copilot and similar large language models would automate underwriting, speed claims processing, and unlock efficiency. In reality, these tools are general-purpose pattern-matchers trained on historical text. When faced with a risk class that did not exist five years ago—AI infrastructure—they have nothing to match against. They hallucinate plausible-sounding answers. They miss tail risks. They cannot reason about correlated failures in systems they have never seen. Meanwhile, every regulator in Europe, including the FCA and the PRA, is now scrutinising whether firms deploying AI to make risk decisions can actually explain what the AI is doing and why. The EU AI Act and ICO guidance make it clear: opaque AI is a compliance liability, not an asset. AIG's capacity squeeze is not a shortage of capital. It is a shortage of rigorous, explainable underwriting tools designed specifically for unprecedented risk.

Trovix's view is simple and unpopular: generic AI cannot solve this problem. What UK insurers need is domain-specific AI that understands insurance logic, not just language. That means tools built by people who know underwriting, claims validation, and regulatory reporting—not by teams optimising for engagement metrics. Trovix Sift is built exactly for this: it extracts and validates data from risk submissions and historical claims with the precision and explainability that underwriters and regulators can audit. It does not hallucinate. It does not match patterns to data it has never seen. It extracts structured fact and flags ambiguity. Compare this to Luminance or Harvey, which are powered by transformer-based LLMs and marketed as 'AI-native underwriting tools.' They are faster at summarisation, yes. But speed without rigour is malpractice under the Consumer Duty. Tools like Trovix Aria are built to augment human underwriting judgment with retrieval-grounded knowledge from your own firm's playbooks, case law, and regulatory guidance—not from the internet. That is the difference between a chatbot and a working partner.

If you are a P&C insurer, a broker, or a financial services firm placing risk right now, do three things immediately. First, audit every AI tool you are using in underwriting, pricing, or claims triage. Ask your vendor this question: 'How do you handle novel risks with no historical data?' If the answer is 'we model it' or 'the LLM infers it,' you have a compliance problem. Second, implement explainable, purpose-built data extraction in your intake and validation workflows. The PRA's rules on model governance (SS1/23) and the FCA's Consumer Duty both require you to know exactly how risk decisions are being made. Third, start now on building a proprietary knowledge layer—your own underwriting rules, precedents, and decision logic—so your AI tools reason from your firm's actual experience, not from scraped internet data. This is not quick. It is not cheap. But AIG's warning shows what happens when you bet on generic AI: you run out of capacity to do the work that actually matters.

Source: Bloomberg News