Chris Newman's comments at InsurTech Connect last week capture the moment the insurance industry is actually living through: the move from AI as a tool (like Luminance for document review or Legora for claims triage) to AI as an agent—a system that makes decisions, initiates actions, and operates independently within defined boundaries. That shift is real. It is also premature. The ACORD Solutions Group CEO is right that data quality, source verification, and decision accountability are now the bottleneck. But he understates the problem. UK regulated insurers—especially those writing motor, homeowners, or professional indemnity—operate under FCA Consumer Duty PS22/9 and PRA SS1/23 (operational resilience). When an agentic system denies a claim, delays underwriting, or flags fraud on questionable data, the firm remains liable. The regulator does not care that 'the AI decided it.' This is not a technical problem wrapped in compliance language. It is a compliance problem disguised as technical.

The pattern is becoming clear across insurance, legal, and financial services. Generic large language models and pre-trained agentic frameworks (including Microsoft Copilot in its various guises) are being deployed into regulated environments where they have no visibility into data lineage, no audit trail of reasoning, and no way to prove that a decision was made on verifiable information. When things go wrong—and they will—firms discover that they cannot answer basic questions: Where did this data come from? Who verified it? When was it last refreshed? Was it used in another decision that might have created a conflict? The industry is building on sand. Lloyd's Blueprint Two and the emerging EU AI Act requirements for high-risk systems make this worse, not better. A mid-market insurer using a generic agentic platform to automate underwriting decisions without first establishing data provenance is now building regulatory exposure into their operations.

Trovix's view is direct: agentic AI in regulated firms must start with data. Not models. Not speed. Not automation theater. Data. Before any agent can make a decision, a firm must know with certainty what information exists, where it came from, when it was validated, who is authorised to act on it, and how it connects to other data in the same decision chain. This requires infrastructure—not just tooling. Tools like Harvey or Luminance excel at specific, bounded tasks within known datasets. Agentic systems operating across customer data, third-party feeds, internal systems, and external triggers operate in an entirely different risk class. They require Trovix Sift at the foundation—real-time document intelligence and data extraction that establishes provenance on entry—coupled with Trovix Audit to create an immutable record of what data was used in each decision, by which agent, at what time, with what confidence score. Only then is accountability possible. Only then does the FCA take you seriously.

If you are a mid-market insurer or broker exploring agentic AI right now, do three things immediately. First, audit your current data ecosystem. Map where information lives, who owns it, how it moves between systems, and whether it is logged. Second, do not deploy any agentic system until you have provenance infrastructure in place. Yes, this delays time-to-market. It also prevents regulatory harm. Third, treat data governance as engineering, not compliance theater. Ask your vendors to show you the audit trail, not just the performance metrics. If they cannot prove where a decision came from and why, do not use the system in any regulated decision. The market will learn this the hard way—through enforcement action, civil litigation, or portfolio losses—or it can learn it now through disciplined implementation.

Source: Carrier Management