Anthropic released ten AI agents last week designed to draft pitch decks, review financial statements and flag compliance issues for human review. On the surface, this matters to UK banking, insurance and asset management firms operating under FCA Consumer Duty PS22/9 and PRA SS1/23 — the automation of document-heavy, repetitive work sounds like obvious efficiency. But the story buried in this announcement is more uncomfortable: the industry is shipping agentic AI into regulated environments before anyone has seriously solved how to govern it. Anthropic's agents will make decisions, escalate cases, and produce work product. The question no vendor is answering convincingly is: who is responsible when they get it wrong?

This reflects a pattern we see across the industry right now. Harvey, Luminance and other document AI platforms have trained us to think of AI as a faster document reviewer. Copilot products from Microsoft and others position it as a smarter assistant — still human-controlled, still human-responsible. But agents are different. They make autonomous decisions within defined parameters. They delegate upward. They run unsupervised. The financial services sector is now being offered this capability by every major AI vendor, and the regulatory framework has not caught up. The FCA's AI Principles and the incoming EU AI Act classify many of these agents as 'high-risk systems' requiring documented governance, bias testing, and audit trails. Most firms implementing them are not doing any of that. They are simply turning them on.

Here is Trovix's plain view: you should not deploy agentic AI in a compliance workflow without an AI governance framework in place first. This is not idealism — it is practical liability management. A financial adviser's AI agent that escalates a case to compliance based on a miscalculation is not a time-saving innovation; it is a regulatory risk with your firm's name on it. We see firms choosing between Anthropic, OpenAI and smaller vendors based on model capability. That is the wrong question. The right question is: which vendor gives you the audit trail, the decision explainability, the testing framework and the governance dashboard you need to satisfy the FCA, the PRA and your own auditors? Trovix Audit exists specifically because we watched early agentic deployments fail — not because the AI was weak, but because nobody could prove to a regulator what the AI actually did.

If you are a mid-market law firm, insurer or financial services practice, you have a three-month window before agentic AI becomes table stakes in your sector. Do not spend it evaluating which vendor's agents are fastest. Spend it on governance: document what decisions your agents can make autonomously, what triggers human escalation, how you will audit them, and how you will explain them to the FCA or ICO under GDPR. Then choose your AI vendor based on how well they support that governance, not on the feature list. The firms that move this way will be the ones still operating confidently in two years.

Source: Bloomberg News