Anthropic's announcement in May of AI agents designed to handle financial services workflows is being celebrated as a watershed moment. It is not. What matters to UK regulated firms is not what the technology can do in a lab, but what it can do under FCA supervision, PRA oversight, and SRA governance. Anthropic has unveiled agents that can field tasks. It has not answered the question that keeps compliance officers awake: if an AI agent makes a discretionary decision in mortgage underwriting, investment advisory, or claims triage, who bears the professional liability? The FCA's Consumer Duty PS22/9 does not disappear because the decision-maker is a neural network. Neither does the SRA Code for solicitors. And neither does ISO 42001 certification, which UK firms increasingly need. Anthropic's agents are tools. The accountability gap remains a human problem.

This story is part of a pattern. Anthropic joins Harvey, Legora, Luminance, and others in racing to deploy agentic AI into professional services workflows—especially financial and legal verticals where the stakes and margins are highest. Each vendor is essentially saying the same thing: our agents can think, decide, and act with minimal human intervention. Each one is quietly ignoring that UK regulation does not care what the technology vendor believes. The FRA ISA UK auditing standard explicitly requires auditors to understand the systems they rely on. The EU AI Act, which now influences UK thinking on AI governance, classifies financial services and legal work as high-risk. Lloyd's Blueprint Two mandates that insurers document their AI decision rationale. The pattern is clear: vendors are shipping autonomy; regulators are demanding accountability. The gap between the two is where UK firms will get trapped.

Here is Trovix's honest view. Agentic AI in financial services has its place—but only where human oversight is built into the architecture from the start, not bolted on after launch. The mistake firms make is thinking that buying Anthropic agents (or Harvey, or any other product) solves their workflow problem. It does not. It transfers the problem upstream. You still need to understand what data the agent is learning from. You still need to audit its decisions against FCA Consumer Duty expectations. You still need to explain to the ICO why certain customer records were accessed during an agent's autonomous run. And you still need to document why you chose to trust that agent with that decision, in that context, for that customer segment. Most agentic AI products treat the human-in-the-loop as a feature to be minimized. We treat it as a requirement to be designed. That difference matters. If you are seriously considering deploying AI agents in regulated work, the question is not whether Anthropic's agents are clever. The question is whether your documentation trail will survive an FCA thematic review.

If you run a mid-market law firm, insurance broker, financial adviser, or accountancy practice, here is what you should do right now. Do not wait for the AI vendors to solve the compliance problem. They will not. Instead, map which workflows currently have clear decision criteria and strong audit trails. Those are the only workflows where agentic AI makes sense. For everything else—and that is most of regulated financial and legal work—use AI as a tool to accelerate human decision-making, not replace it. Trovix Sift extracts data and surfaces patterns; a human expert decides. Trovix Aria retrieves relevant knowledge; a fee-earner applies judgment. That is the model that works under UK regulation. Start there. Monitor the FCA's forthcoming guidance on AI governance. And when a vendor tells you their agents can make autonomous decisions in regulated contexts, ask them for the compliance sign-off. If they cannot produce it, they are selling you a liability, not a solution.

Source: Bloomberg News