Anthropic launched ten AI agents this month targeting financial services professionals: tools to draft pitch decks, review financial statements and escalate cases for compliance review. For UK regulated firms operating under FCA Consumer Duty PS22/9 and PRA SS1/23, this headline matters because it signals the industry is moving toward autonomous decision-making in regulated workflows. But there is a critical gap between what these tools can do and what regulators will hold you responsible for. A document drafted by an AI agent is still your document. A compliance case escalated by an algorithm is still your escalation decision. The FCA does not outsource accountability to Claude.

This story fits a wider pattern: every major AI lab is now launching 'agents' designed to operate semi-autonomously in regulated sectors. Harvey did this for legal work. Luminance positioned itself as a document intelligence layer. Microsoft Copilot is embedding itself into financial platforms. But the industry is conflating capability with compliance. An agent that can review a financial statement at machine speed is not the same as one that meets the standards of conduct under COBS 2R or provides an audit trail that satisfies the FRC ISA UK requirements. The vendors are moving fast. The regulators are watching. UK firms are caught in the middle, often treating these tools as if governance can be bolted on later.

Here is Trovix's direct view: agentic AI in regulated sectors needs architectural humility, not architectural ambition. Anthropic's agents will work well for genuinely low-stakes tasks—generating first drafts, flagging obvious patterns, suggesting next steps. But they should not be the decision-maker in compliance review, regulatory interpretation, or anything that touches the ICO UK GDPR, the SRA Code, or the EU AI Act. The honest difference between Trovix and pure agentic approaches is that we build tools—Trovix Sift for data extraction, Trovix Brief for intake automation, Trovix Aria for knowledge support—that augment human judgment rather than replace it. We do not pretend a model output is a compliance decision. An agent that escalates to a human is useful. An agent that escalates and implies the human can now relax is dangerous.

What should a mid-market law firm, insurance broker or financial services practice do right now? Do not wait for the perfect agent. Instead, audit your current workflow: where do humans actually need to make the judgment call, and where is a tool just speeding up legwork? Anthropic's agents are worth testing on low-risk tasks—initial document triage, data extraction from routine submissions, pattern spotting in datasets. But do not deploy them in compliance sign-off, regulatory advice, or anything that creates a written record for the FCA. And if you do test them, document what the agent did, what the human verified, and why. That is not friction. That is governance. If a vendor tells you their agent replaces human oversight in a regulated context, they are selling you compliance theatre. You are responsible for knowing what you bought and why.

Source: Bloomberg News