On 5 May, Anthropic announced 10 AI agents designed to handle banking, insurance and fintech workflows — drafting pitch decks, reviewing financial statements, even escalating cases for compliance review. The market reacted instantly: FactSet and Morningstar stock fell. But the real story matters far more to mid-market UK legal, insurance and financial services firms than to vendor valuations. These agents are being positioned as replaceable parts in compliance workflows. They cannot be. Under FCA Consumer Duty PS22/9 and PRA SS1/23, the firm retains accountability for decisions that affect clients or customers — full stop. No AI agent can delegate that accountability upward to Anthropic's data centre. The question is not whether agents can draft documents or spot patterns in statements. The question is: who answers to the regulator when they get it wrong?

This announcement is part of a two-year pattern we have tracked closely. First came general-purpose LLMs (ChatGPT, Claude, Gemini). Then came specialist products built on those models (Harvey for legal, Luminance for document review, Legora for compliance). Now comes the agent layer — autonomous systems that chain actions together with minimal human intervention. Each wave has promised to shrink the human role. Each has collided with the reality that UK regulated firms cannot shrink accountability. The FCA's 2024 guidance on AI governance, the EU AI Act's focus on human-in-the-loop systems, and the SRA's Technical Standards all point in the same direction: systems that reduce the visible human decision point are systems that regulators scrutinise hardest. Anthropic's agents may be technically sophisticated. But they are arriving into a regulatory environment that is actively hostile to the premise they embody.

Here is Trovix's honest view. Agentic AI will have a real role in financial services and professional firms — but only as an augmentation layer, not an autonomy layer. The difference is crucial. Augmentation means the AI prepares work for a human who retains full decision authority and can explain their reasoning. Autonomy means the AI makes the call, and the human rubber-stamps it. Anthropic's agents are being marketed as the latter. We believe that is regulatory quicksand. The better approach — and the one we have built Trovix Aria to support — is to give fee-earners and compliance officers AI that surfaces insight, flags risk, and prepares draft work, but leaves the certification moment human and visible. When Luminance or Harvey users review AI output, they are not checking a box. They are making a decision they can defend. Trovix Sift uses the same principle: extract and flag, but do not decide. The compliance escalation that Anthropic's agents perform ought to be a human triage call informed by AI pattern-spotting, not an autonomous agent deciding what rises to the board.

If you lead a mid-market firm — law, insurance, accountancy, or financial services — the news from Anthropic should prompt one question now, not in 18 months: what decisions in your critical workflows are currently human, and which could you sensibly hand to an autonomous system under your FCA or PRA operating conditions? The answer for most firms is: fewer than you think. Pitch decks, yes. Compliance escalation, no. Contract redlines, maybe — if a qualified human signs off. Statement analysis for pattern-spotting, yes. Statement analysis for regulatory sign-off, no. Start there. Then ask your AI vendor not whether their system is state-of-the-art (it probably is), but whether they have built in the human decision layer that your regulator will expect to see. If they hesitate, or if they pitch 'efficiency' over 'auditability', walk. Trovix Watch can help you track where FCA and ICO guidance is moving on this. The real competitive advantage over the next 18 months belongs to firms that use AI to make humans faster and better informed, not to make humans redundant.

Source: Bloomberg News