On 5 May, Anthropic announced AI agents designed to handle financial services work: drafting pitch decks, reviewing financial statements, escalating compliance cases. This matters immediately to UK insurers, asset managers and investment firms because the FCA's Consumer Duty (PS22/9) and PRA's operational resilience rules (SS1/23) now require firms to demonstrate that critical decisions rest on human accountability. Anthropic's announcement reflects a genuine capability advance. But it also reveals how the vendor ecosystem is racing ahead of the governance frameworks that regulated firms actually need to implement these tools safely.

This is part of a broader pattern we have watched unfold over eighteen months: the shift from document AI (Harvey, Luminance, Legora) to autonomous decision-making. Document AI is relatively safe because humans remain in control of the output — a lawyer still reads the summary, a compliance officer still makes the call. Autonomous agents are different. They are designed to escalate, decide, and act without human review at every step. The financial services sector is particularly exposed to this transition because (unlike law firms) financial firms do not have a built-in culture of human sign-off on every decision. An insurer or fund manager who adopts an autonomous compliance agent may genuinely believe it is now 'handling' compliance reviews — when in fact it is creating a ghost layer of decision-making that sits between the firm and its regulators.

Trovix's view is straightforward: autonomous agents in regulated financial services are not yet mature enough to operate without explicit human checkpoints, and vendors marketing them as 'handle compliance' are being irresponsible. The EU AI Act will classify financial decision-making as high-risk from 2025 onwards, and the UK will follow. Compliance will require documented human review of agent outputs, audit trails for every escalation, and explicit oversight of the training data and objectives that guide the agent. Anthropic's agents may be technically sound, but they are being positioned as a shortcut to that governance work — and they are not. Firms considering agentic AI should instead ask: Where is the human checkpoint? Who signs off? What does the FCA expect to see in my file? Products like Trovix Aria and Trovix Watch address the real problem differently — they augment human judgment rather than replace it, and they leave a defensible audit trail.

If you are a mid-market financial services firm or insurer, here is what to do right now: resist the temptation to adopt Anthropic's agents (or similar tools from Claude, OpenAI, or others) without first mapping where human sign-off actually sits in your compliance and underwriting workflows. Do not let a vendor demo convince you that an agent can 'escalate and handle' a decision. Test the agent on non-critical work first — draft documents, data extraction, initial document classification — where humans retain editorial control. Document your governance decision in writing: this agent has these powers, this agent requires human review here, this output is audit-trailed there. Then share that with your FCA relationship manager or auditor. The firms that will win over the next two years are not those that deploy the most advanced agents. They are those that deploy agents honestly, with governance that regulators actually believe.

Source: Bloomberg News