Anthropic released ten AI agents last month designed to handle financial services work — pitch decks, statement review, compliance escalation. The news hit hard enough to rattle FactSet and Morningstar shares. For UK regulated firms, the headline is seductive: automation of grinding document work. The reality is more complex. These tools promise to escalate compliance issues, but escalation is not the same as review. FCA Consumer Duty PS22/9 and PRA SS1/23 both demand that firms understand the decisions affecting their clients and regulatory obligations. An autonomous agent that flags 200 items for human review has not solved compliance — it has shifted liability and created a new audit trail problem. Mid-market financial services firms, law firms and accountancies cannot simply bolt an agent into their workflow and call it governance.

This story is part of a pattern we have watched develop over 18 months: vendors chasing scale and spectacle in agent design while underestimating regulatory and operational friction. Harvey and Legora proved earlier that document automation could work in legal. But they scaled through partnership and integration with existing workflows — not by selling autonomous agents. Luminance built document intelligence the hard way: training on thousands of real files, embedding risk taxonomy, and keeping humans in control of the decision. Anthropic's approach is different. It is general-purpose and broad. That breadth is a marketing advantage. It is also a governance weakness. Financial services is not a domain where general-purpose works well. Compliance is granular. Client outcomes depend on detail. Regulators expect explainability. An AI agent that drafts a pitch deck is one problem. An AI agent that reviews a financial statement for regulatory issues and escalates is another entirely.

Here is Trovix's honest view: autonomous compliance agents fail where human judgment matters most. The FCA does not care how many documents an agent reviewed. It cares whether your firm knew what it was doing. When you deploy an agent in compliance review, audit or document assessment, you are creating a new layer between the client and the decision-maker. You are also creating a new vector for bias, hallucination and context loss. The alternative is not to reject agents. It is to deploy them where they add real value without creating regulatory theatre. That means: agents for triage and categorisation, not final judgment; agents for flagging patterns, not autonomous escalation; agents that augment human expertise, not replace it. Trovix Sift uses document intelligence differently. It extracts and structures information humans need to make decisions. It does not pretend to be the decision. Trovix Aria works similarly — it retrieves and contextualises information for fee-earners. Neither tool pretends to autonomy where governance and accountability demand human control.

What should a mid-market insurance firm, law practice or accountancy do right now? Do not wait for a vendor to tell you their agent is 'FCA-compliant' or 'audit-ready'. Those words mean nothing. Instead: audit your compliance workflows and identify where humans are making the critical judgment calls. Those are the places agents cannot go, or should not. Look for upstream work — intake, triage, data extraction, flagging duplicates, categorising documents — where agents can add real speed without creating liability. If you are reviewing vendor claims about autonomous compliance agents, ask for their explainability design, not their accuracy scores. Ask how they handle edge cases and regulatory exceptions. Ask who is liable when the agent gets it wrong. If the vendor cannot answer clearly, the agent is not ready for your regulated business. Anthropic's announcement is technically impressive. But impressive technology in the wrong governance context becomes a risk, not a tool.

Source: Bloomberg News