Anthropic's announcement in May that Claude can now draft pitch decks, review financial statements and escalate compliance cases is being treated as a watershed moment for Wall Street. For UK regulated firms — the law practices, insurers, asset managers and accountancies that actually need to satisfy the FCA Consumer Duty, SRA Code, PRA SS1/23 and ICO UK GDPR — it should be treated as a warning sign. These agents can perform discrete tasks at scale. What they cannot do is own responsibility for the decisions they make. When a Claude agent flags a transaction for compliance review, who has actually reviewed it? When it drafts a client pitch, who verifies it against the FRM rules? When it extracts data from a financial statement, who validates the extraction? The regulatory framework doesn't recognize 'the AI flagged it' as evidence of due care. It recognizes documented human judgment.

This matters because we're watching the industry split into two camps. One camp — Harvey, Luminance, and now Anthropic — is building general-purpose agents that can operate across multiple document types and workflows. They're powerful. They're also designed on the assumption that regulated firms will somehow retain meaningful human oversight while processing at AI speed. The other camp, where better outcomes actually happen, recognizes that automation in regulated environments requires three things: first, task-specific tools that humans can understand and audit; second, clear data lineage so you can prove where every extracted fact came from; third, integration with your actual control framework, not layered on top of it. This is harder to market than 'Claude handles your compliance review,' but it's what the FCA, SRA and PRA actually want to see.

Trovix's view is straightforward: agents are only as trustworthy as the data they consume and the processes they feed into. A general-purpose AI agent that can handle ten different financial services tasks is a liability multiplier unless it sits within an architecture that isolates each task, validates each output against source documents, and leaves a trail a regulator can follow. If Anthropic's agents are being deployed to 'escalate cases for compliance review,' someone still has to do the reviewing — and if the reviewing human is working blind to how the agent selected those cases, you've created a gap between the illusion of control and actual control. We've seen this pattern before: Harvey launched to great fanfare in litigation; the firms getting real value from Harvey are the ones treating it as a research assistant, not as a lawyer replacement. The same will be true here. Trovix Sift and Trovix Aria work differently — they're built to make the human decision-maker smarter about what they're looking at, not to remove the human from the decision.

If you're a mid-market firm considering this, ask yourself three things before you start. First: can I audit every decision this agent makes in a way that would satisfy a regulator or client? If the answer is 'not yet,' you need tooling upstream that validates the inputs before the agent touches them. Second: does this agent integrate with my existing control workflows, or am I bolting it on? Bolted-on automation creates compliance gaps because your controls don't see it. Third: what happens when the agent is wrong about a financial statement or a compliance flag? Can I trace why it was wrong and fix the training? If not, you're optimizing for speed at the cost of auditability. The most dangerous moment for a regulated firm is when AI feels reliable enough to trust but your controls aren't actually aligned with what it's doing. Don't let general-purpose agents create that moment.

Source: Bloomberg News