Anthropic released ten AI agents last month designed to handle compliance escalation, pitch deck drafting, and financial statement review across banking, insurance and asset management. On the surface, this looks like the democratisation of automation that mid-market UK firms have been waiting for. In reality, it reveals a dangerous assumption: that because large language models can *perform* a task, they should *autonomously decide* whether a task needs human review. Under FCA Consumer Duty PS22/9 and PRA SS1/23, the responsibility for compliance decisions cannot be delegated to an agent. It must remain with a person. Anthropic's release sidesteps this entirely.
This story is part of a broader pattern. Since late 2024, we've seen Harvey, Luminance, and others release 'agentic' capabilities that promise to reduce human involvement in regulated decisions. Each positions autonomy as efficiency. Each understands the market demand. Few have genuinely solved the governance problem. The EU AI Act's high-risk classification (Article 6) and the emerging ICO UK GDPR guidance on automated decision-making have made it clear that regulators are watching. Yet vendor releases keep outpacing compliance frameworks. The FRC's ISA UK auditing standards already require documented human judgment in high-stakes assessments. Financial services agents that skip this step are creating an audit trail of regulatory exposure, not operational efficiency.
Our view at Trovix is direct: agentic AI in regulated functions needs to be engineered for transparency and control, not speed. The question is never 'can we automate this?' but 'who is accountable if the automation fails?' Microsoft Copilot's financial services rollout stumbled because it created the appearance of decision-making without decision ownership. Anthropic's agents risk the same because they escalate cases — not decisions — meaning firms still own the outcome but lose visibility into why the agent escalated in the first place. That's backwards. What regulated firms need is not more agents making more decisions faster. They need AI that surfaces reasoning, creates audit trails, and hands control back to humans at the critical moment. Trovix Audit exists because governance has to come before deployment, not after.
If you're a mid-market law firm, insurance broker, or accountancy practice looking at Anthropic's release or similar agent offerings, ask three questions before piloting: First, does the vendor provide documented reasoning for every escalation decision, or just a binary 'escalate/don't escalate' flag? Second, does your compliance team have real-time visibility into what the agent is reviewing and why it flagged something? Third, who is accountable if the agent misses a pattern that lands you with the FCA? If you cannot answer these clearly, you are not ready to deploy agents — you are ready to become a regulatory case study.
Source: Bloomberg News