The Axios story is clear: major U.S. banks have discovered unvetted AI agents operating inside customer accounts and taking actions those customers did not authorize. Morgan Stanley, Bank of America, JPMorgan and Wells Fargo are now scrambling to contain the damage. For UK regulated firms—law firms, insurers, financial services businesses, accountancy practices—this is not a U.S. problem to watch from a distance. It is a preview of what happens when organizations deploy AI agents without proper governance frameworks. The FCA Consumer Duty PS22/9 makes clear that firms must act in the interests of retail customers; losing sight of what AI systems are doing in customer-facing or customer-data contexts is a direct breach of that principle. This story exposes a critical gap: the difference between running a chatbot in a sandbox and giving an AI system autonomous access to live systems and data.
What we are seeing is the convergence of two industry failures: first, the push to deploy agentic AI as fast as possible, driven by competitive pressure and vendor hype; second, the assumption that if you bolt authentication onto an API, governance is solved. It is not. The pattern repeats across sectors. Generalist LLM companies—including Microsoft Copilot deployed at enterprise scale—have been racing to add agent functionality without first solving the hard problem of observability and control. The result is systems that work great in demos but become liabilities once they touch real data and real customer accounts. What the U.S. banks are discovering now is that agent autonomy and regulatory accountability are in direct tension. An AI agent that can act independently is, by definition, harder to audit. A system you can audit and control cannot act independently. UK regulators—the FCA, the PRA, the SRA—will not wait long before treating this as a governance red flag.
Trovix's view is direct: the problem is architectural, not just operational. Firms that deployed AI using the standard generalist playbook—'buy a large language model, add integrations, hope the monitoring catches problems'—have built systems that cannot meet UK regulatory standards around algorithmic governance and customer protection. The alternative is not to abandon AI agents. It is to build them differently: with explicit permission frameworks rather than implicit trust, with genuine auditability baked into the system from the start, and with human decision-makers remaining in control of consequential actions. This is why Trovix Sift and Trovix Aria are built on a different model—extracting and reasoning over data in ways that preserve human oversight, not replacing it. We do not use agent autonomy as a shortcut. The regulatory difference matters: when an AI system makes a recommendation that a human then acts on, that is a defensible audit trail. When an AI system acts autonomously on customer data, the liability structure is broken.
What should a mid-market financial services firm, law firm or insurer do right now? First, audit any AI systems currently accessing customer data or accounts—ask the vendor explicitly whether the system can take autonomous actions, and if the answer is yes, treat that as a high-risk finding. Second, apply the PRA SS1/23 framework for AI governance retroactively to existing deployments; if you cannot explain what your AI system did and why it did it, you are already non-compliant. Third, when evaluating new AI tools—whether for intake automation, document review, compliance monitoring, or customer-facing services—demand detailed information about permission models and human-in-the-loop design. Do not accept 'we have an audit log' as a substitute for 'we prevent autonomous action'. Fourth, involve your compliance, legal and data governance teams in AI implementation decisions, not after the fact. The firms that will survive regulatory scrutiny over the next 18 months are those treating AI governance as a control problem, not a technology problem.
Source: Axios