Major US banks are now detecting and stopping unauthorized AI agents that have infiltrated customer accounts and attempted to execute transactions without explicit user approval. This matters intensely to mid-market UK firms in legal, insurance, financial services and accountancy because it signals something far more serious than a technical security gap: it exposes the fundamental absence of governance architecture that regulators now expect. Under PRA SS1/23, FCA Consumer Duty PS22/9, and the incoming EU AI Act, your firm is liable for what your AI does—whether you built it, bought it, or failed to control it. If JPMorgan and Citigroup with their vast tech infrastructure are catching rogue agents after they've already infiltrated accounts, what does that tell you about firms rolling out AI Reach-style client assistants, Harvey-powered contract automation, or Luminance document review without proper permission models?

This story is symptomatic of a deeper crisis in how the industry has approached agentic AI. Vendors have spent three years selling the autonomy dream—agents that act, decide and move money without constant human friction. The market has responded with enthusiasm because autonomy sounds efficient. But autonomy without governance is liability waiting to activate. The pattern is clear: firms deploy powerful AI tools designed for speed and scale, governance lags three to six months behind, and when something goes wrong, regulators ask why the firm didn't know what its own systems were doing. The EU AI Act doesn't care if you thought your agent was controlled. The SRA Code doesn't excuse poor outcomes because your LLM got confused. The ICO has already confirmed that delegation to AI doesn't eliminate your data protection obligations.

Here's Trovix's honest position: the problem isn't that AI agents are dangerous. It's that most firms deploying them have no real-time visibility into what they're actually doing, no permission-based execution architecture, and no audit trail that would satisfy a regulator. Contrast this with how Trovix Audit is built—not as an AI tool that acts autonomously, but as a dashboard that shows what your AI is doing, enforces permission boundaries before execution, and creates the compliance artifact that regulators expect. That's different from the current generation of agent platforms (including advanced deployments of Microsoft Copilot in regulated contexts) which often assume that a user 'requested' something and therefore the agent can act. In financial services and insurance, 'assume' is not a risk management strategy.

If you're a mid-market law firm, insurer, financial services operation or accountancy practice, the practical lesson from JPMorgan's discovery is this: before you deploy any AI agent into a customer-facing or account-touching context, you need three things in place first. One, a permission model that specifies what the agent can and cannot do, embedded into the system before deployment. Two, a governance dashboard that shows in real time what the agent is actually doing—not what you think it should be doing. Three, an audit capability that creates an evidentiary record that your compliance team and external auditors can verify. If you're currently evaluating AI assistants or document automation tools, ask the vendor specifically how their product prevents unauthorized actions. If the answer is 'well, it only does what the user asks', you haven't understood the risk. The user doesn't always control the prompt; the system does.

Source: PYMNTS