JPMorgan's autonomous payment agents sound futuristic. The reality is blunt: 34% of firms are already breached by rogue autonomous actions. UK regulated firms need governance-first thinking, not velocity-first hype.
Agentic AI  Trovix BriefFinancial Services · Legal Tech

JPMorgan and Hyundai Card are testing AI agents that execute business payments without human intervention. This matters to UK regulated firms because the FCA Consumer Duty (PS22/9) and PRA SS1/23 require that any system making financial decisions must be auditable, explainable, and immediately stoppable. What JPMorgan is doing is not inherently wrong—but the data buried in this story is alarming: unauthorized autonomous agent actions were already the third most common AI security breach in 2026, affecting 34.1% of firms reporting such incidents. For a mid-market law firm processing client funds, an insurer settling claims, or a financial services firm executing trades, that breach rate is not acceptable.

This story reflects a wider rush to deploy agentic AI—systems that make decisions and take actions without waiting for human approval. The industry logic is obvious: speed, cost reduction, 24/7 operation. But the security data reveals that speed without governance is a liability masquerading as innovation. We are seeing two distinct approaches emerge. One camp (represented by JPMorgan's move, and mirrored in some enterprise AI platforms like Microsoft Copilot Pro's autonomous mode) prioritizes velocity and assumes governance will catch up later. The other camp—which includes regulatory-first thinking in Lloyd's Blueprint Two and the emerging EU AI Act high-risk classification of autonomous financial systems—insists that governance must be built in from day one, not bolted on after a breach occurs.

Here is Trovix's straight take: autonomous agents in regulated firms need three things that most AI implementations currently lack. First, they need continuous audit trails that show not just what the agent did, but why it made that decision—the reasoning must be extractable in real time, not reconstructed after a problem. Second, they need human checkpoints at the right moments, not at every moment. A solicitor using Trovix Brief to ingest a client file does not need approval for each data extraction—but a claims system autonomously settling a £50,000 payout does. Third, they need isolation: an agent failure in one workflow must not cascade into other systems. Most cloud-native AI platforms do not build this isolation by default because it adds latency and cost. We do, because regulated firms do not have the luxury of assuming that the next breach will be someone else's problem.

What should a mid-market regulated firm do right now? Do not wait for JPMorgan's agents to reach production before thinking about your own governance framework. If you are considering any autonomous system—whether for payment processing, claims settlement, underwriting, or due diligence—start with this: (1) map every decision point where an agent could act, and ask which ones genuinely need human judgment removed; (2) audit your current AI implementations against the ICO UK GDPR guidance on automated decision-making, because most firms are running systems that would not survive an FCA skilled person review; (3) build your governance model before you scale the agent. The firms that will thrive are those that treat AI agents as regulated entities, not tools.

Source: CNBC

Related Trovix product:

Trovix Brief →Book a demo →