FIS and Anthropic's new banking agent entering commercial use is a milestone in enterprise AI deployment—but the Forbes story gets at something most vendors won't admit: when an AI agent makes a compliance mistake, regulators will hold the bank accountable, not Anthropic. This matters urgently for UK regulated firms because it exposes a fantasy that's driven too much AI adoption. You cannot buy your way out of duty of care by purchasing an off-the-shelf agent, no matter how capable. The FCA's Consumer Duty (PS22/9) and the PRA's expectations on operational resilience (SS1/23) make this crystal clear. Your firm remains the decision-maker. The AI is the tool. That distinction is not semantic—it's regulatory law.

What's happening now is a reckoning. For the past two years, financial services, insurance, legal, and accountancy firms have treated generative AI and AI agents as magic compliance accelerators—bolt them on, reduce headcount, ship faster. The industry is learning the hard way that agents amplify your risk profile if governance isn't bulletproof. Harvey, Legora, and Luminance have all succeeded because they solved a narrower problem with human-in-the-loop architecture, not because they replaced human judgment. Agents that operate without explicit human verification gates or audit trails are now being stress-tested in regulated environments. FIS's launch doesn't prove agents work in banking; it proves that when they do, the institution deploying them has done the governance work first, not after.

Here's Trovix's view: your AI governance framework must be operational before you deploy any agent—not parallel to it, not afterward. This means documented decision trees showing when and why the AI is authorised to act, automated audit logging that captures every decision and its reasoning, and a clear escalation path when confidence scores drop or ambiguity appears. If your governance exists only in a folder of PowerPoint slides reviewed by compliance once a year, you are not ready for agents. We've seen firms use Copilot or other general-purpose tools and assume they've 'got AI governance' because they ticked a box on the vendor's checklist. That's not governance—that's liability deferral. Real governance means Trovix Audit sits alongside deployment, tracking how the AI is being used, what it's being asked to do, and where human override is happening. It also means Trovix Watch keeping you ahead of incoming regulatory signals about AI—because FCA, ICO, and Lloyd's Blueprint Two requirements are hardening monthly.

For mid-market legal practices, insurers, financial services firms, and accountancy practices right now: audit your current AI deployments this week. Ask: does a human always review before this AI output becomes a client decision or regulatory submission? Is there a decision log? If the answer is no to either, you have a governance gap. Before you deploy or expand any agent—whether it's for claims assessment, contract review, compliance checking, or financial reporting—map out exactly who is accountable for wrong decisions, what the audit trail will capture, and how you will prove to the FCA or ICO that you tested it properly. FIS's agent is good technology. But the story isn't about the technology. It's about liability. Don't let vendors tell you otherwise.

Source: Forbes