Anthropic's announcement in May 2026 of pre-built AI agents for major banks and full Microsoft 365 integration reveals something the industry has been building toward quietly: financial services firms are moving from supervised AI tools to autonomous decision-making systems. This matters urgently to mid-market UK legal, insurance, and financial services firms because the FCA's AI governance expectations—set out in PS22/9 and reinforced by PRA SS1/23—assume human review at decision points. Anthropic's infrastructure play positions Claude as an 'operating layer' for Wall Street. That phrase should make your compliance team sit up. Operating layers make decisions without waiting for human approval. The FCA, the SRA, and the FRC expect you to know exactly what your AI is doing, why it did it, and to be able to explain it to a client or regulator. Autonomous agents make that exponentially harder.

This is not the first time we have seen vendors push capability ahead of governance. But this move is different in scale and intent. Previous waves—Harvey in legal due diligence, Luminance in document review, Microsoft Copilot across enterprises—built tools that augment human work. They sit in the lane marked 'decision support'. Anthropic and others are now building systems intended to execute work autonomously: portfolio analysis, trade research, risk scoring, compliance screening. The EU AI Act recognises this shift by placing autonomous systems in a higher risk category. The UK has not yet formally adopted the same taxonomy, but the direction is clear. The gap between what technology can do and what regulation permits is widening. Firms that wait for regulatory clarity will be waiting a long time.

Trovix's view is direct: autonomous agents in regulated environments need three things that most pre-built platforms do not provide by default. First, they need an audit trail that regulators can actually follow—not logs buried in vendor infrastructure, but transparent, timestamped evidence of input, reasoning, output, and override. Second, they need governance that lives inside the firm, not delegated to vendor dashboards. That is why we built Trovix Audit—because FCA Consumer Duty PS22/9 makes your firm accountable, not Claude. Third, they need to preserve the human decision-maker as the legal and moral agent, not merely as a rubber-stamp. The Anthropic model works for Goldman Sachs, which has the compliance infrastructure of a small government. It does not work for a 40-person insurance broker or a 15-lawyer firm. The vendors are not selling you the same product twice; they are selling institutional-grade AI to institutions and hoping mid-market firms will accept lower governance as a trade-off for speed.

What should you do? First, do not adopt any autonomous agent platform without asking your vendor: can you run your own audit trail, in your own environment, that survives a compliance review? If the answer is 'we have dashboards', push back. Second, request a demo where the system makes a material decision and then you override it—and you see exactly what changed in the reasoning and audit trail as a result. Third, involve your in-house counsel and compliance team now, not after deployment. Do not let engineering timelines drive governance decisions. The FCA's expectations around AI governance—transparency, accountability, human oversight—have not changed. They have just become harder to meet as systems get smarter. The firms that will thrive in this transition are those that build governance into the purchase decision, not tacked onto it afterward.

Source: Fortune