What Does It Take to Govern AI That Can Act on Its Own?

Agentic AI is helping FIs to improve the quality and reduce the cost of financial crime investigations.

However, agentic AI can go beyond the capabilities of a traditional model – it can plan, select tools, retrieve data, and recommend or take action across a multi-step investigation. Specific guidelines for agentic are needed, but they don't yet exist.

Hawk and Oliver Wyman have taken the existing supervisory principles for traditional AI and extended them for agentic – thus giving FIs a framework for safe implemention.

What the Report Covers

Key points:

• How to extend existing model risk management guidance – including OCC, the Wolfsberg Group, and FATF principles – to agentic systems
• A dual-lens materiality framework separating inherent business risk from design-controlled factors like autonomy and tool access
• The enterprise agent assurance layer needed to register, test, observe, challenge, constrain, and stop every material agent
• A decision-rights ladder mapping agent autonomy to the human approval and oversight each tier requires

With a governed agentic AI approach, every investigation agent operates within a built-in charter, with clear permissions, and a complete audit trail – so oversight keeps pace with what the agent actually does.

Instead of a one-time model approval, institutions get continuous, runtime assurance, with human oversight calibrated to each agent's real risk rather than a single review standard applied across the board.

That makes it possible to meet examiners' expectations for explainability, monitoring, and governance evidence, while retaining full accountability and intervention rights even where the underlying agentic capability is vendor-sourced.

To download your copy of the 34-page framework, please complete the form.