Supervising the Unsleeping: Overseeing Agentic AI When it Cannot be Fully Human

The EU AI Act mandates human oversight for high-risk AI systems. But oversight did not start with GenAI. Financial services had a head start with mature model risk management and validation culture due to the regulatory scrutiny and risk it faces. Here, oversight grew up in traditional data science and MRM: with dashboards and “management by exception” alerts routing anomalies to humans for review. This paradigm has been strained by GenAI, where open-ended workflows with probabilistically generated unstructured data breaks the tidy oversight metrics. Manual human review has hit the walls of scale, speed and consistency. Now, agentic AI exacerbates this further: agents plan, act and iterate at machine pace. This makes human “watch-every-output” oversight impossible, and we lose the focus point of human accountability. We unpack the core paradox: agentic aims to automate, but its autonomy creates failure surfaces that demand detailed governance, risking the benefits it promises. We lay out current best practices for governing agents: defense-in-depth controls, human-in-command checkpoints and automated monitors that prove control at scale.

What you will learn:

• A concise history of human oversight: from quantitative, thresholder “management by exception” to agentic AI with oversight as part of the system. 
• Banking’s regulatory driven enhancements: mature MRM/validation, immutable audit trails and AI-assisted evaluation. 
• Concrete governance patterns for oversight of agentic AI.