AI agent governance

The starting point is a named owner for every agent and an inventory of all agents in use, the same way banks keep a model inventory. Then come limits: which tools and data an agent may use, which amounts it may move and when a person has to approve. Every action is logged, so the bank can reconstruct a decision months later. Bain describes one bank that shuts its agents down automatically after a set time.

The chain of steps creates a specific risk: each single action may be allowed, while the sequence produces a result nobody approved. Agent governance therefore reviews whole workflows and tests them before an agent is given more autonomy.

Singapore's Infocomm Media Development Authority published the Model AI Governance Framework for Agentic AI. The Model AI Governance Framework for Agentic AI has four parts: assess and bound the risks up front, for example by limiting autonomy, tools and data access; make humans meaningfully accountable, with approvals at significant checkpoints; apply technical controls and roll agents out step by step; and tell end users when agents are used. KASIKORN Business-Technology Group, the technology arm of the Thai bank KBank, says in the factsheet that it has begun deploying agents across the bank and uses the framework for its governance.

In the United States, the observations on AI agents from FINRA describe six kinds of agents at member firms, from conversational and software development agents to trade and AML surveillance and trade execution. FINRA notes that agents may lack the transparency and predictability that supervisory practices assume.

The EU AI Act has no category for agents; the duties follow the use case, so an agent that scores consumer credit is high-risk and needs human oversight. The principles from BaFin for algorithms in decision-making ask for clear responsibility in each phase of a model, and its guidance on ICT risks of AI applies DORA to the tools an agent reaches.

For a German bank, agent governance is therefore an extension of what it already runs: model risk management, the ICT third-party register and the four-eyes principle. The general framework for these controls is AI governance in banking.

AI agent governance is the framework of owners, limits, approvals and logs that controls what autonomous AI agents may do in an organization, from design to retirement.

AI governance controls models and their outputs, while agent governance adds control over actions: which tools an agent calls and which transactions it may start without a person.

No law is written for agents alone, and Singapore's voluntary framework is the most specific guidance so far; in the EU the AI Act and DORA apply to agents through their use case and the systems they reach.