HKMA’s ART framework moves bank AI governance toward agent registries and intervention controls

The Hong Kong Monetary Authority has introduced an ART framework for responsible use of artificial intelligence in banking. ART stands for Accountability, Resilience and Trust and combines current supervisory guidance, a planned training-data platform, further sandbox testing and a future discussion paper on agentic AI governance.[1][2]

The current circular and annexes apply to all authorised institutions. The separate agentic-governance work explores intent-anchored accountability, a registry of material AI agents and intervention mechanisms for a forthcoming discussion paper.[1][2][4]

ART connects governance to operating resilience

Agentic systems change the control problem

A conventional model may produce a score, forecast or draft for a person to review. An agentic system can plan and execute a sequence of tasks using data, tools and workflows. That increases the number of decisions occurring between the original instruction and the final customer or market outcome.[1][4]

HKMA says more than 60% of use cases in the second GenAI Sandbox cohort explored agentic AI. Examples are described as automated planning and execution of multi-step tasks within defined boundaries and controls. The figure shows industry direction; it is not evidence that 60% of Hong Kong banks have deployed autonomous agents in live customer processes.[1][3]

A material-agent registry would solve an inventory problem

Banks cannot govern an AI agent they cannot identify. A registry of material agents could connect the system’s purpose, owner, permitted tools, data access, model dependencies and escalation route. The word “material” will need a workable threshold: a customer-facing payments agent and an internal drafting assistant do not create the same operational or conduct risk.[4]

The registry concept also needs lifecycle controls. An agent can change when its model, prompts, tools, data sources or permissions change. Treating registration as a one-time project record would miss the system that actually operates after updates.[4][2]

Intent-anchored accountability needs a chain of evidence

Agentic systems execute several intermediate actions, so a bank needs evidence linking the original approved intent to the plans and tool calls that followed. That does not require every technical step to be exposed to every customer, but it does require the institution to reconstruct why the system acted and which control permitted it.[4]

For example, a bank could authorise an agent to investigate a failed payment while retaining separate approval for releasing funds. A useful test would change the agent’s tool permissions, check whether an attempted action is blocked, pause the workflow and reconstruct what happened from the records. This is an illustrative control test, not a reported sandbox result.[4][2]

The training-data platform addresses a separate dependency

HKMA is working with Cyberport and the University of Hong Kong on an AI training-data-sharing platform with data-quality assurance as a design principle. Shared or synthetic data can broaden testing, but quality, provenance, permitted use and representativeness still determine whether a model behaves reliably in a bank’s target population.[1]

The operating question is whether greater autonomy remains traceable to an approved purpose, enforced permissions and a usable intervention process. The forthcoming discussion paper will give banks the next opportunity to test the registry’s scope and the evidence expected around those controls.[1][2][4]

Report an error or submit an update