What happened
The Hong Kong Monetary Authority published new supervisory guidance on responsible adoption of artificial intelligence on 15 October. It replaces the High-level Principles on AI that had stood since November 2019, and it organises expectations around three pillars the regulator calls ART: accountability, resilience and trust.
The sandbox results explain the emphasis. More than 60% of the use cases in the second Generative AI Sandbox cohort involved AI agents, and half of those were multi-agent applications.
The expectations are practical rather than philosophical. Each AI use case needs clear ownership and a place on an up to date inventory of approved applications. Oversight moves from human in the loop to human on the loop, which the HKMA frames as continued accountability and monitoring rather than approval at every step. On resilience, banks should hold lineage across models, data, tools, infrastructure and third-party providers, and keep corrective measures ready, including kill switches. On trust, the regulator points at the agent harness, the boundaries that translate risk appetite into what an agent may do, and at validation under extreme scenarios.
A discussion paper on agentic AI governance comes next, covering intent anchored accountability, a registry of material AI agents and intervention mechanisms.
Why this is a GRC story
The inventory becomes the control of record. Clear ownership per use case plus an up to date inventory of approved applications turns a policy statement into an auditable list. Most firms can produce that list for the models they bought. Fewer can produce it for the agents a business unit assembled on a low code tool, which is where the sandbox cohort was heading.
Human on the loop changes what evidence looks like. If a person no longer approves each action, the evidence of oversight is monitoring: thresholds, logs, alerts and the record of what was stopped. A signed approval ticket stops demonstrating control. A monitoring report does.
A kill switch needs a rehearsal. Naming it is the easy part. Like a recovery objective, it needs an owner, a defined trigger and a test, otherwise it exists only in a document.
Lineage is third-party risk for agents. Tracing a decision back through a model, a data set, a tool and a provider is agent supply chain mapping, and it is the capability banks will lack the first time a supervisor needs an answer within a week.
Intent anchored accountability is a design artefact. Writing down what an agent is for, and tracing its permissions back to that purpose, happens at build time. Risk usually meets the design months later, so the practical question is whether compliance sits in the room when the agent is scoped.
What to watch
The discussion paper will matter more than the guidance, because it will decide what counts as a material AI agent. Without a definition, a registry is either unworkable or a full time job.
For anyone outside Hong Kong banking, the pattern is worth tracking. Intent, lineage and intervention are the same three questions other regulators are circling.
The first step is small. Count the AI agents in production, then pick one and try to explain its tool permissions in a single sentence. If that takes a meeting, the inventory is not ready.
Attribution: Analysis based on Hong Kong Monetary Authority and related public reporting. This article is original commentary, not a repost of the source material.