Project SAGE: Who Gives an AI Agent Authority to Act?

An AI agent may find a sensible next step for your business. That does not establish its authority to take it. AI agent governance means deciding what a system may do, on whose behalf, within which limits, and who answers for the outcome.

That question sits at the centre of Hong Kong’s Project SAGE. It also matters to how we think about SenangAI at Kognosys: helping a business understand its problems, risks and opportunities, then supporting action under the business’s own rules.

What Project SAGE is exploring

On 2 November 2026, the Hong Kong Monetary Authority launched Project SAGE, short for Securing Agentic Finance with Governance and Enablement.

Its scope includes potential uses such as everyday payments, procurement and treasury management. Alongside those applications, HKMA will explore agent identity and delegated authority, interoperability, and legal and regulatory questions including liability and consumer protection.

Delegated authority means permission to act on someone else’s behalf. Knowing which agent sent a request is one question. Knowing whether its owner authorised that particular request is another.

SAGE is an exploratory initiative, rather than a completed rulebook. Separately, HKMA confirmed that it published a discussion paper on banks’ use of agentic AI on the same day, seeking industry feedback on governance.

The distinction matters. Businesses should understand the direction of supervisory work without presenting proposals as settled requirements.

The business problem starts before payment

Consider a fictional distributor. An AI system notices that one product is selling slowly while a related product sells well. It proposes a bundle that could improve stock movement and margin.

That could be useful. But several different decisions sit behind the suggestion.

Can the system see the supplier cost? Is that cost current? Has it included delivery charges and existing customer discounts? Can it prepare an offer, change a price, contact customers or commit stock?

These questions involve different permissions and different people. A sales manager might approve a campaign while finance sets a minimum margin and operations confirms available stock.

Calling the whole sequence “an AI workflow” does not remove those responsibilities. The workflow needs to preserve them.

This is why the difference between an AI agent and a chatbot becomes important once tools can change business records or make commitments. A helpful explanation and an authorised action carry different consequences.

AI agent governance needs to reach the action

For SenangAI, the design question is how to carry the business’s rules through the entire journey from information to action. I would assess that journey through five practical questions.

What evidence supports the recommendation? A proposed bundle should show the relevant sales, stock and cost records. If sources disagree or information is missing, the uncertainty should be visible before someone approves it.

Whose access is being used? A request made for a salesperson should respect that person’s authorised scope. Connecting multiple systems should not quietly give every user access to everything those systems contain.

What action is permitted? Reading a record, drafting an offer and changing a price are separate capabilities. Each needs a clear boundary. Instructions written in a chat window should not be the only thing preventing an unauthorised change.

Who owns the decision? Approval should belong to someone with authority over the commitment. A button labelled “approve” is weak control if any employee can press it or cannot see what will happen next.

What happens when execution goes wrong? The business needs a record of what was attempted and completed, a way to suspend further action, and a recovery process. Stopping an agent does not reverse a payment or recall a message already sent.

These are evaluation criteria for the systems we build and deploy. They should be demonstrated in the actual workflow, including refusals and failures.

Some expectations already exist

It would be a mistake to read SAGE’s exploratory status as meaning that AI governance starts sometime in the future.

In remarks on 16 October, HKMA described its published supervisory guidance as calling for clear ownership of each AI use case and an up-to-date inventory of approved applications. It also emphasised ongoing monitoring, defined boundaries and response measures, including ways to suspend agents where appropriate.

The same remarks said HKMA would explore a centralised supervisory registry. An organisation’s own inventory and a possible sector-wide registry serve different purposes. A business can identify its systems and owners today while a regulator considers wider arrangements.

Those remarks concern banking supervision in Hong Kong. They do not establish a new general requirement for every Malaysian business. The practical questions about ownership and authority are still useful when assessing an agent connected to your own systems.

What this means for SenangAI

Our intended direction for SenangAI is broad business intelligence and action: helping people identify issues, trace causes, assess risks and find opportunities across the information their organisation is permitted to use.

That ambition makes governance part of the product design. If a system can connect a sales opportunity to stock, customer history and financial information, it also needs to respect the boundaries around each source and each possible action.

The aim is to make the evidence, uncertainty, proposed next step and responsible decision-maker clear. Where execution is enabled, the permitted action and approval conditions need to be clear too.

These are design goals. Each deployment needs validation of its integrations, access rules, approvals, logs and recovery arrangements before the business relies on them. This article makes no claim of HKMA endorsement, SAGE certification or independently verified compliance for SenangAI.

For a business starting now, choose one useful workflow and map it from request to result. Identify the records it reads, the changes it can make, the approvals it needs and the person who can stop it. Then test a request it should refuse and an action that fails halfway through.

That gives an enterprise AI roadmap something concrete to build on. AI agent governance becomes meaningful when the business can show how authority is granted, enforced and withdrawn during real work.

Quick answers

Is Project SAGE a new AI law?

HKMA announced an initiative to explore agentic finance applications and supporting governance arrangements. The launch announcement does not establish a universal AI agent licensing system.

What is the difference between agent identity and authority?

Identity establishes which agent is making a request. Authority establishes whose permission it carries and whether that permission covers the requested action.

Does every AI action need human approval?

The business should decide according to the action’s risk and consequences. A narrowly defined routine action may be permitted within limits, while a significant financial or customer commitment may need approval from a named owner.

How does this relate to SenangAI?

It informs our design direction: evidence people can inspect, access boundaries, clear decision ownership and controlled action. The controls and capabilities available in a particular deployment need to be verified for that deployment.