Hong Kong’s financial regulators are moving the city’s generative AI experiment from small-scale banking trials toward a wider test of how autonomous systems could be used across finance. The opening Sandbox++ group includes 36 selected use cases involving banks, securities firms, insurers, pension institutions, and technology partners. SCMP reported that the projects will examine tasks including customer onboarding, payments, claims, and customer service, with trials expected later this year.
The cohort is significant because it brings four regulators into a single framework. The Hong Kong Monetary Authority, Securities and Futures Commission, Insurance Authority, and Mandatory Provident Fund Schemes Authority jointly announced the selection on August 27. The group chose 36 use cases from nearly 100 proposals, drawing in 30 financial institutions and 27 technology partners. Named participants include Ant Bank (Hong Kong), WeChat Pay Hong Kong, HSBC Life (International), and Manulife Hong Kong.
The program is not a blanket authorization to put AI agents in charge of financial decisions. It is a controlled testing environment designed to show where systems can safely assist, automate, or coordinate parts of a process. That distinction matters. Finance is full of tasks where speed and personalization are valuable, but it also has strict duties around identity checks, customer treatment, privacy, operational resilience, and accountability. The sandbox will test whether agentic AI can deliver practical benefits without weakening those safeguards.
From GenAI Experiments to Agentic Financial Workflows
The first generation of commercial AI tools in finance has often been used for summaries, document drafting, chat interfaces, internal search, and coding assistance. Agentic AI suggests a broader role. Instead of merely generating an answer, an agent can be set up to pursue a goal across several steps, retrieve information, pass a case to another system, or prepare an action for human approval. That could make onboarding smoother, help route a payment issue, organize an insurance claim, or guide a customer through a complex service journey.
The word “agentic” should not obscure the limits of the current experiment. The selected projects are tests, not mass-market rollouts. Regulators will want to understand when a system can act with limited autonomy, when it must escalate a decision, and how an institution can reconstruct what happened if a customer disputes an outcome. Those questions are especially important in areas such as payments and insurance, where a seemingly minor error can have direct financial consequences.
Cyberport’s Artificial Intelligence Supercomputing Centre will manage the platform for the trials, according to the joint statement cited by SCMP. That adds an infrastructure component to the policy effort. A local test environment can help participating institutions assess model performance and data controls without immediately building separate AI systems from scratch. It can also make it easier for authorities to observe common challenges across sectors rather than treating every new use case as an isolated project.
Hong Kong has already been exploring how AI affects the financial system, but Sandbox++ is broader in both regulatory reach and use-case ambition. EastFrontier previously covered Hong Kong’s rise in AI-related privacy complaints, a reminder that public trust can be damaged if adoption runs ahead of safeguards. The new sandbox is therefore as much about testing accountability as it is about testing technical capability.
Why Financial Regulators Want a Shared Testing Ground
Financial innovation often crosses institutional boundaries. A customer may use a payment service connected to a bank account, an insurer, a pension provider, and an external technology platform. If each sector develops AI rules and experiments in isolation, gaps can emerge exactly where data and responsibility pass from one party to another. A joint sandbox can make those edges visible earlier.
The 36 selected projects also create a way for regulators to compare approaches. One institution may use an AI system to identify missing information during onboarding. Another may use a related tool to organize claims documentation or support call-center staff. The objective is not to force every firm into the same technology design. It is to develop a clearer picture of which controls, audit trails, testing procedures, and human-review arrangements work across different types of financial activity.
That approach matters for China-linked technology companies active in Hong Kong. Ant Bank (Hong Kong) and WeChat Pay Hong Kong are part of digital ecosystems with large user bases and extensive experience in payments. Their participation makes the program relevant not only to traditional financial institutions but also to the platforms through which many customers encounter digital services. Hong Kong’s value as a testing venue comes partly from this mix of international banks, mainland-connected platforms, and local regulators.
The initiative arrives as mainland companies continue to move AI tools from demonstration to enterprise use. EastFrontier’s examination of ByteDance’s Doubao Work showed a parallel shift in business software, where agents are being designed to consolidate information and help workers complete multi-step tasks. Finance is a tougher destination because the consequences of a bad recommendation or mistaken action are more immediate. That makes Hong Kong’s sandbox a useful stress test for the agent concept.
The Real Test Is Governance, Not Just Performance
A polished demonstration can make an AI agent look capable, but regulated deployment requires more than an impressive interface. Institutions will need to determine which data an agent can access, which systems it can connect to, and when it can do more than provide a suggestion. They will need controls that prevent a model from making unsupported claims, revealing confidential information, or taking an action outside its mandate.
Model reliability is only one part of that work. Financial companies must also address cyber risk, third-party dependencies, model updates, and the possibility that a malicious input could redirect an automated workflow. The program’s value will depend on whether it produces reusable answers to those operational questions. A sandbox that only confirms that AI can summarize a form will not change much. One that helps participants establish robust ways to test, monitor, and govern agentic systems could shape how adoption proceeds across the city.
This opening group is therefore an early institutional signal, not a final verdict on agentic AI. Hong Kong’s authorities are trying to create room for banks and technology partners to experiment while keeping a close view of the risks. The selected 36 projects will reveal whether the city can translate AI enthusiasm into supervised financial products and processes that customers can trust.