The policy signal: AI growth with seven areas of risk governance
Hong Kong's 2026 Policy Address sets out a seven-pronged strategy for managing AI risks. The areas cover AI-enabled crime, protection of children and young people, an AI ethics framework, application safety, liability for AI-related accidents, safety guidelines for AI agents, and the impact of AI on employment. The government also proposes a new AI Efficacy Enhancement Team within the Digital Policy Office and a Commissioner for AI. Together, these measures signal that AI governance is becoming an ongoing, cross-department management discipline rather than a one-off technology project.
The same policy agenda still promotes experimentation. The expanded Generative AI Sandbox++ is intended to cover securities, wealth management, insurance and the Mandatory Provident Fund, while adoption is planned across health, legal services, construction, transport, welfare, tourism and support for small and medium enterprises. The practical message for business is not to stop testing. It is to make sure each test has controls proportionate to its potential impact, an accountable owner and evidence that can be reviewed later.
The government's stated guardrails for its own use of AI agents provide another useful reference point. Agents are limited to standardised and repetitive internal tasks; they should not make decisions that directly affect the public or replace professional judgement. High-risk actions require human oversight. Confidential or personal data should not be sent to external platforms, and deployment should be preceded by risk assessment, access controls, isolated execution and operation logs. These are sensible design clues for companies, although they should not be mistaken for a complete set of rules already binding every private organisation.
What this changes for Hong Kong businesses
First, responsibility can no longer sit only with the IT department. A marketing team using generative AI for copy, a service team deploying a chatbot, or a human-resources team screening candidates can create privacy, intellectual-property, bias, accuracy, consumer-protection and reputation risks at the same time. Each use case needs a business owner, supported by legal, compliance, privacy, cybersecurity and technical specialists. Cross-functional ownership turns broad principles into workable decisions.
Second, the level of risk should determine the level of approval. The Privacy Commissioner's ethical AI guidance recommends a risk-based approach supported by accountability, human oversight, transparency and interpretability, data privacy, fairness, beneficial AI, and reliability, robustness and security. It also describes a green, amber and red style of use-case classification. A low-risk productivity task can follow standard controls; an amber use case needs stronger assessment and monitoring; a high or unacceptable risk may need to be stopped. This is more usable than a single company-wide answer of either 'AI allowed' or 'AI banned'.
Third, supplier contracts and evidence retention matter more. When an organisation uses an external model or cloud service, it should know whether submitted data is used for training, where it is stored, who can access it, how long it is kept and how the system will be revalidated after major updates. If output affects employment, credit, insurance, professional advice, customer rights or safety, the company should retain the relevant prompt, model or service version, output, human review and final decision. Those records support audits, but they are equally important for finding the cause and limiting harm when something goes wrong.
Days 1–30: find out what the business is already using
Start with an AI use-case inventory. For each entry, record the department, purpose, model or tool, types of data, people affected, use of the output, accountable owner and supplier. Do not count only formally procured systems. Include employee-created generative AI accounts, browser extensions, automated meeting notes and AI features bundled into office software. Without a reasonably complete inventory, management cannot see where the most important exposures are.
A rapid first classification can use three questions. Does the input contain personal, confidential or contract-restricted information? Could the output directly affect a person's rights, opportunities, money or safety? If the output is wrong, can a qualified person reliably detect and correct it before harm occurs? A 'yes' to any question means the tool should not be treated as an ordinary low-risk productivity aid.
While the inventory is being built, issue short interim rules. For example, prohibit uploads of non-public commercial data, customer records, identity documents, health information or complete contracts to unapproved tools. Require users to verify quotations, figures and professional conclusions. Clearly list the outputs that must receive human review. The policy should contain concrete examples and a list of approved tools; abstract warnings alone are difficult for staff to apply.
Days 31–60: build controls, testing and meaningful human oversight
After the inventory, define a written approval path for amber and red use cases. Each assessment should specify permitted and prohibited data, expected performance, a minimum test set, checks for bias and error, the reviewer's qualifications, escalation routes and stop conditions. Public-facing generative AI should be challenged with prompt injection, data leakage, unsafe content, fabricated information and failure scenarios, not just normal conversations.
Human oversight must be capable of changing the result. Reviewers need enough time, source information and authority to reject an AI recommendation. Interfaces should not present model output as certain. Recruitment, performance management, credit, insurance, medical, legal and other high-impact processes need explicit human decision points and an appeal route. Where appropriate, affected people should also be told what role AI plays in the process.
At the same time, review suppliers' data-processing terms, security evidence, subcontracting, incident-notification deadlines, model-update policy and exit arrangements. If a provider cannot explain data flows or supply adequate access and operation logs, reduce the scope of use or move to a service tier with stronger administrative controls. A popular product is not automatically suitable for sensitive work.
Days 61–90: make governance repeatable
The final stage turns a one-time review into a management system. A cross-functional AI governance group can review new use cases, incidents, complaints, major model changes and high-risk exceptions each month. Give the group measurable indicators, such as the number of unapproved tools discovered, the human-override rate, error rates, incident-response time and training completion. Retest significant use cases on a defined cycle and whenever the data, model or intended use changes.
Create an AI incident process that defines when to stop a system, preserve evidence and notify management, suppliers, customers or regulators. Exercises should cover personal-data leakage, fabricated professional advice, discriminatory outcomes, unauthorised automated actions and third-party service failure. Governance documents, training records and technical logs should connect, enabling the company to demonstrate that controls operate in practice rather than existing only on paper.
AIPPA's conclusion is that the most valuable preparation is not an attempt to predict every future rule. It is a clear, traceable and accountable process for using AI. That foundation supports safe experimentation and makes it easier to respond when Hong Kong publishes more detailed ethics, agent-safety or liability measures. This article is a practical analysis of public official information available on 2 October 2026 and is not legal advice. Regulated firms should confirm sector-specific expectations with their regulator and professional advisers.