
What it is: Hong Kong-first, internationally informed
The Guidelines were produced by AIHK's Safety & Governance Council, whose members come from institutions including the University of Hong Kong, HKUST, HKBU and Lingnan University, spanning computer science, data science, cognitive science, public policy and philosophy. The document opens by observing that AI is already embedded in the ordinary activities of Hong Kong organisations — research, drafting, customer service, fraud detection, operational decision-making, software development, financial analysis and content creation — and that a growing number of systems also reach into enterprise systems, invoke tools and take multi-step actions.
The Guidelines are framed as a voluntary adoption framework for organisations operating in or from Hong Kong, covering businesses, public bodies, professional associations, educational organisations and NGOs. They take Hong Kong's existing law and official frameworks as their baseline — including the Personal Data (Privacy) Ordinance and the Digital Policy Office's Ethical AI Framework — while benchmarking against practice in the EU, the US, the UK, Mainland China, Japan, South Korea and Singapore.
Three risk tiers: controls proportional to impact
The core of the framework is a proportionality principle: the strength of controls should match the impact of each use case. Applications fall into three tiers — baseline, elevated and high-impact. An internal drafting tool may sit at baseline; significant financial transactions, contractual commitments, sensitive customer communications, and high-value, irreversible or externally consequential actions fall into the high-impact tier, demanding stronger human oversight.
Risk depends less on the tool than on its connections and authority. The Guidelines give the example of an assistant that answers questions in a controlled environment and may be baseline risk — until it is connected to customer records, given access to confidential information, or authorised to send external communications, at which point its risk becomes elevated or high-impact. An agent that only retrieves information from a limited internal knowledge base may be low risk; one that can make payments, change customer records, access sensitive data or coordinate actions across systems is high-impact.
Know Your Agent: verify before you grant
For AI agents, the Guidelines introduce Know Your Agent (KYA): identify an agent, its accountable owner and the party it represents, and verify its approved purpose, delegated authority and permissions before granting access or allowing action. The logic: a conventional drafting assistant produces text for a person to review, whereas a connected agent may use that text to initiate a workflow, update a system or send a communication — and the two warrant different governance.
Agent safeguards also include approval gates for consequential actions, continuous monitoring, and tested stop-and-recovery controls. According to Fintech News Hong Kong, the Guidelines require human approval before an agent takes hard-to-reverse actions such as payments, contractual commitments or external communications, and oversight should cover the agent's entire workflow rather than only the content it produces.
Board accountability and a 90-day starting plan
On roles, the Guidelines assign responsibility across the AI value chain: boards and senior management carry oversight duties, organisations should appoint an accountable executive to coordinate AI adoption, and a central AI use register should be established. Vendor due diligence is recommended, with the explicit point that outsourcing procurement does not shift the organisation's own responsibility.
Implementation follows a staged roadmap. Stage One — establishing visibility and accountability — should normally be achievable within the first 90 days of a formal AI adoption programme: appoint the accountable executive, establish a governance forum, publish approved-tool and employee-use rules, launch the inventory, and identify material use cases. Over the following three to six months, organisations introduce risk-tiered intake and approval workflows, classify material systems, conduct vendor due diligence, and set testing and monitoring requirements.
International alignment: ISO/IEC 42001 and labelling
The Guidelines also benchmark against ISO/IEC 42001:2023, which the document describes as the world's first international standard for an AI management system, covering AI-specific risk assessment, performance measurement and lifecycle management. Organisations are encouraged to align with it voluntarily, and to label AI-generated content. Fintech News Hong Kong reported AIHK founder and chairman Neil Tan as saying the goal is informed, accountable risk-taking rather than zero risk, so organisations can scale AI responsibly and meet the expectations of customers and regulators.
It is worth noting what the Guidelines are not: they are voluntary and carry no legal force. Their role is to help organisations build practical structures on top of existing law, and to prepare for legislation that may come later. For Hong Kong businesses, the document's value is its concreteness — a three-tier risk model, KYA checks and a 90-day starting plan are components that can be lifted directly into internal policy. This article is based on AIHK's public document and published reporting and is not legal advice; regulated industries should confirm specific requirements with their regulators and professional advisers.