Practical Guide
AI policy template for a Hong Kong company
An AI policy template for a Hong Kong company should answer the question an employee faces at 4pm: can I put this material into this tool, and who checks the result? If the document only says "use AI responsibly", it will not settle that decision. Start with a short rule staff can apply during real work, then attach the detailed controls your business needs.
This is a working starting point, not a substitute for your firm's privacy, security or legal review. The right boundaries depend on the data you hold, the tools you have approved and the decisions your staff make.
AI policy template for Hong Kong: the one-page version
Put the policy owner and version date at the top. Then fill in the brackets before sharing it. A blank field is a decision still to be made, not permission for everyone to improvise.
A policy only works when someone can use it to decide what to do with the file on their desk.
1. Purpose and scope. "This policy applies to [teams and roles] when using [approved AI tools] for company work. The policy owner is [role]. Last reviewed: [date]." State whether contractors and personal devices are covered. Point to existing data and information security policies instead of writing a conflicting set of rules.
2. Permitted tasks. "Staff may use the approved tools to [draft from cleared notes, summarise permitted documents, prepare internal outlines]. Staff may not use them to [make final client, employment or regulatory decisions] without the named approval process." Describe tasks, not a vague promise to "explore AI".
3. Approved tools and accounts. "Use only [named products and account types]. Do not move company work into a personal account or a new plug-in without [owner] approving it." Include the actual link to the internal approved-tool list. A product name alone is not enough if the free, business and enterprise accounts have different controls.
4. Input boundary. "Do not enter [client secrets, personal data, credentials or restricted files] unless [the specific tool and use] has been cleared by [privacy/security owner]. Use fictional or redacted material for practice." Define your data classes in language employees already recognise. A deal name removed from a document may still leave enough detail to identify the client.
5. Output checks. "The person using the output checks facts, numbers, sources, names and confidential material against the original. [Role] approves anything sent to a client or used for a consequential decision." The reviewer needs access to the source. Asking the same chatbot if it made a mistake is not a sign-off.
6. Records and retention. "Save approved final work in [system]. Follow [retention rule] for prompts and generated files. Do not keep unnecessary copies in a personal folder or chat history." State what staff should record when AI materially shapes a decision or external document. The answer may differ between a disposable outline and a board paper.
7. Mistakes and escalation. "If restricted data is entered into an unapproved tool, or an AI output causes a material error, stop using that output and tell [named contact/channel] promptly. Do not delete evidence or try to quietly fix a sent document." Your incident team decides the next steps. Staff need a reporting route they can find in seconds.
8. Review and changes. "[Owner] reviews the approved tools and this policy every [interval] and after a material tool, data or process change. Changes are communicated through [channel]." A dated owner is more useful than a policy that nobody is assigned to maintain.
Those eight lines fit on one page when you remove the explanations. Keep the explanations in a staff guide, with examples from your own work. The one-page policy should remain easy to scan.
What the Hong Kong privacy guidance actually asks you to consider
The PCPD checklist for employee use of generative AI recommends that organisations specify permitted tools, permissible tasks, input and output handling, security and what happens when the rules are breached. It is a checklist for developing an internal policy, not a ready-made approval of any particular product.
The PCPD's AI Model Personal Data Protection Framework calls for governance, risk assessment and human oversight proportionate to the risk. In the policy above, that becomes a named owner, a data boundary and a review step. A high-impact use needs more than the same sign-off as a low-stakes internal outline.
Hong Kong's Personal Data (Privacy) Ordinance applies to personal data handled through AI. The PCPD's summary of the six Data Protection Principles covers collection, accuracy and retention, use, security, openness, and access and correction. Get the people who already own privacy and security to adapt the template to your obligations and industry rules. Do not call a copied one-page document "compliant" because it contains these headings.
Test the policy on three real decisions
Before circulating the page, ask three people to use it. Give them ordinary scenarios with fictional details.
First, an account manager wants to summarise a client's email thread in a public chatbot. Can they tell whether the tool and account are approved and whether the thread contains restricted material? If the answer is unclear, the tool list and input rule need work. My guide to whether ChatGPT is safe for work data explains why the account and the material both matter.
Second, an analyst drafts a proposal from a cleared brief and a pricing spreadsheet. Can they name the source checks and the person who approves the final promise? Use the five checks before an AI draft goes to a client as the test. A correct paragraph can still contain last year's fee or a scope nobody agreed.
Third, someone pastes a confidential attachment into the wrong account. Do they know where to report it without searching a shared drive for twenty minutes? If the reporting route is vague, rewrite it before launch. A policy should make the right next action obvious when someone has already made a mistake.
The policy does not teach the workflow
A staff member can follow every rule and still produce poor work. They may have approved access but no idea how to frame a useful task, separate facts from guesses or review a fluent answer. That is why I would pair the page with a short exercise on a cleared example from the team's actual week. The free prompt library offers a starting format; the exercise tells you whether the team can apply it.
For a Hong Kong firm, I would start with one recurring document, name its source files and final owner, and test the review before broadening the list of permitted tasks. If you want help shaping that working session, see how I approach corporate AI training or talk through the first workflow with me.
Put the rules into one real workflow.
Bring an approved document and the people who own its review. We can build a practical exercise around your team's actual work.
Discuss corporate AI training