當一間香港公司要為 AI 訂規矩，會議桌上通常會出現三份文件：個人資料私隱專員公署的《人工智能（AI）：個人資料保障模範框架》、ISO/IEC 42001，以及美國 NIST 的 AI 風險管理框架。三份都值得讀，三份都假設公司已經有一份內部文件：自己的 AI 使用政策。When a Hong Kong company sets the rules for AI, three documents usually land on the meeting table: the Privacy Commissioner's Model Personal Data Protection Framework, ISO/IEC 42001, and the NIST AI Risk Management Framework. All three are worth reading, and all three assume the company already has an internal document: its AI use policy.
三份文件處理的是三個層次。模範框架是香港的建議做法，涵蓋 AI 策略與管治、風險評估與人為監督、模型定製與系統管理，以及與持份者的溝通，對應的是《個人資料（私隱）條例》[1][3]。ISO/IEC 42001 是一套管理系統標準，要求機構建立、實施、維持並持續改進一個 AI 管理系統，用的是「策劃、執行、檢查、改進」的循環[4]。NIST 的框架把風險管理工作分成四個功能：治理、對應、量度、管理[5]。三者都可以幫公司想清楚要管甚麼，但三者都沒有提供一份可以直接發給同事的文件。The three documents work at three levels. The Model Framework is Hong Kong recommended practice, covering AI strategy and governance, risk assessment and human oversight, model customisation and system management, and communication with stakeholders, in line with the Personal Data (Privacy) Ordinance[1][3]. ISO/IEC 42001 is a management system standard: it asks an organisation to establish, implement, maintain and continually improve an AI management system, using a plan, do, check, act cycle[4]. The NIST framework divides the risk work into four functions: govern, map, measure and manage[5]. Each one helps a company decide what to manage. None of them supplies a document a company can hand to its staff.
那份文件就是 AI 使用政策，而它不需要長。三頁 A4 足夠寫完五個決定：誰可以用哪些工具、可以輸入甚麼資料、輸出怎樣核實與標示、裝置與事故怎樣處理、由誰負責與何時檢視。私隱專員公署在 2025 年 3 月發表的《僱員使用生成式 AI 的指引清單》，把這五項寫成機構制定內部政策時應涵蓋的內容[2]。三個框架的分別、五個部分的範例句子、一份填好的例子與一張對應表，就是公司要寫的第一份文件；政策的限制，與它管不到的部分，在最後一部分講清楚。That document is the AI use policy, and it does not need to be long. Three A4 pages are enough for five decisions: who may use which tools, what data may be entered, how output is checked and labelled, how devices and incidents are handled, and who is responsible and when the policy is reviewed. The Privacy Commissioner's checklist for employees using generative AI, published in March 2025, turns these into the content an organisation's internal policy should cover[2]. The differences between the three frameworks, sample wording for the five parts, a filled example and a mapping table are the first document the company has to write; its limits, and what it cannot cover, come at the end.
摘要Summary
- 企業 AI 治理的三個常見起點各有用途：私隱專員公署的《人工智能（AI）：個人資料保障模範框架》（2024 年 6 月）給香港的建議做法，ISO/IEC 42001:2023 給一套可審核的管理系統要求，美國 NIST 的 AI 風險管理框架把工作分成治理、對應、量度、管理四個功能[1][4][5]。The three usual starting points for corporate AI governance do different jobs. The Privacy Commissioner's Model Personal Data Protection Framework (June 2024) sets out Hong Kong recommended practice, ISO/IEC 42001:2023 sets out auditable management system requirements, and the NIST AI Risk Management Framework divides the work into four functions: govern, map, measure and manage[1][4][5].
- 三個框架都假設公司已有一份內部 AI 使用政策。私隱專員公署在 2025 年 3 月發表的《僱員使用生成式 AI 的指引清單》，列出這份政策應涵蓋的五項內容：獲准使用的範圍、個人資料保障、合法及合乎道德的使用與偏見預防、數據安全，以及違反政策的後果[2]。All three frameworks assume the company already has an internal AI use policy. The Privacy Commissioner's checklist for employees using generative AI, published in March 2025, lists the five decisions that policy should cover: the approved scope of use, personal data protection, lawful and ethical use and bias prevention, data security, and the consequences of breaching the policy[2].
- 三頁 A4 可以寫完五個決定：誰可以用哪些工具、可以輸入甚麼資料、輸出怎樣核實與標示、裝置與事故怎樣處理、由誰負責與何時檢視。Three A4 pages can carry five decisions: who may use which tools, what data may be entered, how output is checked and labelled, how devices and incidents are handled, and who is responsible and when the policy is reviewed.
- 每個部分都有一個可以照抄的範例句子與一個字數預算，總長度控制在一千五百字以內，附錄另計。Each part comes with sample wording that can be adapted and a word budget, keeping the policy under 1,500 characters plus appendices.
- 五個部分可以逐一對回三個框架：哪一項對應模範框架的四個範疇、哪一項對應 ISO/IEC 42001 的管理系統要求、哪一項對應 NIST 的四個功能。政策是三個框架的共同底層，本身不構成第四個框架[1][4][5]。Each of the five parts maps back to the three frameworks: which one corresponds to an area of the Model Framework, which one to an ISO/IEC 42001 management system requirement, and which one to a NIST function. The policy is the common layer under all three, and it is not a fourth framework[1][4][5].
- 政策是底線；員工培訓、技術控制與監管機構的指引仍然決定它是否有效。私隱專員公署在 2026 年 8 月就代理式 AI 另發補充指引，說明這個題目仍在移動[6]。A policy is the floor. Staff training, technical controls and regulator guidance still decide whether it works, and the Privacy Commissioner issued separate supplementary guidance on agentic AI in August 2026, which shows the subject keeps moving[6].
三個框架各自回答一個問題Each framework answers a different question
三份文件經常被放在同一個問題下比較，但它們回答的問題不同。表一比較三者。模範框架問的是：一間在香港使用 AI 的機構，怎樣做才符合《私隱條例》[1][3]。ISO/IEC 42001 問的是：機構要建立怎樣的 AI 管理系統，才可以被審核與持續改進[4]。NIST 的框架問的是：一間機構用甚麼步驟去識別、量度與管理 AI 風險[5]。同一間公司可以三個都用，而答案不會互相矛盾，因為它們在三個層次上工作。The three documents are often compared as answers to one question, but they answer different questions. Table 1 compares them. The Model Framework asks how an organisation using AI in Hong Kong can meet the Personal Data (Privacy) Ordinance[1][3]. ISO/IEC 42001 asks what kind of AI management system an organisation needs so that it can be audited and continually improved[4]. The NIST framework asks what steps an organisation takes to identify, measure and manage AI risk[5]. One company can use all three, and the answers do not conflict, because the three work at three levels.
| 框架Framework | 發佈者與日期Publisher and date | 性質What it is | 對公司的用處What it gives a company | 它不提供甚麼What it does not give |
|---|---|---|---|---|
| 《人工智能（AI）：個人資料保障模範框架》Model Personal Data Protection Framework | 個人資料私隱專員公署，2024 年 6 月[1]Privacy Commissioner for Personal Data, June 2024[1] | 模範框架，四個範疇的建議措施；法律要求來自《私隱條例》[1]A model framework of recommended measures in four areas; the legal requirement comes from the Ordinance[1] | 把 AI 治理連到《私隱條例》的六項保障資料原則，涵蓋採購、風險評估、人為監督與溝通[1][3]Connects AI governance to the six data protection principles of the Ordinance, covering procurement, risk assessment, human oversight and communication[1][3] | 一份現成的內部政策；其他風險（知識產權、偏見、員工行為）要公司自己補A ready internal policy, and it leaves other risks such as intellectual property, bias and staff conduct to the company |
| ISO/IEC 42001:2023ISO/IEC 42001:2023 | ISO/IEC，2023 年 12 月[4]ISO/IEC, December 2023[4] | 管理系統標準，要求建立、實施、維持與持續改進 AI 管理系統；認證由獨立機構自願申請[4]A management system standard requiring an AI management system to be established, implemented, maintained and continually improved; certification is voluntary and issued by an independent body[4] | 一套管理系統的骨架：政策、角色、風險、控制、監察與改進[4]The skeleton of a management system: policy, roles, risk, controls, monitoring and improvement[4] | 現成的政策文本；認證需要時間、紀錄與外部審核，ISO 不自行發出認證Ready policy text; certification takes time, records and an external audit, and ISO does not certify organisations itself |
| NIST AI 風險管理框架NIST AI Risk Management Framework | 美國國家標準與技術研究院，2023 年 1 月[5]US National Institute of Standards and Technology, January 2023[5] | 自願的風險管理框架，四個功能[5]A voluntary risk management framework with four functions[5] | 用治理、對應、量度、管理四個功能組織風險工作[5]Organises risk work through govern, map, measure and manage[5] | 認證；香港法律的對應Certification, and a mapping to Hong Kong law |
NIST 的網頁在 2026 年 11 月顯示，AI RMF 正在修訂；四個功能的劃分是 1.0 版本（2023 年）的分法[5]。In November 2026 the NIST page shows that the AI RMF is being revised; the four functions are the 1.0 version's (2023) division of the work[5].
先寫政策，後選框架Write the policy before choosing a framework
公司最容易卡住的地方，是把「選哪個框架」當成第一步。三個框架的比較可以之後再做，因為無論選哪一個，公司都要先有一份文件，寫清楚同事實際上可以做甚麼。私隱專員公署的《僱員使用生成式 AI 的指引清單》就是把這份文件拆成五項內容：獲准使用的範圍、個人資料的保障、合法及合乎道德的使用與偏見預防、數據安全，以及違反政策的後果[2]。五項都是可以執行的決定，所以它們可以用三頁寫完。The place companies stall is treating "which framework" as the first step. The comparison can wait, because whichever framework a company picks, it needs one document that states what colleagues may actually do. The Privacy Commissioner's checklist for employees using generative AI divides that document into five parts: the approved scope of use, personal data protection, lawful and ethical use and bias prevention, data security, and the consequences of breaching the policy[2]. All five are decisions, which is why three pages are enough.
這份文件的測試很簡單：一位新同事讀完之後，能不能回答五條問題。第一，我可以用哪個工具。第二，甚麼資料不可以輸入。第三，輸出要怎樣核實、對外發出時要不要標明。第四，出了事找誰、幾時要找。第五，這份文件由誰負責、下次幾時檢視。五條問題答得出來，政策就可用；答不出來，寫得再長都只是立場書。The document has a simple test: after reading it, can a new colleague answer five questions? First, which tools may I use. Second, what data must I never enter. Third, how is output checked and when does it need a label before it goes out. Fourth, who do I tell when an incident happens, and by when. Fifth, who is responsible for this policy and when is it next reviewed. If the five answers are clear, the policy works; if they are not, more pages only make it longer.
三頁的分工是固定的。第一頁寫目的、適用對象與獲准工具，並把工具清單放在附錄，因為工具會換，政策不應該每次換工具都改版。第二頁寫資料規則：哪些資料可以輸入、哪些不可以、輸出可以存放在哪裏、對外發出前要怎樣核實與標示。第三頁寫裝置與權限、事故報告、違反的後果、負責人與檢視日期。每頁大約四百至六百字，全文控制在一千五百字以內，其餘放在附錄：獲准工具清單、事故聯絡人、培訓紀錄表。The three pages have fixed jobs. Page one carries the purpose, who the policy applies to and the approved tools, with the tool list in an appendix, because tools change and a policy should not need a new version every time one does. Page two carries the data rules: what may be entered, what may not, where output may be stored, and how it is checked and labelled before it goes out. Page three carries devices and permissions, incident reporting, the consequences of a breach, the responsible person and the review date. Each page runs to about 400 to 600 characters, the whole policy stays under 1,500 characters, and the rest goes into appendices: the approved tool list, the incident contact and a training record form.
負責人可以是一個人，不必一開始就是一個委員會。模範框架在管治架構一節建議成立 AI 管治委員會或類似組織[1]，對四十人的公司，先由一位營運或資訊科技主管具名負責、每季與管理層檢視一次，已經足以讓政策有人回答問題；公司長大或客戶開始問審核，才把委員會正式化。政策上寫的是職位，換人時不必改版。The responsible person can be one person; a committee is not needed on day one. The Model Framework recommends an AI governance committee or similar body in its governance structure section[1]. For a 40-person company, naming one operations or IT lead and reviewing with management once a quarter is enough for the policy to have someone who answers questions about it; formalise the committee when the company grows or customers start asking for assurance. The policy names the role, so a change of staff does not force a new version.
三頁政策的五個部分The five parts of a three-page policy
五個部分對應私隱專員公署清單的五項內容，清單把五個部分拆成 15 項檢查，順序也照清單排：先說誰可以用甚麼，再說可以輸入甚麼，然後是輸出怎樣處理、裝置與事故怎樣管，最後是後果、負責人與檢視[2]。表二逐部分列出要決定的事、一段可以改寫的範例句子與字數預算。範例句子是起點；公司要按自己的《員工手冊》、合約與實際工具改寫。The five parts follow the five content areas of the Privacy Commissioner's checklist, which breaks them into 15 checked items, in the same order: who may use what, then what may be entered, then how output is handled, how devices and incidents are managed, and finally consequences, the responsible person and review[2]. Table 2 lists, for each part, the decision it carries, a sample sentence that can be adapted, and a word budget. The sample wording is a starting point; adapt it to the company's staff handbook, contracts and actual tools.
| 部分Part | 要決定的事The decision | 範例句子Sample wording | 字數預算Word budget |
|---|---|---|---|
| 一、適用範圍與獲准工具1. Scope and approved tools | 誰受政策約束、哪些工具獲准、哪些用途獲准、政策適用於公司哪個實體Who the policy binds, which tools are approved, which uses are permitted, and which company entity it covers | 「本政策適用於所有以公司名義使用生成式 AI 的同事、實習生與承辦商。獲准使用的工具列於附錄一，只可使用公司帳戶；未經資訊科技部批准，不得以個人帳戶處理公司資料。」"This policy applies to every colleague, intern and contractor who uses generative AI in the company's name. The approved tools are listed in Appendix 1 and may be used only through company accounts; no company data may be processed through a personal account without IT approval." | 200 至 300 字，工具清單放附錄200 to 300 characters, with the tool list in an appendix |
| 二、個人資料與敏感資料2. Personal and sensitive data | 哪些資料可以輸入、哪些不可以、輸出可以存放在哪裏、保留多久What may be entered, what may not, where output may be stored, and how long it is kept | 「禁止將客戶及員工的個人資料、未公開的合約與財務資料輸入未經批准的生成式 AI 工具。經批准的工具只可處理已去識別化的資料；輸出須存放於公司指定位置，不得存於個人裝置。」"Customer and employee personal data, unpublished contracts and financial information must not be entered into an unapproved generative AI tool. Approved tools may process de-identified data only, and output must be stored in the company's designated location, not on personal devices." | 300 至 400 字300 to 400 characters |
| 三、核實、標示與合法使用3. Checking, labelling and lawful use | 事實與數字怎樣核實、何時標明 AI 協助、偏見與歧視怎樣處理、哪些用途一律禁止How facts and figures are checked, when AI assistance is labelled, how bias and discrimination are handled, and which uses are prohibited | 「對外發出的內容，同事須自行核實事實、數字與引文，並在適當位置標明由 AI 協助生成。涉及歧視、偏見或可能違法的輸出，須停止使用並向事故聯絡人報告。」"Before anything goes out, colleagues must check the facts, figures and quotations themselves and label AI assistance where it applies. Output that is discriminatory, biased or possibly unlawful must be stopped and reported to the incident contact." | 250 至 350 字250 to 350 characters |
| 四、裝置、權限與事故報告4. Devices, permissions and incidents | 哪些裝置可以使用、哪些同事可以使用、憑證與保安設定、事故種類與通報時限Which devices and staff may use the tools, credentials and security settings, and what counts as an incident and when it must be reported | 「只可在公司提供或已登記的裝置上使用獲准工具，並須開啟多重驗證。發現資料外洩、誤輸個人資料、異常輸出或懷疑違法的輸出，須於同一工作日通知事故聯絡人。」"Approved tools may be used only on company-issued or registered devices with multi-factor authentication enabled. A data leak, personal data entered by mistake, an anomalous output or a possibly unlawful output must be reported to the incident contact on the same working day." | 300 至 400 字300 to 400 characters |
| 五、後果、負責人與檢視5. Consequences, responsible role and review | 違反政策的後果、負責職位、檢視周期、培訓與紀錄怎樣保存The consequences of a breach, the responsible role, the review cycle, and how training and records are kept | 「本政策由營運總監負責，每六個月檢視一次，或在相關指引更新時提前檢視。違反本政策的後果依《員工手冊》處理。全體同事須於入職及每次修訂後完成一次培訓並記錄。」"The operations director is responsible for this policy, which is reviewed every six months, or earlier when the relevant guidance changes. Breaches are handled under the staff handbook. Every colleague completes one training session at induction and after each revision, and the record is kept." | 200 至 300 字200 to 300 characters |
五個部分加起來大約一千三百字，放進三頁 A4 仍然寬裕。第四部分的三件事最容易漏：一是「同一工作日」的時限，沒有時限的通報等於沒有通報；二是保安設定的要求，工具預設未必符合公司需要；三是「異常輸出」的定義，寫得太窄，同事遇到奇怪的答案時會以為不用上報。第二部分則要把保留期寫出來，因為《私隱條例》要求個人資料的保存時間不超過達致原來目的所需的時間[3]。The five parts add up to about 1,300 characters, which still fits comfortably into three A4 pages. Three details in part four are the easiest to leave out: a reporting deadline, because a duty to report with no deadline is not a duty; the security settings, because a tool's defaults may not meet the company's needs; and a definition of "anomalous output", because one drawn too narrowly leaves colleagues thinking a strange answer is not worth reporting. Part two has to state the retention period, because the Ordinance requires personal data to be kept no longer than is necessary for the purpose it was collected for[3].
一份填好的例子A filled example
以下是一間虛構的香港貿易公司，四十人，售賣家居用品，客戶包括本地零售商與少量海外買家。它用公司帳戶的 ChatGPT Team 與 Microsoft 365 內的 Copilot 寫報價、整理會議紀錄與翻譯產品頁。這份政策用來示範五個部分怎樣落筆，與任何客戶無關。The company below is fictional: a 40-person Hong Kong trading firm selling household goods to local retailers and a few overseas buyers. It uses ChatGPT Team on a company account and Copilot inside Microsoft 365 to draft quotations, summarise meetings and translate product pages. The policy shows how the five parts are written; no client is involved.
第一頁寫目的、適用對象與獲准工具。「本政策適用於所有以公司名義使用生成式 AI 的同事、實習生與承辦商。獲准使用的工具為 ChatGPT Team（公司帳戶）與 Microsoft 365 Copilot；未經資訊科技主管批准，不得使用其他工具，亦不得以個人帳戶處理公司資料。」工具清單放在附錄一，寫明版本、帳戶類型與批准日期。負責人寫職位：營運總監。Page one carries the purpose, who it binds and the approved tools. "This policy applies to every colleague, intern and contractor who uses generative AI in the company's name. The approved tools are ChatGPT Team on a company account and Microsoft 365 Copilot; no other tool may be used without the IT lead's approval, and no company data may be processed through a personal account." The tool list sits in Appendix 1 with the plan, account type and approval date. The responsible role is written as a position: operations director.
第二頁寫資料規則。「客戶與供應商的個人資料（姓名、地址、電話、訂單內容）、未公開的採購價與合約條款，不得輸入生成式 AI 工具。產品規格、已公開的產品資料與去識別化的銷售數字可以使用。」輸出存放於公司的 SharePoint 專案資料夾，不得下載到個人裝置；保留期跟隨專案資料夾的既定政策，一般為兩年。核實規則寫得具體：報價單的價錢與庫存，同事要對回系統才可發出；對外使用的產品描述，由同事核對規格與法規字眼，並在檔案內標明由 AI 協助起草。Page two carries the data rules. "Personal data of customers and suppliers (names, addresses, phone numbers, order details), unpublished purchase prices and contract terms must not be entered into a generative AI tool. Product specifications, published product information and de-identified sales figures may be used." Output is stored in the company's SharePoint project folder and must not be downloaded to personal devices; retention follows the project folder policy, normally two years. The checking rule is concrete: colleagues check prices and stock against the system before a quotation goes out, and for product descriptions used externally they verify specifications and regulated wording, marking the draft as AI-assisted in the file.
第三頁寫裝置、事故、後果與檢視。「只可在公司提供的筆記本電腦或已登記的裝置上使用，並須開啟多重驗證。發現資料外洩、誤輸個人資料、異常輸出或懷疑違法的輸出，須於同一工作日通知營運總監。」事故例子寫得出來，同事才認得：把客戶名單貼進對話、把未公開的合約貼進 Copilot、工具輸出一個不存在的產品編號、輸出帶有歧視性內容。違反後果依《員工手冊》；政策每六個月檢視一次，或在私隱專員公署更新指引時提前檢視；全體同事入職時完成一次培訓，修訂後再完成一次，紀錄由人力資源保存。Page three carries devices, incidents, consequences and review. "Approved tools may be used only on company-issued laptops or registered devices with multi-factor authentication. A data leak, personal data entered by mistake, an anomalous output or a possibly unlawful output must be reported to the operations director on the same working day." The examples are written out so colleagues recognise an incident: a customer list pasted into a chat, an unpublished contract pasted into Copilot, a product code that does not exist, output containing discriminatory language. Breaches follow the staff handbook. The policy is reviewed every six months, or earlier when the Privacy Commissioner updates the guidance. Every colleague completes training at induction and again after each revision, and HR keeps the record.
這份政策沒有寫「本公司重視人工智能倫理」之類的段落，因為那句話不會改變任何一個決定。它做了五件事：指定工具、劃定資料界線、寫明核實與標示、規定通報時限、指名負責人與檢視日期。附錄一、二、三分別放工具清單、事故聯絡人與培訓紀錄表，三頁正文之外另計。公司的行業與規模不同，答案會不同，五個部分不變。This policy has no paragraph saying the company values AI ethics, because such a sentence would not change a single decision. It does five jobs: names the tools, draws the data line, sets the checking and labelling rule, fixes a reporting deadline, and names the responsible role and the review date. Appendices 1, 2 and 3 hold the tool list, the incident contact and the training record form, and sit outside the three pages. A different industry and size will produce different answers; the five parts stay the same.
五個部分怎樣對回三個框架How the five parts map to the three frameworks
政策是三個框架共同假設存在的那份文件，五個部分都可以對回框架的具體章節。表三的左邊是政策，右邊三欄分別是私隱專員公署清單、模範框架與 ISO/IEC 42001 的管理系統元素，最後一欄是 NIST 的四個功能[1][2][4][5]。對應的作用是讓公司知道：寫政策同時是三個框架的起點。The policy is the document all three frameworks assume exists, and each of the five parts maps back to specific framework sections. Table 3 has the policy on the left and, on the right, the Privacy Commissioner's checklist, the Model Framework, the management system elements of ISO/IEC 42001, and the four NIST functions[1][2][4][5]. The point of the mapping is that writing the policy is the starting point of all three.
| 政策部分Policy part | 私隱專員公署清單PCPD checklist | 模範框架Model Framework | ISO/IEC 42001ISO/IEC 42001 | NIST AI RMFNIST AI RMF |
|---|---|---|---|---|
| 一、適用範圍與獲准工具1. Scope and approved tools | 獲准使用生成式 AI 的範圍[2]Approved scope of generative AI use[2] | 第一部 1.1 AI 策略、1.3 管治架構[1]Part I, 1.1 AI strategy and 1.3 governance structure[1] | 管理系統的政策、角色與權責[4]Management system policy, roles and responsibilities[4] | 治理（GOVERN）GOVERN |
| 二、個人資料與敏感資料2. Personal and sensitive data | 保障個人資料私隱[2]Personal data protection[2] | 第三部 3.1 為定製及使用 AI 準備數據[1]Part III, 3.1 preparing data for AI customisation and use[1] | 資料與風險控制[4]Data and risk controls[4] | 對應（MAP）MAP |
| 三、核實、標示與合法使用3. Checking, labelling and lawful use | 合法及合乎道德的使用與偏見預防[2]Lawful and ethical use and bias prevention[2] | 第二部 2.2 決定人為監督的程度[1]Part II, 2.2 determining the level of human oversight[1] | 運作控制與人為監督[4]Operational controls and human oversight[4] | 量度（MEASURE）MEASURE |
| 四、裝置、權限與事故報告4. Devices, permissions and incidents | 數據安全[2]Data security[2] | 第三部 3.3 AI 系統的管理與持續監察[1]Part III, 3.3 managing and continuously monitoring AI systems[1] | 事故處理與持續監察[4]Incident handling and ongoing monitoring[4] | 管理（MANAGE）MANAGE |
| 五、後果、負責人與檢視5. Consequences, responsible role and review | 違反政策或指引[2]Breach of the policy or guidance[2] | 第一部 1.3 管治架構、1.4 培訓及加強認識[1]Part I, 1.3 governance structure and 1.4 training and awareness[1] | 角色、能力、管理審查與改進[4]Roles, competence, management review and improvement[4] | 治理（GOVERN）GOVERN |
怎樣選，取決於公司要應付的是誰。客戶與監管機構主要關心個人資料，先做政策與清單，再把模範框架的四個範疇當作檢查表，是最省時的一條路[1][2]。企業客戶或投標要求供應商證明有管理系統，才值得投入 ISO/IEC 42001 的認證工程，因為它要的是持續紀錄與外部審核，一份文件只是起點[4]。公司想建立一套可重複的風險流程，NIST 的四個功能可以直接當作風險登記冊的結構[5]。三條路都可以由同一份三頁政策開始。Which route to take depends on who the company has to satisfy. Where customers and regulators care most about personal data, the fastest route is the policy and the checklist, with the Model Framework's four areas used as a checking list[1][2]. Where an enterprise customer or a tender requires proof of a management system, the investment in ISO/IEC 42001 certification is justified, because it asks for ongoing records and an external audit; one document is only the start[4]. Where the company wants a repeatable risk process, the four NIST functions can serve as the structure of a risk register[5]. All three routes can start from the same three-page policy.
政策之外，公司還要處理它管不到的部分。受監管行業的合規要求，由各自的監管機構決定；跨境資料轉移、知識產權與僱傭合約的責任，屬於法律與合約範圍；工具的技術控制（例如資料外洩防護、租戶設定與日誌）由資訊科技部門處理[3]。政策寫得清楚，是讓這三部分有人跟進的起點。Beyond the policy, a company has to handle what the policy cannot. Compliance requirements for regulated industries are set by each regulator; cross-border data transfer, intellectual property and employment contract liability sit with the law and the contracts; the technical controls on the tools, such as data loss prevention, tenant settings and logs, belong to IT[3]. A clear policy names who picks up each of the three areas.
我們的位置與政策的限制Where we stand, and the limits of a policy
在企業 AI 培訓的入門課程裏，內部 AI 使用指引是其中一項交付物。我們的做法是與客戶一起，用他們實際使用的工具、實際的資料流與實際的批核流程，寫成一份可以發給同事的文件；五個部分就是我們用的骨架。同一個行業、不同規模的公司，答案會不同：一間貿易公司最在意報價與客戶名單，一間專業事務所最在意客戶機密與文件版本，一間學校最在意學生資料與家長同意。An internal AI use guide is one of the deliverables of our corporate AI training. We write it with the client, using the tools they actually run, the data that actually moves and the approvals they actually use, so that it becomes a document they can send to colleagues; the five parts are the skeleton we use. The answers differ by industry and size: a trading company cares most about quotations and customer lists, a professional firm about client confidentiality and document versions, a school about student data and parental consent.
實際推行時，最常見的失敗是政策由一個部門寫，其他部門不知道。人力和資訊科技各自有一份規則，財務有自己的審批，前線同事收到三份文件，最後只記得一句「不要亂用」。三頁政策的價值在於它逼公司把五個決定放在同一頁上，由同一個人負責，並且每次改版都通知同一批同事。私隱專員公署的清單把培訓、支援隊伍與反饋機制列為支援措施，三項都寫進推行計劃，政策才有人跟[2]。In practice, the most common failure is a policy written by one department that the others never hear about. HR and IT each have their own rules, finance has its own approvals, and frontline colleagues end up with three documents and one memory: don't use it carelessly. The value of a three-page policy is that it forces the five decisions onto the same few pages, under one responsible person, with the same group told about every revision. The Privacy Commissioner's checklist lists training, a support team and a feedback channel as supporting measures; put all three into the rollout and the policy has someone following it[2].
政策也有它做不到的事。第一，它改變不了行為，除非同事受過訓練、遇過例子、知道找誰問。第二，它代替不了技術控制，工具本身的權限、租戶設定與日誌，要由資訊科技部門處理。第三，它不構成法律意見，涉及個人資料、僱傭、知識產權與合約責任的判斷，要由公司的法律顧問處理。這三件事講清楚，政策的承諾才不會超出它能交付的範圍。A policy also has limits. First, it does not change behaviour unless colleagues have been trained, have seen examples and know whom to ask. Second, it does not replace technical controls: the tool's permissions, tenant settings and logs are IT's work. Third, it is not legal advice: questions about personal data, employment, intellectual property and contract liability belong with the company's legal adviser. Naming those three limits keeps the policy's promise inside what it can deliver.
框架與指引會更新。私隱專員公署在 2026 年 8 月就代理式 AI 發表補充指引，說明模範框架的原則一般仍然適用，但代理式 AI 的自主程度帶來額外風險[6]。所以第五部分寫的是檢視周期：每六個月一次，或在相關指引更新時提前。政策是一份活的文件，這句話要寫在文件裏，才有下一次檢視。Frameworks and guidance change. In August 2026 the Privacy Commissioner published supplementary guidance on agentic AI, saying that the Model Framework's principles generally still apply while the greater autonomy of agentic AI brings extra risks[6]. Part five therefore sets a review cycle: every six months, or earlier when the relevant guidance changes. A policy is a living document, and that sentence has to be in the document for there to be a next review.
