香港企業不是沒有用 AI，而是把「有工具在用」當成「流程已經改變」。Hong Kong companies are not failing to use AI. They are treating "a tool is in use" as "the process has changed".這兩件事之間隔著四個層級，每升一級，誰負責、資料碰到甚麼、以及「正確」是甚麼意思都會改變。用上一級的方法去管下一級，通常會出事。 Four levels sit between those two statements, and each one changes who is accountable, what data is touched, and what "correct" even means. Managing one level with the habits of the level below it is how projects fail.
個人資料私隱專員公署（PCPD）在 2026 年 1 月展開第三輪合規檢查，5 月公布結果：60 間受查機構中，57 間（95%）在日常營運使用 AI，較 2025 年上升 15 個百分點；其中 29 間（51%）同時使用三個或以上 AI 系統[1]。使用已經發生，問題不只在「有沒有用」。In January 2026 the Office of the Privacy Commissioner for Personal Data (PCPD) began its third round of compliance checks and published the findings in May: of 60 organisations reviewed, 57 (95%) used AI in day-to-day operations, up 15 percentage points on 2025, and 29 of those 57 (51%) ran three or more AI systems[1]. Use has happened. The question is no longer only whether.
把眼光放遠一點，麥肯錫 2025 年的全球調查顯示 88% 的企業至少有一個職能經常使用 AI，但只有 39% 能在企業層面指出 EBIT 的影響[2]。RAND 訪談 65 位有五年以上經驗的資料科學家與工程師後指出，按部分估算，超過 80% 的 AI 項目失敗，失敗率是其他 IT 項目的兩倍[3]。Zoom out, and McKinsey's 2025 global survey found 88% of organisations using AI regularly in at least one function, but only 39% reporting any EBIT impact at the enterprise level[2]. RAND interviewed 65 data scientists and engineers with at least five years of experience, and noted that by some estimates more than 80% of AI projects fail, twice the failure rate of IT projects that do not involve AI[3].
這幾個數字不矛盾。它們描述同一個現象：使用 AI 很容易，改變一個流程很難。前者是工具問題，後者是層級問題。These figures do not contradict one another. They describe one phenomenon: using AI is easy, changing a process is hard. The first is a tool question. The second is a level question.
摘要Summary
- 採用不是主要瓶頸。私隱專員公署審查的 60 間機構中，95% 在日常營運使用 AI，51% 同時跑三個或以上系統[1]。缺的是由「有人用」走到「流程已改」。Adoption is not the main bottleneck. Of the 60 organisations reviewed by the PCPD, 95% use AI in daily operations and 51% run three or more systems[1]. What is missing is the step from use to a changed process.
- 五級由個人使用到代理執行。每升一級，權限、資料與責任都改變；不是同一件事做得更好。The five levels run from individual use to agentic execution. Each step changes permissions, data and accountability. It is not the same job done better.
- 大多數公司應該停在 L3 或 L4。L1 與 L2 要做好管治，L3 選一個流程做出可量度的結果，L4 需要知識庫與評估集，L5 需要權限設計與審計。Most companies should stop at L3 or L4. Get L1 and L2 governed, take one process to L3 with a measurable result, earn L4 with a knowledge base and an evaluation set, and treat L5 as an engineering and audit problem.
- 香港的難點不只一個：中小企佔 98%、採用已近普及但流程未改、廣東話與繁體中文的表現差距、PDPO 與行業監管、政策推力，以及同一個用例面對不同監管者。Hong Kong's constraints do not have a single cause: SMEs make up 98% of enterprises, adoption is near-universal but processes have not changed, Cantonese and Traditional Chinese still lag, PDPO sits alongside sector regulators, policy momentum pushes buying, and the same use case meets different regulators.
- 不要用「有幾多人用」量度成效。用一個流程的周期時間、錯誤率與單位成本。使用率只證明員工找到了工具。Do not measure adoption by headcount. Measure one process by cycle time, error rate and cost per transaction. Usage only proves that staff found a tool.
五個層級一覽The five levels at a glance
這不是成熟度評分，也不是要每間公司爬到最高級。分級的標準只有一個：當 AI 出錯時，誰要承擔後果、後果有多貴。從這個角度看，L1 與 L5 的距離不是「懂不懂得用」，而是「有沒有人授權它行動，以及有沒有人能證明它做過甚麼」。This is not a maturity score, and not every company should climb to the top. There is one test behind the tiers: when the AI is wrong, who carries the consequence, and how expensive is it. From that angle, the distance between L1 and L5 is not sophistication. It is whether someone has authorised the system to act, and whether anyone can prove what it did.
由上至下：自主程度與出錯後果遞增Top to bottom: autonomy and the cost of error increase
-
L1
個人使用Individual use
員工用 ChatGPT、Gemini 等處理文件、寫電郵、翻譯Staff use ChatGPT, Gemini and similar tools to draft, summarise and translate
-
L2
團隊工具Team tooling
M365 Copilot 或 Workspace 內建的助手，出現在電郵、文件、會議M365 Copilot or Workspace assistants inside email, documents and meetings
-
L3
流程自動化Process automation
發票核對、查詢分流、報告草稿等固定流程自動完成Invoices, enquiry triage and report drafts handled by a defined workflow
-
L4
企業知識系統Company knowledge
以內部文件回答問題，答案附來源段落Answers drawn from internal documents, with the source paragraph attached
-
L5
代理執行Agentic execution
代理自行退款、更新記錄、發出採購單Agents issue refunds, update records and raise purchase orders
圖一：五個層級的分界，不在技術難度，而在「誰授權它行動」與「出錯後能否復原」。Figure 1: the levels are divided not by technical difficulty but by who authorises the action and whether an error can be undone.
香港的五個結構性難點，加上一個政策因素Five structural constraints in Hong Kong, plus one policy factor
同一套技術，放在香港會遇到不同的阻力。這些不是「文化問題」，而是可以指出來的結構條件；它們解釋了為甚麼同一個層級，在香港比其他市場更難通過。The same technology meets different resistance in Hong Kong. These are not cultural generalities. They are structural conditions that explain why the same level is harder to clear here than in other markets.
1. 九成八的企業是中小企1. Ninety-eight per cent of enterprises are SMEs
2026 年 6 月，香港約有 36 萬家中小企，佔企業總數 98% 以上，聘用約 120 萬人[14]，相當於約 45% 的私營機構僱員[15]。這意味著最常見的情況是：決定買工具的人，也是唯一會設定權限、清理資料、寫提示詞的人。沒有資料團隊，沒有專責的風險管理，也沒有人有時間做評估集。這不是資源不足的修辭，而是直接影響 L3 與 L4 能否成立的條件。In June 2026 Hong Kong had about 360,000 SMEs, more than 98% of all business establishments, employing about 1.2 million people[14], roughly 45% of the private-sector workforce[15]. The commonest situation is therefore this: the person who decides to buy the tool is also the only person who will set permissions, clean the data and write the prompts. There is no data team, no dedicated risk function, and nobody with time to build an evaluation set. That is not a rhetorical point about resources. It is a direct constraint on whether L3 and L4 can exist at all.
2. 採用已近普及，但採用不等於改變流程2. Adoption is close to universal, but adoption is not process change
私隱專員公署的第三輪合規檢查（2026 年 5 月公布）發現，95% 的受查機構在日常營運使用 AI，51% 同時使用三個或以上系統[1]。另一項調查的樣本更接近一般企業：2025 年訪問 350 位香港在職人士，三分之一表示公司尚未部署任何 AI 用例[13]。兩個數字不矛盾，因為樣本不同：前者是被抽查的機構，後者包括仍在觀望的公司。全球數字則顯示，78% 的 AI 使用者自帶工具上班，中小企更達 80%[5]；52% 的人不願承認自己用 AI 處理重要工作[5]。香港的分別是，這裡的公司更小，所以一次資料外洩的相對代價更大。The PCPD's third round of compliance checks, published in May 2026, found 95% of the organisations reviewed using AI in daily operations, with 51% running three or more systems[1]. A different sample looks closer to the average company: a 2025 survey of 350 professionals in Hong Kong found that one third said their company had not deployed any AI use case[13]. The two figures do not contradict each other, because the samples differ: one is a set of reviewed organisations, the other includes companies still watching. Globally, 78% of AI users bring their own tools to work, rising to 80% at small and medium-sized companies[5], and 52% are reluctant to admit using AI for their most important tasks[5]. The difference in Hong Kong is that the companies are smaller, so the relative cost of one leak is larger.
3. 廣東話與繁體中文，沒有一個現成的基準3. Cantonese and Traditional Chinese have no ready-made benchmark
香港的工作語言是廣東話，文件是繁體中文與英文混合，還夾雜大量本地專有名詞。2025 年在 NAACL 發表的評估指出，廣東話大型語言模型的表現仍然落後於普通話與英文[16]。這不是說模型不能用，而是說外國基準的分數不能直接當作你的分數。你要自己寫一組問題：真實的客戶電郵、真實的內部通告、真實的投訴記錄。沒有這一組問題，L4 的「準確」無法驗證。The working language in Hong Kong is Cantonese, the documents are a mixture of Traditional Chinese and English, and both are full of local terminology. A benchmark published at NAACL in 2025 found that Cantonese large language models still lag behind Mandarin and English[16]. That does not mean the models are unusable. It means a score from an overseas benchmark is not your score. You need your own question set: real client emails, real internal notices, real complaint records. Without that set, "accurate" at L4 cannot be verified.
4. 沒有一條統一的 AI 法例，但要求沒有比較少4. There is no single AI statute, but there is no shortage of requirements
香港沒有單一的 AI 法例，義務來自《個人資料（私隱）條例》的六項保障資料原則[7]、行業監管與政府指引。PCPD 在 2024 年 6 月發表的模型框架要求四件事：制定 AI 策略與治理、進行風險評估與人工監督、管理模型自訂與系統實作、與持份者溝通[6]。第 33 條（跨境轉移）至今仍未生效，但公署的指引與 DPP1、DPP3、DPP4 仍然適用[8]；涉及內地個人資料，另有粵港澳大灣區標準合同[9]。2026 年 3 月，政府在立法會表示已成立跨部門工作組檢視現有法例，但尚未決定個別問題是否以立法處理[20]。對企業而言，這代表兩件事：第一，沒有「等法例出台才做」這個選項；第二，你的合規文件要能同時應付私隱專員與行業監管者。Hong Kong has no single AI statute. The obligations come from the six Data Protection Principles in the Personal Data (Privacy) Ordinance[7], from sector regulators and from government guidance. The PCPD's model framework, published in June 2024, asks for four things: an AI strategy and governance, risk assessment and human oversight, management of model customisation and system implementation, and engagement with stakeholders[6]. Section 33, on cross-border transfer, is still not in force, but the Commissioner's guidance and DPP1, DPP3 and DPP4 still apply[8]. Where personal information from the mainland is involved, the GBA Standard Contract is the route[9]. In March 2026 the Government told the Legislative Council that an inter-departmental working group is reviewing existing legislation, but that no decision has been made on whether individual issues should be handled by legislation[20]. For a company, that means two things: waiting for a statute is not a plan, and your compliance file has to satisfy both the Privacy Commissioner and your sector regulator.
5. 同一個用例，不同監管者5. The same use case, a different regulator
金管局的 GenAI Sandbox 在 2024 年 8 月啟動，首批 15 個用例來自 10 間銀行與 4 間技術夥伴[11]。證監會 2024 年 11 月的通函要求高級管理層對 AI 語言模型的整個生命週期負責，並把投資建議與投資研究列為高風險用例[10]。同一套客戶查詢助手，放在銀行、券商與貿易公司，需要的人工監督與文件完全不同。企業要先回答「誰是我們的監管者」，而不是「哪個模型最好」。The HKMA's GenAI Sandbox launched in August 2024, and its first cohort of 15 use cases came from 10 banks and four technology partners[11]. An SFC circular in November 2024 makes senior management responsible across the model lifecycle and treats investment recommendations and investment research as high-risk use cases[10]. The same client-enquiry assistant, placed in a bank, a broker and a trading company, needs completely different human oversight and documentation. The first question is not which model is best. It is who regulates us.
6. 政策推力會製造「不買就落後」的壓力6. Policy momentum creates pressure to buy before thinking
2025 年施政報告把 AI 定為香港的核心產業，提出「強化基建、應用導向」，並在 2026 年設立 10 億元的香港人工智能研發院[12]。這些投資解決的是供給側問題，不會自動解決你的流程問題。壓力之下，企業容易由「問題是甚麼」跳去「買哪一個」。Gartner 在 2025 年指出，超過四成的代理式 AI 項目會在 2027 年底前取消，原因包括成本上升、價值不清與風險控制不足；市場上數以千計的代理供應商，估計只有約 130 間具備真正的代理能力[18]。The 2025 Policy Address made AI a core industry, set out a strategy of strengthening infrastructure and promoting application, and earmarked HK$1 billion for a Hong Kong AI Research and Development Institute in 2026[12]. Those investments address the supply side. They do not fix your process. Under that pressure, companies jump from "what is the problem" to "which one do we buy". Gartner said in 2025 that more than 40% of agentic AI projects will be cancelled by the end of 2027, citing escalating costs, unclear value and inadequate risk controls, and estimated that only about 130 of the thousands of agentic AI vendors are real[18].
這六點合起來，令香港的 AI 導入呈現一個形狀：L1 與 L2 來得很快，L3 之後很慢。不是因為員工不願意學，而是因為 L3 之後要求的是流程、資料與權限的整理，而這些工作沒有捷徑，也不能外判給模型。Together, these six points give Hong Kong's AI adoption its shape: L1 and L2 arrive quickly, and everything after L3 is slow. It is not because staff are unwilling to learn. After L3, the work is the unglamorous sorting of processes, data and permissions, and no model can do that for you.
逐級拆解：每一級實際是甚麼，會卡在哪裡Level by level: what each one is, and where it stalls
門檻不是建議，而是最低要求；未通過就升上去，問題只會在下一個層級以更貴的方式出現。The gate is not a suggestion. It is the minimum. Skip it and the same problem reappears one level up, at a higher price.
L1 個人使用：問題不是工具，是邊界L1 Individual use: the problem is the perimeter, not the tool
實際樣子。員工用個人帳戶的 ChatGPT、Gemini 或 DeepSeek，把會議記錄、報價單、客戶投訴貼進去，請 AI 寫回覆、摘要、翻譯。這不是偷偷摸摸，而是最理性的做法：工具免費、即時、好用，而公司沒有提供替代品。What it looks like. An employee opens ChatGPT, Gemini or DeepSeek on a personal account, pastes in meeting notes, a quotation or a customer complaint, and asks for a reply, a summary or a translation. This is not subterfuge. It is the most rational option available: the tool is free, immediate and useful, and the company has not offered an alternative.
為何會卡住。這是邊界問題。PDPO 的 DPP4（保障資料第 4 原則：保安措施）要求資料使用者採取切實可行的保安措施[7]；把客戶資料貼進消費者版本的服務，等於把控制權交出去，而公司通常連「貼過甚麼」都不知道。第二個問題是核實：證監會的通函直接指出，AI 語言模型會自信地給出錯誤答案，使用者傾向不加批判地接受[10]。Why it stalls. This is a perimeter problem. DPP4, the security principle, requires a data user to take all practicable steps to protect personal data[7]. Pasting client data into a consumer service hands over control, and the company usually cannot even say what was pasted. The second problem is verification: the SFC circular states plainly that AI language models can produce confident but wrong answers, and that users tend to accept the output without critical evaluation[10].
香港的情況。私隱專員公署的檢查發現，24 間經 AI 收集或使用個人資料的機構，全部容許員工在工作中使用生成式 AI；其中 71% 已制定內部政策，另外 21% 計劃制定[1]。換句話說，批准往往先於政策。在中小企，批准一個工具可能只需要老闆一句話，但沒有人負責設定資料保留、權限與離職後的存取。結果是「有批准，但沒有控制」。The Hong Kong version. The PCPD found that all 24 organisations collecting or using personal data through AI allowed employees to use generative AI at work; 71% had an internal policy, and a further 21% planned one[1]. Permission, in other words, tends to arrive before policy. In an SME, approving a tool can take one sentence from the boss, but nobody owns retention settings, permissions or access after someone leaves. The result is approval without control.
L2 團隊工具：買了授權，不等於改了流程L2 Team tooling: a licence is not a changed process
實際樣子。公司買了 Microsoft 365 Copilot 或 Google Workspace 的 Gemini，助手出現在 Outlook、Word、Teams 與 Docs。會議後自動整理行動項，在電郵串中找回上個月的承諾，把一份 30 頁的標書摘要成半頁。What it looks like. The company buys Microsoft 365 Copilot or Gemini in Google Workspace, and the assistant appears inside Outlook, Word, Teams and Docs. It turns a meeting into action items, finds last month's promise in an email thread, and reduces a 30-page tender to half a page.
為何會卡住。授權買了，流程沒有改。如果員工仍然手動整理一次會議記錄，節省不會出現；如果沒有人重新設計會議的輸入與輸出，Copilot 只是多了一個按鈕。第二個問題是權限：助手會找到使用者本來有權限的資料。如果共用資料夾的權限從未整理，它會把不該出現的文件帶到眼前。這不是模型幻覺，而是權限設計的鏡子。Why it stalls. The licences are bought and the workflow is unchanged. If staff still rewrite the minutes by hand, nothing is saved; if nobody redesigns what goes into and comes out of a meeting, Copilot is one more button. The second problem is permissions. The assistant surfaces what the user already has access to. If shared drives have never been tidied, it will bring documents into view that should not be there. That is not a hallucination. It is a mirror held up to the permission design.
香港的情況。20 至 200 人的公司，通常由一位 IT 同事兼顧系統、設備與供應商，沒有人負責知識管理或使用者體驗。雙語文件令提示詞與輸出更不穩定：同一份文件有中文版、英文版與一個「最後更新」不明的版本。採購按人頭計，部門為了「有」而買，不是為了「用」而買。The Hong Kong version. In a company of 20 to 200 people, one IT colleague usually covers systems, devices and vendors. Nobody owns knowledge management or user experience. Bilingual documents make prompts and outputs less stable: the same document exists in Chinese, in English, and in a third version whose last update is unknown. Licensing is bought per head, and departments buy to have it rather than to use it.
L3 流程自動化：多數工作在例外，不在模型L3 Process automation: most of the work is in the exceptions
實際樣子。收到發票後自動讀取、對採購單、標示差異；客戶查詢自動分流到售前、售後或投訴；每週報告自動起草，由人修改後發出。這一步開始，AI 不再只是「幫我寫」，而是「幫我完成一件事」。What it looks like. An invoice is read, matched against a purchase order and flagged where it differs. A customer enquiry is routed to sales, support or complaints. A weekly report is drafted automatically and edited by a person before it goes out. From this point, AI is not helping someone write. It is completing a task.
為何會卡住。多數失敗不在模型，在例外。真實的香港辦公室有手寫表格、傳真、掃描 PDF、WhatsApp 語音，以及「呢張單係上個月嘅」。如果一個流程有 15% 的例外，自動化其餘 85% 仍然有價值，但前提是你知道例外去哪裡、由誰接手、多久要處理完。第二個問題是整合：舊 ERP 沒有 API，或者只有原廠可以改。RAND 把「基建不足」列為 AI 項目失敗的五大原因之一[3]。Why it stalls. Most failures here are not model failures. They are exceptions. A real Hong Kong office has handwritten forms, faxes, scanned PDFs, WhatsApp voice notes and invoices dated last month. If 15% of a process is exceptional, automating the other 85% is still worth doing, but only if you know where the exception goes, who picks it up and how quickly it has to be cleared. The second problem is integration: the ERP has no API, or only the original vendor can change it. RAND lists inadequate infrastructure as one of the five leading causes of AI project failure[3].
香港的情況。流程往往沒有文件，知道流程的人就是瓶頸，而這個人通常也是最有價值、最沒有時間的人。供應商按「接入系統」報價，但實際工作量的八成在清理資料與處理例外，這部分很少寫在報價單上。The Hong Kong version. The process is often undocumented. The person who knows it is the bottleneck, and is usually the busiest and most valuable person in the room. Vendors quote for "system integration", but about 80% of the actual effort goes into cleaning data and handling exceptions, which rarely appears on the quotation.
L4 企業知識系統：文件不是知識庫L4 Company knowledge: documents are not a knowledge base
實際樣子。內部助手回答「年假怎樣計」、「這份合約的賠償上限是多少」、「上次投標的技術方案在哪裡」，答案附上文件名稱與段落。使用者不再需要知道文件放在哪個資料夾，但需要知道答案是否仍然有效。What it looks like. An internal assistant answers "how is annual leave calculated", "what is the liability cap in this contract" and "where is the technical proposal from the last tender", with the document and paragraph attached. The user no longer needs to know which folder holds the answer, but does need to know whether the answer is still valid.
為何會卡住。知識庫不是文件夾。文件有版本、有草稿、有互相矛盾的政策；如果檢索找到 2022 年的版本，答案會很流暢，但錯誤。第二個問題是權限：HR 政策可以全公司看，員工手冊的附錄可能不行。RAG（檢索增強生成）系統必須在檢索時套用權限，而不是在介面上把連結隱藏；後者只是把門關上，資料仍然在裡面。Why it stalls. A knowledge base is not a folder. Documents have versions, drafts and policies that contradict each other. If retrieval finds the 2022 version, the answer will be fluent and wrong. The second problem is permission. An HR policy may be open to everyone; an appendix to the staff handbook may not be. A retrieval-augmented generation (RAG) system has to apply permissions at retrieval time, not hide a link in the interface. Hiding the link closes the door and leaves the data inside.
香港的情況。繁體中文與英文混合，加上本地專有名詞（強積金、ORSO 職業退休計劃、差餉、地契、《公司條例》），令通用模型的語義匹配不穩定；廣東話口語查詢又是另一層。NAACL 的評估顯示廣東話表現仍有差距[16]，所以評估集必須由本地使用者寫，不能由供應商提供。The Hong Kong version. Traditional Chinese mixed with English, plus local terms such as MPF, ORSO, rates, Crown leases and the Companies Ordinance, makes semantic matching unstable in general-purpose models. Spoken Cantonese queries add another layer. The NAACL benchmark found a persistent gap in Cantonese[16], which is why the evaluation set has to be written by local users, not supplied by the vendor.
L5 代理執行：由「說錯話」變成「做錯事」L5 Agentic execution: from saying the wrong thing to doing the wrong thing
實際樣子。代理不單回答，還會行動：回覆客戶並退款、更新 CRM、發出採購單、安排面試時間。它需要的不只是模型，還有身份、權限、工具與一個可以回復原狀的方法。What it looks like. The agent does not only answer. It acts: replies to a customer and issues a refund, updates the CRM, raises a purchase order, schedules an interview. It needs more than a model: it needs an identity, permissions, tools and a way to undo what it did.
為何會卡住。風險的性質改變了。OWASP 2025 年把「過度代理權限」（LLM06）與提示注入（LLM01）列為大型語言模型應用的主要風險[17]。一個擁有退款權限的代理，會被一封惡意電郵指示；一個能改記錄的代理，出錯之後難以回復。證監會要求高級管理層在整個模型生命週期負起責任[10]；金管局則以沙盒方式，讓銀行在受控環境測試用例[11]。這些要求不是文書工作，而是要求你在出事時說得出：誰批准、權限是甚麼、日誌在哪。Why it stalls. The nature of the risk changes. The 2025 OWASP list puts excessive agency (LLM06) and prompt injection (LLM01) among the top risks for LLM applications[17]. An agent with refund permission can be instructed by a malicious email; an agent that can edit records is hard to roll back. The SFC makes senior management accountable across the model lifecycle[10], and the HKMA runs a sandbox so banks can test use cases in a controlled environment[11]. These are not paperwork requirements. They require you to answer, after an incident: who approved this, what were the permissions, and where are the logs.
香港的情況。多數中小企沒有 24 小時的保安團隊，代理由供應商託管，權限往往一次過開到最大，因為這樣「先試得到」。這個做法在試點沒有問題，在生產環境就是把公司的操作權交給一個會被人說服的系統。The Hong Kong version. Most SMEs have no 24-hour security team, the agent is hosted by a vendor, and permissions are often opened to the maximum because that is what makes the pilot work. In a pilot that is acceptable. In production it hands the company's operating authority to a system that can be talked into things.
五個重複出現的失敗，跨越所有層級Five failures that repeat at every level
2025 年一項訪問 350 位香港在職人士的調查，把減慢 AI 規模化的原因排了序：資料質素與可得性 58%、成本與回報 35%、私隱與安全 29%、資源與技能 25%[13]。這些是症狀。把它們放回層級，會看到五個更根本的失敗模式，其中四個直接對應 RAND 的訪談結果[3]。A 2025 survey of 350 professionals in Hong Kong ranked what slows AI at scale: data quality and availability 58%, cost and return 35%, privacy and security 29%, and resources and skills 25%[13]. Those are symptoms. Put them back against the levels and five deeper failure patterns appear, four of which map directly onto RAND's findings[3].
| 失敗模式Pattern | 你聽到的說法What you hear | 真正的原因What is actually happening | 一個測試One test |
|---|---|---|---|
| 問題定義錯Wrong problem | 「AI 好似唔係好幫到手」"AI doesn't seem to help much" | 用示範檔案設計，用真實例外驗收Designed on demo files, judged on real exceptions | 用 100 個真實案例重跑Re-run on 100 real cases |
| 資料與權限Data and permissions | 「AI 唔準」"The AI is inaccurate" | 文件過時、版本混亂、權限從未整理Stale documents, version chaos, permissions never tidied | 抽 20 份文件，問誰保證它最新Pull 20 documents and ask who guarantees they are current |
| 技術先行Technology first | 「人哋有，我哋都要有」"Everyone has one, so should we" | 政策壓力與供應商推銷代替了問題Policy pressure and vendor selling replaced the problem | 先寫一個不用 AI 的方案，比較成本與風險Write a no-AI option first and compare cost and risk |
| 基建與整合Infrastructure | 「試點成功，但上不到生產」"The pilot worked, but it never shipped" | 舊系統沒有 API，審批與權限要重做No API, and approvals and permissions have to be rebuilt | 問 IT：上線要開幾個權限、誰批Ask IT how many permissions production needs, and who signs |
| 期望與量度Expectation and measurement | 「無法證明回報」"We can't prove the return" | 沒有基線，只量使用率；員工看不到對自己有何好處No baseline, only usage is measured; staff cannot see what changes for them | 今天先量一次周期時間與錯誤率Measure cycle time and error rate today |
失敗率的分歧值得說明。MIT 的 NANDA 在 2025 年的報告指出，95% 的企業從生成式 AI 得到零回報，只有 5% 的試點能產生可觀價值[4]；麥肯錫的調查則有 39% 受訪者表示見到企業層面的 EBIT 影響[2]。兩個數字差距這麼大，因為問的問題不同、樣本不同、定義也不同。NANDA 看的是「有沒有可量度的損益影響」，樣本是 300 多個部署與 52 間機構的訪談[4]；麥肯錫問的是受訪者自己的判斷[2]。The disagreement over failure rates is worth explaining. MIT's NANDA report in 2025 found that 95% of organisations were getting zero return from generative AI, with only 5% of pilots extracting material value[4]. McKinsey's survey found 39% of respondents reporting EBIT impact at the enterprise level[2]. The gap is wide because the questions, samples and definitions differ. NANDA looked for measurable profit-and-loss impact across more than 300 deployments and 52 organisational interviews[4]; McKinsey asked respondents for their own assessment[2].
兩個數字指向同一件事：瓶頸在規模化，不在使用權。對企業的意義不是「95% 會失敗」，而是「試點成功，不能預測生產環境的成敗」。試點通常只有三份文件、一個熱心的同事、沒有例外；生產環境有幾千份文件、五個部門、以及一個不願意改變的供應商。Both figures point to the same thing: the bottleneck is scaling, not access. The lesson for a company is not "95% will fail". It is that a successful pilot does not predict a successful production system. A pilot has three documents, one enthusiastic colleague and no exceptions. Production has a few thousand documents, five departments and a vendor who does not want to change anything.
所以在香港，順序比選型重要。先決定你要解決的是哪一個層級的問題，再決定用哪一個模型。下一節把合規要求放進同一個順序裡，因為在香港，合規不是最後一步的檢查，而是決定一個層級能否成立的前提。In Hong Kong, sequence matters more than product choice. Decide which level you are solving for, then choose the model. The next section places the compliance questions in that same sequence, because in Hong Kong compliance is not a final check. It is a precondition for whether a level can exist at all.
合規地圖：四個問題，決定你要準備甚麼The compliance map: four questions that decide what you need
香港的合規不是一個勾選框，而是一條路徑。四個問題按順序問，答案會決定你要準備的文件與控制措施。同一個 AI 功能，答案不同，要求可以相差很遠。政府亦發布了《香港生成式人工智能技術及應用指引》，把治理、風險評估、資料保護與安全寫成可操作的做法[19]。Compliance in Hong Kong is not a checkbox. It is a path. Ask these four questions in order, and the answers decide which documents and controls you need. The same AI feature can attract very different requirements depending on the answers. The Government has also published the Hong Kong Generative AI Technical and Application Guideline, which turns governance, risk assessment, data protection and security into operational steps[19].
- 1涉及個人資料？Personal data?
PDPO 六項保障資料原則；PCPD 模型框架的四個範圍PDPO's six Data Protection Principles; the PCPD framework's four areas
- 2資料會跨境？Does data cross a border?
第 33 條未生效，但 DPP1、DPP3、DPP4 仍適用；內地資料走 GBA 標準合同Section 33 is not in force; DPP1, DPP3 and DPP4 still apply; mainland data uses the GBA Standard Contract
- 3受監管行業？A regulated sector?
金管局與證監會的額外要求；高風險用例要更高程度的人工監督Extra HKMA and SFC expectations; high-risk cases need more oversight
- 4會自行採取行動？Will it take action?
最小權限、人工批准門檻、不可修改日誌、停用開關Least privilege, approval thresholds, immutable logs, a kill switch
四個答案，加上 AI 清單、風險評估、供應商條款與事故程序，就是你的合規檔案。Four answers, plus an AI inventory, a risk assessment, vendor terms and an incident procedure, are your compliance file.
圖二：合規路徑。問題的順序比答案重要，因為第一個「是」已經決定了後面要準備的文件。Figure 2: the compliance path. The order of the questions matters more than the answers, because the first "yes" already decides which documents are needed.
| 要留下的文件Keep on file | 內容What goes in it | 為甚麼需要Why it matters |
|---|---|---|
| AI 清單AI inventory | 每個工具或系統的用途、擁有者、資料類別、供應商、所在層級Purpose, owner, data category, vendor and level for each tool or system | 沒有清單，出事時找不到範圍Without it, you cannot scope an incident |
| 風險評估Risk assessment | 個人資料影響、錯誤後果、受影響人數、人工監督程度Personal-data impact, cost of error, number affected, level of human oversight | PCPD 框架要求以風險為本[6]The PCPD framework asks for a risk-based approach[6] |
| 供應商條款Vendor terms | 資料用途、是否用於訓練、儲存地、再轉移、刪除、審計、事故通知Use, training, location, onward transfer, deletion, audit, breach notice | DPP4 與跨境要求不會因外包而消失[8]DPP4 and cross-border duties do not disappear with outsourcing[8] |
| 評估紀錄Evaluation record | 100 條真實問題、通過標準、上次測試日期、誰簽署One hundred real questions, pass criteria, last test date, who signed | 證明「準確」不是感覺，而是可重複的測試It turns accuracy from an impression into a repeatable test |
| 人工監督Human oversight | 誰覆核、多久一次、例外如何升級、誰可以停用Who reviews, how often, how exceptions escalate, who can switch it off | L4 與 L5 的核心控制，也是監管者最先問的問題The core control at L4 and L5, and the first question a regulator asks |
| 事故程序Incident procedure | 誰負責、多久通報、如何通知受影響人士、如何回復Who leads, how fast to notify, how to contact affected people, how to recover | 出事時才寫，通常已經太遲Written during an incident, it is written too late |
如果你需要一個現成的管理系統模板，ISO/IEC 42001 是可認證的 AI 管理系統標準[21]，NIST 的 AI 風險管理框架提供另一套常用語言[22]。兩者都不是香港法例，但可以讓內部文件有一致的結構，也方便供應商對照。If you need a ready-made management-system template, ISO/IEC 42001 is a certifiable AI management system standard[21], and NIST's AI Risk Management Framework offers another common vocabulary[22]. Neither is Hong Kong law, but both give internal documents a consistent structure and make vendor comparisons easier.
三個常見的誤解Three common misunderstandings
「模型在海外，所以 PDPO 不適用。」PDPO 管的是資料使用者對個人資料的控制，不是伺服器放在哪裡。你把資料交給誰、為甚麼、保留多久，仍然是你的事[7]。"The model is overseas, so PDPO does not apply." PDPO governs the data user's control over personal data, not the location of the server. Who you give the data to, why, and for how long remains your responsibility[7].
「我們沒有出售資料，所以沒有問題。」DPP1 關乎收集目的與相稱性，DPP3 關乎使用限制。把客戶資料貼進一個用於訓練的免費服務，即使沒有金錢交易，也可能是超出原本目的的使用。"We do not sell data, so there is no issue." DPP1 is about purpose and necessity at collection; DPP3 is about limiting use. Pasting client data into a free service that trains on it can be a use beyond the original purpose, even with no money changing hands.
「客戶同意了就可以。」同意必須是自願、具體、知情的。在僱傭關係裡，員工的同意尤其難成立，因為拒絕的成本太高。較穩妥的做法，是減少收集、限制用途、做好保安，而不是依賴一張同意書。"The client consented, so it is fine." Consent has to be voluntary, specific and informed. In an employment relationship it is especially hard to establish, because the cost of refusing is high. The safer route is to collect less, limit the use and secure what you hold, rather than rely on a consent form.
成本：三種成本，只有一種會出現在報價單上Cost: three kinds, and only one appears on the quotation
企業通常只比較授權費。真正的成本分三層：固定、變動、隱藏。香港的特別之處是量小、人工貴，所以隱藏成本佔的比例更高：一個錯誤由人手補救的時間，往往比模型的費用貴。Companies usually compare licence fees. The real cost has three layers: fixed, variable and hidden. Hong Kong's particular shape is small volumes and expensive labour, so the hidden layer is proportionally larger. The staff time spent correcting an error usually costs more than the model that made it.
| 成本類別Layer | 包括甚麼What it covers | 常見低估What gets underestimated |
|---|---|---|
| 固定成本Fixed | 整合、評估集、權限整理、訓練、流程重寫Integration, evaluation set, permission clean-up, training, process redesign | 資料清理與權限整理的時間，通常等於整合的兩倍Data and permission clean-up usually takes twice as long as integration |
| 變動成本Variable | 授權、token、儲存、支援Licences, tokens, storage, support | 用量增長、模型升級、匯率與價格調整Volume growth, model upgrades, exchange rates and price changes |
| 隱藏成本Hidden | 覆核時間、錯誤重做、維護、供應商管理Review time, rework, maintenance, vendor management | 假設 AI 出錯不需要人手補救Assuming an AI error needs no human repair |
用一個假設例子說明比例。一間 150 人的公司做 L2，並把一個文件流程推到 L3：40 個授權，假設每個每月 HK$300，一年是 HK$144,000；流程每月處理 8,000 份文件，每份約 1,500 個輸入 token 與 300 個輸出 token，一年約 1.73 億個 token。這兩個數字都容易計。難計的是第三格：如果每 1,000 份文件有 5 份要人手覆核，一年就是 480 份，每份 10 分鐘，等於 80 小時。如果每份要 30 分鐘，就是 240 小時。A hypothetical example shows the proportions. A company of 150 people runs L2 and takes one document process to L3: 40 licences at an assumed HK$300 a month is HK$144,000 a year. The process handles 8,000 documents a month at roughly 1,500 input tokens and 300 output tokens each, which is about 173 million tokens a year. Those two figures are easy to calculate. The third is not: if five documents in every 1,000 need manual review, that is 480 reviews a year, or 80 hours at ten minutes each. At 30 minutes each, it is 240 hours.
以上是假設，不是報價。重點不是總數，而是第三格：多數香港項目的預算，是在隱藏成本爆掉，而不是在授權費。簽約前先問：出錯時，誰用幾多時間補救？These are assumptions, not a quotation. The point is not the total but the third layer. Most Hong Kong projects blow their budget on hidden cost, not on licences. Before signing, ask who spends how many hours repairing an error.
資助方面，香港企業最常問的是「有沒有錢」。持續進修基金為 18 歲或以上的合資格人士提供最多 HK$25,000 的學費發還，課程必須在可獲發還款項名單內，出席率至少七成，評核至少五成[23]。科技券（TVP）自 2024 年 12 月 31 日之後停止接受新申請[24]，所以用 TVP 假設去編預算的計劃，需要重新計算。資助條款會變，申請前應以官方最新公布為準。On funding, the question Hong Kong companies ask most is whether money is available. The Continuing Education Fund reimburses up to HK$25,000 for eligible adults aged 18 or over, provided the course is on the reimbursable list, attendance is at least 70% and the assessment mark is at least 50%[23]. The Technology Voucher Programme stopped accepting new applications after 31 December 2024[24], so any budget built on TVP has to be recalculated. Funding terms change; check the current official position before applying.
90 天與 12 個月：一個可執行的次序Ninety days and twelve months: a sequence you can run
90 天的目的不是完成轉型，而是取得一個可比較的基線，以及一個有依據的決定。沒有基線的試點，最後通常只會變成一場示範。The point of ninety days is not to finish a transformation. It is to get a comparable baseline and a decision with evidence behind it. A pilot with no baseline usually ends as a demonstration.
| 階段Stage | 時間When | 做甚麼What you do | 完成標準Done when |
|---|---|---|---|
| 0 清單0 Inventory | 第 1 至 2 週Weeks 1–2 | 誰在用甚麼、碰到甚麼資料、由誰付費Who uses what, what data it touches, who pays | 一張表：工具、人數、資料類別、層級One table: tool, headcount, data category, level |
| 1 管治1 Governance | 第 3 至 4 週Weeks 3–4 | 一頁政策、核准清單、60 分鐘訓練One-page policy, approved list, 60-minute session | 八成員工說得出哪些資料不可輸入80% of staff can name the data that must never be entered |
| 2 基線2 Baseline | 第 5 至 6 週Weeks 5–6 | 選一個流程，量一次現狀Choose one process and measure it as it is | 周期時間、錯誤率、單位成本各一個數字One number each for cycle time, error rate and unit cost |
| 3 試點3 Pilot | 第 7 至 10 週Weeks 7–10 | 用 100 個真實案例跑，不用示範檔案Run 100 real cases, not demo files | 達到預先設定的錯誤預算The pre-agreed error budget is met |
| 4 決定4 Decision | 第 11 至 12 週Weeks 11–12 | 擴大、修正或停止Scale, fix or stop | 書面決定，寫明負責人與下一個門檻A written decision naming the owner and the next gate |
之後的 12 個月，最多做三個流程。每多一個流程，就多一組例外、一個供應商與一條權限路徑。到 L4 之前，先有一組 100 條的評估問題；到 L5 之前，先通過上一節的八項門檻。這個次序聽起來慢，但它比「先買再算」快，因為它避免了兩次重做。For the following twelve months, take on no more than three processes. Each additional process brings another set of exceptions, another vendor and another permission path. Before L4, have the 100-question evaluation set. Before L5, clear the eight gates in the previous section. The sequence sounds slow. It is faster than buy-first-and-see, because it avoids doing the work twice.
八條問題，判斷一個方案值不值得做Eight questions that decide whether a proposal is worth doing
你不需要成為技術專家，也可以問出關鍵問題。任何一條答不出來，都是一個風險，而不是一個細節。You do not need to be a technical expert to ask the questions that matter. Any one of these eight without an answer is a risk, not a detail.
| 問題Question | 為甚麼要問Why it matters |
|---|---|
| 這個系統哪一部分是模型，哪一部分是規則？Which part is the model, and which part is rules? | 分不清，通常代表你買的是包裝。Gartner 估計真正具備代理能力的供應商只有約 130 間[18]。If the vendor cannot separate them, you are usually buying packaging. Gartner estimates only about 130 vendors are genuinely agentic[18]. |
| 我們的資料會用來訓練嗎？可以關閉嗎？Will our data train the model, and can that be switched off? | 這是 DPP3 使用限制的問題，不是合約細節。This is a DPP3 use-limitation question, not a contract detail. |
| 資料存放在哪個司法管轄區？誰可以存取？Where is the data stored, and who can access it? | 跨境與 DPP4 保安要求，會決定你要不要做額外評估[8]。Cross-border location and DPP4 security decide whether extra assessment is needed[8]. |
| 權限做到文件級，還是資料夾級？Are permissions per document or per folder? | L4 的成敗在這一行。資料夾級權限，通常意味著答案會越權。This one line decides L4. Folder-level permissions usually mean answers reach too far. |
| 你的評估集是甚麼？有沒有廣東話與繁體中文案例？What is your evaluation set, and does it include Cantonese and Traditional Chinese? | 廣東話表現仍有差距[16]，通用基準不能代替你的案例。Cantonese still lags[16]; a general benchmark is not your case set. |
| 模型更新時，怎樣通知與回歸測試？When the model changes, how are we notified and retested? | 輸出會隨版本漂移。沒有回歸測試，今天的準確不等於下個月的準確。Output drifts between versions. Without regression testing, accurate today does not mean accurate next month. |
| 如果我們停止使用，資料與設定可以帶走嗎？If we leave, can we take the data and configuration with us? | 退出成本是風險的一部分，也是供應商集中度的問題[10]。Exit cost is part of the risk, and part of the concentration problem[10]. |
| 現在有多少客戶在生產環境用這個功能？How many clients run this function in production today? | 試點與生產是兩個產品。要求可聯絡的參考客戶。A pilot and a production system are two different products. Ask for a reference you can call. |
採購上，有一個簡單的做法：要求一個兩星期的付費試點，用你的 100 個真實案例，而不是用供應商的示範檔案。試點要寫明通過標準、錯誤預算與退出條件。願意接受這個條件的供應商，通常也願意在出事時負責。In procurement, there is one simple move: ask for a two-week paid pilot using your 100 real cases, not the vendor's demo files. Write the pass criteria, the error budget and the exit conditions into the pilot. A vendor willing to accept those terms is usually a vendor willing to be accountable when something goes wrong.
我們怎樣教這一套How we teach this
我們的企業培訓《生成式人工智能與商業自動化》用客戶自己的流程做教材。第一天結束時，學員手上有一個可以用的工具，而不是一本筆記。這樣做的好處不是「有成品」，而是學員在同一個下午就遇到例外、權限與資料質素的問題，而這些問題在課堂上可以問。Our corporate programme, Generative AI and Business Automation, uses the client's own processes as the teaching material. By the end of the first day, participants have a working tool rather than a notebook. The benefit is not the artefact. It is that they meet the exceptions, permissions and data-quality problems in the same afternoon, while there is still someone to ask.
課程的重點不是模型清單，而是三個動作：把流程寫出來、把資料整理到可用、把權限與人工覆核設好。這三件事，正好對應 L3 與 L4。課程亦會說明哪一級不值得做：如果一個流程每月只發生十次，自動化的回報通常不足以支持維護成本。The programme is not a tour of models. It is three actions: write the process down, get the data to a usable state, and set the permissions and human review. Those three map directly onto L3 and L4. The programme also covers the levels not worth doing: if a process happens ten times a month, automation rarely earns enough to cover its maintenance.
如果貴公司正在決定由哪一級開始，可以先看企業 AI 方案的課程內容與安排；如果只是想核對本文的判斷，也歡迎把你們的流程寫給我們，我們會指出哪一格最可能出問題。If you are deciding where to start, the corporate AI programme page sets out the content and format. If you only want a second opinion on the judgement here, send us the process and we will tell you which box is most likely to break.
企業最常問的問題Questions companies ask most
香港企業應該由哪一級開始？Which level should a Hong Kong company start at?
先確認 L1 的管治是否存在。私隱專員公署的檢查顯示，受查機構幾乎全部在用 AI，但政策與訓練的覆蓋仍不一致[1]；如果員工已在用未經核准的工具，第一步不是升級，而是畫清邊界。之後選一個流程，由 L2 或 L3 開始。First check whether L1 is governed. The PCPD's checks show that almost every organisation reviewed is using AI, but policy and training coverage still varies[1]. If staff already use unapproved tools, the first step is to set the perimeter, not to upgrade. Then choose one process and start at L2 or L3.
是否一定要用最大、最新的模型？Do we have to use the largest, newest model?
不是。先用能通過你的評估集的最小模型。較小的模型通常更便宜、更快，也較容易控制資料流向。只有當評估顯示它做不到，才升級。模型選擇是一個測試問題，不是一個品牌問題。No. Start with the smallest model that passes your evaluation set. Smaller models are usually cheaper, faster and easier to keep data flows under control. Upgrade only when the evaluation shows they cannot do the job. Model choice is a test question, not a brand question.
中小企沒有人手，怎樣做？What can an SME do without spare headcount?
把範圍縮到一個流程、一個擁有者、90 天。不要同時做三個部門。香港 98% 的企業是中小企[14]，這個限制是常態，不是例外；能在 90 天內交出一個可比較的結果，已經比多數試點走得遠。Shrink the scope to one process, one owner and 90 days. Do not run three departments at once. Ninety-eight per cent of Hong Kong enterprises are SMEs[14]; the constraint is normal, not exceptional. Delivering one comparable result in 90 days already puts you ahead of most pilots.
用 ChatGPT 處理客戶資料可以嗎？Can we put client data into ChatGPT?
取決於你用的是哪一種帳戶與條款，以及你有沒有合法依據。消費者版本通常會用資料改善服務；企業版本可以關閉訓練，但仍然要符合 DPP1、DPP3 與 DPP4[7]。最安全的做法，是不要把可識別個人的資料輸入未經批准的服務。It depends on the account, the terms, and whether you have a lawful basis. Consumer versions commonly use data to improve the service; business versions can switch training off, but DPP1, DPP3 and DPP4 still apply[7]. The safe rule is not to put identifiable personal data into an unapproved service.
香港有沒有 AI 法例？Does Hong Kong have an AI law?
沒有一條統一的 AI 法例。義務來自《個人資料（私隱）條例》的六項保障資料原則、行業監管（例如金管局與證監會）與政府指引[20]。政府已成立跨部門工作組檢視現有法例，但尚未決定個別問題是否以立法處理[20]。所以「等立法」不是一個可行的計劃。There is no single AI statute. Obligations come from the six Data Protection Principles in the PDPO, from sector regulators such as the HKMA and SFC, and from government guidance[20]. An inter-departmental working group is reviewing existing legislation, but no decision has been made on whether individual issues should be handled by legislation[20]. Waiting for a statute is not a plan.
什麼時候需要成立 AI 治理委員會？When do we need an AI governance committee?
當你進入 L3，或任何一個流程會影響客戶、金錢或員工記錄時。委員會不需要大，但要有三個角色：業務擁有者、IT 或保安負責人，以及能說「不」的風險或法務代表。PCPD 的框架把 AI 治理列為第一個範圍[6]。When you reach L3, or when any process can affect a client, money or an employee record. The committee does not need to be large, but it needs three roles: the business owner, IT or security, and a risk or legal representative with the authority to say no. The PCPD framework puts AI governance in its first area[6].
如何量度 AI 的回報？How should we measure the return?
導入前先量一次周期時間、錯誤率與單位成本；導入後用同一組指標比較。不要用自報的「節省了多少小時」，因為它通常不包括覆核、重做與維護。麥肯錫的調查顯示，只有 39% 的企業能在企業層面指出 EBIT 影響[2]，部分原因就是基線與量度方法沒有定義好。Measure cycle time, error rate and unit cost before implementation, then compare on the same three metrics afterwards. Do not rely on self-reported hours saved, which usually exclude review, rework and maintenance. McKinsey found only 39% of organisations could point to EBIT impact at the enterprise level[2], partly because the baseline and the measurement were never defined.
代理（agents）現在可以用嗎？Can we use agents now?
可以，但限於低風險、可回復、有上限的動作，而且要有兩人批准、日誌與停用開關。Gartner 預計超過四成的代理項目會在 2027 年底前取消，原因多數是價值不清與風險控制不足[18]。先問「它做錯時，我們怎樣復原」，再問「它有多聰明」。Yes, but only for low-risk, reversible actions with a cap, two-person approval, logs and a kill switch. Gartner expects more than 40% of agentic projects to be cancelled by the end of 2027, mostly because value is unclear and risk controls are inadequate[18]. Ask how you undo an error before asking how smart the agent is.
