一間公司決定開始做 AI。第一次會議，有人提議買一套工具，有人提議把由報價到收款的整條流程自動化。六個月之後，工具已經買下，流程圖畫過兩版，日常的做法沒有改變，月結仍然要人手追。A company decides to start with AI. In the first meeting someone proposes buying a tool, someone else proposes automating the whole run from quotation to cash collection. Six months later the licences are bought, the process map has been drawn twice, daily work has not changed, and the month-end close still needs someone chasing.
兩組數字說明為甚麼會這樣。McKinsey 在 2026 年的全球調查發現，八成受訪者說 AI 提升了自己的生產力，而企業層面見到 EBIT 影響的只有 37%，與一年前相同[2]。香港的情況相近：私隱專員公署在 2026 年 5 月完成的循規審查，60 間受查機構中 57 間（95%）在日常營運使用 AI；使用 AI 的機構之中，29 間（51%）同時使用三個或以上系統[1]。Two sets of numbers explain why. In McKinsey's 2026 global survey, 80 percent of respondents said AI had improved their own productivity, while only 37 percent attributed any EBIT impact to AI at the enterprise level, the same share as a year earlier[2]. Hong Kong shows the same pattern: in the compliance review completed by the Privacy Commissioner in May 2026, 57 of the 60 organisations reviewed (95 percent) used AI in daily operations, and 29 of the organisations using AI (51 percent) ran three or more systems[1].
起點的選擇比工具與決心都重要。一個部門的自動化可以由一條已經可以量度的流程開始：它有排隊、有觸碰次數、有重做比率。以下四個切入點，每個配一個公司可以從自己系統讀到的指標、一個前置條件與一個風險；然後是一個 30 日試行，讓選擇在一個月內被證明或被否決。The choice of starting point matters more than the tool and the resolve. Automation in one department can begin with a process that can already be measured: it has a queue, a touch count and a rework rate. Each of the four entry points below carries an indicator a company can read from its own systems, a precondition and a risk. A 30-day pilot then proves the choice or kills it within a month.
摘要Summary
- 部門自動化的成敗在起點。買一套工具，與把一整條跨部門流程自動化，兩者都未量過現況；一個部門的自動化應該由一條已經可以量度的流程開始。The starting point decides the outcome. Buying a tool and automating an entire cross-department process both begin without a measurement of the current state. Automation in one department should begin with a process that can already be measured.
- 四個切入點是財務與應付賬款、銷售與報價、人力資源與入職、營運與客戶查詢。每個配一個可以由公司自己系統讀到的指標、一個前置條件與一個風險。The four entry points are finance and accounts payable, sales and quotation, human resources and onboarding, and operations and customer enquiries. Each one comes with an indicator a company can read from its own systems, a precondition and a risk.
- 分工的形狀固定：AI 讀取、擷取、配對、起草與標示例外；人批准、決定、議價與簽名。用應付賬款做示範，四個步驟逐行寫清楚哪一步交給 AI、哪一步留給人。The division of work has a fixed shape. AI reads, extracts, matches, drafts and flags exceptions; a person approves, decides, negotiates and signs. Accounts payable serves as the worked example, with four steps showing what goes to AI and what stays with a person.
- 30 日試行由兩星期的紀錄取基線，抽 10 至 20 個同類個案，前後量四個數字：周期時間、觸碰次數、重做比率、排隊時間。停手規則與第 30 日的決定在開始前寫死。A 30-day pilot takes its baseline from two weeks of records, samples 10 to 20 cases of the same type, and measures four numbers before and after: cycle time, touch count, rework rate and queue age. The stop rule and the day-30 decision are fixed before the pilot starts.
- 三個不應該開始的地方：由多個系統串起來的跨部門端到端流程、紀錄仍然在紙張與收件箱的流程、以及第一次試行就選上不可逆或受監管的決定。Three places not to start: a cross-department end-to-end process stitched from several systems, a process whose records still sit on paper or in inboxes, and a first pilot on a decision that is irreversible or regulated.
- 我們的做法是一條流程、一組固定個案、一起取基線；量不到的試行我們不接。代價寫在我們怎樣做那一節：基線有時要在建置之前先花兩星期。Our method is one process, a fixed case set and a baseline taken together; we turn down pilots we cannot measure. The cost is stated in the section on our approach: a baseline sometimes takes two weeks before any build begins.
四個切入點Four entry points
選一個部門，第一眼看的通常是有幾多人在做重複工作。這個數字容易數出來，卻答不到一個月之後要答的那條問題：做好了沒有。可以量度的流程有三個特徵：一個看得見的排隊、一個數得到的觸碰次數、一個記錄下來的重做比率。以下四個切入點都有這三個特徵，分別在量度成本與出錯的代價。Choosing a department usually starts with counting how many people do repetitive work. That number is easy to produce and does not answer the question one month later: did it work? A process that can be measured has three features: a visible queue, a countable number of touches, and a rework rate recorded somewhere. All four entry points below have them, and they differ in the cost of measurement and the cost of an error.
| 部門Department | 可量度指標Measurable indicator | 前置條件Precondition | 風險Risk |
|---|---|---|---|
| 財務／應付賬款Finance / accounts payable | 每張發票由收到到入帳的時間；支付批次之前要人手補的張數Time from invoice receipt to posting; the number of invoices a person has to complete before a payment run | 發票進入一個固定郵箱或系統；供應商與採購單有現行版本Invoices arrive in one fixed mailbox or system; supplier records and purchase orders have current versions | 付款放行不可逆，放行留給人Releasing a payment cannot be undone, so a person releases it |
| 銷售／報價Sales / quotation | 由收到查詢到報價發出的時間；同一個客戶改了幾多版Time from enquiry received to quotation issued; how many versions one customer goes through | 查詢有單一入口；產品與價格表有版本Enquiries have a single entry point; products and the price list carry versions | 折扣與條款是承諾，由人批Discounts and terms are commitments, so a person approves them |
| 人力資源／入職Human resources / onboarding | 由接受 offer 到第一日的日數；新人第一週重複問的次數Days from an accepted offer to the first day; how often a new joiner asks the same question in week one | 入職清單每一步有負責人；合約與政策是現行版本Every onboarding step has a named person; contracts and policies are current versions | 個人資料與僱傭合約受法規管，權限要分清Personal data and employment contracts are regulated, so access has to be separated |
| 營運／客戶查詢Operations / customer enquiries | 首次回覆時間；重開工單比率；一宗查詢轉了幾多次First response time; the rate of reopened tickets; how many times one enquiry is transferred | 查詢集中到一個收件箱或工單系統；回覆範本可以讓同事修改Enquiries land in one inbox or ticketing system; reply templates can be edited by the team | 投訴、賠償與關係由人接Complaints, compensation and relationships stay with a person |
應付賬款的好處是個案同質。一張發票的起點是收件時間，終點是入帳日期，中間的動作在會計系統裡有紀錄。多數公司的發票流程有五至六個觸碰點：收件、分類、配對、查詢、入帳、付款。按規則做的收件、分類與配對可以交給 AI；查詢要談，入帳要核，付款要放行，後面三步都有人。第一次試行只做一件事：把收件到入帳的時間量出來，再看自動讀取與配對之後，同一個時間縮短了幾多。Accounts payable benefits from case homogeneity. An invoice starts at the receipt timestamp and ends at the posting date, and the actions between them leave a record in the accounting system. Most companies run five or six touch points: receipt, classification, matching, query, posting and payment. Receipt, classification and matching follow rules and can go to AI; a query is a conversation, a posting needs a check and a payment needs a release, so a person stays in each of the last three. The first pilot does one thing: measure receipt-to-posting time, then see how far that time falls once reading and matching are automated.
銷售與報價的個案沒有財務那麼同質。新客戶與舊客戶、標準產品與改裝，處理方法不同，所以前置條件寫得最嚴：查詢要有一個入口，價格表要有一個版本。少了這兩樣，量到的時間反映的是混亂。交給 AI 的是背景研究、報價草稿與跟進提示；折扣、條款與成交由人決定。指標除了時間，還要記住一個數字：報價發出之前改了幾多版。改版次數比時間更早顯示問題，因為它指出資料在哪一步對不上。Sales and quotation cases are less homogeneous than finance ones. A new customer and a repeat customer, a standard product and a modified one, are handled differently, which is why the precondition is strictest here: one entry point for enquiries and one version of the price list. Without those two, the time measured reflects confusion. Background research, quotation drafts and follow-up prompts go to AI; discounts, terms and closing the deal stay with a person. Alongside time, record one more number: how many versions a quotation goes through before it leaves. The version count shows a problem earlier than time does, because it points to the step where the data does not agree.
人力資源與入職的價值在時間的固定性。由接受 offer 到第一日，工作有明確的起點與終點，中間多數是行政：合約、帳戶、設備、簡介、排程。指標有兩個：日數，與新人第一週重複問問題的次數。第二個數字可以直接數：同一個問題在幾多個新人身上出現過。入職資料包、政策問答與排程可以交給 AI；合約、薪酬與個別安排留在人手裡。個人資料的處理要在設計這一步寫清楚，包括哪些欄位不進入 AI 系統。Human resources and onboarding gain from their fixed timeline. From an accepted offer to the first day, the work has a clear start and end, and most of it is administrative: the contract, accounts, equipment, induction and scheduling. Two indicators matter: the number of days, and how often new joiners ask the same question in their first week. The second one can be counted directly: how many new joiners ask it. The onboarding pack, policy questions and scheduling can go to AI; the contract, pay and individual arrangements stay with a person. How personal data is handled has to be written at the design stage, including which fields never enter an AI system.
營運與客戶查詢最容易看到效果，也最容易被示範帶錯。示範給你看一條寫得漂亮的回覆草稿；真正要量的是首次回覆時間、重開工單比率與轉介次數。三個數字都在工單系統裡。分流、起草與知識庫問答可以交給 AI；投訴、賠償、議價與關係由人接。前置條件是一條紀律：所有查詢進同一個入口，包括 WhatsApp。做不到這一點，試行量到的是同事的記憶力。Operations and customer enquiries show a result fastest, and a demo can mislead most easily here. A demo shows a well-written reply draft; what needs measuring is first response time, the reopened-ticket rate and the number of transfers. All three sit in the ticketing system. Routing, drafting and knowledge-base answers can go to AI; complaints, compensation, negotiation and relationships stay with a person. The precondition is a discipline: every enquiry enters through one point, WhatsApp included. Without that, the pilot measures how well the team remembers.
哪兩個步驟交給 AI，哪兩個留給人Which two steps go to AI, which two stay with a person
用應付賬款做一次示範。同一條流程、四個步驟：前兩個以 AI 為主，後兩個以人為主。界線畫在可以重做與不可以重做之間。一份文件讀錯可以再讀一次；一筆錢付出去要追回來，成本高得多。The worked example comes from accounts payable. One process, four steps: the first two belong mostly to AI, the last two mostly to a person. The line runs between work that can be redone and work that cannot. A document read wrongly can be read again; a payment sent out costs far more to recover.
| 步驟Step | 交給 AITo AI | 留給人Stays with a person | 為甚麼Why |
|---|---|---|---|
| 收件與讀取Receipt and reading | 從固定郵箱讀取發票，擷取供應商、發票號碼、金額與到期日Read invoices from the fixed mailbox and extract the supplier, invoice number, amount and due date | 核對新供應商，補上缺採購單的個案Check a new supplier and complete the cases with no purchase order | 讀取是重複動作；新供應商要對名單Reading is repetitive; a new supplier has to be checked against the list |
| 配對與標示Matching and flagging | 把發票、採購單與收貨紀錄三方配對，標示對不上的項目Match the invoice, purchase order and goods receipt, and flag what does not agree | 決定差異怎樣處理：補文件、扣款或退貨Decide how a difference is handled: obtain the document, deduct, or return the goods | 配對按規則；差異處理是判斷Matching follows rules; handling a difference is a judgement |
| 入帳與對賬Posting and reconciliation | 把已配對的發票入帳，起草月結對賬表Post the matched invoices and draft the month-end reconciliation | 簽對賬，回答審計的問題Sign the reconciliation and answer the auditor's questions | 入帳可以更正；簽名與責任跟人A posting can be corrected; the signature and the responsibility stay with a person |
| 付款放行Payment release | 準備付款批次與付款清單Prepare the payment batch and the payment list | 批核並放行付款Approve and release the payment | 款項一出去，追回的成本高Once money leaves, recovery is expensive |
這張表有兩個用途。第一，寫進試行的範圍：交給 AI 的步驟是建置範圍，留給人的步驟是驗收條件。第二，寫進交接文件：一個月之後接手的人看這張表，就知道哪一步由系統做、哪一步要簽名。The table has two uses. The first is the pilot scope: the steps going to AI are the build, and the steps staying with a person are the acceptance conditions. The second is the handover document, so whoever takes over a month later can see which step the system runs and which step needs a signature.
人仍然在流程裡，位置移後了：由做每一步，變成看例外與簽名。這個分工要寫在設計裡，因為它決定試行要量哪四個數字。一份文件讀得快了幾多、少了幾多個人手動作、重做幾多、等幾久，四樣都跟這張表有關。People remain in the process, further back: instead of doing every step, they read exceptions and sign. That division belongs in the design, because it decides which four numbers the pilot measures. How much faster a document moves, how many manual actions disappear, how much is redone, and how long it waits all follow from this table.
30 日試行怎樣設計How to design the 30-day pilot
基線由兩星期的紀錄取，不由記憶取。多數流程已經有紀錄：發票的收件與入帳日期、工單的開啟與關閉時間、報價單的版本、電郵的時間戳。抽 10 至 20 個同類個案，逐個計時間與觸碰次數，把定義寫進同一份檔案。這份檔案在試行前確認一次，30 日之內不再改；要換工具，先寫進紀錄，第 30 天再算。The baseline comes from two weeks of records, not from memory. Most processes already leave records: invoice receipt and posting dates, ticket open and close times, quotation versions, email timestamps. Sample 10 to 20 cases of the same type, time each one and count its touches, and write the definitions into one file. That file is confirmed once before the pilot and left alone for the 30 days; a change of tool is recorded and counted on day 30.
個案集要同類。同一個供應商類別、同一種查詢、同一級職位，才算同一類。10 至 20 個夠看到方向；100 個不會更清楚，只會更慢。試行期間新出現的個案照樣記錄，但不混入比較。The case set has to be homogeneous. The same supplier category, the same kind of enquiry, the same job level: only then is it one type. Ten to 20 cases are enough to show a direction; 100 will not be clearer, only slower. New cases that appear during the pilot are recorded but kept out of the comparison.
要量四個數字。周期時間是一個個案由起點到終點的工時；觸碰次數是一宗個案經過幾多個人手動作；重做比率是要重做或退回的比率；排隊時間是個案在等、未有人處理的時間。四個數字各有一個定義與一個取數位置，寫在基線那一份檔案裡。培訓之後量哪幾個指標，用的是同一套方法，另一邊寫在〈培訓成效怎樣量度〉。Four numbers are measured. Cycle time is the working time from the start of a case to its end. Touch count is how many manual actions one case passes through. Rework rate is the share that has to be redone or returned. Queue age is the time a case spends waiting, before anyone works on it. Each number has one definition and one place it is read from, written into the baseline file. Measuring the effect of training follows the same approach from the other end; the article on how training outcomes are measured sets out that side.
| 指標Indicator | 怎樣量How it is measured | 從哪裡取數Where the number comes from |
|---|---|---|
| 周期時間Cycle time | 起點時間戳到終點時間戳的工時Working time between the start timestamp and the end timestamp | 發票的收件與入帳日期、工單的開啟與關閉、報價的版本紀錄Invoice receipt and posting dates, ticket open and close times, quotation version records |
| 觸碰次數Touch count | 一宗個案經過的人手動作數The number of manual actions one case passes through | 系統的處理紀錄、審批紀錄、轉介紀錄System handling records, approval records, referral records |
| 重做比率Rework rate | 要重做或退回的個案佔全部個案的比例The share of cases that had to be redone or returned | 退回紀錄、重開紀錄、修訂版本Return records, reopening records, revision versions |
| 排隊時間Queue age | 個案在等、未有人處理的時間Time a case spends waiting with nobody working on it | 工單的等待狀態、收件箱的收件與首次處理時間Ticket waiting status, the inbox receipt and first handling times |
停手規則在開始前寫死。第 30 天，如果四個數字都停在原來的位置，或者例外比率高過開始時定的上限，就停，並把原因寫下來。寫原因比寫結論有用：常見的三個原因是流程未改、工具未進入流程、個案集換了。第 30 日的決定只有三個：繼續、重新設計、停止。三個都可以接受；含糊不可以。判斷只看四個數字，不看回報率；四個數字由你的系統紀錄取，不由我們補寫。The stop rule is fixed before the start. On day 30, if all four numbers sit where they began, or the exception rate is above the ceiling set at the start, the pilot stops and the reason is written down. Writing the reason is more useful than writing a conclusion: the three common reasons are that the process did not change, the tool did not enter the flow, and the case set changed. The day-30 decision has three options: continue, redesign, stop. All three are acceptable; a vague answer is not. The judgement rests on the four numbers, not a return promise; the numbers are taken from your system records and we do not fill them in.
三個不應該開始的地方Three places not to start
以下三個起點，看起來都像自動化，實際上會令第一次試行量不到一個可用的數字。The three starting points below all look like automation. Each one leaves the first pilot without a usable number.
一、多個系統串起來的跨部門流程。One: a cross-department process stitched from several systems. 由收到查詢到收款，中間有報價、訂單、送貨、發票與收款，每一段由不同部門、不同系統記錄。這種流程的基線要跨系統拼接，而拼接的部分通常靠人手補。試行一開始就要處理資料對不上，量到的差異說不清是自動化的效果還是拼接的誤差。Gartner 預測，超過四成代理式 AI 項目會在 2027 年底前取消，原因包括成本上升、商業價值不清與風險控制不足[3]；價值不清，多數由範圍太大開始。The quotation, the order, delivery, the invoice and the payment sit between an enquiry and the cash, each recorded by a different department in a different system. The baseline for such a process has to be stitched across systems, and the stitching is usually done by hand. The pilot starts by fixing data that does not agree, and the difference it measures cannot be separated from the error of the stitching. Gartner predicts that over 40 percent of agentic AI projects will be cancelled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls[3]. Unclear value usually begins with a scope that is too large.
二、紀錄仍然在紙張、WhatsApp 與口頭批准的流程。Two: a process whose records still sit on paper, in WhatsApp and in verbal approvals. 有些流程的紀錄全部在紙上，或者散在對話與記憶裡。這種流程要先做一件事：把紀錄做出來。一個表格、一個時間戳、一個集中的收件箱，任何一項都可以是第一步。跳過這一步，試行的第一個星期會用來爭論發生過甚麼事，量到的數字沒有人信。Some processes keep every record on paper, or scattered across conversations and memory. Such a process needs the record first. A form, a timestamp or a single inbox, any one of the three can be the first step. Skip it, and the pilot's first week is spent arguing about what happened, and nobody trusts the number that comes out.
三、第一次試行就選上不可逆或受監管的決定。Three: a first pilot on a decision that is irreversible or regulated. 信貸審批、法律意見、醫療判斷與付款放行，錯一次的成本高，而監管者與審計要看到控制。這些決定可以是後期階段：由人保留決定權，AI 做準備工作。第一次試行選上它們，代價最高，學到的最少，整個計劃還可能在風險評估那一關停下來。Credit approval, legal advice, medical judgement and payment release carry a high cost when wrong, and regulators and auditors need to see the control. These decisions can be a later phase: the decision stays with a person, and AI does the preparation. Choosing one for the first pilot costs the most, teaches the least, and can stop the whole plan at the risk assessment.
三個起點有一個共同問題：試行完結時，你說不出一個前後對比的數字。可以量度，是第一次試行唯一的入場條件。The three share one problem: at the end of the pilot, there is no before-and-after number to state. Being measurable is the only entry condition for a first pilot.
我們怎樣做Our approach
我們的做法是一條流程、一組固定個案、一起取基線。第一次見面，我們問三件事：這條流程的起點與終點在哪裡、過去兩星期的紀錄在誰手上、哪一個數字現在最難看。三個答案都有，才排得出一個 30 日試行。Our method is one process, a fixed case set, and a baseline taken together. At the first meeting we ask three questions: where the process starts and ends, who holds the last two weeks of records, and which number looks worst today. With all three answers, a 30-day pilot can be laid out.
試行之後才建置。我們做的部門自動化由一條流程開始，每一項都對應一個可量度的指標。先界定量哪個數字，再決定建甚麼；報價在範圍界定之後才出，範圍寫得清楚，後面的爭議自然少。The build comes after the pilot. Our department automation starts from one process, and every item maps to a measurable indicator. Decide which number is measured first, then decide what to build; the quotation follows the scope definition, and a clear scope leaves less to argue about later.
弱點有兩個。第一，量不到的試行我們不接，有些工作我們會推掉，包括一些看起來很急的項目。第二，基線要兩星期，建置之前先花時間在紀錄上；對想即時見到工具的公司，這兩星期像沒有進展。我們仍然這樣做，因為基線一旦跳過，第 30 天就沒有一個可以信的對比。There are two weaknesses. First, we turn down pilots we cannot measure, which means declining some work, including projects that look urgent. Second, the baseline takes two weeks of records before any build; to a company that wants to see a tool now, those two weeks look like no progress. We keep the two weeks because a skipped baseline leaves nothing trustworthy to compare against on day 30.
我們不簽發回報率證書，也不替你寫數字。報告由你的同事簽；我們提供定義、抽樣方法與紀錄格式。理由很實際：一份由供應商寫的回報，在你自己的預算會議上說服力最低。We do not issue a return-on-investment certificate and we do not write your numbers. Your colleague signs the report; we provide the definitions, the sampling method and the record format. The reason is practical: a return written by the supplier carries the least weight in your own budget meeting.
同一個指標放在不同層級，量到的內容不同。一個部門的流程未穩定，先做流程自動化，量到的只是混亂期。每個層級的難題與通過門檻，見〈企業 AI 導入的五個層級〉。The same indicator means a different measurement at each level. Run process automation in a department whose process is not yet stable, and the measurement reflects the instability. The constraints and gates at each level are set out in the five levels of AI adoption in Hong Kong.
