一場培訓之後，手上有出席名單、滿意度分數，以及幾句「都有用」。三個月後再問同一個部門，答案多數是「有在用 AI」。這些話都可能是真的，但它們都回答不了同一個問題：工作有沒有改變。After a training day there is an attendance list, a satisfaction score and a few sentences about how useful it was. Ask the same department three months later and the answer is usually that people are using AI. All of that can be true, and none of it answers the question that matters: did the work change?
香港的數字說明為甚麼要問這一條。Workday 在 2025 年 11 月訪問 3,200 名全職員工，全部任職於年收入一億美元以上的機構，並且日常使用 AI。香港的結果是：92% 說 AI 每週為他們節省一至七小時；58% 每週要花一至兩小時修正、改寫與核對 AI 的產出；只有 30% 能穩定拿到幾乎不用修改的結果[1]。The Hong Kong figures show why the question matters. Workday surveyed 3,200 full-time employees in November 2025. All worked at organisations with annual revenue above US$100 million, and all used AI day to day. In Hong Kong, 92% said AI saves them one to seven hours a week, 58% spend one to two hours a week correcting, rewriting and checking what it produced, and only 30% consistently get output that needs little revision[1].
「節省了幾多小時」聽起來中立，但它只算了一半的帳：修訂的時間留在外面，而修訂正是培訓應該減少的那一部分。要判斷一場培訓有沒有用，看的是四件落在工作上的事：一條流程走得快了幾多、同事自己交得出幾多件可用的東西、有幾多人接得住別人做的模板、以及有幾多個案仍然要人處理。這四件事都可以在培訓前後各量一次，而且都不必問當事人覺得怎樣。So “hours saved” is not a neutral number. It leaves the rework outside the calculation, and rework is the part training is meant to reduce. Whether a course worked shows up in four things on the work itself: how much faster one process runs, how many usable outputs the team produces on its own, how many people can pick up a template someone else built, and how many cases still need a person. Each can be measured before and after, and none of them asks anyone how they felt.
摘要Summary
- 「有用」評的是課室，工作要另外量。柯氏四層評估把反應放在第一層，模型的說明寫明第一層不足以判斷行為有否轉移[2]。高分與改變之間沒有必然關係。“Useful” is a verdict on the room. The work has to be measured separately. The Kirkpatrick model puts reaction at Level 1, and its own description says Level 1 is not enough to tell whether behaviour transfers[2]. A high score and a change in the work do not necessarily arrive together.
- 四個指標：一個具名流程的周期時間；第一週內同事不靠導師交出的可用產出；由第二個人接手的模板與提示；仍然要人處理的例外比率。四個都在工作上留下一份紀錄。Four indicators: cycle time on one named process; usable outputs the team produces unaided in the first week; a template or prompt that a second person takes over; the share of cases that still need a person. Each leaves a record in the work itself.
- 量法只有一條規則：培訓前量一次基線，第 30 天用同一個定義再量一次。兩次都要在系統裡留下檔案，因為一年之後沒有人記得當時的條件。There is one rule of method: take a baseline before the course, then measure again on day 30 against the same definition. Both readings have to sit in a system, because a year later nobody remembers the conditions.
- 最容易被高估的數字是「節省了幾多小時」。香港 58% 的員工每週花一至兩小時修正 AI 的產出[1]；這些時間不入帳，那個節省只是一個上限。The most over-stated number is hours saved. In Hong Kong, 58% of employees spend one to two hours a week correcting what AI produced[1]. With those hours left out of the account, the saving is only a ceiling.
- 少過十個人，不要量比率。樣本小的百分比，一個人的變動就是三十三個百分點。看兩三件具體產出，加一次覆核。With fewer than ten people, do not measure a ratio. In a small sample, one person moving changes the percentage by thirty-three points. Look at two or three concrete outputs and one review.
「有用」是一個關於課室的答案“Useful” is an answer about the room
一個四點六分的培訓，可以改變一個部門的做法，也可以完全沒有留下任何東西。分數與結果之間沒有固定的關係，因為分數問的是感受，結果要問行為。A course rated 4.6 out of 5 can change how a department works, and it can also leave nothing behind. There is no fixed relationship between the score and the result, because the score asks about a feeling and the result asks about behaviour.
柯氏四層評估把這件事分開講：第一層是反應，第二層是學習，第三層是行為，第四層是結果。模型的說明寫得清楚：第一層不足以判斷行為有否轉移；第三層要看的，是同事回到自己的環境之後，有沒有做出那幾個關鍵動作，以及有沒有人跟進[2]。企業培訓最常停在第一層與第二層之間，因為後兩層要等，而且要其他部門配合。The Kirkpatrick model separates the question into four levels: reaction, learning, behaviour and results. Its own description is explicit. Level 1 is not sufficient to tell whether behaviour transfer will happen; Level 3 looks at whether people perform the critical behaviours once they are back in their own environment, and whether anyone supports and checks them[2]. Corporate training usually stops between the first and second level, because the next two require waiting and require other departments to cooperate.
反應仍然有用。一整班人說內容與自己的工作無關，是一個要即日處理的訊號，因為內容錯了，後面的量度都會白做。它用來篩選，未到結論。Reaction still has a use. A whole class saying the content has nothing to do with their work is a signal to act on the same day, because the content is wrong and everything measured afterwards is wasted. It is a filter, not a conclusion.
自評的偏差有實測數字。METR 在 2025 年做了一個隨機對照實驗：16 名資深開源開發者，246 個真實工作項，准用 AI 的一組平均慢了 19%。他們事前預計快 24%，做完之後仍然認為自己快了 20%[3]。同一個團隊在 2026 年再訪問 349 名技術工作者，自報的價值提升中位數是 1.4 至 2 倍，報告同時寫明，公開的自評數字一向高於實地實驗的結果[4]。The bias in self-assessment has been measured. In 2025 METR ran a randomised controlled trial: 16 experienced open-source developers, 246 real work items, and the group allowed to use AI took 19% longer on average. They had forecast a 24% speed-up, and afterwards still believed they had been 20% faster[3]. The same team surveyed 349 technical workers in 2026 and found a median self-reported value gain of 1.4 to 2 times, while noting that public self-reports have consistently run above results from field experiments[4].
一個人對自己效率的判斷，可以與實際相差四十個百分點。一間公司對一場培訓的判斷，如果只靠同一批人的判斷，誤差不會更小。A person's judgement of their own speed can miss by forty percentage points. A company's judgement of a course, resting on the same people's judgement, will not be more accurate.
四個指標，每一個都在工作上留下一份紀錄Four indicators, each leaving a record in the work
四個指標指向同一件事：改變有沒有發生在流程裡。它們不需要問卷，也不需要等年結；需要的是一條具名的流程、一組真實個案，以及一個願意在培訓前花半天的人。The four point at one thing: whether the change reached the process. None of them needs a questionnaire or a year-end close. What they need is one named process, a set of real cases, and someone willing to spend half a day before the course.
| 指標Indicator | 用甚麼單位What you count | 從哪裡取數Where the number comes from |
|---|---|---|
| 周期時間Cycle time | 一條具名流程由收到需求到交付的工時Working hours for one named process, from request received to delivered | 工單系統、電郵時間戳、報價紀錄Ticketing system, email timestamps, quotation records |
| 可用產出Usable output | 第一週內無需導師協助、另一個人可以直接使用的件數Pieces produced in week one without help from the trainer, that another person can use as they are | 交付檔案、審批紀錄Delivered files, approval records |
| 重用Reuse | 由第二位同事接手的次數The number of times a second colleague takes it over | 共用提示庫與模板的使用紀錄Usage records in the shared prompt library and templates |
| 例外比率Exception rate | 仍然要人處理的個案佔全部個案的比例The share of cases that still need a person | 抽 20 至 30 個個案，逐個分類Sample 20 to 30 cases and classify each one |
先確認那條流程在哪一級。個人在用工具、團隊共用工具、流程已自動化，三種情況量到的東西不同；香港企業導入 AI 的五個層級把每一級的難題分開講，選錯一級，量度也會跟著錯。First place the process on the right level. One person using a tool, a team sharing a tool, and an automated process are three different things to measure; the five levels of AI adoption in Hong Kong sets out what stalls at each one, and measuring against the wrong level produces the wrong number.
1. 周期時間：選一條流程，寫下起點與終點1. Cycle time: one process, a written start and end
「報價單由收到客戶需求到發出」是一條可以量的流程：起點是客戶電郵的時間戳，終點是報價單發出的時間。量十個個案的工時，取平均，不要量感覺。十個個案已經足夠看到一個方向，一百個不會更清楚。“Quotation from customer request to issue” is a process you can measure: the start is the timestamp on the customer's email, the end is when the quotation goes out. Record the hours on ten cases and take the average. Ten is enough to show a direction; a hundred will not make it clearer.
陷阱在同時改動。培訓之後同一個流程還換了模板、加了審批，周期時間的變化就不能只歸功於培訓。要麼一次只改一件事，要麼在紀錄裡寫明同時改動了甚麼，讓讀數字的人自己打折。The trap is changing two things at once. If the same process also got a new template and an extra approval step after the course, the change in cycle time cannot be credited to the course alone. Either change one thing at a time, or write down what else changed so that whoever reads the number can discount it.
2. 可用產出：把「可用」在培訓前寫死2. Usable output: define “usable” before the course
培訓後第一週，數一數有幾件東西是同事自己做出來、而且另一個人可以直接拿去用。不需要導師在旁，不需要再改三次。這個數字通常比預期低，而低正是它有用的原因：它把「上過堂」與「交得出」分開。In the week after the course, count how many pieces the team produced on its own that someone else could use as they are. No trainer at the elbow, no third round of edits. The number is usually lower than expected, and that is what makes it useful: it separates having attended from having delivered.
可用的標準要在培訓前寫好，例如「另一位同事可以在十分鐘內完成同一件事，不必問原作者」。標準寫得寬，數字會漂亮，然後沒有用途。Write the standard down beforehand, for example that another colleague can complete the same task within ten minutes without asking the person who made it. A loose standard produces a flattering number and nothing else.
3. 重用：第二個人用得着才算3. Reuse: it counts when a second person uses it
同一個提示、模板或流程，由一個沒有參與建立它的人再用一次。這是最難造數的指標，也是最難假造的。建立者自己用十次，只證明他懂得用；第二個人用一次，才證明它被交了出去。A prompt, template or workflow is used again by someone who had no part in building it. This is the hardest indicator to dress up and the hardest to fake. The builder using it ten times proves the builder knows how. A second person using it once proves it was handed over.
紀錄不必另外做表。共用檔案本身就有開啟與編輯的紀錄，看那份紀錄就夠。要留意的是同一批人互相重用，那仍然是一個小圈子的熟練度，未到團隊的能力。No extra log is needed. The shared file already records opens and edits, and that record is the evidence. What to watch for is the same few people reusing each other's work: that is still one small circle's fluency, not a team capability.
4. 例外比率：還有幾多個案要人接手4. Exception rate: how many cases still need a person
抽 20 至 30 個個案，逐個問：這一個最後由人處理了嗎？比例下降，代表流程真的接手了一部分；比例不變，代表 AI 只是加快了第一步，人仍然由頭做到尾。Sample 20 to 30 cases and ask of each one: did a person have to finish it? A falling share means the process genuinely took over part of the work. An unchanged share means AI only sped up the first step while a person still carried the case from beginning to end.
這個指標最容易在定義上出錯。例外的定義要在基線那次寫死：哪幾類算例外、哪幾類不算、邊界個案怎樣處理。分類一旦放鬆，比率自然會跌，跌的是數字，工作沒有變。This is the indicator most easily broken by its own definition. Fix the definition at the baseline: which categories count as exceptions, which do not, and how borderline cases are handled. Loosen the categories later and the rate falls on its own, without the work changing.
一次 30 天的量度：基線、第一週、第 30 天One 30-day measurement: baseline, first week, day 30
30 天是一個折衷。太短，改變未出現；太長，其他事情會混進來：換人、換系統、旺季。一個月足夠看到第一輪使用，也足夠記得當時的條件。Thirty days is a compromise. Shorter and the change has not appeared; longer and other things join the picture, such as a departure, a system migration or the busy season. A month is long enough for the first round of use and short enough to remember the conditions.
- 前 7 天7 days before基線Baseline抽十個真實個案，量周期時間與例外數，把「可用」與「例外」的定義寫進同一份檔案。Ten real cases: record cycle time and exceptions, and write the definitions of “usable” and “exception” into the same file.
- 開課日Course day培訓The course課程用同一條流程的真實個案設計，示範的就是那十個個案的做法。The course is designed around the same process and its real cases, and works on how those ten cases were handled.
- 第 1 至 7 天Days 1–7第一週First week數可用產出，導師不介入；把做成的模板放進共用位置，寫一句使用方法。Count usable outputs without the trainer stepping in, and put the finished template in the shared location with one line on how to use it.
- 第 8 至 30 天Days 8–30接手Takeover找一個沒有參加建立的人用同一個模板做一件真事，看他需要問幾多次問題。Have someone who did not build the template use it on one real task, and note how many questions they ask.
- 第 30 天Day 30覆核Review另抽十個同類個案，用同一組定義再量一次，比較四個數字，把沒有動的原因寫下來。Take ten new cases of the same kind, measure them against the same definitions, compare the four numbers, and write down why any of them did not move.
圖一：五個時點。只有基線與第 30 天是量度，中間三個時點做的是把改變推出去。Figure 1: five points in time. Only the baseline and day 30 are measurements. The three in between are where the change is pushed out into the work.
基線一定要在培訓前量。培訓後補做的基線，是用結果去猜起點，兩者都會錯。沒有現成數據的公司，可以用系統自己的紀錄重建：電郵時間戳、工單的開啟與關閉時間、報價單的版本紀錄。抽十個就夠，不必全量。The baseline has to be taken before the course. A baseline reconstructed afterwards uses the result to guess the starting point, and gets both wrong. A company without the data to hand can rebuild it from records it already keeps: email timestamps, when a ticket opened and closed, the version history of a quotation. Ten cases is enough; a full extract is not needed.
第 30 天另抽十個同類個案：不要挑最近最忙的，也不要用培訓時示範過的那幾個。同一類、同一組定義、同一個抽樣方法，三樣都要對得上，否則比較的基礎不見了。On day 30, take ten new cases of the same kind. Do not pick the busiest week, and do not reuse the cases used as examples in the course. Same category, same definitions, same sampling method: all three have to line up, or there is nothing to compare.
如果第 30 天的數字沒有動，先把原因寫下來，再決定下一步。常見的三個原因是：流程沒有改、工具沒有到位、培訓教的動作與流程無關。三個原因指向三種不同的處理，混在一起看，只會得出「培訓沒有用」這一個結論，而這個結論幫不到任何人。If the number has not moved on day 30, write down why before deciding what to do. Three causes cover most cases: the process did not change, the tool was not ready, or the course taught actions unrelated to the process. Each points to a different fix. Blended together they produce one conclusion, that the training did not work, and that conclusion helps nobody.
三個該丟的指標，以及它們可以留下來做甚麼Three indicators to drop, and what each can still do
多數企業培訓報告的第一頁都有這三個數字。它們之所以留下，因為即日有數、容易比較、不需要其他部門配合。錯處不在於假，在於與結果之間沒有穩定的關係。These three numbers sit on the first page of most corporate training reports. They survive because they are available the same day, easy to compare and require no cooperation from anyone else. Their fault is not that they are false. It is that they bear no stable relationship to the result.
| 指標Measure | 它量到甚麼What it actually measures | 為甚麼不足以說明成效Why it cannot show the work changed |
|---|---|---|
| 滿意度分數Satisfaction score | 參加者對課室的即時反應The room's immediate reaction | 同一個分數可以出現在完全沒有改變的部門，也可以出現在改變了的部門The same score appears in departments that changed nothing and in departments that changed plenty |
| 出席率Attendance | 幾多人坐在房間裡How many people were in the room | 香港 88% 的機構中，不足一半職位已按 AI 能力調整[1]：坐滿人，流程仍然一樣In 88% of Hong Kong organisations, fewer than half of roles have been updated for AI[1]: a full room, an unchanged process |
| 自評信心Self-rated confidence | 參加者覺得自己學會了What participants believe they learned | 自評與實測可以相差四十個百分點：METR 的開發者自評快了 20%，實測慢了 19%[3]Self-assessment can miss by forty percentage points: METR's developers believed they were 20% faster while measuring 19% slower[3] |
不要一次過把它們刪掉。三個數字各有一次合法的用途，只要放回它自己的位置。Do not delete all three. Each has one legitimate use, as long as it goes back to its own place.
滿意度分數留在內容檢查：一整班人說內容與自己的工作無關，就要改課程，不必等下一個季度。出席率留在成本計算：每人每小時的成本是一個採購數字，回答的是「這次要幾多錢」，與「有沒有用」無關。自評信心可以留一句，前提是旁邊放一件實物，讓讀的人自己對照。Keep the satisfaction score as a content check: a whole class saying the content has nothing to do with their work means the course needs changing, not next quarter's review. Keep attendance in the cost line, where cost per person-hour answers what the session cost and says nothing about whether it worked. Self-rated confidence can stay as a single line, provided a real object sits beside it for the reader to compare against.
把它們從「成效」那一欄移到「過程」那一欄，同一份報告的結論就會改變。三個數字仍然在，只是不再假裝自己回答了成效的問題。Move them from the results column to the process column and the same report reaches a different conclusion. The three numbers are still there; they simply stop pretending to answer the question about results.
四種情況，先不要量Four situations where measuring is wasted
量度有成本：抽個案、跑系統紀錄、開會對定義。有些情況下，這些成本換不到一個可以用的數字，換到一個誤差很大的數字。Measurement has a cost: sampling cases, pulling system records, agreeing definitions in a meeting. In some situations that cost buys nothing usable, or buys a number with a wide error bar.
少過十個人。比率是雜訊。三個人之中有一個做得到，是 33%，也是一個。人數少的時候，看兩三件具體產出，加一次覆核，比一個百分點可靠。Fewer than ten people. A ratio is noise. One out of three is 33%, and it is also one person. In a small group, two or three concrete outputs plus one review beat a percentage point.
工具還在換。流程的周期時間會被工具遷移吃掉，量到的是一個混亂期。等工具穩定兩星期，再量基線。The tool is still changing. A migration eats the cycle time, so what gets measured is a period of disruption. Let the tool settle for two weeks, then take the baseline.
流程本身不穩定。同一件事每個星期做法都不同，周期時間量到的是流程，培訓不在其中。先把流程寫下來，有固定做法，再談量度。The process is not stable. If the same task is done differently every week, cycle time measures the process and the course is not in it. Write the process down, get one way of doing it, and measure after that.
對象是整間公司，而沒有一條共同流程。全公司普及課的價值在共同語言，單一流程的改善要靠部門。先選一個部門、一條流程，量度才有落點。The audience is the whole company and there is no shared process. A company-wide onboarding course is worth its cost in shared language; improving a single process happens in a department. Pick one department and one process, and the measurement has somewhere to land.
四種情況有一個共同的替代做法：不要量比率，量一件實物。一件由同事自己做完、另一個人接得住的東西，在這些情況下比一個百分點可靠，也更容易在下一次會議上講清楚。All four have the same fallback: do not measure a ratio, measure one real object. Something a colleague finished unaided and someone else picked up is more reliable than a percentage point in these situations, and easier to explain at the next meeting.
我們在量度上做到哪一步How far we go with the measurement
課程由一條流程開始。開課前兩星期，我們與你選定一條流程與十個個案，把基線量好，然後才寫課程內容。三個培訓方向（企業 AI 培訓）都用這個做法，分別在於深度與人數。A course starts with a process. Two weeks before the session, we choose one process and ten cases with you, take the baseline, and only then write the content. All three tracks of corporate AI training work this way, and differ in depth and group size.
我們留下一個共用的提示庫與模板。重用這個指標需要有一件東西可以被重用；沒有共用的檔案，第三個指標永遠量不到。課後的一次跟進諮詢用來對第 30 天的數字，不對課程的評分。We leave a shared prompt library and templates behind. The reuse indicator needs an object that can be reused, and without a shared file it can never be measured. The follow-up session after the course is for the day-30 numbers, not for rating the course.
我們不做的事：不簽發回報率證書，也不替你寫數字。回報屬於你自己的帳目，我們提供方法與紀錄的格式，數字由你的系統出。這一點寫在四項服務的範圍裡，也寫在報價單之前。What we do not do: issue a return-on-investment certificate, or write the figures for you. The return belongs in your own accounts. We supply the method and the format for the records; the numbers come out of your systems. That is stated in the scope of the four services and again before any quotation.
如果你已經做過一輪培訓而沒有基線，可以由系統紀錄重建一次。重建的基線會在報告裡註明是重建的：它仍然有用，但它可以被引用到甚麼程度，讀的人有權知道。If a round of training has already happened without a baseline, it can be rebuilt from system records once. The rebuilt baseline is labelled as rebuilt in the report: it is still useful, and whoever reads it is entitled to know how far it can be quoted.
企業最常問的問題Questions companies ask most
只有一日培訓，量得到嗎？Can a one-day course be measured?
量得到，但要把範圍縮到一條流程、十個個案。一日課程不足以改變整間公司的做法，足以改變一條流程的第一步。把「可用」與「例外」的定義在開課前寫好，第 30 天另抽十個同類個案，用同一組定義再量一次，就有一個可以比較的結果。Yes, if the scope narrows to one process and ten cases. A one-day course will not change how a whole company works, and it can change the first step of one process. Write the definitions of usable and exception before the session, take ten new cases on day 30 and measure them against the same definitions, and there is a result to compare.
沒有基線數據，可以由哪裡補？There is no baseline. Where do we get one?
由系統自己的紀錄重建：電郵時間戳、工單的開啟與關閉時間、報價單的版本紀錄、共用檔案的編輯紀錄。抽十個個案，取平均，並在報告上註明這是重建的基線。重建的數字仍然可以用來比較，只是不能與當時同步量度的數字混在一起講。Rebuild it from records the system already keeps: email timestamps, when tickets opened and closed, quotation version history, edit history on shared files. Take ten cases, average them, and mark the figure as a rebuilt baseline. It can still be compared; it cannot be presented as if it had been captured at the time.
四個指標要全部做嗎？Do we need all four indicators?
不需要。多數公司由兩項開始：周期時間與例外比率，因為兩者都可以由現有系統出數，不必另外做表。共用提示庫建立之後，再加重用。可用產出最花時間，通常在第二輪才做，因為它需要一個寫得夠窄的「可用」標準。No. Most companies start with two, cycle time and exception rate, because both come out of systems they already run. Add reuse once a shared prompt library exists. Usable output takes the most preparation and usually comes in the second round, because it needs a narrow enough definition of usable.
量一次之後，多久再量？After the first measurement, how often should we repeat it?
第 30 天之後，按季度再量一次，用同一個抽樣方法與同一組定義。第一年量四次，夠看到趨勢；每次都換定義，四次數字加起來會比一次更不可靠。如果流程本身在改，季度那一次要註明同時改動了甚麼。After day 30, measure quarterly using the same sampling method and the same definitions. Four readings in the first year are enough to show a trend; changing the definitions each time makes four readings less reliable than one. If the process itself is changing, note what else changed alongside the quarterly figure.
滿意度分數還有用嗎？Is the satisfaction score useless now?
有用，用在內容檢查。一整班人給低分而理由是「與我的工作無關」，這是一個要即刻處理的訊號，因為內容錯了，後面量甚麼都沒有意義。它的用途到這裡為止：它告訴你課室的反應，答不到工作的改變。It is useful as a content check. A whole class scoring low and saying the content has nothing to do with their work is a signal to act on immediately, because the content is wrong and anything measured afterwards is meaningless. That is the end of its use: it reports the room, and cannot answer what happened to the work.
你們會替我們寫回報率報告嗎？Will you write the return-on-investment report for us?
不會。我們提供方法、定義與紀錄的格式，並在第 30 天與你一起對數字；數字本身由你的系統出，報告由你的同事簽。這樣做的原因很實際：回報要經得起你自己的審計，一份由供應商寫的回報，在預算會議上說服力最低。No. We provide the method, the definitions and the format for the records, and we go through the numbers with you on day 30. The figures come out of your systems and your own colleagues sign the report. The reason is practical: a return has to survive your own audit, and a return written by the supplier is the least persuasive document in a budget meeting.
