一間 18 人的香港公司上完一輪 AI 培訓,用四星期試了一條報價與合規查核的流程。共用資料夾裡放著一份舊費率表與一份修訂版,附加費文件用跨頁合併儲存格排版。第一版 AI 助手偶爾引用舊價,偶爾把一個港口的附加費接到另一個港口。A Hong Kong client with 18 staff finished a round of AI training and then ran one quotation and compliance task for four weeks. The shared folder held an old rate card and a revised one, and the surcharge documents used merged cells across pages. The first version of the AI assistant sometimes cited the old price and sometimes attached one port's surcharge to another port.
我們改了兩件事:為每份文件加上「生效中」或「已歸檔」的標籤與生效日期,搜尋時先按日期過濾;把文件按版面解析成結構化資料,每個答案附上來源檔案與頁碼。同一條任務的平均處理時間由 38 分鐘降至 11 分鐘,兩個數字都包含同事最後核對的時間[3]。兩星期後,18 人裡有 14 人每星期至少用三次[3]。We made two changes: every document carries an "active" or "archived" label with an effective date and retrieval filters on it, and the documents are parsed by layout into structured data with each answer pointing back to its file and page. Average handling time on that one task fell from 38 minutes to 11, and both figures include the colleague's check at the end[3]. Two weeks later, 14 of the 18 were using it at least three times a week[3].
這間公司的課程內容與其他客戶差不多。真正決定它會不會留下來的,是第一個工作流程:一件每星期重複、有文件可以對答案、最後有人核對的工作。培訓之後停下來的情況,我們見到的次序相當固定:第一個流程選得太大;最後的檢查沒有指定由誰做;第一個用的人不知道例外在哪裡。三件事都可以在開課前決定。The course content for this company was much like any other. What decided whether the work survived was the first workflow: a task that repeats every week, has a document to check the answer against, and ends with a person checking it. Where training stops, the order we see is fairly fixed: the first workflow was too wide, the check at the end was never assigned, or the first person using it did not know where the exceptions sit. All three can be settled before the course starts.
四件事決定這條流程會不會留下來:第一個工作流程的三個條件、第一班人的三個問題、三十日紀錄的四個欄位,以及政府 5,000 萬元計劃與公司自己試行的分別。每個數字都連著它的基數,我們做錯的地方也寫下來。Four questions decide whether the workflow survives: the three conditions for a first workflow, three questions about who to send first, the four fields of a 30-day record, and how the HK$50 million public programme differs from a company's own trial. Every figure carries its base, and the places we got it wrong are written down too.
摘要Summary
- 培訓之後工作停下來,通常因為第一個工作流程選得太大,或者最後的檢查沒有指定由誰做[3]。Work stops after training because the first workflow was too wide, or because the check at the end was never assigned[3].
- 一個 18 人的報價流程試了四星期:文件加上生效標籤、答案附上來源頁碼之後,同一任務的平均處理時間由 38 分鐘降至 11 分鐘,兩個數字都包含核對時間[3]。One quotation workflow ran for four weeks with 18 staff: after documents carried effective-date labels and answers carried their source page, average handling time on that task fell from 38 minutes to 11, both figures including the check[3].
- 兩星期後,18 人裡有 14 人每星期至少用三次[3]。Two weeks later, 14 of the 18 were using it at least three times a week[3].
- 第一個工作流程要同時符合三個條件:同一件事每星期重複、有一份文件可以對答案、最後有一個人核對。A first workflow meets three conditions at once: the task repeats every week, a document exists to check the answer against, and a person checks it at the end.
- 第一班人由一位熟悉例外的同事配一位打字快的同事;22 人的工作坊裡,最實用的一個審批流程由一位 50 歲、不懂寫程式的營運主管在 45 分鐘內建成[3]。The first cohort pairs one colleague who knows the exceptions with one who types fast; at a workshop for 22 managers, the most practical approval workflow was built in 45 minutes by a 50-year-old operations head with no coding background[3].
- 三十日紀錄只有四個欄位,第 30 天的會議只作一個決定:保留、收窄或停止。The 30-day record has four fields, and the meeting on day 30 makes one decision: keep it, narrow it or stop it.
- 政府在 2026 至 27 年度撥款 5,000 萬元推行「全民 AI」普惠計劃,目標在今明兩個財政年度舉辦超過 200 項活動、預計惠及約 5 萬人次[1][2]。The Government put HK$50 million into the AI for All scheme in 2026/27, aiming at more than 200 activities and about 50,000 attendances across this and the next financial year[1][2].
培訓之後,為甚麼工作停了下來?Why does the work stop after the training?
我們見到的原因有三個,次序相當固定:第一個工作流程選得太大;最後的檢查沒有指定由誰做;第一個用的人不知道例外在哪裡。課程內容很少是原因:我們交付過的課程,內容差別不大,留下來的比例卻差別很大[3]。We see three reasons, in a fairly fixed order: the first workflow was too wide, the check at the end was never assigned, or the first person using it did not know where the exceptions sit. Course content is rarely the reason: the courses we have delivered differ little in content and differ a great deal in how much of the work survived[3].
這間公司把三件事都遇上了。舊費率表與修訂版同時放在共用資料夾,助手看不出哪一份有效;附加費文件的合併儲存格令行與行對不上,答案把一個港口的附加費接到另一個港口;同事要核對一份答案,得由頭讀一次。三件事的性質相同:文件本身沒有告訴機器自己是哪一個版本,也沒有一條路徑由答案回到原文。This company met all three. An old rate card and a revised one sat in the same shared folder and the assistant could not tell which one applied; merged cells in the surcharge documents broke the row-to-row mapping and the answer attached one port's surcharge to another; and a colleague checking an answer had to read the whole thing again. All three have the same cause: the document did not say which version it was, and nothing led from the answer back to the page.
| 出錯的地方What went wrong | 我們改的地方What we changed | 之後的紀錄What the record showed |
|---|---|---|
| 助手引用舊費率表的價錢,因為兩份文件在同一個資料夾裡。The assistant cited the old rate card, because both versions sat in one folder. | 每份文件加上「生效中」或「已歸檔」的標籤與生效日期,搜尋時先按日期過濾[3]。Every document carries an active or archived label with an effective date, and retrieval filters on the date first[3]. | 搜尋只會取生效中的版本,舊價留在檔案裡而不進入答案。Retrieval only takes the active version, and the old price stays in the file without entering the answer. |
| 附加費表用跨頁合併儲存格,港口與附加費的行對不上。The surcharge tables used merged cells across pages, so rows and ports did not line up. | 按版面解析成結構化資料,港口、附加費與有效期逐行對應[3]。Layout-aware parsing turns the pages into structured data, with port, surcharge and validity matched row by row[3]. | 一個港口的附加費不再接到另一個港口,出錯時可以指出是哪一行。One port's surcharge no longer lands on another, and an error can be traced to a row. |
| 答案看不出出自哪一頁,同事核對時要由頭讀一次。The answer did not show which page it came from, so the check meant reading it again. | 每句附上來源檔案與頁碼,核對的人可以直接跳到那一頁[3]。Each sentence carries its source file and page, so the person checking can jump straight to it[3]. | 核對由「再讀一次」變成「跳到那一頁看兩個數字」。The check changed from reading it again to opening the page and looking at two figures. |
38 分鐘變成 11 分鐘,靠的是三處改動一起生效,而 11 分鐘裡仍然有一個人在看。抽走最後的核對,時間會再短幾分鐘,價錢會錯,第四星期就不會有人再用。留住流程的是那個核對,所以第一個流程要選一件錯得起、也查得出的工作。The fall from 38 minutes to 11 came from the three changes together, and those 11 minutes still contain a person looking. Take the check away and the task gets a few minutes faster, the prices go wrong, and by the fourth week nobody uses it. The check is what keeps the workflow alive, which is why the first task should be one where a mistake is both survivable and findable.
第一個工作流程要符合甚麼條件?What conditions does the first workflow have to meet?
三個條件要同時成立:同一件事每星期出現、有一份文件可以對答案、最後有一個人核對。三個都成立,這條流程值得做第一個;缺一個,第 30 天就會停,而停的原因與課程無關。Three conditions have to hold at once: the task repeats every week, a document exists to check the answer against, and a person checks it at the end. With all three, the workflow is worth doing first; with one missing, it stops by day 30, and the reason has nothing to do with the course.
每星期重複,才有一個月前後可以比較的時間,也才有足夠的個案。一個月做一次的流程要等一季才看得出分別,等到第三個月,問題通常已經不了了之。文件帶來一個對錯的標準:報價有費率表,合規有條款,採購有批核紀錄,答案可以逐項對。最後的核對帶來責任:核對的人在紀錄上簽名,出錯時知道找誰,流程才有機會被修正。Weekly repetition is what gives a month of comparable before-and-after, and enough cases to compare. A task that runs once a month takes a quarter before a difference shows, and by the third month the question has usually been dropped. A document brings a standard for right and wrong: quotations have rate cards, compliance has clauses, procurement has approval records, and an answer can be checked line by line. The check brings responsibility: the person checking signs the record, an error has a name attached to it, and the workflow can be corrected.
| 候選流程Candidate workflow | 每星期重複Repeats weekly | 可以對答案的文件Document to check against | 誰做最後的核對Who checks it | 判斷Verdict |
|---|---|---|---|---|
| 報價與合規查核,每一張報價都要對費率表與條款。Quotation and compliance checks, where every quote is matched to a rate card and the terms. | 每星期都有,旺季更多。Every week, and more in a busy season. | 費率表與條款文件齊備,價錢與條文可以逐項對。The rate card and the terms are on file, so price and clause can be checked line by line. | 報價的同事在發送前核對一次。The colleague sending the quote checks it once. | 適合做第一個,錯誤查得出,代價有限。Good first choice: errors are findable and the cost of one is limited. |
| 每月管理報告,把各部門的數字整理成一份。The monthly management report, which pulls departmental figures into one document. | 每月一次,一個月只有一個樣本。Once a month, which gives one sample a month. | 數字在系統裡,寫法有一套慣例。The figures sit in the system and the write-up follows a house convention. | 部門主管在提交前審閱。The department head reviews it before it goes up. | 適合第二個,重複次數太少。Better as the second: it repeats too rarely. |
| 客戶查詢回覆,把查詢分類並起草第一封回覆。Customer enquiries, sorted by type and drafted into a first reply. | 每星期數十封,旺季更多。Dozens a week, and more in a busy season. | 回覆對錯由客戶決定,內部只有語氣與條款的標準。The customer decides what is right, and internally only tone and terms are set. | 同事在發送前讀一次。A colleague reads it before sending. | 可以做,先把查詢分類,再選其中一類。Workable if the enquiries are sorted first and one type is chosen. |
| 會議紀錄,把一場會議的內容變成可以跟進的清單。Meeting minutes, turning a meeting into a list that can be followed up. | 每星期都有幾場。Several every week. | 錄音可以對,取捨有一套慣例。The recording can be checked against, and the write-up follows a convention. | 會議主持人在發出前確認。The chair confirms it before it goes out. | 可以做,先把格式固定,再處理內容。Workable if the format is fixed before the content. |
| 履歷篩選,把申請人按職位要求排序。CV screening, ranking applicants against the job requirements. | 招聘季節每星期都有。Every week during a hiring season. | 篩選標準要自己寫,寫得不清楚就沒有對錯可言。The screening standard has to be written first, and an unclear one leaves nothing to check against. | 招聘同事在面試前看一次。The hiring colleague reviews it before interviews. | 最後才做,標準寫清楚之前會放大偏見。Do it last: before the standard is written, it amplifies bias. |
十分鐘可以做完這個判斷。三個條件都答「是」,下星期開始試;兩個答「是」,把任務收窄一半,例如只做其中一個客戶或其中一種報價;一個答「是」,換一件工作。收窄比加大容易,而且第二個月可以加回來。把流程加大到覆蓋整個部門,是我們見過最常見的第一個錯誤。The judgement takes ten minutes. Three yes answers mean starting next week; two mean cutting the task in half, for one client or one kind of quote; one means choosing another task. Narrowing is easier than widening, and the second month can add work back. Widening the workflow to cover a whole department is the most common first mistake we see.
第一班應該找誰?Who should be in the first cohort?
找一個熟悉例外的人,配一個打字快的人,兩個人一組。熟悉例外的人寫得出規矩,打字快的人試得出工具的邊界,兩個人一起做,第一條流程才會在兩星期內成形。這個組合還有第二個好處:一個星期內就會有人開始改,改動留在共用的檔案裡,第二個人接手時不用再問。Take one person who knows the exceptions and pair them with one who types fast. The first writes the rules, the second finds where the tool breaks, and together they get a first workflow into shape within two weeks. The pairing has a second advantage: someone starts changing the workflow within a week, the changes stay in the shared file, and the next person does not have to ask again.
22 位管理人員的工作坊裡,我們原本預期年輕同事領先。最實用的一個審批流程,由一位 50 歲、不懂寫程式的營運主管在 45 分鐘內建成,因為他先寫下二十年來的客戶例外與合規底線,再把每一條變成一個判斷[3]。年輕同事打字更快,但未必分得出一個流暢的答案與一個正確的答案。工作坊之後我們改了兩件事:第一小時用來審核輸出,第二小時才生成內容;每一組配一位熟悉業務的同事與一位熟悉工具的同事。At a workshop for 22 managers we expected the younger staff to lead. The most practical approval workflow was built in 45 minutes by a 50-year-old operations head with no coding background, because he first wrote down twenty years of customer exceptions and compliance red lines and then turned each one into a decision[3]. Younger staff typed faster, yet could not always tell a fluent answer from a correct one. Two changes followed that workshop: the first hour is spent auditing output and only the second hour generates content, and every group pairs someone who knows the business with someone who knows the tool.
「以前以為學 AI 係要我哋呢啲唔識寫程式嘅人由頭學做 IT;上完先發現,其實係將我做咗十幾年嘅判斷經驗『寫成規矩』教識部機,經驗反而變得更值錢。」"I thought learning AI meant people like us, who cannot code, starting over as IT. What I found was that I was writing more than a decade of judgement into rules and teaching them to the machine, and the experience became worth more."
一位企業培訓班上的營運主管,匿名引述,同意書日期為 2026 年 10 月 29 日[3]。A senior operations executive in a corporate training cohort, quoted anonymously, with consent dated 29 October 2026[3].
三個問題可以代替直覺。第一個問他平日要處理多少例外,第二個問他會不會每天指出工具答錯的地方,第三個問他願不願意在第 30 天改一次流程。Three questions do the work of intuition. The first asks how many exceptions the person handles in a normal week, the second whether they already point out where a tool is wrong, and the third whether they would change the workflow once at day 30.
| 問題Question | 答案指向What the answer points to | 為甚麼Why it matters |
|---|---|---|
| 他平日要處理多少例外?How many exceptions does this person handle in a normal week? | 例外愈多,他愈寫得出規矩。The more exceptions, the more rules the person can write. | 例外是流程裡最難寫的部分,工具要靠這部分才做得對。Exceptions are the hardest part of a workflow to write down, and they decide whether the tool gets it right. |
| 他會不會每天指出工具答錯的地方?Does this person already point out where a tool is wrong, most days? | 會,他自然成為那個核對的人。If yes, this is the person who will do the check. | 核對需要一個願意說「這句錯了」的人,紀錄才不會停在簽名。The check needs someone willing to say an answer is wrong, or the record becomes a signature. |
| 他願不願意在第 30 天改一次流程?Would this person change the workflow once at day 30? | 願意,流程會隨著例外更新。If yes, the workflow keeps up with the exceptions. | 第 30 天的改動通常很小,例如加一個欄位或改一句提示,之後的半年靠它。The day-30 change is usually small, one field or one sentence in a prompt, and the next six months rest on it. |
三個答案落在「多」、「會」、「願意」,第一班就有了一個會用的人;第一個問題的答案是「少」,就先換人,換工具解決不了。這三條是我們從自己的班觀察到的,寫成檢查表方便使用,沒有經過外部驗證;它們提高機會,沒有保證。Three answers of "many", "yes" and "willing" give the first cohort someone who will use it; a first answer of "few" means changing the person first, and changing the tool will not fix it. These three come from our own cohorts, written down as a checklist for use, with no external validation: they raise the likelihood and offer no guarantee.
三十日裡要記錄甚麼?What gets recorded over 30 days?
四星期的試行只記四個欄位:日期與人、任務、時間、核對結果。每次使用一行,一行填完不超過一分鐘,第 30 天才有數據可以看。The record for a four-week trial has four fields: the date and the person, the task, the time, and what the check found. One row per use, and a row takes under a minute to fill, and by day 30 there are numbers to look at.
| 欄位Field | 記甚麼What goes in it | 為甚麼要它Why it is there | 一個例子An example |
|---|---|---|---|
| 日期與人Date and person | 哪一位同事、哪一天,寫名字或員工編號都可以。Which colleague on which day, as a name or a staff number. | 沒有名字的紀錄看不出是全部人在用,還是兩個人在用。A record without names cannot tell a whole team from two people. | 10 月 8 日,營運部同事。8 October, an operations colleague. |
| 任務Task | 哪一條流程、哪一個客戶或案件,寫一個可以對上的編號。Which workflow and which client or case, as an identifier that can be matched. | 同一條流程在不同客戶上的時間差很遠,混在一起看沒有意思。One workflow takes very different times on different clients, and mixing them says nothing. | 報價查核,客戶甲。Quotation check, client A. |
| 時間Time | 由開始到完成,包括最後核對的時間。From start to finish, including the final check. | 不計核對的時間會令流程看起來比真實快,第 30 天就無法作決定。Leaving the check out makes the workflow look faster than it is, and day 30 then has nothing to decide on. | 14 分鐘,其中 3 分鐘核對。14 minutes, 3 of them on the check. |
| 核對結果What the check found | 改了甚麼、改了哪一類,例如價錢、日期、條款或語氣。What was changed and of what kind, such as a price, a date, a clause or the tone. | 錯誤的類別決定下一步改文件、改提示,還是改流程。The kind of error decides whether the next change is to the document, the prompt or the workflow. | 改了兩個價,兩個都出自舊費率表。Two prices changed, both from the old rate card. |
不記登入次數。登入次數會隨通知、會議與假期上落,與工作有沒有做完無關。也不記「有沒有用過 AI」,那個問題的答案永遠是「有」。第 30 天只看兩個數字:仍然在使用的人數,與同一條任務的中位時間,兩者都連著核對。18 人的公司在第 14 天量到 14 人,在那兩個星期裡,每人每星期至少用三次[3];中位時間由 38 分鐘走到 11 分鐘,同一個起點與終點都寫在紀錄裡。Logins are not recorded. They rise and fall with notifications, meetings and leave, and they say nothing about whether the work got done. Neither is "has anyone used AI", a question whose answer is always yes. Day 30 looks at two numbers: how many people are still using it, and the median time on the same task, both including the check. The company with 18 staff showed 14 users at day 14, each of them using it at least three times a week[3], and the median time moved from 38 minutes to 11, with the same start and end written into the record.
第 30 天開一次會,會議只作一個決定。出席的人是做核對的同事與可以批准下一筆支出的人。保留,代表第二個月加一件同類工作;收窄,代表把任務切一半,例如只做一種報價;停止,代表這一條流程完結,換下一條,工具與檔案留在原位。三個決定都比「再觀察一個月」有用,觀察一個月的結果通常是第 60 天還在觀察。One meeting on day 30 makes one decision, and the people in the room are the colleague who does the check and whoever can approve the next spend. Keep means adding one more task of the same kind in month two; narrow means cutting the task in half, for one kind of quote; stop means the workflow ends, the next one starts, and the tool and the files stay where they are. All three beat watching for another month, which usually ends with the same meeting on day 60.
試行之前先寫一頁,頁上有六個欄位Write one page before the trial, with six fields on it
一頁紙寫清楚六個欄位,試行就少了很多爭執。這一頁由提出試行的人寫,由批准支出的人簽,之後的紀錄與第 30 天的會議都對著它。One page that settles six fields removes most of the arguments from a trial. The person proposing the trial writes it, whoever approves the spend signs it, and the record and the day-30 meeting both work from it.
| 欄位Field | 寫甚麼What goes in it | 沒有它會怎樣What happens without it |
|---|---|---|
| 流程的起點與終點Where the workflow starts and ends | 寫兩個動作,例如由收到報價要求到回覆客戶。Name two actions, for example from receiving a quote request to replying to the client. | 範圍會在第三個星期自己長大,加到未經同意的步驟。The scope grows on its own in week three, into steps that were never agreed. |
| 十個個案The ten cases | 寫明哪十個,包括一個旺季個案與一個例外。Name the ten, including one from a busy week and one exception. | 紀錄會混入不同難度的個案,時間的比較失去意義。The record mixes cases of different difficulty, and comparing times stops meaning anything. |
| 基線The baseline | 現在的平均時間與出錯次數,由系統或人手量一次。Today's average time and error count, measured once from the system or by hand. | 第 30 天沒有起點可以比,改善講不出來。Day 30 has no starting point, so the improvement cannot be stated. |
| 核對的人The person who checks | 一位同事的名字與職位,以及他會看哪一項。One colleague's name and role, and the one item they will look at. | 核對會落到「大家」身上,最後由誰做都不清楚。The check lands on everyone, and it is unclear who does it. |
| 第 30 天的決定The day-30 decision | 保留、收窄或停止,並寫明誰有權決定。Keep, narrow or stop, and who has the authority to decide. | 試行沒有終點,第二條流程永遠不會開始。The trial has no end, so the second workflow never starts. |
| 不做的範圍What is out of scope | 哪些客戶、哪些任務、哪些資料不在這次試行內。Which clients, tasks and data are outside this trial. | 例外會變成慣例,資料的界線也會模糊。Exceptions become the norm and the data boundary blurs. |
第六個欄位最常被跳過,也最常在前兩星期出事。資料的界線可以照〈企業 AI 治理〉的三頁政策寫,試行的範圍寫在這一頁,兩份文件放在一起看。The sixth field is skipped most often and causes trouble in the first fortnight. The data boundary can follow the three-page policy in corporate AI governance, and the scope of the trial belongs on this page, with the two documents read together.
這一頁紙不需要軟件,一份文件就可以。寫完之後,課程的內容由它長出來,第 30 天的會議也對著它開。The brief needs no software, and one document is enough. The course content then grows out of it, and the day-30 meeting runs against it.
政府資助的課程與公司自己的試行有甚麼分別?How do publicly funded courses differ from a company's own trial?
資助的課程回答「由誰教」,公司的試行回答「同事會不會用」。兩者的紀錄不同:一邊數活動與人次,另一邊數第 30 天仍然在用的同事。兩件事可以同時安排,很多公司用同一筆培訓預算做完兩件。A funded course answers who teaches it, and a company's trial answers whether colleagues use it. The records differ: one counts activities and attendances, the other counts the colleagues still using the workflow on day 30. Both can run at once, and many companies do both on one training budget.
2026 至 27 年度的《財政預算案》撥款 5,000 萬元推行「全民 AI」普惠計劃,由數碼港、香港科技園公司與香港生產力促進局三個機構協作推行,目標在今明兩個財政年度舉辦超過 200 項活動與課程,預計惠及約 5 萬人次[1][2]。三個機構的分工寫得清楚:數碼港負責學生、家長、長者與弱勢社群的基礎普及,香港科技園負責大學生、青年研究員與專業人士的進階應用,生產力促進局負責中小企與在職人士的實務提升[1]。The 2026/27 Budget put HK$50 million into the AI for All scheme, run jointly by Cyberport, the Hong Kong Science and Technology Parks Corporation and the Hong Kong Productivity Council, aiming at more than 200 activities and courses across this and the next financial year and reaching about 50,000 attendances[1][2]. The division of work is set out clearly: Cyberport takes basic reach for students, parents, older people and disadvantaged groups, the Science Park takes advanced application for university students, young researchers and professionals, and the Productivity Council takes practical uplift for small and medium enterprises and working people[1].
個人的進修另有持續進修基金:年滿 18 歲的香港居民在修畢基金課程後可以申領上限 25,000 元,基金課程涵蓋人工智能,政府表示暫時沒有計劃另設「AI 專項」資助。僱員再培訓局升格為技能提升局之後,會繼續提供涵蓋人工智能應用的技能為本課程[1]。Individuals have a separate route through the Continuing Education Fund: a Hong Kong resident aged 18 or over can claim up to HK$25,000 after completing a fund course, the fund's courses include AI, and the Government says it has no plan for a separate AI stream. With the Employees Retraining Board upgraded to a skills authority, it will keep running skills-based courses that cover AI applications[1].
| 推行機構Body | 主要對象Main audience | 公司可以怎樣用How a company can use it |
|---|---|---|
| 數碼港負責基礎普及[1]。Cyberport covers basic reach[1]. | 對象是學生、家長、長者與弱勢社群。Students, parents, older people and disadvantaged groups. | 安排未接觸過人工智能的同事建立基本認知,之後再談流程。Give colleagues who have never touched AI a basic footing before any workflow is discussed. |
| 香港科技園負責專業轉化[1]。The Science Park covers professional conversion[1]. | 對象是大學生、青年研究員與專業人士。University students, young researchers and professionals. | 安排技術團隊上進階課,並用應用比賽測試一條真實流程。Send the technical team to advanced classes and use an application contest to test one real workflow. |
| 生產力促進局負責實務提升[1]。The Productivity Council covers practical uplift[1]. | 對象是中小企與在職人士。Small and medium enterprises and working people. | 安排營運、銷售與行政同事上實務課,課後一星期開始試行。Send operations, sales and administrative colleagues to practical sessions and start the trial a week later. |
課程之後仍然要有試行,因為課程的紀錄停在出席。報名之前問三個問題:課程會不會留下共用的檔案或模板、哪一條任務可以在課後一星期開始試、誰會在第 30 天看紀錄。三個問題都有答案的課程,會令第一條流程在課程結束後一星期內開始,那才是它值回票價的地方。A course still needs a trial after it, because a course's record stops at attendance. Three questions before booking: does the course leave a shared file or template behind, which task can start in the week after it, and who will look at the record on day 30. A course that answers all three brings the first workflow forward by a week, and that is where its fee earns its place.
我們交付到哪裏,停在哪裏What we carry, and where we stop
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。三個培訓方向都用這個做法,分別在於深度與人數,服務範圍寫在〈企業 AI 培訓〉的四項服務裡。課後留一個共用提示庫與模板,第 30 天的一次跟進用來對兩個數字,不對課程評分。Two weeks before a course, we choose one workflow 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. Afterwards a shared prompt library and templates stay behind, and the day-30 follow-up is for the two numbers, not for rating the course.
數字由你的系統出。我們不簽發回報率證書,也不替你寫數字,只提供方法與紀錄的格式。四個欄位的表可以照抄,中位時間的算法與三個不該用的指標寫在〈企業 AI 培訓怎樣量度成效〉。四星期之後要加第二條流程,先看〈部門自動化由哪個部門開始〉的四個切入點,再決定加在哪裏。The numbers come out of your systems. We issue no return-on-investment certificate and write no figures for you, and we supply the method and the format for the record. The four-field table can be copied as it stands, and the median-time method and the three indicators to drop are set out in how to measure AI training. Before a second workflow in week four, which department to automate first sets out four entry points, and that is where the choice belongs.
我們做錯的地方也寫下來。一個約 30 人、六節 90 分鐘的學校課程,早期大約三分之一的連線嘗試失敗,因為麥克風互相干擾,Wi-Fi 同時連線也吃力。我們改成按鍵說話、單向耳機、前半節只做文字練習、分組排隊,並加一個本機版本;之後的失敗大幅減少[3]。環境是流程的一部分,公司與課堂都一樣:一個在會議室行得通的設定,換到課室就會露底。The places we got it wrong are written down too. In a school programme of about 30 students across six 90-minute sessions, about a third of connection attempts failed early on, because microphones picked each other up and the Wi-Fi struggled with the concurrency. We moved to push-to-talk with one-way headsets, kept the first half of the lesson to text, staggered the queue and added a local fallback, and the failures fell sharply after that[3]. The room is part of the workflow, in a classroom and in an office alike: a setting that works in a meeting room shows its limits in a classroom.
數字來自一條測試過的任務與一次四星期的試行,客戶的名稱、行業與月份都沒有寫;一段引述有同意書,日期是 2026 年 10 月 29 日;全頁沒有收入或利潤的數字;三個問題與四個欄位是我們的做法,未經外部驗證,22 位管理人員的觀察來自一次工作坊,寫成可以照做的檢查表,能引用的範圍就到這裏。The figures come from one tested task and one four-week trial, with the client's name, industry and month left out; one quotation carries consent dated 29 October 2026; no revenue or profit figure appears anywhere; and the three questions and four fields are our own practice, with no external validation, drawn from one workshop of 22 managers and written as a checklist a reader can follow. That is as far as the numbers can be quoted.