專題深挖入職階梯正在收窄：AI 勞動市場衝擊的三組數字與一份企業清單The narrowing entry ladder: three sets of numbers on AI's labour impact, and one checklist for employers

香港勞工及福利局 2026 年 9 月公布《人力推算》中期更新，指本港整體人力短缺由 18 萬收窄至 13 萬，主要來自輸入人才與人工智能應用的增加。同一份報告給出三個更值得留意的數字：約兩至三成職位（主要為文書）將因 AI 應用重組；適合應屆畢業生的職位空缺較 2022 年少約六成；約 7 萬名 45 至 54 歲、沒有大專學歷的文書支援人員面臨轉型壓力。Hong Kong's Labour and Welfare Bureau published the mid-term update of its Manpower Projection in September 2026, showing the overall shortfall narrowing from 180,000 to 130,000, driven largely by imported talent and rising AI adoption. The same report gives three figures that matter more: roughly 20% to 30% of positions, mostly clerical, will be restructured because of AI; vacancies suitable for fresh graduates are down about 60% against 2022; and about 70,000 clerical support workers aged 45 to 54 without post-secondary education face transition pressure.

報告同時指出，短期出現大規模裁員的風險相對較低，而個別崗位因 AI 應用可提升 20% 至 30% 生產力。把這幾句放在一起讀，結論很具體：企業今天感受到的不是人手過剩，而是同一個編制能夠承接更多工作，於是新人不再被需要；而現職同事因為產出上限提高，反而更難被取代。The report also notes that the risk of large-scale layoffs in the short term remains relatively low, while AI adoption can lift productivity by 20% to 30% in certain roles. Read together, the conclusion is concrete: what a company feels today is not surplus labour but a fixed headcount absorbing more work, so newcomers are no longer required, while incumbents become harder to replace because their output ceiling has risen.

高盛研究報告（2026 年 9 月 21 日）把 AI 對招聘的影響畫成一張地圖：美國客服中心就業較趨勢低 39%，加拿大低 33%，德國低 27%；軟件出版、管理顧問與廣告服務亦出現類似收縮，而藍領就業幾乎沒有變動。報告的判斷是，招聘轉弱幾乎完全集中在知識型服務業——也就是那些「現成自動化工具已經存在」的行業。A Goldman Sachs research report dated 21 September 2026 maps where AI is actually affecting hiring: call-centre employment sits 39% below trend in the United States, 33% below in Canada and 27% below in Germany, with software publishing, management consulting and advertising showing similar contractions, while blue-collar employment has barely moved. The report's conclusion is that hiring weakness is concentrated almost entirely in knowledge-service industries — the ones where labour-automating tools already exist.

這張地圖的價值在於它的窄。如果 AI 真的如部分預測那樣全面取代人手，我們應該看到各行各業同時收縮；實際上收縮集中在少數幾個職能，而且集中程度高得可以直接點名。對企業而言，這意味著風險評估不需要做得很大，只需要做得準。The value of that map is how narrow it is. If AI were displacing workers across the board as some forecasts suggest, we would see simultaneous contraction in every sector; instead it clusters in a handful of functions, tightly enough to name. For an employer, that means risk assessment does not need to be broad, only precise.

美國裁員公告數據顯示，AI 作為理由的熱度在 2026 年 8 月出現轉折。Challenger, Gray & Christmas 的報告指，AI 由 3 月至 7 月連續五個月成為最常被引用的裁員理由，8 月回落至第四位，該月涉及 3,462 人；自 2023 年至 2026 年 7 月，被引用的裁員公告累計 184,538 宗。同期每周新申領失業救濟人數降至 196,000（截至 9 月 12 日一週），為 1969 年以來其中一個最低讀數。US job-cut announcements show a turning point in August 2026. Challenger, Gray & Christmas reported that AI, the most-cited reason for five consecutive months from March to July, fell to fourth place in August with 3,462 cuts; cumulative announcements citing AI from 2023 to July 2026 reached 184,538. Over the same period weekly initial unemployment claims fell to 196,000 for the week ended 12 September, among the lowest readings since 1969.

史丹福數碼經濟實驗室以 ADP 實際薪酬紀錄為基礎的分析則提供另一個角度：在 AI 曝露度最高的職業中，22 至 25 歲年輕人的就業水平較同儕低約 19%，而這個差距「幾乎全部來自招聘減少，而非裁員」。換句話說，沒有人被解僱，只是原本會開的初級職位沒有開。裁員數字與招聘數字量度的是兩件不同的事，只看前者，會以為一切如常。Analysis by Stanford's Digital Economy Lab, built on ADP payroll records rather than surveys, offers another angle: employment for 22- to 25-year-olds in the most AI-exposed occupations sits about 19% below that of less-exposed peers, and that gap runs almost entirely through reduced hiring rather than layoffs. In other words, nobody was dismissed — the entry-level roles that would have been posted simply were not. Redundancy data and hiring data measure different things, and watching only the former suggests nothing has changed.

既然衝擊出現在招聘端，企業內部的監察指標就要跟著搬位置。第一，初級職位的申請與錄取比例：如果這個數字持續上升，代表市場上同類人才供過於求，你的議價能力正在改善，但也代表你正在失去培養未來主管的渠道。第二，同一工序的產出除以人手：這是最直接反映工具是否已經承接工作的比率。If the impact shows up on the hiring side, internal monitoring has to move there too. First, the ratio of applications to offers for junior roles: if it keeps rising, similar talent is in surplus and your bargaining position is improving — but you are also losing the channel that used to produce future managers. Second, output divided by headcount for the same process: the most direct measure of whether a tool has genuinely absorbed the work.

第三，職位說明書的更新頻率。當一份職務說明超過一年未修訂，通常意味著裡面已經有一部分工序實際上由工具完成，但責任仍在一個人身上。第四，可以對外披露的 AI 使用清單：監管機構在歐洲與內地都開始要求企業說明在哪些環節使用 AI，這份清單既是合規材料，也是內部管理的基礎。這四項都不需要新系統，用現有數據就可以開始做。Third, how often job descriptions are revised. A description untouched for over a year usually means part of the work is in fact done by a tool while accountability still rests with a person. Fourth, a disclosable list of where AI is used: regulators in both Europe and mainland China are now asking companies to state this, and the same list serves internal management. None of the four requires a new system; existing data is enough to start.