過去兩年，AI 行業在問「模型能做什麼」。下一個五年，真正的問題只有一個：你能讓它被做成嗎。 For two years, the AI industry asked: what can the model do? For the next five, only one question matters: can you get it done?
問題在「AI 實施鴻溝」——從 benchmark 上能做到企業裏真的被做成之間的距離。這個鴻溝需要一門獨立學科。骨架可以從醫學界 20 年前走過的 Implementation Science 裏搬過來，但必須為 AI 的高速迭代、角色模糊、監管漂移做改造。我們押注：2027 年第一個「AI Implementation」諮詢品牌出現；2028 年大學開 AI 實施科學課程；2030 年成為正式學科。 The problem is the AI Implementation Gap—the distance between what AI can do on a benchmark and what actually gets done inside an enterprise. This gap requires its own discipline. The skeleton can be borrowed from medicine's twenty-year-old Implementation Science, but it must be rebuilt for AI's pace, role ambiguity, and regulatory drift. Our prediction: by 2027, the first "AI Implementation" consulting brand emerges. By 2028, universities offer AI Implementation Science courses. By 2030, it becomes a recognized discipline.
模型每代都在漲。企業裏的項目還在原地。中間這條鴻溝，模型公司不解決，諮詢公司沒工具。 Models keep getting smarter. Enterprise projects don't move. The model companies won't close this gap. Consultancies don't have the tools.
MIT NANDA: 95% 的企業 GenAI 試點零可衡量回報；69% 的企業發現員工在用公司禁止的「影子 AI」。95% of enterprise GenAI pilots show zero measurable return. 69% of enterprises detect "shadow AI" use.
McKinsey: 88% 用上了 AI；僅 39% 拿到企業級利潤；多數改善 <5%。88% use AI somewhere; only 39% see enterprise-level profit; most see <5% improvement.
Gartner: 到 2026 年，60% 缺乏 AI-ready 數據的項目會被放棄。By 2026, 60% of GenAI projects without AI-ready data will be abandoned.
買模型的錢只佔總成本的 15-20%。剩下 80% 是「落地税」——數據清洗、工作流改造、champion 招募、合規審查、監管迭代、組織牴觸。這些沒一項出現在供應商的報價單上。 The model accounts for 15-20% of total cost. The other 80% is the Landing Tax—data cleaning, workflow redesign, champion recruitment, compliance review, regulatory iteration, organizational resistance. None of it appears on the vendor's invoice.
Klarna: 2024 年初與 OpenAI 合作上線 AI 客服，一個月處理 230 萬次對話。承諾 4000 萬美元利潤提升。2025 年 5 月 CEO 承認「步子邁太大」，退回人機混合。Partnered with OpenAI on customer service. 2.3M conversations/month. $40M profit lift promised. May 2025: CEO admits going too far; rolls back to human-AI hybrid.
McDonald's: 2021 年與 IBM 做 AI 點單，覆蓋 100+ drive-thru。顧客點一份薯條，AI 加 260 塊麥樂雞。三年後撤掉。2021 IBM partnership on AI voice ordering at 100+ drive-thrus. Customer asks for fries; AI adds 260 chicken nuggets. Pulled after three years.
這兩家虧的都是落地的錢，不是模型的錢。 Both lost on the landing, not on the model.
20 年前醫學界遇到過一模一樣的事：大量臨牀試驗證明藥有效，但醫生不按指南開藥，病人不按處方吃藥。一條循證實踐從論文走到臨牀平均要 17 年（Balas & Boren, 2000；Morris et al., 2011）。醫學界的解法叫 Implementation Science。今天還活着，120 多個研究中心、幾千篇頂會論文。 Twenty years ago, medicine faced the same problem: clinical trials proved drugs worked, but doctors didn't follow guidelines and patients didn't follow prescriptions. Translating evidence from paper to bedside took 17 years on average (Balas & Boren, 2000; Morris et al., 2011). Medicine's answer was Implementation Science. The discipline lives—120+ research centers, thousands of peer-reviewed papers.
我們要的，不是照抄醫學。是把骨架留下來，在 AI 的異質性之上重建。 We don't copy medicine. We keep the skeleton and rebuild on AI's heterogeneity.
| 維度Dimension | 原版 CFIROriginal CFIR | AI 版改造AI Adaptation |
|---|---|---|
| 創新本身Innovation | 干預的屬性Properties of intervention | + 迭代速率（你賣的是一個正在變的東西）+ Iteration velocity |
| 外部環境External | 外部政策External policy | + 監管漂移頻率（合規線月級變化）+ Regulatory drift frequency |
| 內部環境Internal | 組織文化Org culture | 數據資產、流程兼容性、工具棧Data assets, workflow fit, tooling |
| 人People | 個體特徵Individual traits | + 角色重疊度（員工不是不想用 AI，是不想用你給的那一個）+ Role overlap |
| 實施過程Process | 計劃與評估Planning & evaluation | 試點 → 灰度 → 全量 → 維護，全程持續調優Pilot → gradual → full → maintain |
上線後用 RE-AIM 維護框架管 5 個維度：Reach / Effectiveness / Adoption / Implementation / Maintenance。 Post-launch, RE-AIM runs five dimensions: Reach / Effectiveness / Adoption / Implementation / Maintenance.
| 2026 | AI 部署失敗進入財報披露季AI deployment failures enter earnings disclosure season |
| 2027 | 第一個「AI Implementation」諮詢品牌出現First "AI Implementation" consulting brand emerges |
| 2028 | 大學開 AI 實施科學課程Universities offer AI Implementation Science courses |
| 2029 | AI 實施師薪資超過一線 AI 工程師AI Implementation Specialists out-earn frontline AI engineers |
| 2030 | AI 實施學成為正式學科AI Implementation Science becomes a recognized discipline |
從「benchmark 上能做」到「企業裏真的被做成」中間的距離。The distance between what AI can do on a benchmark and what actually gets done inside an enterprise.
benchmark 越高，企業越容易高估，摔得越狠。The higher the benchmark, the more enterprises overestimate, the harder they fall.
從「AI 能做」到「AI 被做成」中間要交的税。The tax paid between what AI can do and what AI gets done.
「這是我賭上接下來五年職業方向的一件事。」 "This is the bet I'm placing on my next five years."
電郵僅用於發送報告與學科進展，不分享給第三方。 Your email is used only for the report and discipline updates. We don't share with third parties.
讀完這份報告，下一步：診斷你企業的 AI 實施狀況。 Next step: diagnose your enterprise AI implementation.
預約診斷Book