01 ·

你是誰？先選一個鏡頭Who are you? Pick a lens

INSEAD MBA '26Kellogg exchangeISO/IEC 42001 LAISO/IEC 23894ISO/IEC 27001 LACISA (exam passed)ISO/IEC 27701BS 10012ISO/IEC 27799NVIDIA AI Infra & OpsJHU AI for CybersecurityMcKinsey ForwardEU AI Act

[email protected] · linkedin.com/in/sy8888 · Singapore / Taipei

AIOM Outcome Cockpit可以實際操作的子專案：Monday brief、結果 vs baseline、震動偏差、KPI 象限、生命週期。這裡用的是示範資料。A working sub-project: Monday brief, outcome vs baseline, vibration deviation, KPI quadrant, lifecycle. Sample data.

開啟 ↗Open ↗

接著看：AI outcome 要怎麼控制Below: how an AI outcome is controlled↓

AI outcome 怎麼控制：就靠一個迴路How an AI outcome is controlled: one loop

五個節點，每一個都對到一章。治理跟資安是第三個節點，不是放在附錄。自動播放會停在這個鏡頭最在意的那個節點。Five nodes, one chapter each. Governance and security are the third node, not an appendix. Autoplay stops on the node this lens cares about most.

這一路是怎麼學來的Where it was learned, in order

從左到右照時間走：先是台大做醫學影像的機器學習，再來是 Deloitte 的資安跟 Zero Trust，然後 iKala 的企業 AI，2025 到 2026 才是 INSEAD，還有芝加哥 Kellogg 的交換。Left to right, in time order: machine learning for medical imaging at NTU, then cyber and Zero Trust at Deloitte, then enterprise AI at iKala, and only in 2025–26 INSEAD and the Kellogg exchange in Chicago.

02 · KPI 方法論The method

從結果往回推的五個問題Five questions,
asked backwards from the result

五個問題，順序不能調換。AI 要到第四題才會出現。Five questions, always in this order. AI only shows up at question four.

每一步都有它存在的理由，但「必要」不代表「只能這樣做」。先問這一步產出的是什麼，再問最直接的路是哪一條。Every step exists for a reason. But "necessary" doesn't mean "only this way." Ask what it produces, then ask for the most direct route.

1 · 它會產出什麼？1 · What does it produce?日期、數量、門檻這種的 → 用規則就好。分類、預測 → 靠資料跟標籤。判斷、生成 → 這時候才用模型。A date, a count, a threshold → a rule. Classify or predict → data and labels. Judge or generate → then a model.

2 · 最直接的路是什麼？2 · The most direct route?先算一下不用新技術的做法要花多少錢。沒有這個基準，什麼 AI 方案看起來都划算。First cost out the no-new-tech option. Without that baseline, every AI option looks worth it.

3 · 多付的錢買到了什麼？3 · What does the extra cost buy?要往上爬一階，就要講得出數字會多動多少、風險要多扛多少。講不出來就先不要上去。Going up a rung has to buy a number I can state, and a risk I can state. If I can't, I stay down.

這不是標準，是問問題的順序。越下面越便宜，先從下面問起。It's not a standard, it's an order of questions. Cheaper at the bottom, so start there.

先從數字開始，找到產生這個數字的流程，再找出最值得改的那一步，挑最便宜又改得動它的方法。然後決定星期一誰要開始做不一樣的事。Start with the number. Find the process behind it. Find the one step worth changing. Pick the cheapest thing that changes it. Then decide who does something different on Monday.

03 · 每週一早上的一頁The page I read on Monday morning

Outcome Dashboard

一頁只看一個數字，其他內容都是在解釋它為什麼會動。示範資料用的是捷運手扶梯的預測性維修。One number per page. Everything else explains why it moves. Illustrative data: metro escalator maintenance.

這一章只是靜態示意；完整互動版放在This chapter is a static sketch; the working version is the AIOM Outcome Cockpit ↗

KPI vs 承諾KPI vs commitment

採用（leading）Adoption (leading)

舊表單還有 35% 的人在用：舊流程不退場的話，數字就會彈回去。35% still on the old sheet. If it doesn't retire, the number bounces back.

Baseline: Jan–Jun avg, SOW §4.2Source: CMMS + sensor gateway, client system of recordReviewed weekly with client示範資料Illustrative data

#	承諾（SOW）Promised (SOW)	做出來的（我驗證）Built (verified by me)	狀態Status	行動Action
M1	12 個月的感測跟維修歷史資料都進到 pipeline 裡12 months of sensor + maintenance history in the pipeline	9 個月；2024 Q1 的檔案格式不一樣，還沒解析9 months; 2024 Q1 files in a different format, unparsed	Drift	我自己寫了解析腳本把資料補齊，下週再重跑 baselineWrote the parser myself; baseline re-run next week
M2	模型 v1 precision ≥ 0.70 on holdoutModel v1 precision ≥ 0.70 on holdout	0.74，holdout 的定義已經核對過0.74, holdout definition checked	On promise	—
M3	警示會自動建立 CMMS 工單Alerts auto-create CMMS work orders	現在是每天匯出 CSV，再用人工建單Daily CSV export, manual ticketing	Drift	IT 那邊的 API 權限卡了 6 週；先用 n8n 寫進去，等 API 開通之後再切換過去Client IT API access stuck 6 weeks; n8n writes for now, switch when API opens
M4	20 台先試行，技師也訓練完了20 units piloted, technicians trained	20 台已上線，18 位技師都訓練完了20 live, 18 technicians trained	On promise	—
M5	舊的 Excel 巡檢表要退場Old Excel inspection sheet retired	還在做，有 35% 還在用In progress, 35% still using it	In progress	跟站務主管約時間談，把退場日期定下來Meeting station managers to set a retirement date

Built 那一欄，是我自己跑過、看過、打開過之後才寫上去的，不是從狀態報告抄來的。I write the Built column after I've run it, read it and opened it myself. Not from a status report.

升級紀錄（一定會附上解法）Escalation log (always with a fix)

W	問題Issue	建議解法Recommended fix	狀態Status
19	CMMS API 權限延誤，M3 會影響到 Q2 的數字CMMS API delay will hit the Q2 number	n8n 先暫代寫入，兩週就能上n8n as interim writer, live in two weeks	Open
21	兩站的技師不信警示，還是手動覆核Two stations don't trust alerts, re-check manually	先拿前 10 次警示的命中紀錄給他們看Show them the hit record of the first 10 alerts	Open
15	Baseline 期間有碰到一次電力中斷Baseline window included a power outage	照 SOW 的 exclusions 先剔掉，再重算一次Excluded per SOW, baseline recalculated	Closed

證據鏈跟下一個機會Evidence trail & next opportunity

機會Opportunity	為什麼是現在Why now	估值Est.
電梯模組，同一個站點Elevator module, same stations	同一套感測跟工單流程可以重複用Same sensor + work-order flow reusable	SGD 180K
第二線，20 台Second line, 20 units	試行 vs 未試行的對照已經成立Pilot vs control comparison established	SGD 220K

每個月的 KPI 是客戶的 Facilities 出的，不是我們自己算的Monthly KPI issued by the client's Facilities team, not computed by us
避開的停機損失，用客戶財務那邊的每小時成本換算Avoided downtime valued at the client's own hourly cost

哪個 KPI 先做？看三個分數：影響力、業務衝擊（越低越好）、導入成本（越低越好）。你可以改分數、拉權重，排名會即時更新。Which KPI first? Three scores: impact, disruption (lower is better), cost (lower is better). Change a score, drag a weight, the ranking updates.

Net Gain = Impact × 20 · Investment Cost = (w_D·Disruption + w_C·Cost) ÷ (w_D+w_C) × 20 · Priority = w_I·Net Gain − (1−w_I)·Investment Cost · ROI% = Priority ÷ Investment Cost × 100

影響力權重Impact weight w_I 0.50

衝擊性權重Disruption weight w_D 0.50

成本權重Cost weight w_C 0.50

#	候選 KPICandidate KPI	影響Impact	衝擊Disrupt	成本Cost	Net Gain	Inv. Cost	Priority	ROI%

左上角是 Net Gain 高、Investment Cost 低的，先做這些。氣泡越大代表 ROI 越高，藍色的是前兩名。Top-left is high Net Gain, low Investment Cost: do first. Bigger bubble, higher ROI; blue marks the top two.

04 · 每一個承諾的數字，都有一張卡One card per committed number

KPI CardKPI Card

交付物是我交給你的東西；成果是你的數字真的有動。SOW 裡這兩個都要寫進去。A deliverable is what I hand you. An outcome is your number moving. The SOW needs both.

一條線上有三個點：起點、承諾、現在。One line, three marks: where we started, what we promised, where we are.

Baseline 120 h用上半年的平均，在 SOW 裡簽認。沒有 baseline，你就沒辦法說它「有動」。First-half average, signed in the SOW. No baseline, no "it moved."

Target 84 h · Q4−30%。而且要連續一季都達標才算數。−30%. It has to hold for a full quarter.

Guardrails誤報要 ≤ 15%，計畫性維修最多只能 +10%。不可以用多排維修的方式把停機「消滅」掉。False alerts under 15%, planned maintenance up no more than 10%. You can't fix downtime by scheduling more downtime.

如果 leading 指標畫不到 KPI，KPI 又畫不到錢，那它就是虛榮指標。If I can't draw a line from the indicator to the KPI, and from the KPI to money, it's a vanity metric.

客戶不肯承諾數字怎麼辦？Client won't commit a number? 那我先承諾量測方法。四週把 baseline 建起來，再一起定數字。Then I commit to the measurement. Four weeks for a baseline, then we set the number together.

每月非計畫停機時數Unplanned downtime · h / monthMRT escalators · 範例example

Business question: 停機造成多少服務中斷，還有多少緊急出勤成本How much service disruption and emergency call-out cost does downtime create
Metric & formula: Σ 非計畫停機分鐘 ÷ 60，每月計算，計畫性維修不算進去Σ unplanned downtime minutes ÷ 60, monthly, excluding planned maintenance
System of record: CMMS 工單的時間戳記；由客戶的 Facilities 部門持有CMMS work-order timestamps; owned by the client's Facilities team
Grain & window: 每台 / 每月；每月出報，每季判定Per unit / per month; monthly report, quarterly judgement
Baseline: Jan–Jun 平均 120h，在 SOW §4.2 簽認過Jan–Jun average 120h, signed in SOW §4.2
Target & deadline: 84h · Q4 月底Q4 month-end
Exclusions: 電力中斷、人為破壞、供應商召回Power outage, vandalism, vendor recall
Guardrails: 誤報率 ≤ 15%；計畫性維修的時數增加不能超過 10%False-alert rate ≤ 15%; planned maintenance hours up no more than 10%
Leading indicators: 警示在 24h 內處理掉的比例、工單是從模型來的比例Alerts handled within 24h; share of work orders from the model
Owner (client): Head of Station Facilities
Attribution rule: 試行的 20 台 vs 沒試行的 20 台，做對照20 piloted units vs 20 untouched units

05 · 當 KPI 是一段爬坡，而不是一個點When the KPI is a climb, not a point

漸進式 KPI：爬梯子The progressive KPI:
climbing the ladder

有些成果就像爬梯子。你得先爬上一階，才看得到下一階，所以 KPI 要跟著梯子調。Some outcomes are a climb. Each rung shows you the next one, so the KPI moves with the ladder.

示範用的是 Zero Trust 的 Identity 支柱，照 CISA ZTMM 分四級。Example: the Identity pillar of Zero Trust, on CISA's four levels.

梯子是怎麼讓 KPI 動起來的How the ladder keeps the KPI honest

每一階都有一張 KPI card。量的是這一階的條件做到了多少。One KPI card per rung. It measures how much of that rung is done.
只承諾下一階。爬到了之後，再來估下一階。合約就用 change order 一階一階往上加。Commit to the next rung only. Re-estimate when you get there. Extend by change order.
護欄是用來防偷爬的。不能改定義，也不能把風險評得比較低。Guardrails stop cheating. No redefining levels, no rating risk down.
用在哪：監管成熟度、Zero Trust、AI 治理、AI 採用都可以。Works for: regulatory maturity, Zero Trust, AI governance, AI adoption.

美國分行：先量固有風險，再決定每個領域要爬到哪一級。CAT 在 2025 年退場，換成 NIST CSF 也一樣用。US branches: measure inherent risk first, then set the level each domain has to reach. CAT retired in 2025; the same logic works on NIST CSF.

固有風險檔案：這家分行為什麼需要爬到這一階Inherent risk profile: why this branch has to climb this high

技術與連線類型Technologies & connection types

服務通路Delivery channels

線上 / 行動產品跟技術服務Online/mobile products & tech services

組織特性Organizational characteristics

外部威脅External threats

這家分行整體是 Moderate。風險越高，就要爬得越高。This branch: Moderate. Higher risk, higher climb.

領域 3 一年裡的四個 gateDomain 3, four gates in a year

Q1	Q2	Q3	Q4
Evolving 那一階的陳述 100% 都答「是」，而且證據查得到100% of Evolving statements "Yes", evidence on file	Intermediate 做到 40%，每個缺口都有 owner 跟日期40% of Intermediate; every gap has an owner and a date	完成 80%，內稽抽測也通過了80%; internal-audit sampling passed	100%，有獨立驗證，會向董事會跟主管機關報告100%; independently verified; reported to board and regulator

我爬過的另一座山：在 Deloitte 的時候起草的四構面 ESG 成熟度模型，拿了 Deloitte 創新首獎，後來變成台灣證交所 ESG 評鑑的基礎。Another mountain I've drawn: the four-dimension ESG maturity model I drafted at Deloitte won the firm's innovation first prize and became the basis of the Taiwan Stock Exchange's ESG evaluation.

同一套梯子拿來套在 AI 上。每一階都有自己的 KPI 跟 gate。The same ladder, applied to AI. Each rung has its own KPI and gate.

這裡放兩張參考圖。FSC 的 Zero Trust 指引，參考的就是 CISA 跟美國國防部的分級方式。Two reference figures. The FSC Zero Trust guidelines drew on CISA's and the US DoD's levels.

Implementing the CISA Zero Trust Maturity Model: seven steps from assess and benchmark to partner with expert — Implementing the CISA Zero Trust Maturity Model · reference figure (Napkin)

Zero Trust maturity journey from Traditional to Initial, Advanced and Optimal drawn as a mountain climb — Zero Trust Maturity Journey · reference figure

06 · 治理本身就是交付物，不是附錄Governance is a deliverable, not an appendix

AI 模型治理：三件事，每一件都要有證據AI model governance:
three things, each with evidence

要把 AI 安全、透明地接進系統，靠的是三根柱子：監督、稽核軌跡、風險評分。我拿做了六年的那些標準，把這三件事變成客戶驗收得了的東西。Integrating AI securely and transparently stands on three pillars: oversight, audit trails, risk scoring. I turn them into things a client can accept, using the standards I've worked with for six years.

ISO/IEC 42001 LAISO/IEC 23894ISO/IEC 27001 LAISO/IEC 27701BS 10012ISO/IEC 27799CISA (exam passed)EU AI Act

每個 use case 開工前先評一次，有改動就再評一次。分數會決定要不要有人在迴圈裡、要不要往上報到管理層。Score every use case before we start and again after any change. The score decides whether a human stays in the loop and whether it goes to management.

Inherent Risk = Likelihood (1–5) × Impact (1–5) · Residual Risk = Inherent × (1 − Control effectiveness) · Gate: Residual ≥ 10 → 一定要有 HITL + 管理層簽核HITL required + management sign-off；5–9.9 → 關鍵決策 HITL + 抽測HITL on key decisions + sampling；< 5 → 監控monitor

EU AI Act 的分級是另外一條線，不能拿分數去換：只要落在高風險清單裡（信用評估、關鍵基礎設施的安全元件、醫療器材），就直接走高風險的義務。ISO/IEC 23894 管的是方法，EU AI Act 管的是門檻。這裡用的是示範資料。The EU AI Act tier is a separate line, not something a score can buy down: anything on the high-risk list (creditworthiness, safety components of critical infrastructure, medical devices) takes the high-risk obligations outright. ISO/IEC 23894 supplies the method; the EU AI Act supplies the threshold. Sample data.

治理不是擺在結尾的那一章。八步裡面每一步都有一樣治理產出，第八步的 evidence pack 就是把這些全部裝訂起來。Governance isn't the last chapter. Every one of the eight steps produces one governance artefact; the evidence pack in step eight is just those, bound together.

07 · 我以前是怎麼做的How I have done it

Proposal → SOW → 上線驗收Proposal → SOW → delivery

在 Deloitte 是先談 proposal，簽約之後再寫 SOW，人天由我來算。在 iKala 則是 PoC 先跑，收簽約金再加訂閱。兩邊一樣的是：proposal 裡的每一塊，都要變成可以驗收的東西。Deloitte: negotiate the proposal, sign, then write the SOW. I priced in person-days. iKala: PoC first, signing fee plus subscription. Either way, every block of the proposal has to become something you can verify.

我對照了兩個外部的基準：一個是 McKinsey 轉型辦公室的 L0–L5 閘門，另一個是 SaaS 客戶成功那套 value realization。我的八步跟它們都對得上，差別只在我把「簽字」單獨寫成一步。Two outside benchmarks: McKinsey's L0–L5 transformation gates, and SaaS value-realization practice. My eight steps line up with both. The difference is that I make the signature its own step.

−70%一開始估的價值，到 L5 真的做出來的時候平均會掉七成，所以 pipeline 要有目標的 3 倍Initial value estimates fall ~70% by L5, so the pipeline needs 3× the targetMcKinsey, 18 APAC transformations

~30%只有三成的轉型有達到原定目標Only about 30% of transformations hit their intended targetsMcKinsey

4每個 initiative 大概四個里程碑最剛好：有早期預警，又不會變成微管理About four milestones per initiative: early warning without micromanagingMcKinsey

2–4每個帳戶定 2–4 個 SMART KPI，每一個都要帶 baseline、資料來源、target 跟 deadline2–4 SMART KPIs per account, each with baseline, data source, target, deadlineValue realization practice

我已經在做的Already in my method

KPI 要先把 baseline、資料來源、target、deadline、owner 寫清楚（KPI card）KPI with baseline, data source, target, deadline, owner (the KPI card)
量測在 onboarding 的時候就接好，不要事後才去找資料（instrumentation 本身就是交付物）Instrument during onboarding, don't hunt for data afterwards (instrumentation is a deliverable)
每週看 tactical、每月看 KPI、每季做 business review（這就是 dashboard 的節奏）Weekly tactical, monthly KPI pulse, quarterly review (the dashboard cadence)
把採用率當 leading 指標，看 activation、使用頻率跟深度Adoption as a leading indicator: activation, frequency, depth
要報給管預算的人，不是只報給使用者Report to the budget owner, not only the users

我從基準學到、要加進去的東西What I'm taking from the benchmarks

一開始的估算要打折：每個候選 KPI 的預估價值先乘以 0.3，再來排優先順序Discount early estimates: multiply each candidate's estimated value by 0.3 before ranking
每週每個 initiative 都要有一個具體動作，不用等到里程碑One concrete action per initiative every week, not just at milestones
砍指標：對結果貢獻不到 0.01% 的指標就合併掉Prune metrics: anything under 0.01% of impact gets folded in
一個 owner 最多扛三個 initiative，不然會變成 20% 的人做 80% 的事Cap owners at three initiatives, so 20% of people don't carry 80% of the load
QBR 之前先把資料對清楚，不要在主管面前吵數字Reconcile the data before the QBR, never in front of the executive

來源：Sources: McKinsey, Keeping transformations on target · Umbrex, Value Realization & Success Management

先把目標設好，再看現況，收尾一定要有人簽字。上面是 Evlos 的五個階段，藍框是我多加的三步。Target first, then the current state. And someone signs at the end. Evlos phases on top; the blue ones are the three steps I add.

最常斷在 5 跟 6 之間：上線慶祝完，舊表單其實還開在第二個螢幕上。所以退場日期要寫進 SOW，也要放上 dashboard。It usually breaks between 5 and 6. We celebrate go-live and the old sheet is still open on the second screen. So the retirement date goes in the SOW and on the dashboard.

Proposal 裡的東西In the proposal	SOW 裡要變成什麼Becomes, in the SOW	我過去的做法My past practice
價值主張、預期效益Value proposition	KPI card 上要有：數字、baseline、target、deadline、owner、量測方法KPI card: number, baseline, target, deadline, owner, method	板階 roadmap 上的每個 initiative，都有 named owner 跟 KPIEvery initiative on a board roadmap had a named owner and KPI
Deliverables 清單Deliverables list	每一項都配 acceptance criteria：怎樣才算交件、誰來簽收、幾天內要回覆Each with acceptance criteria: what counts, who signs, response days	政府案的話，期初 / 期中 / 期末報告各自都有審查會跟修正期Government: inception / interim / final reports, each with review and revision window
時程Timeline	Milestone 搭配 gate，過不了就不往下走Milestones with gates	Zero Trust：assessment 過了，才會進到 implementationZero Trust: implementation only after the assessment gate
範圍Scope	In / out of scope / client obligationsIn / out of scope / client obligations	資料、權限、人、決策時限，這些都寫成客戶的義務Data, access, people and decision deadlines as client obligations
報價Pricing	付款跟 acceptance 綁在一起；outcome-based 的部分要寫爭議處理條款Payment tied to acceptance; dispute clause for any outcome-based part	按人天計價（Deloitte）；bundle 的成本跟毛利模型（iKala）Person-day pricing (Deloitte); bundle cost and margin model (iKala)
假設Assumptions	Assumptions & dependenciesAssumptions & dependencies	資料拿得到、API 開得通、使用者可以來參加訓練Data accessible, API available, users available for training
—	Change control	政府案就走公文變更；企業案就走 change orderOfficial correspondence for government; change orders for enterprise

08 · 交付這件事是在哪裡學的Where the delivery discipline came from

從政府到銀行到企業 AIGovernment, banks,
enterprise AI

政府案教會我文件，銀行教會我稽核跟跨境，企業 AI 教會我 on-site 跟續約。Government taught me paperwork. Banks taught me audit and cross-border work. Enterprise AI taught me on-site builds and renewals.

創業：Encorphin 安可啡 · AIQ，當醫師的第二個助理The startup: Encorphin · AIQ, the doctor's second assistant

門診拖得太晚，醫生一直重複問同樣的病史。我們訪了 100 多位醫師、跑了 40 多家診所，確認瓶頸是在「聽跟問」這件事上，才開始寫第一行程式。Clinics ran late and doctors kept asking the same history questions. We interviewed over 100 doctors and visited 40-plus clinics to confirm the bottleneck was listen-and-ask, before the first line of code.

1 · 候診Waiting room病人平板問卷Patient tablet intake按科別設計的專科問卷，病人到診之前先填好。Specialty questionnaire, answered before the consult.

2 · AIQ一頁摘要One-page summary把主訴、病史、風險旗標整理成一頁，讓醫師看得完。Chief complaint, history and risk flags on one page a doctor can read.

3 · 診間Consult醫師看診Doctor consults少問重複的問題，多做一點決策；摘要是可以編輯的。Fewer repeated questions, more decisions; the summary is editable.

4 · 病歷Record進病歷跟追蹤Into the record醫師確認過才會進病歷；資料再回饋到問卷。Enters the record only after the doctor confirms; data feeds back into the questionnaire.

4 + 2家醫院 + 診所上線hospitals + clinics live

208 → 183分鐘，胸腔科一節門診（−12%）min per chest-medicine session (−12%)

−67%病人等候時間，試點科別patient waiting time, pilot department

3人的開發團隊，我自己帶；U-Start 第一階段 NT$50 萬-person dev team I led; U-Start phase one NT$500K

這是我第一次學到：先找到瓶頸，再把工具放進去。後來每個案子都是這樣開始的。SaMD 的資料隱私跟醫療資訊安全（後來的 ISO 27799、27701），我也是從這裡開始在意的。First time I learned it: find the bottleneck, then place the tool. Every engagement since has started the same way. Caring about SaMD data privacy and health-information security (later ISO 27799, 27701) started here too.

encorphin.com ↗

第一次自己帶團隊：衛生福利部的標案（2021）First time leading a team: a Ministry of Health and Welfare tender (2021)

第一次帶 5 個人。政府案教會我一件事：交付物不等於成果，每一條都會有人查。My first team of five. Government work taught me that a deliverable isn't an outcome, and every line gets checked.

留下來的習慣：承諾寫得越具體，驗收的時候就越不會吵。The habit I kept: the more specific the promise, the quieter the acceptance.

民間企業跟跨境：併購之後的全行合規審查Private sector and cross-border: a post-acquisition bank-wide compliance review

中信買下泰國第五大銀行之後，第一次做的全行合規審查。兩套制度、兩個主管機關、兩種語言。CTBC had just bought Thailand's fifth-largest bank. First bank-wide compliance review. Two operating models, two regulators, two languages.

我把 KYC 跟交易監控從開戶到申報整條流程都對映一遍，看分行實際上怎麼做、SOP 跟系統設定有沒有對得上。I mapped KYC and transaction monitoring end to end, and checked whether what the branch did, what the SOP said and what the system was set to actually agreed.

海外案子怎麼做的細節，當面聊。The overseas details are better in person.

海外專案：美國分行的 AML/KYC 跟 IT 控制審查Overseas delivery: AML/KYC and IT control reviews for US branches

兆豐在美國的四個分行：矽谷、舊金山、洛杉磯、芝加哥。做了十次以上的 AML/KYC 跟 IT 控制審查。Mega Bank's four US branches: Silicon Valley, San Francisco, LA, Chicago. Ten-plus AML/KYC and IT control reviews.

同一套評分套在總行跟每一個分行，結果才比得出來。成熟度那章的成熟度爬坡就是從這裡來的。One scoring rubric across head office and every branch, so results actually compare. The maturity climb in the maturity chapter comes from here.

官股民營：資策會（2022–2023，18 個月）Government-affiliated, privately run: III (2022–2023, 18 months)

民營的組織，但有官方的背景。用民間的速度做事，用公部門的紀律把證據留下來。Private organisation, public mandate. You work at private-sector speed but keep public-sector paperwork.

18 個月。有點像新加坡那種國家級機構，兩邊都要顧到。Eighteen months. A bit like Singapore's national bodies: you serve both sides.

從一個案子做到 US$4M：台灣金融業第一個 Zero Trust 計畫From one project to US$4M: Taiwan's first Zero Trust programme in financial services

一個諮詢案，後來變成五大銀行的多年合約。US$4M 已經簽了，路線圖排到 2028。One advisory project became multi-year accounts with the five largest banks. US$4M signed, roadmaps through 2028.

先跟客戶把驗證得了的成果鎖定下來，再跟主管機關一起寫指引，讓它變成整個產業的 baseline。評估、建置、維運全部包在一起。I locked a verifiable outcome with the client, then co-wrote the guidelines with the regulator so they became the whole sector's baseline. Assessment, build and maintenance in one bundle.

企業 AI：on-site 建置跟續約（iKala, Dec 2024 – Jan 2026）Enterprise AI: on-site builds and renewal (iKala, Dec 2024 – Jan 2026)

12 個月開出 US$2M+ 的 pipeline，自己親手成交了 6 家。還跟 FDE 一起進客戶現場。A US$2M+ pipeline in twelve months, six accounts closed myself. On site with the engineers.

下一章講的是怎麼做的。The next chapter shows how.

09 · On-site 跟 FDE 一起做On site with the forward-deployed engineer

Nexus one-boxNexus one-box

就是一台機器：裡面有 GPU、私有 LLM、RAG、agent，專門給不能上公有雲的客戶。Leadtek 出硬體，iKala 出模型、負責把東西做出來；夥伴圈有 NVIDIA、Google、Qualcomm、Wistron。能上雲的客戶就走混合架構。六個月就出貨。One box: GPUs, a private LLM, RAG and agents, for clients who can't use public cloud. Leadtek brought the hardware, iKala owned the model and delivery; NVIDIA, Google, Qualcomm and Wistron in the partner circle. Clients who could use cloud got a hybrid design. Shipped in six months.

One-box單機、少量的 GPU。單一部門的知識問答跟 agent。第一張單：醫療健檢中心，大約 US$100K。Single unit, few GPUs. One department's Q&A and agents. First deal: a health-screening provider, ~US$100K.

Rack-mounted multi-GPU看模型大小跟同時使用的人數來加 GPU。第二單是資安服務商，AI 伺服器 + WAF + 軟體 bundle，大約 US$600K。GPUs added for model size and concurrent users. Second deal: a security services firm, AI server + WAF + software bundle, ~US$600K.

Cabinet好幾台伺服器，要配合客戶機房的空間跟電力。規則是：模型大小 × 預期併發 × 機房條件，再回推出價格。Several servers, matched to the client's data-centre space and power. Rule: model size × expected concurrency × room conditions, then back into price.

工程師管 pipeline 跟檢索。我管的是放進去的東西，還有答案到底對不對。The engineer owns the pipeline and retrieval. I own what goes in, and whether the answers are right.

AI 答錯了？先問它是落在哪個桶子。每個桶子的 owner 跟修法都不一樣。AI got it wrong? First question: which bucket. Each one has a different owner and a different fix.

content~70%
來源沒寫到或寫錯 → 我來補內容、改 metadataSource missing or wrong → I fix content and metadata

retrieval
有寫到但沒抓到 → FDE 去調 chunking、索引、參數Written but not found → FDE tunes chunking, index, parameters

prompt
有抓到但答錯方向 → 一起把 prompt 改掉Found but misdirected → we fix the prompt together

model
前面三個都對還是錯 → 那就換模型或加規則。這種最少見。All three right and still wrong → swap model or add rules. Rarest.

比例是經驗值，示意用的。Split is an experience-based illustration.

10 · 一個知識型的生成系統是怎麼被規格化的How a knowledge-based generation system gets specified

AI 使用手冊生成系統AI user-manual
generation system

iKala 2024 年的提案：用 50 份手冊建知識庫，一鍵就能生成新手冊的初稿。我從這案子帶走四件事：驗收指標要先寫、gate 要有停損、資料條件要寫進範圍、方案要有階梯。An iKala proposal from 2024: build a knowledge base from 50 manuals, generate a first draft in one click. I took four things from it: metrics before build, gates with a stop, data readiness in scope, and an options ladder.

三種導入策略，我們選了 ShaperThree adoption strategies; Shaper was chosen

TTaker直接拿市面上的 GenAI 方案接進既有的工作流程，不做客製。Consume off-the-shelf GenAI via API inside existing workflows; no customisation.

SShaper針對場景把既有模型做部分客製或微調，再整合進工作流程。這個案子是自架開源模型（Self-host），需要的時候才 fine-tune，用 Docker 容器交給客戶，部署在 Azure 上。Partly customise or fine-tune an existing model for the scenario and integrate it. Here: self-hosted open-source model, fine-tuned only if needed, delivered as a Docker container on Azure.

MMaker從頭訓練全新的模型。成本跟人才門檻都最高。Train a new model from scratch. Highest cost and talent bar.

選 Shaper 的理由跟方法論那章講的一樣：多付的錢，要買得到講得出來的東西。Same reason as the method chapter: the extra money has to buy something you can say out loud.

範圍怎麼收How the scope was bounded

試行標的Pilot target	工業電腦這條產品線，50 份使用手冊（含系列）；成效確認了，才會擴到其他產品線Industrial-computer line, 50 user manuals (incl. series); other lines only after results
資料就緒條件Data readiness	請客戶給 Word 原檔，不收 PDF 掃描檔，這樣 OCR 的誤差才不會跑進訓練集Client supplies Word source files, not scanned PDFs, so OCR error never enters the training set
模型跟所有權Model & ownership	用開源模型；要不要 fine-tune，看初步產出跟需求貼不貼近；用 Container 的形式交件Open-source model; fine-tune only if first outputs miss the need; delivered as a container
部署Deployment	Microsoft Azure；用 Azure AD 做登入跟權限Microsoft Azure; Azure AD sign-in and permissions
產出形式Output	一鍵就產出整份初版手冊，使用者還是可以自己編修One-click first draft of the whole manual; the user keeps editorial control

虛線左邊是 AI 模型，右邊是軟體工程。這條線其實就是合約裡講的誰負責什麼。Left of the dashed line is the AI model, right is the software. That line is who owns what in the contract.

資料怎麼走How data moves

使用者在 UI 上建立新手冊：先定義範圍、章節、規格，再上傳參考文件User creates a manual in the UI: scope, chapters, specs; uploads reference files
Backend 把上傳的文件跟勾選的需求組成 prompt，再送給 AI SystemBackend composes the prompt from uploads and selections and sends it to the AI system
AI System 去比對 Vector DB，再依關聯性把索引值回傳回來The AI system matches against the vector DB and returns relevance-ranked indexes
Backend 靠索引去 Relational DB 把文件段落撈出來；完整的文件跟歷史手冊就放在 Cloud StorageBackend pulls passages from the relational DB by index; full files and past manuals live in cloud storage

這張圖為什麼重要Why this diagram matters

索引會回到 Relational DB 去取原文，所以每一段生成的內容都追得到來源：這就是 audit trail 的雛形。Indexes go back to the relational DB for the original text, so every generated passage traces to a source: the seed of an audit trail.
Azure AD 擋在最前面，權限決定你看得到哪些手冊：知識來源的治理，從登入就開始了。Azure AD sits in front; permissions decide which manuals are visible. Knowledge-source governance starts at sign-in.
模型用容器交給客戶，部署在客戶自己手上，不綁 API 供應商。The model ships as a container; the client owns the deployment and is not tied to an API vendor.

上面這條是文件進來的路，下面這條是問題進來的路。Top row: how documents come in. Bottom row: how a question comes in.

Pre-processing 做什麼What pre-processing does

照目錄（TOC）來切文件，每一章就是一個可以檢索的單位Split by table of contents; one retrievable unit per chapter
建 metadata：規格（Spec）、元件（Components）、產品線Build metadata: specs, components, product line
先跑 Parser、OCR、版面分析、表格結構辨識，再 embedding 進 Vector DBParser, OCR, layout analysis, table-structure recognition, then embed into the vector DB

Retrieval 跟生成Retrieval and generation

Query analyze：先分詞、去掉停用詞，再理解結構跟意圖Query analysis: tokenise, drop stop-words, read structure and intent
Task dispatch 把任務分給對應的 executor；Multi-way recall 從好幾個模型跟來源把結果召回來；Re-rank 再依關聯性排序Task dispatch routes to the right executor; multi-way recall pulls from several models and sources; re-rank orders by relevance
LLM 把召回的 chunk 轉成人看得懂的草稿，盡量保留原本的格式；每一段都附上前三名的替代內容The LLM turns recalled chunks into a readable draft in the original format, with the top three alternatives per passage

先寫 User Story，再把它翻成 Prompt Story。這一步就是把「產品經理要什麼」轉譯成「模型要做什麼」。User story first, then a prompt story. That's the translation from what the PM wants to what the model does.

驗收指標在建置之前就先寫好Acceptance metrics written before the build

	第一階段 · 8/2Stage 1 · Aug 2	第二階段 · 10/11Stage 2 · Oct 11
F1-score	≥ 0.5	0.6 – 0.7
MRR	—	≥ 0.3
資料Data	50 份拿來訓練；每個產品線另外留 3 份做線下驗證；用各產品線的平均來判定50 for training; 3 more per product line held out for offline validation; judged on per-line averages
評估方式Method	用線下評估（短時間內可以多跑幾次），不做 A/B 跟滿意度調查（時間長、又有主觀偏誤）Offline evaluation (fast iteration), not A/B or satisfaction surveys (slow, subjective)
停損Stop rule	第一階段沒過，可以決議要不要繼續專案If Stage 1 fails, the client may decide whether to continue

階段跟交付物Phases and deliverables

0	啟動：做 AI VOC 驗證、畫架構圖稿、把 SOW 定稿、開啟動會議Kickoff: AI voice-of-customer validation, architecture draft, SOW final, kickoff
1	系統設計包含：需求訪談、wireframe、UI、軟體架構、DB schema、API spec、資料清洗、基礎模型的整合驗證；另外做一章內容生成的 demoDesign: interviews, wireframes, UI, software architecture, DB schema, API spec, data cleaning, base-model integration check; one-chapter generation demo
2	開發測試：模型調整跟 prompt 測試、AI 架構施作、模型封裝、API 整合、SIT 報告Build & test: model tuning and prompt tests, AI architecture build, model packaging, API integration, SIT report
3	驗收結案：Azure 環境、UAT、正式上線、文件交齊、驗收單簽好Acceptance: Azure environment, UAT, go-live, document handover, signed acceptance

四個方案，越左邊的越完整Four scope options, fuller to the left

一完整系統1 Full systemUI + 模型 + 串接 + 環境；管理層可以真的評估出成效UI + model + integration + environment; management can measure real effect

二縮減軟工2 Reduced SW拿掉登入跟封面頁；標準章節預設是英文Drop sign-in and cover pages; standard chapters default to English

三模型 + API3 Model + API前端客戶自己做；模型跟 API 是完整的Client builds the front end; model and API complete

四模型 PoC4 Model PoC只驗核心章節；沒有 API；成效要客戶自己推估Core chapters only; no API; client estimates the effect itself

有疑慮的話就先做簡化版，之後再按階段加約。用 change order 一階一階往上爬。If in doubt, start simple and add contracts stage by stage. Climb by change order.

提案裡的做法In the proposal	對應到我的方法In my method
F1 跟 MRR 的門檻在建置之前就寫進驗收標準，而且是有日期的F1 and MRR thresholds written into acceptance before build, with dates	Define the number upfront；KPI card 的 metric、target、deadlineDefine the number upfront; the KPI card's metric, target, deadline
第一階段沒過的話，客戶可以決定不要再繼續If Stage 1 fails the client may stop	Gate 不是延後再補，是停損；信任是從敢寫停損開始的A gate is a stop, not a deferral; trust starts with daring to write the stop
只收 Word 檔，不收 PDF 掃描檔Word only, no scanned PDFs	資料就緒的條件寫進 client obligations；專案實績那章四個桶子裡的 content 問題，在這裡就先堵住Data readiness as a client obligation; the content bucket from the track-record chapter is blocked here
照 TOC 切塊、建 Spec / Components 的 metadata、索引要能回到原文Split by TOC, build spec/component metadata, index back to source	Nexus 知識層怎麼切塊，還有六個 metadata 欄位；來源都追得回去Nexus knowledge-layer chunking and six metadata fields; source traceability
User Story → Prompt StoryUser story → prompt story	Spec & build literacy：要能把業務需求翻成模型要做的 Input / Process / Output，才有辦法驗工程師的計畫Spec & build literacy: translating business need into input / process / output is what lets me verify an engineer's plan
四個方案像階梯，先做簡單的，再照階段加約Four-option ladder; start simple, extend by stage	Find the smarter path；也就是成熟度那章講的「只承諾下一階」Find the smarter path; the maturity chapter's "commit to the next rung only"
先做 50 份、一條產品線，成效確認了再擴50 manuals, one product line, scale only after results	照 Validate → Launch → Scale 的順序走；Scale 是過了 gate 之後的事Validate → Launch → Scale; scale comes after the gate

來源：iKala AI Lab 提案 v1.0，2024 年 6 月。圖是我自己重畫的；金額跟客戶沒有公開。Source: iKala AI Lab proposal v1.0, June 2024. Diagrams redrawn by me; pricing and client withheld.

11 · 找更快、更便宜的路Finding the faster, cheaper path

兩次對「用 AI」這件事說不Two times I said no
to "use AI"

軍品倉儲Defence warehousing

C

國防客戶想要一套 AI 倉儲系統來管彈藥跟軍品。需求書上真的就寫著「AI 驅動」。A defence client wanted an AI warehouse system for ammunition and supplies. The brief literally said "AI-driven".

A

前兩次開會只講流程。效期就是日期，安全庫存就是門檻，補貨就是規則。沒有一件事需要預測。我跟他們說：你們不需要 AI。我們就做了資料模型、規則引擎，還有安全的 dashboard，全部 on-prem。The first two sessions were all about the process. Expiry is a date. Safety stock is a threshold. Reorder is a rule. Nothing needed predicting. So I told them: you don't need AI. We built a data model, a rules engine and a secure dashboard, all on-prem.

R

只花一小部分的時間跟成本，就拿到一套能用、也能稽核的系統。我的規則是：答案如果是日期、數量或門檻，那就是 workflow。AI 是留給需要判斷的時候用的。A working, auditable system in a fraction of the time and cost. My rule since: if the answer is a date, a count or a threshold, it's workflow. AI is for when the answer is a judgement.

完整版Full story

我留了一扇門：等資料乾淨一年之後，再來談需求預測。角色權限跟日誌從第一天就做進去，因為在封閉網路裡面，稽核軌跡比模型還重要。I left one door open: once the data had been clean for a year, we could talk about forecasting. Role-based access and logging went in from day one, because inside a closed network the audit trail matters more than any model.

震動感測器 → 手扶梯Vibration sensors → escalators

C

感測器客戶以為現在的 AI 可以直接預測機械手臂故障。我們收了震動資料，也收了第一筆錢，團隊就準備從 FFT 從頭做起。A sensor client believed today's AI could predict robotic-arm failures out of the box. We had the vibration data, we'd taken a first payment, and the team was about to build FFT features from scratch.

A

我問了兩件事。第一，物理 AI 平台都已經有了，為什麼還要用手工建？第二，更大的問題是：我們有很多震動資料，但幾乎沒有標好的故障。沒有故障，就沒有東西可以預測。所以我回去提了一個有標籤的案子：捷運手扶梯。I asked two things. One: physics-AI platforms already exist, why are we hand-building this? Two, the bigger one: we had lots of vibration data and almost no labelled failures. No failures, nothing to predict. So I went back with a case that had labels: metro escalators.

R

成交大約 US$150K，做出來的是一個有 baseline 的預測維修模型。機械手臂留到第二階段再做。學到一課：最聰明的路很少是更好的模型，而是把問題問對。Closed at about US$150K, delivered a predictive-maintenance model with a real baseline. The robotic arm became phase two. The lesson: the smartest path is rarely a better model. It's a better question.

完整版Full story

我提的是 NVIDIA PhysicsNeMo 這類有物理資訊的代理模型，也問過能不能用物理模型去生成合成的故障樣本。這個提議本身沒有被採用，但它把真正的問題挖出來了：標籤不夠。What I raised was physics-informed surrogate modelling, NVIDIA PhysicsNeMo and the like, including whether physics models could generate synthetic failure samples. That route wasn't taken, but asking is what exposed the real problem: no labels.

12 · 這些經驗怎麼接到一個 outcome 角色上How the experience lands in an outcome role

對應，跟前 90 天The mapping,
and the first 90 days

左邊是我做過的，右邊是這個角色要的。點一下右邊，就會看到連線。Left: what I've done. Right: what the role needs. Tap the right side to see the lines.

Deloitte →

iKala →

前 90 天First 90 days

Day 1–30接手兩個還在進行中的案子，把每個案子的 KPI card 重做一遍，找出沒有 baseline 或沒有 client owner 的，然後把 dashboard 立起來。Take over two live engagements, redo each KPI card, find the ones with no baseline or no client owner, stand up the dashboard.

Day 31–60每個案子都做一次 built vs promised 稽核，自己親手驗證至少一個技術 milestone（跑一次 pipeline、看一次 eval set）。有 drift 的，就帶著解法往上報。One built-vs-promised audit per engagement; personally verify at least one technical milestone (run the pipeline, read the eval set). Escalate drift with a fix attached.

Day 61–90每個案子都寫 evidence pack 跟下一個機會，至少要有一個真的進到提案；再把 KPI card 跟 dashboard 的模板變成整個團隊共用的。Evidence pack and next opportunity per engagement, at least one into proposal; make the KPI card and dashboard templates team-wide.

13 · 開工前的釐清Before any build starts

我會問客戶的十二個問題Twelve questions
I ask a client

第一次見面就問。答案會決定整合怎麼做、交付物怎麼切。每一題都附上我為什麼要問，還有什麼樣的答案算警訊。I ask these in the first meeting. The answers decide how we integrate and how we cut the delivery. Each one comes with why I ask, and which answer is a red flag.