Session 2 Pre-Class Notes and Case Discussion Guide
Session 2 課前講義與個案討論引導
Theme: Agentic Workflows and Shifting Pricing Units
主題：Agentic Workflows and Shifting Pricing Units
Main case: Should We Deploy a Gen AI Salesbot?
主個案：Should We Deploy a Gen AI Salesbot?
Through PulsePoint’s dilemma, this page examines how firms restructure workflows, decide what should be delegated to AI, govern instability, and rethink value capture when software turns into agentic execution.
本頁以 PulsePoint 的兩難為主軸，帶你檢視企業如何重組工作流程、決定哪些任務交給 AI、管理輸出不穩定性，並在軟體走向代理執行時，重新思考價值如何被捕捉與定價。
Course logic: begin with PulsePoint’s business dilemma, move into AI’s impact on the funnel and workflow, then extend the discussion into agentic workflow design, governance, pricing, and a blueprint-based lab.
課程邏輯：先從 PulsePoint 的商業兩難出發，接著進入 AI 對行銷漏斗與工作流程的衝擊，再延伸到代理式工作流設計、治理、定價，以及以藍圖為核心的 Lab 實作。
Clarify PulsePoint’s business essence, value proposition, and the difference between human B2B selling and bot-based selling.
釐清 PulsePoint 的商業本質、價值主張，以及真人 B2B 銷售與 Bot 銷售之間的關鍵差異。
Use prompts directly, compare outputs, and observe how funnel logic, customer touchpoints, and brand risk change when AI speaks to customers.
直接操作提示詞、比較輸出結果，觀察當 AI 直接面對客戶時，行銷漏斗、顧客接觸點與品牌風險如何改變。
Decide what to automate, where humans must stay in the loop, how to deploy safely, and how chatbot logic differs from agentic AI.
決定哪些任務適合自動化、哪些環節仍需人類介入、如何安全導入，以及 Chatbot 與 Agentic AI 的本質差異。
Move from job unbundling to agentic rebundling, then build a workflow blueprint with triggers, inputs, tools, outputs, review gates, and evaluation logic.
從工作解構走向代理式重組，再透過 Triggers、Inputs、Tools、Outputs、Review Gates 與評估機制，完成一張工作流藍圖。
Do not rush into the AI decision. First clarify what PulsePoint actually sells, where customization creates cost, and why the sales team and the largest client do not read “efficiency” in the same way.
不要急著進入 AI 導入決策。先釐清 PulsePoint 真正賣的是什麼、客製化如何形成成本，以及業務團隊與最大客戶為何不會用同一種方式理解「效率」。
Case setup: PulsePoint is framed as a large B2B digital marketing firm with heavy customization and a high cost structure. The CTO wants GenAI to lower delivery cost, while sales-side actors worry that automation may destroy trust, nuance, and account-specific judgment.
個案設定：PulsePoint 被設定為一家規模龐大的 B2B 數位行銷公司，具有高度客製化與高成本結構。CTO 希望利用生成式 AI 降低交付成本，但業務端擔心自動化會破壞信任、脈絡判斷與客戶關係中的細膩度。
Is it media inventory, customized content, sales enablement, account intelligence, or workflow relief for the client?
它賣的是媒體版位、客製內容、銷售支援、帳戶洞察，還是替客戶減少協調摩擦的工作流能力？
Break the business into four layers: target customer, workflow outsourced by the client, degree of customization, and cost-to-serve. The key issue is whether AI removes cost only—or also changes the value proposition.
把業務拆成四層：目標客戶、客戶外包出去的工作流程、客製化程度，以及服務成本。真正的問題是：AI 只是在降低成本，還是同時改變了價值主張？
Where do trust, negotiation, empathy, and exception handling still matter?
信任、談判、同理心與例外處理，究竟在哪些地方仍然不可被取代？
Bot-friendly clients often seek speed, self-service, and lower ambiguity. Bot-resistant clients often want memory, reassurance, political sensitivity, and account-level judgment—especially when the stakes are high.
偏好 Bot 的客戶通常追求速度、自助服務與低模糊性；排斥 Bot 的客戶通常重視記憶感、安心感、政治敏感度與帳戶層級判斷，尤其在交易風險高時更是如此。
Q: What, specifically, is PulsePoint “doing” as a USD 2 billion B2B digital marketing company? What core value does it create for enterprise clients, and how are customization and cost structurally connected?
Q：作為一家年營收 20 億美元的 B2B 數位行銷公司，PulsePoint 具體到底在「做什麼」？它為企業客戶創造了什麼核心價值？高度客製化與高成本結構又是如何被綁在一起的？
Use this lens: Treat PulsePoint as a bundle of outsourced B2B tasks rather than a generic “marketing company.” Ask what uncertainty the client is paying to reduce: targeting uncertainty, message-fit uncertainty, coordination burden, or sales-conversion uncertainty.
建議鏡片：不要把 PulsePoint 當成抽象的「行銷公司」，而要把它看成一組被企業客戶外包出去的 B2B 任務。請問自己：客戶付費，是為了降低哪一種不確定性？是受眾判斷的不確定性、訊息貼合度的不確定性、協調成本，還是成交轉換的不確定性？
Q: What is the largest gap between human B2B selling and bot selling? Which client profiles may actively want bot service, and which profiles may react like Orion’s CEO Tyrell Durant and push back hard?
Q：在 B2B 情境中，真人銷售與 Bot 銷售最大的差異是什麼？哪一類客戶可能主動想要 Bot 服務？哪一類客戶會像最大客戶 Orion 的 CEO Tyrell Durant 一樣強烈反彈？
Use this lens: Segment clients by relationship intensity, switching cost, political sensitivity, and the need for tailored reassurance. The more the sale depends on trust repair, negotiation, and reading ambiguity, the harder it is to replace the human seller cleanly.
建議鏡片：請依照關係強度、轉換成本、政治敏感度，以及是否需要客製化安撫來分層客戶。越是仰賴信任修復、談判與模糊情境判讀的交易，就越難以用 Bot 乾淨地取代真人業務。
This stage moves from the case into direct experimentation. Use AI as a lab object: ask it about the funnel, test prompt sensitivity, then evaluate what those differences imply for brand consistency and client-facing deployment.
這一段從個案走向直接實驗。請把 AI 當作實驗對象：先問它行銷漏斗會怎麼變，再測試提示詞敏感度，最後評估這些差異對品牌一致性與對客導入代表什麼。
How will generative AI and AI chatbots affect the traditional marketing funnel?生成式AI與AI Chatbot將如何影響傳統行銷漏斗？
Prompt A: Recommend 2 birthday dinner venues in Songshan, Taipei.提示詞 A： 推薦2個在台北松山的生日聚餐地點的建議
Prompt B: Recommend two birthday dinner venues in Songshan, Taipei.提示詞 B： 推薦兩個在台北松山的生日聚餐地點的建議
Q: When AI content generation and chatbots become common, how do awareness, consideration, search, evaluation, and conversion change? Record an AI answer, then critique it.
Q：當 AI 內容生成與聊天機器人普及後，Aware、Appeal、Ask、Act、Advocate 或傳統漏斗中的認知、考慮、搜尋、評估與轉換，會如何變化？請先記錄 AI 的回答，再提出你自己的批判。
Discussion angle: AI often compresses search, evaluation, and first response into a single conversational interface. The funnel becomes less linear, more recursive, and increasingly mediated by recommendation systems or agents rather than by brand-owned webpages alone.
討論方向：AI 會把搜尋、評估與第一輪互動壓縮到同一個對話介面裡。漏斗因此變得不那麼線性，更加循環，也更常由推薦系統或代理來中介，而不是只靠品牌自有網站承接流量。
Q: Open two new conversations, use Prompt A and Prompt B, then compare whether the recommended venues, tone, and reasoning stay identical. What does any difference imply when a Salesbot faces customers directly?
Q：請在兩個全新視窗分別輸入 Prompt A 與 Prompt B，比較推薦地點、語氣與推理是否完全一致。若存在差異，這對直接面對客戶的 Salesbot 代表什麼？
Q: If output is stochastic, how can the firm govern AI so B2B selling remains stable, consistent, and confidential?
Q：如果 AI 的輸出具有隨機性，企業要怎麼規範它，才能讓 B2B 銷售維持穩定、一致，並避免洩漏商業機密？
Governance lens: combine system prompts, data whitelists, grounding/RAG, tool restrictions, confidence thresholds, review gates, and human-in-the-loop approvals. Stability is not achieved by prompting alone; it requires workflow design.
治理鏡片：把系統提示詞、資料白名單、Grounding／RAG、工具權限限制、信心門檻、審核節點，以及 Human-in-the-loop 結合起來。穩定性不是靠一句 prompt 就能得到，而是靠工作流程設計。
This stage reframes the question. The issue is not only “deploy or not deploy.” It is also which tasks should move to AI, what authority an agent receives, how rollout sequencing works, and who captures the value when models are external.
這一段重新定義問題。真正的題目不只是「要不要導入」，而是哪些任務應該交給 AI、代理得到哪些權限、導入順序如何安排，以及當底層模型來自外部供應商時，價值最終被誰拿走。
Good AI candidates include draft generation, data summarization, CRM preparation, and first-pass classification. Human judgment still matters for trust repair, pricing exceptions, political reading, and final commercial commitment.
適合交給 AI 的任務包括：草稿生成、資料統整、CRM 前處理與第一輪分類。人類判斷仍然關鍵的環節則包括：信任修復、價格例外、政治情勢判讀，以及最終商業承諾。
If you are CEO Jeannie Weiss, you must state a position clearly: full deployment, delayed deployment, or conditional deployment. Your reasoning should address cost pressure, innovation anxiety, and reputation risk together.
如果你是 CEO Jeannie Weiss，必須明確表態：全面導入、暫緩導入，或條件式導入。理由不能只談成本，還要同時處理創新焦慮與商譽風險。
A Q&A chatbot mainly speaks. A sales agent acts: it calls tools, changes records, triggers processes, and therefore changes the firm’s permission architecture, logging requirements, and rollback obligations.
一個問答型 Chatbot 主要在「說話」；一個 Sales Agent 則會「行動」：它調用工具、改寫紀錄、觸發流程，因此也改變了企業的權限架構、稽核需求與回滾責任。
If PulsePoint builds on external foundation models, margin may migrate downward to the model layer. At the same time, once agents deliver outcomes directly, the pricing logic may shift from software seats to measurable results.
如果 PulsePoint 建立在外部基礎模型之上，利潤可能被底層模型層吸走。同時，當代理可以直接交付結果時，定價邏輯也可能從軟體席位轉向可衡量的成果。
The session’s theory is built around Sangeet Paul Choudary’s Reshuffle. AI should not be treated merely as an efficiency tool. It is a force that separates tasks inside a job, redistributes them across humans and machines, and eventually changes the structure of value creation.
本次理論以 Sangeet Paul Choudary 的《Reshuffle》為主軸。AI 不能只被理解為提升效率的工具，而是一股把職位內部任務拆開、重新分配給人與機器，並進一步改變價值創造結構的力量。
| Traditional bundled job傳統被打包的工作 | What AI can unbundleAI 可先解構的部分 | What remains distinctly human仍高度仰賴人類的部分 |
|---|---|---|
| A B2B seller collects information, interprets calls, drafts responses, and updates systems end to end.一位 B2B 業務員通常從頭到尾處理資料收集、通話理解、回覆草稿與系統更新。 | AI can turn long, unstructured conversations into structured fields, summaries, next actions, and searchable knowledge.AI 可以把長篇、非結構化的對話轉成結構化欄位、摘要、下一步行動與可搜尋知識。 | Humans still define stakes, read hidden context, and decide what matters commercially or politically.人類仍然要判斷利害關係、讀懂隱性脈絡，並決定商業或政治上真正重要的是什麼。 |
| Routine communication, first-draft writing, and low-level review are embedded inside the same role.例行溝通、草稿撰寫與初步審查，過去都被包在同一個職位裡。 | AI can generate routine drafts, classify intent, and handle repetitive first-pass work at scale.AI 可以大量生成例行草稿、辨識意圖，並接管可重複的第一輪工作。 | Humans still own tone calibration, negotiation, relational repair, and final accountability.人類仍負責語氣校準、談判、關係修復與最終責任承擔。 |
| The job description looks like one role, but it is actually a bundle of heterogeneous tasks.職位描述看起來像單一角色，其實內部是一組異質任務的打包。 | AI decouples tasks at the task level, which is why work becomes fragmented before it is rebuilt.AI 在任務層級把工作解離開來，所以工作會先碎片化，再進入重組階段。 | Humans integrate across exceptions, ambiguity, and ethics once the workflow fragments.當工作流程碎片化後，人類的價值在於整合例外、模糊與倫理判斷。 |
What is the first capability jump that makes job unbundling possible?
讓工作能被拆解的第一個能力躍遷是什麼？
AI’s early strength is to extract structure from messy language, calls, notes, and emails. That removes a large share of information-processing work from the original role.
AI 最早期也最顯著的強項，就是從凌亂的語言、通話、筆記與 Email 中抽出結構。這等於先把大量資訊處理工作從原本職位中剝離出去。
Why does workflow fragmentation happen before workflow redesign?
為什麼工作流程一定會先碎裂，再進入重組？
Once AI handles repetitive information work, decision rights detach from execution. The role stops being one seamless human job and becomes a series of handoffs, rules, and exception points.
當 AI 接手重複性的資訊工作後，決策權與執行權就開始分離。原本看似連續的人類工作，會變成一連串交接、規則與例外節點。
In classic SaaS, the client buys software seats and still needs employees to operate the system. Under agentic rebundling, the client increasingly buys completed work or measurable operational output.
在傳統 SaaS 模式下，客戶買的是軟體席位，仍需要員工親自操作。到了代理式重組階段，客戶越來越可能買到的是「完成的工作」或可衡量的營運輸出。
The firm that controls the end-to-end solution, tool orchestration, and review logic gains the moat. The deeper the workflow integration, the harder the system is to swap out.
掌握端到端解決方案、工具編排與審核邏輯的企業，才會形成護城河。工作流程整合得越深，系統就越難被替換。
The value of an agentic workflow is not “a smarter chatbot.” It is a governed chain that links context, tools, permissions, and human review into one end-to-end operating logic.
Agentic Workflow 的價值，不只是「更聰明的聊天機器人」，而是一條被治理過的鏈條：它把情境、工具、權限與人類審核連成一個端到端的營運邏輯。
Define the agent by objective: is it trying to raise conversion, calm post-sale complaints, or qualify leads? The role definition shapes reasoning and tone.
先用目標來定義代理：它是要提升轉換、安撫售後客訴，還是做線索篩選？角色設定會直接影響推理路徑與回應語氣。
Limit what the agent can read. Use grounding or RAG so responses depend on authorized enterprise data, not on free-floating model memory.
先限制代理可以讀什麼，再透過 Grounding 或 RAG，讓回答建立在被授權的企業資料上，而不是漂浮的模型記憶。
Write the forbidden zones explicitly: no autonomous pricing changes, no unsupervised external statements, no access to protected financial data unless authorized.
把禁區寫死：不可自行改標準定價、不可未審核對外發言、不可讀取未授權的敏感財務資料。
Agentic execution requires traceability. The firm needs logs, escalation history, and visible decision points so errors can be audited and rolled back.
代理式執行一定要可追溯。企業需要留下日誌、升級紀錄與可見的決策點，才能在出錯時追查與回滾。
When discount levels, refunds, or commercial commitments exceed a threshold, the agent should pause and hand control back to a human approver.
當折扣、退款或商業承諾超過一定門檻時，代理就應該暫停動作，把決策權交回給人類審核者。
If the customer is angry, anxious, or signaling distrust, the goal is no longer efficient throughput. Human empathy becomes part of the control system.
當顧客出現憤怒、焦慮或強烈不信任時，目標就不再只是效率。此時人類同理心本身就是控制系統的一部分。
Rare situations and weak-confidence outputs require escalation. The value of the human shifts from routine execution to exception handling.
罕見情境與低信心輸出需要升級處理。人類的價值因此從例行執行，轉向例外管理與風險承擔。
Once AI acts autonomously, humans are no longer only operators. They become reviewers, risk owners, escalation targets, and designers of the workflow boundary.
當 AI 具備自主執行能力後，人類不再只是操作員，而會變成審核者、風險擔責者、升級接管者，以及邊界設計者。
When agents directly complete tasks, “per seat” pricing weakens. The new unit can become resolved tickets, booked meetings, completed policies, or another measurable outcome.
當代理直接完成任務時，「按席位付費」的邏輯會鬆動。新的定價單位可能變成：成功排解的工單、成功預約的會議、完成核保的保單，或其他可衡量成果。
If AI absorbs data gathering, drafting, and first-pass communication, how will future senior talent accumulate judgment? Career paths have to be redesigned.
若 AI 吸收了資料蒐集、草稿撰寫與第一輪溝通，未來的資深人才要怎麼累積判斷力？組織勢必要重新設計職涯路徑。
Buying many AI tools is not the same as transformation. The defensible asset is the firm-specific workflow that ties tools, data, and review logic together.
買很多 AI 工具不等於完成轉型。真正難以複製的資產，是把工具、資料與審核邏輯綁成一體的企業專屬工作流程。
Not the firm that blindly buys the newest model, but the firm that defines human-machine boundaries precisely, governs exceptions well, and rebundles fragmented tasks into coherent workflows.
不是盲目追逐最新模型的企業，而是那些能精準劃分人機邊界、妥善管理例外，並把碎片任務重新組成連貫工作流程的企業。
Synthesis: workflow redesign changes the unit of work, the unit of pricing, and the unit of organizational capability at the same time.
總結：工作流程的重組，會同時改變工作的單位、定價的單位，以及組織能力的單位。
Choose one customer-visible touchpoint, decompose the workflow behind it, then design a blueprint for human-AI division of labor. The point is not “use AI everywhere,” but “assign labor, permissions, and review gates deliberately.”
請選定一個顧客可見的接觸點，拆解其背後工作流程，再設計一張人機分工藍圖。重點不是「到處放 AI」，而是有意識地分配勞務、權限與審核節點。
How does the customer search for a product or solution? Where does discovery friction appear?
顧客如何尋找商品或解決方案？探索摩擦出現在什麼地方？
What happens when the customer reads a post, newsletter, or personalized recommendation?
顧客閱讀社群貼文、EDM 或個人化推薦內容時，背後工作流程如何運作？
How are returns, complaints, and usage questions handled? Where is the current coordination burden highest?
退換貨、客訴與使用教學怎麼被處理？目前協調負擔最高的地方在哪裡？
Think about checkout, guided selling, or cart conversion. Where do people lose time or context?
思考結帳、導購或購物車轉換流程。人在哪裡最容易失去時間或上下文？
| Element核心元素 | What your team must define你的小組必須定義什麼 |
|---|---|
| TriggerTrigger（觸發條件） | What event starts the agent? A customer message, an abandoned cart, a delayed shipment, or something else?什麼事件會啟動代理？顧客發問、購物車放置過久、包裹延遲，還是其他事件？ |
| InputsInputs（輸入資料） | Which contextual variables are required, such as purchase history, loyalty status, inventory, or service history?需要哪些情境資料，例如購買紀錄、會員等級、即時庫存或服務歷史？ |
| ToolsTools（調用工具） | Which APIs, databases, email systems, pricing tools, or internal systems is the agent allowed to call?代理可以調用哪些 API、資料庫、Email 系統、定價工具或內部系統？ |
| OutputsOutputs（輸出結果） | What concrete action comes out: a recommendation, a discount code, a rescheduled booking, or a routed case?最後輸出什麼具體行動：推薦內容、折扣碼、重新安排的預約，還是被轉派的案件？ |
| Review GatesReview Gates（審核節點） | At which point must the AI stop and hand control to a human due to risk, anger, or uncertainty?在哪個節點上，AI 必須因風險、情緒或不確定性而停下來，把控制權交還給人類？ |
A workflow map alone is not enough. Once AI begins to act, the team must define role, permissions, data access, success metrics, failure modes, and accountability. This is the difference between a demo and an enterprise design.
只有流程圖還不夠。當 AI 開始行動時，小組必須補上角色、權限、資料存取、成功指標、失效模式與責任歸屬。這正是「Demo」與「企業級設計」的差別。
State the agent’s goal orientation clearly: sales assistant, complaint-calming specialist, search guide, or something else.
清楚定義代理的目標導向：它是銷售助理、安撫客訴的專家、搜尋導購助手，還是其他角色？
State what the agent is forbidden to do: alter standard prices, issue public statements, or access protected data without approval.
清楚寫出禁止事項：不可自行調整標準定價、不可對外發布聲明、不可未授權存取敏感資料。
Specify exactly what data is allowed, such as product history and sizes, and what is blocked, such as credit-card data or full legal identity.
明確規定可讀取的資料，例如購買品項與尺寸；以及禁止讀取的資料，例如信用卡資訊或完整法定姓名。
Avoid vanity metrics. Use first-contact resolution, conversion rate, or NPS-like service improvement measures.
避免虛榮指標。請使用首度解決率、轉換率，或類似 NPS 的服務改善指標來衡量成效。
Predict the worst cases: hallucinated product specs, circular argument with the customer, wrong compensation, or impossible recommendations.
預先預測最糟情況：例如幻覺式產品規格、與客戶無限爭執、錯誤賠償承諾，或根本不可執行的建議。
Decide who takes accountability, which department receives the case, and how the customer is rescued seamlessly after AI failure.
先決定誰要承擔責任、由哪個部門接手，以及當 AI 失敗後，要如何讓顧客無縫地被人類救援。
[Agent Workflow Blueprint Template] 1. Touchpoint: 2. Customer-visible friction: 3. Trigger: 4. Inputs: 5. Tools: 6. Outputs: 7. Review gates: 8. Agent role: 9. Forbidden actions: 10. Data whitelist: 11. KPI: 12. Worst failure mode: 13. Escalation owner:【Agent Workflow Blueprint Template】 1. 接觸點： 2. 顧客可見的摩擦： 3. Trigger： 4. Inputs： 5. Tools： 6. Outputs： 7. Review Gates： 8. Agent 角色： 9. 禁止事項： 10. Data Whitelist： 11. KPI： 12. 最糟失效模式： 13. 升級接管者：
Workflow Map with Triggers, Inputs, Tools, Outputs, and Review Gates clearly marked.
繳交 Workflow Map，並清楚標示 Triggers、Inputs、Tools、Outputs 與 Review Gates。
Agent Spec covering Roles, Permissions, and the Data Whitelist.
繳交 Agent Spec，內容包含 Roles、Permissions 與 Data Whitelist。
Evaluation Plan covering KPIs, Failure Modes, and Escalation logic.
繳交 Evaluation Plan，內容包含 KPIs、Failure Modes 與 Escalation 機制。
Bring one customer-visible touchpoint for next week. You must label its journey stage clearly—Aware, Appeal, Ask, Act, or Advocate—and identify the largest pain point at that stage.
下週請各組攜帶一個顧客可見的接觸點，並清楚標註它屬於哪一個顧客旅程階段（Aware、Appeal、Ask、Act、Advocate），以及該階段目前最大的痛點。
PulsePoint’s dilemma is not simply whether AI is good or bad. It is about where value sits inside the workflow and which tasks should move first.
PulsePoint 的兩難，不是 AI 好不好，而是價值究竟藏在工作流的哪裡，以及哪些任務應該先移動、哪些不能亂動。
LLM variability is manageable only when the firm redesigns prompts, data access, tools, handoffs, and review logic together.
LLM 的不穩定性，只有在企業同時重設提示詞、資料存取、工具權限、交接方式與審核邏輯時，才真正可管理。
Once AI stops being a text generator and starts acting across systems, the firm must rethink permissions, auditability, build-vs-buy choices, and how outcomes are priced.
當 AI 不再只是文字生成器，而開始跨系統執行任務時，企業就必須重想權限、稽核、Build-vs-Buy，以及成果如何被定價。
The deliverables—workflow map, agent spec, and evaluation plan—force the class to move from abstract debate to operational design.
Workflow Map、Agent Spec 與 Evaluation Plan 三份繳交成果，會把抽象辯論推進到真正的營運設計層次。