The Last Mile Challenge

☘️ Article

• 
• LLM 帶來了 DIY 實驗浪潮，但企業在嘗試之後意識到，要實現大規模、準確，且結合既有業務流程，包含應用程式和數據的 AI，還是需要解決 "last mile" 問題。
• salesforce 認為因此，驅動了其 agentforce 和 data Cloud 的市場需求，進而加速了預訂數字和收入增長
• 對於 CRM 所述的 "last mile" 可理解為企業可以把資料尻去外面串 AI 做 DIY 實驗，但玩一玩會發現：a) 成本過高和耗時, b) 跨部門功能整合串接困難, c) 工作流習慣難更改, d) 產出一致性不好
• OAI 發難之初，滿多說法認為 G 社瀏覽器將消亡，隨後發現沒這麼簡單，因為沒考慮到人們的使用習慣不容易改變；又如有一說 " 未來無軟體 "，需要什麼應用的時候直接 gen 出來就好，這沒考慮成本和一致性標準化的問題
• 未來絕對怎樣怎樣的評論，拿來說書好用吸睛，但實務運作總是漸進
• 如果今天是以單兵作戰的角度來說，AI 可以把本來 " 不會 " 做的事情都弄上 60 分水準。
• 但如果是自己 " 會，且專業 " 的內容，馬上就能抓出一堆問題。
• 這應該是很多人使用上的體驗，能問通識，但專業項目會失真
• 為求準確就需要所謂的 "last mile"，去 tune 出有價值和一致性的產出。
• 總結還是成本：花費是成本、做錯被客訴是成本、土炮瞎搞砸招牌也是成本
• "We bring the last mile. And this is what companies are realizing. They've been experimenting. Many of them -- I was in a conversation yesterday. Somebody said, Miguel, I've been talking to some customers of Salesforce that are building the agentic layer outside Salesforce.
• I'm like, yes, there are many of them that try to do it, try to get the data because you can do it yourself anything. They get the data out of Salesforce, then they build the logic, then they decide that they want to take some actions, but they don't have the capability to execute the thousands of workflows that the customers have already built on Salesforce.
• So basically, by doing that -- and by the way, the humans are working here, and they built this agentic layer here, trying to build a cool interactive conversational UI, but a UI that is disconnected from the humans and it disconnected from the execution and is semi-disconnected with the data because the one -- the moment you move the data out of Salesforce into other data lakes, et cetera, it just becomes obsolete.
• You can do it, but it's expensive.
• So most of these customers have realized that we need -- they need the last mile.
• And they turn to us and they said, "let's do it together. Let's start the agentic transformation journey."
• " We also launched ITSM. I think we're going to have a lot of good discussions in our next earnings around ITSM is booming already.
• We just launched it.
• We won dozens of customers.
• Life Science Cloud was a great example.
• It just quadrupled or I don't know the specific statistic, but it grew a lot during Q3.
• And I just want to mention that this is pretty big because we were doing a lot of business in life science and also medical devices, et cetera.
• And -- but as you know, for many years, we had this great partner that we basically partner and capturing that opportunity.
• We were capturing the outside commercial side, the opportunity.
• They were very focused on the commercial side and the clinical side.
• And then 1.5 years ago, they decided to compete head-to-head with us, and we decided to build solutions to compete with them, and the results have been incredible.
• I mean we announced yesterday Novartis.
• We had announced Pfizer of the top 20.
• Of the top 20, we've already won 5 or 6 and I say 5 or 6 because there's one that is in embargo, but I think there's going to be a press release later today, where we're going to announce another major pharma company going to us.
• So we are gaining market share from Veeva like there is no tomorrow.
• We've already -- in addition to the big 20, which there are still many in the air and they are reviewing, they all want the Salesforce platform.
• They want our agentic capabilities.
• They want our data cloud to unify data across all the different domains in the business.
• And they are afraid to move with a small player that is going to build a new platform for them.
• So many of them want to stay and are staying there's more than 100 life science cloud -- life science customers that have selected Life Science Cloud and are moving off Veeva, and we are just getting started.
• So that's an example of innovation.

✍️ Abstract

「Last Mile」在文中的含義

• 「Last Mile」(最後一哩路) 借用了物流業的術語，比喻將 AI 技術真正落地到商業應用並產生價值 的最後關鍵步驟。
• Last Mile =「讓 AI 不只會『說』(生成內容)，還能精準地在企業既有系統中『做』(執行任務) 的能力」。
  • 簡單來說，如果 AI 模型是「發電廠」，Last Mile 就是把電「接進你家插座並能讓電器運作」的那段複雜工程。

1. 核心定義：從「實驗」到「實戰」的鴻溝

• 在文中，Last Mile 指的是企業在擁有 AI 模型 (如 LLM) 和數據後，要將其轉化為穩定、準確且能執行業務的系統時，所面臨的巨大挑戰。
  • 起點 (0-90%)：企業可以輕易地把資料匯出 (Data Lake)，自己接一個 AI 模型來做問答或生成，這屬於 DIY 實驗階段。
  • Last Mile (最後 10%)：要讓這個 AI 準確執行公司內部的複雜流程 (例如：自動退貨審核、根據合約條款發送報價單、跨部門協調)，且不出錯、符合法規。

2. 為何「Last Mile」這麼難？(文中的痛點)

• 文章指出，許多企業嘗試自己走這最後一哩路 (DIY)，結果發現困難重重：
  • 執行力斷層：外部 DIY 的 AI 系統與公司的核心業務系統 (如 Salesforce) 是「斷開」的。AI 也許知道「該退款」，但它沒有權限或能力去操作公司內部那套複雜的退款系統。
  • 數據即時性失效：當你把資料從 Salesforce 搬出去餵給外部 AI 時，資料就已經變成「舊聞」(obsolete)，不再是即時狀態。
  • 成本與標準化：要自己開發一套能連結所有部門、符合幾千種既有工作流 (Workflows) 的系統，成本過高且難以維護。

3. Salesforce 的解決方案

• Salesforce 在文中強調他們能解決這個問題，因為他們掌握了這「最後一哩路」的關鍵要素：
• Context (情境)：他們擁有即時的 Data + Metadata (資料結構與關聯)。
• Execution (執行)：他們擁有已經建立好的 Deterministic Workflows (確定性的工作流程)。

The Last Mile Challenge

• 圖表描述企業 AI 轉型最後一哩挑戰：Context 為 Data + Metadata，AI 為 Non-Deterministic Reasoning (非確定性推理，指 AI 生成內容多變不固定)，Apps 為 Deterministic Workflows (確定性工作流程，指固定邏輯自動化)，Humans 為 Embedded AI (嵌入式 AI，指 AI 融入人類工作)。
• 企業需解決「Last Mile」成為 Agentic Enterprises (代理式企業，指 AI 代理自主處理業務)，這是 Salesforce 的護城河 (moat，商業術語指競爭優勢)。

LLM DIY 實驗與 Last Mile 問題

• LLM (大型語言模型) 引發 DIY 實驗浪潮，但企業試後發現大規模 AI 需整合既有業務、應用與數據，仍卡在「last mile」問題：成本高耗時、跨部門整合難、工作流習慣不易改、產出不一致。
• 類比：Google 瀏覽器未消亡因使用習慣難變；「未來無軟體」忽略成本與標準化。
• AI 單兵作戰可達通識 60 分水準，但專業領域易失真，需「last mile」調校 (tune) 確保價值與一致性。
• 成本考量：人力花費、客訴、土法煉鋼損品牌皆為成本。

Miguel 訪談引用與客戶洞察

• Miguel 指出客戶試從 Salesforce 導出數據外部建 agentic layer (代理層，指 AI 代理系統)，但無法執行 Salesforce 既有千種工作流，人類操作脫鉤、數據移至 data lake (數據湖，網路術語指大規模儲存多樣數據的倉庫) 即過時，成本高。
• 客戶意識需 Salesforce「last mile」，共同啟動代理轉型。

Salesforce 產品與市場表現

• 推出 ITSM (IT Service Management，IT 服務管理，Salesforce 新雲端解決方案自動化 IT 支援流程)，剛推出即贏數十客戶，下季財報將熱議。
• Life Science Cloud (生命科學雲，Salesforce 為製藥、生醫產業客製 CRM 與數據平台) Q3 成長 4 倍以上，從夥伴 Veeva (Veeva Systems，生命科學 CRM 龍頭) 搶市佔。
• 贏前 20 大藥廠 5-6 家 (如 Novartis、Pfizer)，另有未公布大廠；逾 100 生命科學客戶轉用，強調 Salesforce 平台代理能力與 Data Cloud (數據雲，Salesforce 統一多域數據平台) 優勢，避免小廠風險。