AI落地为什么失败？—95%的企业AI项目死在workflow上

来源BG2 Pod / YouTube嘉宾Ali GhodsiDatabricks CEO、Arvind JainGlean CEO主持人Apoorv AgrawalAltimeter合伙人总时长45分00秒博客日期2025/12/23核心摘要Databricks CEO Ali Ghodsi与Glean CEO Arvind Jain在BG2 Pod上坦诚拆解企业AI落地的真实困境与突破路径。核心论断95%的AI项目失败不是因为技术不行而是因为组织没有把AI嵌入workflow。LLM正在快速commoditize——像加油站一样 interchangeable真正的壁垒是专有数据、workflow integration和agentic系统。两人分享了RBC自动化的金融合规审查、Merck药物发现文献综述、7-Eleven库存预测的真实落地案例也坦承了自己公司内部的失败尝试Glean的AI优先级排序项目、Databricks的custom model尝试。关键洞察企业AI的价值捕获在app layer而非model layerRPA解决了结构化数据的自动化而生成式AI解决了非结构化数据的自动化两者的结合才是enterprise automation的完整图景。一、Consumer AI vs. Enterprise Reality01:00-02:151.1 消费AI与企业AI的根本差异消费AI一个model如ChatGPT服务10亿用户success path清晰企业AI同一个model需要在数千个不同workflow中work每个workflow的context不同核心差异“Consumer AI is about one model serving a billion users. Enterprise AI is about one model serving a billion workflows.[消费级人工智能的核心在于一个模型服务十亿用户企业级人工智能的核心在于一个模型服务十亿个工作流]”企业AI的复杂性security、compliance、governance、data privacy、legacy system integration——这些都是消费AI不需要面对的 思考点one model serving a billion workflows这个比喻精准地捕捉了企业AI的核心挑战。消费AI的scaling是horizontal横向扩展用户企业AI的scaling是vertical纵向深入每个workflow。这是否解释了为什么企业AI公司如Databricks、Glean的revenue per customer远高于消费AI公司二、Why 95% of AI Projects Fail02:15-04:152.1 失败率的数据与真实原因MIT Research95%的企业生成式AI pilot未能交付可衡量的商业价值measurable business value只有5%的AI pilot program实现了快速的revenue acceleration但Arvind Jain的解读“You hear these 95% of projects fail. That’s actually what you want. When you’re actually experimenting with new technology, if all of your projects are failing, that means you’re not trying enough.[你常听说95%的项目都会失败。其实这正是你所希望的。当你真正尝试新技术时如果所有项目都失败了那说明你还不够努力]”2.2 真正的失败原因不是技术是组织不是model不够好——GPT-4/Claude已经在大多数task上足够好不是data不够多——企业有tons of data真正的原因没有把AI embed到workflow中“It’s not just you can just unleash the agents, and it just works. Making AI effective within an organization is a complex engineering challenge that requires deep integration, careful testing, and strong teams[并不仅仅是放手让这些AI Agent去运作它们就能自动生效。要在组织内部有效运用人工智能是一项复杂的工程挑战需要深度集成、仔细测试以及强大的团队].”Ali Ghodsi的补充很多企业把AI当成plug-and-play——buy一个LLM APIexpect magic实际上需要data pipeline[数据管道]、context management[上下文管理]、evaluation framework[评估框架]、human-in-the-loop[人机协同]、continuous iteration[持续迭代]2.3 成功的5%做对了什么他们把AI嵌入到existing workflow中而不是创建new workflow他们 focused onone specific use caseperfected itthen expanded[专注于一个特定的用例将其完善然后进行扩展]他们 invested indata infrastructurebefore AI[在人工智能出现之前就投资了数据基础设施]三、RBC, Merck, and 7-Eleven Use Cases04:15-06:453.1 RBC加拿大皇家银行金融合规审查自动化Problem合规团队需要review thousands of financial documents dailySolutionAI agent自动read、classify、flag异常文档Result处理时间从4小时降至15分钟准确率从人工的85%提升至97%人类reviewer从reader变为validatorKey insightAI没有取代人类而是改变了人类的角色3.2 Merck默克药物发现文献综述Problem药物发现团队需要review millions of scientific papersSolutionAI agent自动summarize、extract key findings、identify patternsResultLiterature review时间从3个月降至2周发现了人类researcher遗漏的3个潜在drug interactionsKey insightAI在read everything上比人类强但在judge what matters上仍需要人类3.3 7-Eleven库存预测Problem8万 SKU的库存管理过度库存和缺货同时存在SolutionAI agent分析sales data[销售数据]、weather[天气]、local events[当地活动]、supplier lead times[供应商交货期]Result库存周转率提升23%缺货率下降40%过期损耗减少15%Key insightAI的价值在于integrating multiple data sources that humans can’t process simultaneously[整合多种人类无法同时处理的数据源] 思考点三个案例的共同点是什么不是AI replaced humans而是AI changed what humans do。RBC的reviewer从reader变validatorMerck的researcher从reader变strategist7-Eleven的manager从data cruncher变decision maker。这是否意味着企业AI的正确narrative不是automation(自动化)“而是augmentation增长”四、What Actually Makes AI Work06:45-08:454.1 三大成功要素要素说明为什么重要Proprietary Data企业独有的数据——客户记录、交易历史、内部文档LLM是commodity但your data is notWorkflow IntegrationAI embed到existing workflow中不创造new workflow用户不需要change behaviorAgentic SystemsAI能自主take action不只是generate text从assistant到executor4.2 Ali Ghodsi的框架Data is the moat“LLMs are like gas stations. They’re everywhere, they’re interchangeable. Your proprietary data is your oil well.[大型语言模型就像加油站。它们无处不在彼此可互换。而你的专有数据就是你的油井]”Workflow is the castle没有workflow integrationAI只是isolated tool不是systemAgents are the armyagents让AI从suggest变为do五、Failed AI Bets at Databricks Glean08:45-11:005.1 Glean的失败AI优先级排序Project让AI自动识别每个员工的top weekly priorities汇总给leadershipWhy it seemed easy“It has all the context inside the company to make it happen[公司内部具备实现这一目标所需的一切条件]”Why it failedPriority是主观的——what’s “important” varies by person, by week, by contextAI无法捕捉隐性知识capture implicit knowledge“我知道这个重要但无法清晰表达为什么”Leadership的expectation与AI的capability存在gapLesson“It actually takes much longer than you know to actually generate success.[实际上要取得成功所需的时间远比你想象的要长得多]”5.2 Glean的另一个失败Custom AI ModelProject为特定product function构建custom AI modelWhy it failed微调成本高于预期维护成本太高基础模型GPT-4/Claude的进步速度超过custom model的迭代速度Lessonreturn to foundation models——less tailored, but more reliable and easier to implement[回归基础模型——虽然定制化程度较低但更可靠且更易于实现]5.3 Databricks的失败过早投入AgenticProject2024年初推出autonomous data agentWhy it failed企业客户not ready——governance、trust、audit trail都不成熟Agent的hallucination在enterprise context中cost太高客户需要human-in-the-loopnot full autonomyLessonenterprise AI需要先证明可靠性prove reliability再赋予自主权grant autonomy 思考点两个CEO坦承失败这本身就是宝贵的signal。很多企业AI的失败不是因为AI不够好而是因为组织没准备好或use case选错了。Glean的priority排序失败揭示了AI在subjective judgment主观判断上的根本性限制——这正是人类judgment的价值所在。六、RPA vs. Generative AI11:00-14:156.1 RPA机器人流程自动化的局限RPA解决的问题结构化数据的自动化固定规则、固定input/output、deterministic[确定性的]例如从A系统copy data到B系统、form fillingRPA的bottleneck每次UI变化都需要重新configure无法处理非结构化数据unstructured dataemail、document、conversation维护成本随流程数量线性增长6.2 生成式AI的互补性生成式AI解决的问题非结构化数据的自动化Email summarization、document extraction、conversation analysis[电子邮件摘要、文档信息提取、对话分析]能理解context、handle variability[上下文、处理变异性]两者结合才是完整图景“RPA handles the structured, repetitive tasks. GenAI handles the unstructured, cognitive tasks. Together, they’re the full stack of enterprise automation.[RPA 负责处理结构化、重复性的任务。生成式人工智能GenAI则负责处理非结构化、需要认知能力的任务。二者结合构成了企业自动化解决方案的完整体系。]”6.3 Ali Ghodsi的预测RPA公司UiPath、Automation Anywhere会被AI-native workflow automation取代不是RPA技术本身被淘汰而是RPA作为独立category会disappear——所有workflow automation都会incorporate AITimeline2-3 years七、Advice for CIOs Planning AI Budgets14:15-16:007.1 Arvind Jain给CIO的建议Rule #1Start with data infrastructure[从数据基础设施开始]如果data is messy, AI will be messyInvest in data cleaning、data governance、data accessibility first[首先应投资于数据清洗、数据治理和数据可访问性]Rule #2Pick one use casemake it workthen expand[选择一个用例先让它正常运行然后再进行扩展]Don’t try to “AI everything” at once[不要试图一下子把“一切都交给AI”]Success breeds success——one win builds organizational confidence[成功会带来更多成功——一次成功就能增强组织的信心]Rule #3Measure outcome, not output[评估结果而非产出]Don’t measure “how many AI models deployed”[不要评估“已部署的人工智能模型数量”]Measure “how much time saved”、“how much revenue increased”、“how many errors reduced”[评估“节省了多少时间”、“收入增加了多少”、“减少了多少错误”]7.2 Ali Ghodsi的补充Budget split[预算分配]建议60% data infrastructure[60% 数据基础设施]20% one use case perfection[20% 某个用例的完善]20% experimentation[20% 实验]Most common mistake把80% budget给AI models20%给data——应该反过来8、AI CapEx and the Revenue Math16:00-18:008.1 AI投资的回报周期Year 1通常是净负债net negative——infrastructure investment、training、failure[基础设施投资、培训、失败]Year 2收支平衡或略有盈余break-even或slightly positiveYear 3复利效应compounding returns——each new use case cheaper than the last[每个新用例的成本都比上一个更低]Ali Ghodsi的比喻“AI investment is like building a factory. You don’t expect ROI in month one. You expect ROI when the factory is running at full capacity.”8.2 收入数学Databricks的数据AI product revenue$1Brun-rate客户采用AI后平台粘性增长3倍以上platform stickiness increases 3xAI customers have 2x higher NRR净收入留存率Net Revenue Retentionthan non-AI customers关键 insightAI不是cost center是retention driver九、The Three Camps of AI18:00-21:009.1 企业AI的价值分层层级代表公司价值捕获持久性模型层OpenAI, Anthropic, Google当前水平较高正在压缩较低——正迅速商品化基础设施层Databricks, Snowflake, AWS当前水平中等正在增长中等——平台锁定应用层Glean, Salesforce, Vertical SaaS当前水平较低呈爆发式增长较高——工作流锁定9.2 为什么App层最终会捕获最多价值Arvind Jain的论点“The value in enterprise AI accrues to the app layer. Models are commodities. Infra is necessary but not sufficient. The company that owns the workflow owns the customer.”类比Model layer Intel芯片——important but not where value accrues[很重要但并非价值产生之处]Infra layer Windows操作系统——necessary platform[必要的平台]App layer Office应用——where users actually work and value is created[用户实际工作并创造价值的地方]9.3 Ali Ghodsi的修正同意App层价值最高但认为Infra层如Databricks是App层的enablerDatabricks的策略成为platform for AI apps——let vertical SaaS companies build on Databricks[让垂直领域 SaaS 公司基于Databricks构建应用]双赢Databricks gets platform revenuevertical SaaS gets AI capability without building infra十、Making AI Useful Inside Enterprises21:00-24:3010.1 Workflow Integration[工作流集成]的深层含义不是add AI button很多企业 mistake AI integration as “add a chatbot to our app”真正的integrationAI invisible地嵌入到every step of workflowEmailAI auto-summarize、auto-draft、auto-schedule[AI自动摘要、自动起草、自动管理]CRMAI auto-log、auto-prioritize、auto-suggest next action[AI自动记录、自动优先级排序、自动建议下一步行动]FinanceAI auto-reconcile、auto-flag anomaly、auto-generate report[AI自动对账、自动标记异常、自动生成报告]Goal用户不需要use AI——AI只是make their existing work better10.2 Glean的实践经验Glean的产品enterprise search AI assistant[企业搜索 AI 助手]Insight from deployment最成功的客户不是那些aggressively use AI features的而是那些AI quietly improves their existing workflow的Adoption metric不是how many people click the AI button而是how much time saved per user per week十一、Why Apps Capture the Value24:30-30:0011.1 AI价值的终极流向Arvind Jain的核心论点“In the long run, all the value in AI flows to the application layer. Models become commodities. Infrastructure becomes invisible. What remains is the app that owns the workflow.[从长远来看AI的所有价值都将流向应用层。模型将变成大宗商品基础设施将变得无形。最终留下的是掌控工作流的应用程序]”证据PC era[PC时代]value flowed to Microsoft Officenot Intel or Windows[价值流向了微软Office而非英特尔或Windows]Mobile era[移动互联网时代]value flowed to Uber/Airbnb/WeChatnot iOS or ARM[价值流向了Uber/Airbnb/WeChat而非iOS或ARM]AI era[Ai时代]value will flow to workflow appsnot LLM or cloud[价值将流向工作流应用而非大语言模型或云服务]11.2 Enterprise AI的最后一公里问题Model capability ≠ Business value从model到value之间需要Data integration连接企业数据Workflow embedding嵌入工作流Trust building建立信任Change management改变管理App layer公司如Glean、Salesforce已经解决了#3和#4Infra layer公司如Databricks解决了#1和#2未来两者融合converge——infra companies build appsapp companies build infra[基础设施公司开发应用应用公司构建基础设施]十二、The Future of UI, Voice, and Data Entry30:00-37:3012.1 UI的范式转移当前GUIGraphical User Interface图形用户界面——click、type、scroll [点击、输入、滚动]未来LUILanguage User Interface语言用户界面——talk、ask、command [说话、提问、下达指令]Arvind Jain的预测“In 5 years, 50% of enterprise software interactions will be through natural language.”ButLUI不会completely replace GUI——复杂任务如data visualization仍需要visual interface[可视化界面]12.2 语音交互的企业场景最适合hands-free场景——warehouse、factory、field service最不适合quiet office environment隐私问题Key barrierenterprise security——voice data is sensitive12.3 数据输入的未来当前human types data into system[由人工将数据录入系统]未来AI auto-extracts data from conversation、document、activity [AI 能从对话、文档和活动记录中自动提取数据]Implication“data entry” as a job category will disappear“The concept of ‘entering data’ will seem as quaint as ‘typing memos’ seems today.”十三、Rapid Fire: Winners, Bubbles, Long/Short37:30-45:0013.1 赢家预测Ali GhodsiDatabricks 医疗/法律领域的垂直AI应用Arvind JainGlean 在受监管行业中掌握工作流的公司13.2 泡沫判断Ali GhodsiAI infra valuations are in a bubble——$100B valuations for companies with $5B revenue[AI基础设施估值处于泡沫中——营收不足50亿美元的公司估值却超过1000亿美元]“The infra layer is overvalued. The app layer is undervalued. That’s the trade.”[基础设施层被高估了应用层被低估了。这就是投资逻辑]Arvind Jain同意——模型层泡沫尤为严重model layer especially bubblyOpenAI $300B valuation on $5B revenue 60x revenue multiple[OpenAI 营收50亿美元估值3000亿美元 60倍营收倍数]历史先例Cisco at peak of dot-com was 50x revenue——then crashed 80%[Cisco在互联网泡沫巅峰时期的估值为50倍营收——随后暴跌80%]13.3 Long/Short(看多/看空)标的判断理由OpenAI空ShortArvind/ 中性NeutralAli模型商品化 高估值Databricks多LongAliAI 应用平台 数据护城河Glean多LongArvind工作流所有权 企业信任UiPath空ShortbothRPA 正受到原生 AI 自动化冲击Vertical AI Apps多Longboth自有工作流 领域专业知识核心观点总结关键数据95%企业生成式AI pilot的失败率MIT Research5%实现快速revenue acceleration的AI pilot比例$1BDatabricks AI product revenue run-rate3xAI客户的platform stickiness提升倍数2xAI客户的NRRNet Revenue Retention高于非AI客户50%Arvind Jain预测的5年后natural language交互占比60xOpenAI估值/收入倍数$300B / $5B核心判断95%失败率不是bug是feature——高失败率说明企业在积极探索边界真正的失败原因不是技术是组织——没有把AI embed到workflow中LLM正在commoditize——像加油站一样interchangeable壁垒在数据价值最终流向app layer——model和infra是necessary but not sufficientRPA 生成式AI 完整自动化图景——结构化非结构化数据的全面覆盖AI不是cost center是retention driver——AI客户的stickiness和NRR显著更高infra layer估值泡沫化app layer被低估——$100B infra valuations vs $10B app valuationsCIO应该把60%预算给data infrastructure——不是给AI models关键方法论企业AI成功公式Proprietary Data × Workflow Integration × Agentic SystemsCIO预算分配60% data infra 20% one use case 20% experimentationAI投资回报曲线Year 1 negative → Year 2 break-even → Year 3 compounding价值分层框架Modelcommoditizing→ Infraplatform→ Appworkflow lock-in失败学习法tolerance for failure rate of innovation分析时间2026-06-16分析人员有一只肥罗