Lio Deploys Autonomous Agents to Challenge Legacy ERP Software Grip

Lio部署自主智能体挑战传统ERP软件主导地位 企业软件老牌巨头无法实现采购自动化，因为其商业模式依赖于出售工作流席位，而非解决记录系统之外发生的繁杂运营异常。 本次对话深入探讨了垂直类AI初创公司如何通过承接静态记录系统之外繁杂的端到端工作，从而超越传统ERP和SaaS巨头。以飞机制造等工业供应链为参考基准，Vladimir Keil解释了多智能体架构如何应对非结构化沟通、复杂的成本工程，以及在间接长尾支出和高风险直接零部件采购中的供应商谈判。Seema Amble概述了智能体自主性的四级分类法，并详细阐述了阻碍传统软件供应商交付实际结果而非售卖软件席位的组织冲突。两人还共同分析了为何企业内部自研AI往往通不过80/20部署测试，以及前向部署工程如何在被忽视的高杠杆企业领域中构建起稳固的垂直软件护城河。 传统的ERP和记录系统仅捕获事后交易记录，这为AI原生初创公司提供了机会，去编排生成这些记录的上游繁杂工作流，涵盖法务、财务、工程以及供应商沟通。 (4:19) 企业对自主智能体的信任是渐进式建立的：采购方最初通过人机协同（human-in-the-loop）监督常规交易，随后将支出上限逐步从1万美元提高至10万美元以上。 (14:01) 企业DIY陷阱：在内部AI原型上达到70%到80%的模型准确率只需20%的工程投入，但由于人类仍必须审核100%的输出，在最终20%的测试框架与异常工程解决之前，自主带来的价值仍为零。 (39:35) 采购具有非对称的财务杠杆效应：由于企业的利润率结构，采购成本每降低1%，给底层损益表（P&L）带来的积极影响相当于营业收入增长10%。 (55:56)

## Background 该视频由知名风险投资机构 a16z 于 2026 年 10 月 2 日在 YouTube 平台发布。本期节目由 a16z 合伙人 Seema Amble 和 Elena Burger 对话采购垂直 AI 初创公司 Lio 的联合创始人兼 CEO Vladimir Keil，深度探讨在传统软件巨头垄断客户、数据和记录系统（System of Record）的背景下，AI 原生企业如何通过接管记录系统之外的复杂隐性工作流与端到端多智能体（Multi-Agent）系统赢得竞争优势。 ## TL;DR 对谈深入剖析了传统 ERP 等记录系统与真实业务运作之间的巨大断层——账目上一行简单的采购数据背后，往往隐藏着成百上千封往来邮件、多方利益博弈及高难度的工程成本测算。与仅能在现有系统表面挂接聊天机器人的 SaaS 巨头不同，新一代垂直 AI 能够通过人在回路（Human-in-the-Loop）渐进式建立企业信任，并在间接采购全自动闭环与直接战略采购的深度博弈中落地。嘉宾指出，企业自研往往会深陷帕累托法则下最后 20% 的工程与异常处理泥潭，而 AI Agent 在供需双边的协同运作，将彻底重塑以实体经济为底色的万亿美元级采购市场。 ## Atmosphere & style 整场对谈兼具顶尖风险投资人的敏锐宏观战略视角与一线创业者的硬核实战工程洞见，语调专业、务实且充满思想交锋。Vladimir Keil 擅长用诸如“制造客机时漏看一封延迟邮件导致数亿美元停工”或“波音买螺栓”等极具画面感且切中痛点的真实案例，将原本枯燥沉闷的供应链后台业务讲得惊心动魄；三位对话者在层层递进的逻辑推演中彼此呼应，既毫不留情地拆穿了巨头“贴皮 AI”与企业盲目自研的原型陷阱，又在终局探讨中展现出以自主智能体重塑物理世界生产力的澎湃野心。 ## Bulletin - [04:19](https://www.youtube.com/watch?v=OTQ-lFsq7zA&t=259s) 传统记录系统（如 SAP 或 Oracle 等 ERP）仅记录成交价格等表面结果，采购领域约 80% 的工作与异常处理均发生在系统之外的海量邮件、会议与电子表格中，这正是端到端 AI Agent 的核心破局点。 - [39:35](https://www.youtube.com/watch?v=OTQ-lFsq7zA&t=2375s) 企业在 AI 原型中达成 70% 的准确率并不等同于 70% 的自动化，残留的 30% 错误率会导致人工必须进行 100% 的逐项核验，甚至制造比原本纯人工流程更多的返工负担。 - [42:05](https://www.youtube.com/watch?v=OTQ-lFsq7zA&t=2525s) 垂直 AI Agent 在复杂企业场景的落地遵循帕累托法则，最后 20% 的工程适配（如跨 ERP 映射、异常分支处理与记忆管理）通常需要消耗企业 80% 的工程心血，导致内部自研往往陷入僵局。 - [44:46](https://www.youtube.com/watch?v=OTQ-lFsq7zA&t=2686s) 采购端在工业制造中掌握极高话语权，能强制制定数据标准与 RFQ 交互规范，使得平台能够从采购端撬动供应商端，演化为双边协同网络并消除交易前多达 5000 个子任务的协同摩擦。 ## Key takeaways - [00:58](https://www.youtube.com/watch?v=OTQ-lFsq7zA&t=58s) 传统软件巨头受限于“记录系统”的孤立属性，且内部在提供“工作流软件”与直接“交付工作成果”之间面临销售和架构层面的组织利益冲突，难以彻底接管端到端业务弧线。 - [04:19](https://www.youtube.com/watch?v=OTQ-lFsq7zA&t=259s) ERP 账目上一笔看似轻巧的 8,000 美元采购记录，背后往往隐藏着长达数周的 3D CAD 建模拆解、500 多封邮件往来以及数十次跨部门会议等庞大隐性成本。 - [06:18](https://www.youtube.com/watch?v=OTQ-lFsq7zA&t=378s) AI 代理的演化分为检索、流程、策略与委托四个层级；现有传统软件多数停留在检索与简单规则流程层，缺乏介入非标业务与战略裁决的高阶判断力。 - [11:10](https://www.youtube.com/watch?v=OTQ-lFsq7zA&t=670s) 企业发票处理等流程中，标准信息提取仅占 20% 的“快乐路径”，而虚假发票、条款冲突与规格偏差等占 80% 的异常处理才是企业真正的核心痛点。 - [16:49](https://www.youtube.com/watch?v=OTQ-lFsq7zA&t=1009s) 企业对 AI 代理的自主谈判授权遵循渐进式信任构建，通常借助人在回路机制从 1 万、2 万美元的小额支出逐步放宽至 10 万美元及以上的高风险决策。 - [16:49](https://www.youtube.com/watch?v=OTQ-lFsq7zA&t=1009s) 大企业长期由于人力匮乏而默认放任 5 万美元以下的长尾支出直接付款，部署 AI 代理在此类真空地带的主动拦截与压测下行风险极低且收益显著。 - [21:00](https://www.youtube.com/watch?v=OTQ-lFsq7zA&t=1260s) 在重工业或数据中心等高精尖物理建设中，关键零部件延误两周的信息若淹没在未读邮件中，可能连锁引发数亿美元停工损失，需要 Agent 基于全球资讯与预测市场进行前瞻性概率推演。 - [26:26](https://www.youtube.com/watch?v=OTQ-lFsq7zA&t=1586s) 单一智能体或 Copilot 无法兼顾法务、财务、成本工程及跨系统执行，必须依靠专精于不同节点且业务顺序严密的多智能体系统才能实现端到端闭环。 - [30:39](https://www.youtube.com/watch?v=OTQ-lFsq7zA&t=1839s) 间接采购追求极致的批量自动化与长尾节流，而涉及数亿美元的直接采购则要求资深工程师深度介入，Agent 主要充当幕后军师或谈判桌上的实时事实核验副驾。 - [34:26](https://www.youtube.com/watch?v=OTQ-lFsq7zA&t=2066s) 通用大模型已在文本与报表处理上高度商品化，但应付成本建模与价格基准测算等核心场景依赖专有内部流水和隐性商业直觉，必须结合垂直微调与以结果为导向的架构攻坚破局。 - [37:58](https://www.youtube.com/watch?v=OTQ-lFsq7zA&t=2278s) 垂直 AI 公司的真正防御性与持久壁垒来源于接管端到端工作所有权并与客户日常运营深度锁定，而非理论上预设的分步护城河。 - [47:39](https://www.youtube.com/watch?v=OTQ-lFsq7zA&t=2859s) 买卖双方虽然在最终价格上存在博弈，但在达成交易前的数千个执行子任务（如填表、核对、加速排期）上诉求完全一致，双边 Agent 协同能大幅消除交易摩擦。 - [49:59](https://www.youtube.com/watch?v=OTQ-lFsq7zA&t=2999s) 针对企业定制化难题，前线部署工程师（FDE）的核心目标是通过将微小业务差异转化为软件自服务功能，最终在当前任务上“自动化掉自己的工作”。 - [55:56](https://www.youtube.com/watch?v=OTQ-lFsq7zA&t=3356s) 采购降本在损益表（P&L）上具备极高的乘数效应，在竞争激烈的行业中，硬性节约 1% 的采购支出对利润的提振往往等同于拉动 10% 的销售额增长。 ## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=OTQ-lFsq7zA&t=0s) 开篇导言与核心观点 如果一家企业要制造一架现代喷气客机，背后往往需要与成千上万家供应商进行协同。在这种高复杂度的供应链中，只要有供应商发来一封邮件称“抱歉，某个特定零部件将延迟两周到货”，而采购团队一旦疏忽漏看了这封邮件，就可能直接引发数亿美元的连锁停工与延误损失。 采购在传统企业架构中长期被局限在一个狭窄的单点操作环节里，但实际上它的触角时刻连接着法务、财务、运营以及跨部门的各类软件系统与团队。对于原生 AI 初创企业而言，真正的战略机遇正在于此——接管并承载整套端到端（End-to-End）的复杂工作流。与此同时，没有任何一家大型企业会在第一天就放手让 AI 代理全权自主处理价格谈判。企业起初既不信任初创公司，也不信任底层技术的可靠性；唯有通过严谨的“人机协同”（Human-in-the-Loop）渐进模式，持续将人工反馈与业务逻辑输入模型，企业才会逐步建立信心，从 1 万美元、2 万美元的小额谈判，一路放宽至 10 万美元乃至更高金额的商业决策。 - 飞机制造的供应链痛点：漏看一封关于零部件延误两周的关键邮件，直接后果可达数亿美元损失。 - 渐进式信任构建：企业对自主谈判代理的授权遵循阶梯式跃升（如 1 万美元 → 2 万美元 → 10 万美元）。 --- ### [00:58](https://www.youtube.com/watch?v=OTQ-lFsq7zA&t=58s) 为什么 AI 原生初创公司依然能击败软件巨头 a16z 合伙人 Seema Amble 在其近期发表的文章《巨头来袭》（*The Incumbents Are Coming*）中提出了一个所有 AI 应用层创业公司都必须面对的尖锐问题：当拥有庞大客户基础、掌握核心业务数据、且系统早已渗透进企业日常运转的传统软件巨头，将前沿大模型直接嫁接到自身产品栈时，AI 初创公司究竟还有何立足之地？ 回顾传统的软件竞争格局，初创企业与现有巨头之间的博弈常被总结为“初创公司在巨头获得创新之前能否拿到分发渠道，而巨头在初创公司获得分发之前能否完成创新”（引用自 a16z 合伙人 Alex Rampell 的名言）。但当下 AI 时代的变量在于，现有巨头似乎能够轻易接入顶级大模型——例如形成所谓的“Claude Force”（Anthropic 的 Claude 叠加 Salesforce），或者各大 SaaS 厂商直接打包大模型功能。在这种巨头兼备分发、数据和熟练用户习惯的压制下，AI 原生初创公司的生存空间究竟在哪里？ Seema 认为，AI 原生初创公司之所以依然具备核心胜算，关键在于传统软件巨头受限于它们所充当的“记录系统”（System of Record），它们只能记录孤立的结果，根本无法完成端到端的工作全貌。以客户流失退款为例：当一位客户在取消订阅后因被扣费而致电客服时，妥善解决这一纠纷绝不仅仅是让 AI 在聊天窗口里回复一句“您被多扣了钱”那么简单。这趟业务闭环必须穿透计费账单、梳理所有聊天记录、并核对法务合同条款。这类上下文分布在各种系统和人员之间，单个记录系统根本无法触及。而 AI 原生初创公司的真正价值，就在于彻底接管这整条业务弧线（End-to-end arc），无论是法务领域从起草文书到开庭审理的全程闭环，还是采购领域的复杂跨部门协同。 --- ### [04:19](https://www.youtube.com/watch?v=OTQ-lFsq7zA&t=259s) ERP 账目上“8000 美元”背后的庞大隐性工作 对于传统记录系统与实际工作之间的巨大脱节，Lio 联合创始人兼 CEO Vladimir Keil（Vlad）以企业采购场景给出了极具画面感的拆解。通常外人谈及采购，脑海中浮现的只是最终的价格数字；而当打开 SAP 等 ERP 系统时，看到的往往也极其简单轻巧——界面上往往就孤零零躺着一行字：“铝材采购：8,000 美元”。 然而，这看似简单的 8,000 美元结果，背后隐藏的工作量却是一座巨大的“业务冰山”。这就像在销售端，创始人拿着一份刚签好的大客户合同向投资人展示“我们又签了一家企业”，表面上只是一张签名页，但其背后实际经历了 30 多场多方利益相关者会议、往来了 500 多封沟通邮件、反复推敲了 20 多个 Excel 模型。 采购端亦是如此。ERP 中最终记录下的 8,000 美元，绝不会展示供应商起初曾强硬反弹并坚持要价 10,000 美元；也不会记录下企业的成本工程师曾连续花费 3 周时间，通过海量 Excel 表格测算与 3D CAD 建模，精细拆解每一个零部件的原料、加工工序与理论底价。绝大部分高价值、高消耗的采购工作，自始至终都发生在 ERP 以及任何既有记录系统之外。 - 表象与现实的反差：ERP 界面记录的只是一行简单的最终成交价（如 8,000 美元）。 - 隐藏在记录之外的协作成本：包含数十次多方会议、500 多封往来邮件、20 余个复杂电子表格模型。 - 工程级评估深度：成本工程师需耗费长达数周进行 3D 几何建模与工艺成本拆解，以对抗供应商最初的溢价（例如 10,000 美元要价）。 --- ### [06:18](https://www.youtube.com/watch?v=OTQ-lFsq7zA&t=378s) AI 代理的四层演进阶梯：从检索到委托决策 过去一年里，许多传统 SaaS 巨头所谓的“AI 战略”，实际上只是简单粗暴地在现有系统上贴上一层对话界面，即所谓的“挂个聊天机器人策略”（Slap on a chatbot strategy）。针对这种流于表面的现象，Seema 在其分析中将企业级 AI Agent 的自主性与判断力演化划分为四个层级：检索代理（Retrieval Agents）、流程代理（Process Agents）、策略代理（Policy Agents）以及委托代理（Principal Agents）。 ```mermaid flowchart LR accTitle: AI 代理的四个能力层级 accDescr: 展示从检索代理到流程代理、策略代理及委托代理的判断力递增路径 subgraph LowJudgment ["规则与数据检索（巨头主阵地）"] A["1. 检索代理<br/>(Retrieval Agent)<br/>查询合同条款与宕机记录"] B["2. 流程代理<br/>(Process Agent)<br/>按规章流程发放既定赔偿"] end subgraph HighJudgment ["复杂业务判断与战略博弈（初创突破口）"] C["3. 策略代理<br/>(Policy Agent)<br/>判定非标场景自由裁量权"] D["4. 委托代理<br/>(Principal Agent)<br/>衡量客户全生命周期战略决策"] end A --> B B --> C C --> D ``` 这四个层级的递进关系可以通过一个“企业客户遭遇服务宕机申请赔偿”的经典案例清晰展现： 1. **检索代理（Retrieval Agent）**：这是现有绝大部分传统软件内置聊天机器人所处的层级。当客户投诉宕机要求赔偿时，检索代理的作用只是从数据库中调取客户的合同条款，核实当时是否确实发生过系统故障，并提取客户基本档案信息进行摘要反馈。它本质上只是系统数据的提取与拼装。 2. **流程代理（Process Agent）**：代理开始具备跨系统推进行动的能力。根据企业现成的《操作手册》，如果记录显示系统在特定日期发生过停机，代理便依据既定规则计算出应赔金额，自动流转到财务系统生成账单抵扣或信用额度。在此阶段，系统虽然执行了动作，但全流程依然是机械遵照既定规章，不存在主观判断。 3. **策略代理（Policy Agent）**：系统开始介入现实中边界模糊的自由裁量领域。比如，客户遭遇的断网停机仅持续了 20 分钟，公司政策中并没有对“20 分钟停机是否构成实质性严重故障”写明非黑即白的界限。此时策略代理需要理解业务语境，评估事件严重程度并代为做出合理的政策适用裁定。 4. **委托代理（Principal Agent）**：代理上升到战略权衡的高度。代理不仅考虑单次故障本身，更会综合权衡该客户的战略价值、全生命周期收益与客户关系维系成本——即便规章流程只支持补偿 500 美元，但考虑到此次故障恰逢客户业务最敏感的窗口期，代理可能会决定主动提供更高额度的商业补偿，以保护长远业务关系。 目前市场上绝大多数巨头推出的 Agent（包括各类辅助签名、从会议纪要提取字段填入 CRM 的工作流代理），充其量仅处于第一层（检索）并刚触及一点第二层（流程），它们根本未曾涉足需要承担实质性业务判断与权衡的高阶层级。 --- ### [09:27](https://www.youtube.com/watch?v=OTQ-lFsq7zA&t=567s) 传统巨头的战略困境与内在掣肘 面对更高层级的业务判断，传统巨头并非不想全力推进，而是深陷一种“既握有巨大优势、又被其内在机制死死束缚”的独特矛盾中。 一方面，巨头拥有绝对的分发优势与天然的客户信任纽带，这使它们追加销售新模块变得极为容易。以 Salesforce 推出 Agentforce 为例，许多大型企业客户几乎不假思索就会选择签约开通，特别是当这类功能几乎是以极低成本甚至免费捆绑赠送的方式提供时。这种牢不可破的分发渠道和品牌背书，看似让巨头占尽了先机。 然而，巨头虽然在口头上和市场宣传中宣称自己正在从检索跨越到流程与策略层，甚至争相与 OpenAI、Anthropic 等前沿基础模型实验室建立战略合作，试图用外部模型的顶级能力来为自己的老旧软件“注入超级动力”；但巨头内在的利益机制、组织惯性以及单一记录系统的底层架构缺陷，却在深层次对其形成了巨大钳制，使其难以真正涉足深水区的核心业务判断。 ### [09:59](https://www.youtube.com/watch?v=OTQ-lFsq7zA&t=599s) 现存巨头的内部冲突：工作流软件与交付结果的天然博弈 在企业级软件市场中，现有行业巨头（Incumbents）即便拥有客户基础、数据和记录系统（System of Record），在迈向高阶 AI Agent 时仍面临深刻的内部组织悖论。Seema Amble 指出，当软件从提供“工作流”（Workflow）转向直接“解决交付工作”（Resolving the work）时，会触发激烈的内部利益冲突。以客户支持场景为例，为人类坐席提供操作界面的工具与端到端彻底解决问题的自主系统，面对的是截然不同的买家，由完全不同的副总裁（VP）与业务线团队负责。这种组织割裂使得巨头在销售激励和产品定位上面临“左右互搏”的经典困局：两个各自背负指标的部门根本无法就销售对象和激励机制达成一致。 与此同时，创业团队与巨头在企业信任的建立逻辑上存在本质差异。巨头天然享有极高的客户基础信任，但这恰恰成了他们的包袱——一旦巨头过早推送尚不成熟、决策粗糙的 Agent 产品，原有的品牌信用就会迅速被侵蚀殆尽，形成“一招不慎满盘皆输”的脆弱性；而对于初创公司而言，信任虽然必须从零开始逐步赢得，却也赋予了它们轻装上阵、快速试错并用新一代架构颠覆市场的空间。 ### [11:10](https://www.youtube.com/watch?v=OTQ-lFsq7zA&t=670s) Agent 的演化象限：从 20% 的快乐路径到 80% 的异常处理 面对 Elena Burger 提出的 Lio 在“检索、流程、策略、委托”（Retrieval, Process, Policy, Principal）四层 Agent 象限中处于何种位置的问题，Vladimir Keil 透露，Lio 的能力实际上横跨了所有层级，具体部署模式完全由任务的复杂度与风险敞口决定。在风险可控的场景中，系统已经能够完全自主运行；而在高风险、大额预算审批中，则严格保留人在回路（Human-in-the-Loop）。值得注意的是，Lio 在发展过程中发现，传统意义上的“流程型 Agent”很快沦为了一项“副线任务”（Side quest）。 以发票处理（Invoice Processing）为例，传统的企业软件逻辑非常机械：从文档中检索提取信息，跨系统核对单据，再将其写入 SAP 或 Oracle 等 ERP 系统中。Vladimir 直言，这套看似占据现有发票软件 100% 市场份额的标准操作，实际上只解决了真实业务场景中 20% 的“快乐路径”（Happy path）；真正耗费大量精力、占据 80% 核心痛点的，是那些令人头疼的异常情况——例如遭遇虚假欺诈发票、条款不一致、规格偏差或非标准结算流程。Lio 敏锐地抓住了这一分水岭，从单纯的信息流转迅速转向深度的“异常处理”，并通过始终保持小步快跑的技术领先性（从三年前仅凭文档检索惊艳客户，到迅速演进至交付端到端自主与长时间运行的长周期 Agent），持续刷新企业客户对新一代工具的认知。 ```mermaid flowchart TD accTitle: Lio Agent Capabilities Progression accDescr: 展现从仅占20%市场需求的标准流程处理向80%复杂异常处理与高阶自主谈判的演进路径 subgraph StandardWork["现有传统软件市场（约占 20% 实际工作）"] A["信息检索与提取"] --> B["跨文档数据核对"] B --> C["回写核心 ERP 系统<br/>(SAP / Oracle)"] end subgraph DeepValue["Lio 核心突破领域（占据 80% 真实痛点）"] C --> D{"路径分流：遭遇非预期异常？"} D -->|"发票造假 / 条款冲突 / 规格偏差"| E["复杂异常处理与策略分析"] D -->|"多方博弈与商业条件重定"| F["多模态深度分析与自主外部谈判"] end ``` ### [14:01](https://www.youtube.com/watch?v=OTQ-lFsq7zA&t=841s) 外部信任与采购谈判：让人工智能代表企业去博弈 让企业接受内部流程自动化是一回事，而放手让 AI Agent 代表企业去与外部第三方供应商博弈谈判，则是完全不同维度的信任挑战。Seema 指出，许多企业过去在使用老牌供应商的自动化工具时深感“被画大饼给坑了”（felt burned by the incumbents），那些产品在实际交付中往往远落后于宣传噱头。为了拆解这一难题，Elena 邀请 Vladimir 深入描摹真实采购场景的复杂版图及其背后的多方博弈。 Vladimir 澄清道，大众认知中的采购往往被简化为个人层面的日常买买买，但企业级采购实际上是“驱动整个经济运转的庞大中枢系统”。它绝非单一维度的敲定价格，而是需要法务、成本工程（Cost Engineering）、财务和运营采购等多个部门协同裁决的交织网络。要在这种错综复杂的业务环境中建立信任，Lio 采取了双轨策略：在技术上，他们在 ChatGPT 爆发前夕就提前数周入局，深度摸清了企业的实际顽疾，凭借扎实的 Eval 评估框架和快速交付能力验证实力；在落地上，他们清醒地意识到没有客户会在第一天就开启“全自动自主谈判”，而是通过透明且无缝的人机协同，为后续的权限下放搭建梯级跳板。 - 采购的多方利益相关者：法律合规核验、成本工程测算、财务预算把控与采购执行团队必须跨职能协同。 - 市场信任赤字：现有客户对传统巨头过去未兑现的自动化承诺抱有天然戒备，新进解决方案必须靠硬核落地来重塑信心。 - 产品化评测机制：通过建立严密的内部评测（Evals），确保模型在涉及多方利益的商业交涉中行为受控。 ### [16:49](https://www.youtube.com/watch?v=OTQ-lFsq7zA&t=1009s) 人在回路与分级授权：从长尾小单到数百万美元复杂决策 为了化解企业对算法失控的担忧，Lio 设计了精密的“人在回路”信任递进机制。即便 Lio 训练出了在宏观层面表现完美的采购谈判 Agent，它依然无法预知某家特定的《财富》前十强（Fortune 10）企业内部隐秘的运作习惯与潜规则。通过在谈判关键节点引入人类反馈，系统得以在实际交互中高频吸收企业的特有知识与决策风格。随着可靠性得到反复印证，客户授予的自动化额度会呈现阶梯式跃升——从最初的 1 万美元、2 万美元小额订单，逐步放宽至 10 万美元级别的自主决策。 对于涉及数百万美元、需要分析复杂 3D 计算机辅助模型（CAD）及精密工程技术图纸的重度采购，Lio 则刻意采取人机协同的长周期运行模式（Long-running agents）：系统自主进行长达数小时的多模态工程与成本数据分析，随后向成本工程师提出专业建议并获取反馈，再进入下一轮商务动作。 在不同金额层级之间，风险收益比呈现出戏剧性的分化： - **小额长尾支出（Tail Spend）的治理黑洞**：由于采购人力严重不足，大型企业在过去通常默认忽略 5 万美元以下的小额支出，甚至直接全额付款不予议价。Vladimir 戏称这是初创企业过去可以利用的一个“漏洞”——“直接给大企业寄一张 4 万美元的账单，他们根本抽不出人手来谈判，大概率会直接付款，除非他们部署了 Lio 的 Agent 来主动拦截并反向压测”。在这类“此前从未有人力去管”的空白地带，部署 Agent 的下行风险无限趋近于零，系统只要挤出任何折扣都是纯粹的增量收益。 - **关系型采购的边界与温度**：商业合作并非单纯冰冷的数字游戏。虽然部分技术或营销、播客类服务在总支出盘子中占比很小，但其背后的战略伙伴关系却至关重要。过度、苛刻地使用算法压榨价格可能会破坏关键商业生态，因此系统必须在冷酷的成本最优与有温度的人际关系维护之间保持精妙的权衡。 ### [19:32](https://www.youtube.com/watch?v=OTQ-lFsq7zA&t=1172s) 关系维护、人在回路与跨职能的复杂谈判 在高端商务合作中，许多采购事项高度依赖长期建立的信任关系。例如在搭建播客演播室时，企业绝不希望随便找个陌生团队来做，因为现有服务商已经深谙机构的运作习惯和录音规范。因此，Lio 在处理这些涉及紧密关系的环节时采用了“人在回路”（Human-in-the-loop）的机制。虽然大部分日常流程可以自主运行，但采购负责人非常在乎沟通时的语气语调（Tone of Voice）和人际关系的维系。 谈判本身绝不仅仅是敲定一个价格数字，它还牵扯到极度复杂的跨部门协作： - **法务（Legal）**：负责合同架构的设计与条款审核。 - **财务（Finance）**：把控付款与回款节奏（Collection Design）。 - **成本工程（Cost Engineering）**：深入分析底层成本结构。 - **采购商务（Procurement Commercials）**：处理商务条款与供应商对接。 参与这些决策的人绝非普通的后勤文员，而是掌握着特定行业、特定公司和特定流程专门知识的高度专业化人才。长线运行的 Lio 智能体正是由这些专业洞察驱动的。正如 Seema Amble 所指出的，采购在历史上曾被视作一个封闭的隔间，但如今它横跨法务、财务、多套软件系统，并将不同领域的通才与专才紧密连接在一起。 --- ### [21:00](https://www.youtube.com/watch?v=OTQ-lFsq7zA&t=1260s) 飞机与数据中心级采购：500 封邮件中藏着的上亿美元危机 当把视角从普通办公用品切换到制造无人机、人形机器人、飞机或建设大型 AI 数据中心时，采购的本质就完全变了。Lio 强调，他们所谈论的采购绝不是买笔记本电脑或铅笔——这些问题早在三年前就已经被轻松解决。在真正重资产、高精尖的物理世界建设中，哪怕极其微小的摩擦，都会产生灾难性的连锁反应。 以制造飞机或数据中心为例，企业需要协调成千上万家供应商并建立工厂。如果其中某一个特定零部件延误两周送达，整个项目的停工损失可能高达数亿美元，甚至导致整体工期被迫后延。 最危险的隐患往往隐藏在看似枯燥的日常事务中： - 供应商发来一封通知：“抱歉，该零件将推迟两周送达。” - 这封关键邮件只是采购经理 Outlook 或 Gmail 收件箱中 500 封未读邮件里的普通一封。 - 一旦采购经理遗漏了这封邮件，上亿美元的损失便立刻铸成。 - 传统记录系统（System of Record / ERP）最终往往只机械地更新一个到货日期（例如从“本周三”变成“下周三”），却完全掩盖了背后可能引发的上亿美元停工代价。 Lio 的智能体所做的不仅是抓取信息，更核心的是评估影响：判断该延误是否会产生实质破坏、影响到底有多大，以及如何在第一时间制定补救方案。 --- ### [24:00](https://www.youtube.com/watch?v=OTQ-lFsq7zA&t=1440s) 物理世界的不可抗力与基于概率的上下文决策 面对物理世界中无法改变的意外（例如货船损坏、海运海峡关闭或货物坠海），AI 智能体无法凭空扭转物理事实，但能够通过全维度的上下文信息进行前瞻性的“概率预测”。 智能体可以通过综合数据评估不同供应商的履约可靠性： - 假设供应商 A 的货物丢失率高达 20%，但价格极为低廉；供应商 B 的丢失率仅为 1%，但价格贵了 10 倍。 - 在一般场景下可能选择供应商 A，但在某些核心部件不可延误的关键节点上，系统会判断多付 10 倍的价格是完全合理的，因为这大幅提升了零件准时且完好运抵的概率。 这种决策依赖于将企业内部与外部世界的上下文全面打通： - **企业内部多利益相关方**：掌握各工厂的实际紧急程度与财务承受底线。 - **外部全球资讯**：实时监控突发新闻事件。 - **预测市场数据**：例如参考 Polymarket 等平台对地缘冲突或航道中断的预测赔率。 - **供需双边动态**：打通买方与卖方之间的信息孤岛。 通过整合内外数据，智能体在风险转化为既成事实之前便完成了概率推演与路线重选，重新定义了企业间（B2B）的交易协同。 --- ### [26:26](https://www.youtube.com/watch?v=OTQ-lFsq7zA&t=1586s) 为什么端到端解决问题必须依靠多智能体系统（Multi-Agent System） Lio 采用了多智能体协同架构。人类现实工作中的采购任务通常涉及 8 个利益相关方、3 个不同职能部门以及至少 5 套不同的软件工具。一个孤立的智能体或单向检索的 Copilot 根本不可能端到端地完成整项工作。 在三年前，Lio 也曾从类似 Copilot 的单点检索工具做起，随后演进到单智能体（Single Agent），但很快便撞上了能力天花板：一个智能体不可能在处理复杂谈判的同时，兼顾合同法律条款的推敲，还要时刻扫描外部新闻动向。只有构建多智能体系统，让各自专精的 Agent 按照严格的业务先后顺序互通信息、协同推演，才能真正实现全流程的端到端自动化。 ```mermaid flowchart TD accTitle: Lio 多智能体端到端采购流程 accDescr: 展示从一线需求提出、多智能体协同评估，到最终履约的端到端采购系统架构。 subgraph FrontEnd["需求发起端"] A["一线工人 / 需求方"] -->|"拍照 / 上传报价单 / Excel"| B["需求解析 Agent"] end subgraph InternalOps["内部协同与库存核查"] B --> C["内部库存 Agent"] C -->|"跨工厂库存调拨查询"| D{"内部是否有货？"} D -- 是 --> E["内部调配并出库"] end subgraph ExternalProcure["外部寻源与商务多智能体网络"] D -- 否 --> F["寻源 Agent (Sourcing)"] F --> G["供应商评估 Agent (外部情报/履约概率)"] G --> H["商务谈判与合同 Agent"] H --> I["法务与财务审批 Agent"] end I --> J["最终订单确认与 ERP 数据写回"] ``` --- ### [28:03](https://www.youtube.com/watch?v=OTQ-lFsq7zA&t=1683s) 波音采购一颗螺栓：从一线随手拍到全流程自动化 以飞机制造巨头（如波音 Boeing）采购一颗特种螺栓为例，这项工作如果交给 Lio 的多智能体系统，已经可以做到全流程自主运行。 传统流程极其繁琐脱节： - 需求往往源自现场施工人员，但他可能每两周才开一次电脑，对 SAP 等复杂的 ERP 软件一窍不通。 - 在采购部门之外，没有任何一线工人关心物料分类（Categories）、总账科目（GL Accounts）或框架协议（Framework Contracts）。 在 Lio 的体系下，一线人员只需拍摄一张螺栓的照片，或者上传一张报价单、一份 Excel 表格，采购部门之外的人员便无需再做任何多余操作。随之，后方的智能体集群立即接管任务： 1. **库存检查智能体**：第一时间检索全系统库存，甚至直接向兄弟工厂询问：“我们能否在内部调拨这批螺栓？” 2. **寻源智能体（Sourcing Agent）**：若内部无法解决，寻源智能体立刻介入，排查已签约的合格内部供应商与潜在的外部供应商，启动后续的比价与采购逻辑。 ### [29:33](https://www.youtube.com/watch?v=OTQ-lFsq7zA&t=1773s) 采购 Agent 的端到端自主闭环与直接采购的引入 在真实的采购场景中，自动化绝不仅仅是触发一条提示词，而是涉及跨系统、跨媒介的漫长链条。一个 Agent 负责起草询价单（RFQ）并通过电子邮件发送给供应商；随后收到的供应商回复往往五花八门——有的内容荒唐无关，有的写在正文里，有的则附带复杂的 PDF 或 Excel 电子表格。系统必须准确提取这些非结构化信息，结合历史价格基准（price benchmarking）识别谈判空间，并交由决策 Agent 判断下一步行动。根据采购场景的性质，后续可以无缝调度不同的执行策略：需要人工把关的战略性谈判、全自主自动化谈判，或是启动反向电子竞价（e-auctions）。 从最初的询价、比价谈判，到订单确认、发货跟踪乃至最终的发票核对，Lio 致力于让多 Agent 系统全流程自主运行，并完整捕获中间的所有业务上下文。这种高度自动化的闭环最先在通用物料采购中跑通，但其真正的商业威力在于向复杂度极高的直接采购（direct procurement）延伸。 ```mermaid flowchart LR accTitle: 端到端采购 Agent 协作流 accDescr: 展示从询价单发送、非结构化报价提取、基准比价到多模式谈判与履约跟踪的完整流程。 subgraph RFQ["询价与数据抓取"] A["生成并发送 RFQ"] --> B["邮件与附件解析<br>(PDF / Excel / 文本)"] end subgraph Decision["分析与决策"] B --> C["价格基准对比"] C --> D{"决策 Agent 路由"} end subgraph Exec["执行与履约"] D -->|"低风险/长尾"| E1["自主谈判 / 竞价"] D -->|"战略性物料"| E2["人机协同谈判"] E1 --> F["订单确认与物流追踪"] E2 --> F F --> G["发票审计入账"] end ``` ### [30:39](https://www.youtube.com/watch?v=OTQ-lFsq7zA&t=1839s) 间接采购与直接采购：自动化与战略协同的分野 直接采购与间接采购在业务逻辑和商业价值上存在根本差异。间接采购的核心诉求是极致的效率与降本自动化，涵盖工厂建设消耗的维修维护配件（MRO）、螺栓等标准件，以及笔记本电脑、铅笔、市场营销服务乃至录制播客所需的演播室搭建。企业通常面临成千上万件“每天需要做上千次但大家一次都不想做”的重复繁琐工作，自动化 Agent 可以全天候接管。更重要的是，Agent 还能介入那些“企业原本由于精力有限而从未涉足、但一旦规模化推进就能产生惊人利润表（P&L）改善”的长尾支出谈判。 直接采购则完全是另一种逻辑。直接物料指的是直接进入最终产品的关键零部件，例如制造飞机、无人机或机器人所必需的核心供应链。这类采购面对的不是 5 万家零散供应商，而是经过极其严格审核的 100 到 2000 家核心供应商。单个供应商名下的采购金额往往高达数亿美元甚至 10 亿美元。在这类核心战略物料上，企业绝不可能直接放手让 Agent 自动敲定交易，而是需要历时三个月的漫长博弈，需要资深工程师深度介入，细致分析铝、原油等大宗原材料价格指数的波动、审查工程图纸的工艺要求并严格核验零件质量。 - **间接采购（Indirect Procurement）**：覆盖 MRO 零部件、办公设备、IT 资产、营销服务等。供应商数量庞大（数以万计），核心目标是通过自主谈判与流程自动化消除琐碎重复劳动，挖掘长尾支出的节流潜力。 - **直接采购（Direct Procurement）**：直接构成产品核心结构的物料（如航空航天、智能硬件关键部件）。供应商集中度高，单笔合约金额达数亿美元级别，极度依赖工程师对技术图纸、材料价格指数及公差质量的深度核算。 ### [32:36](https://www.youtube.com/watch?v=OTQ-lFsq7zA&t=1956s) 幕后智能军师：从深度备战到谈判现场的实时辅助 针对上亿美元级别的直接采购谈判，Elena Burger 提问道：Agent 究竟是作为“台前接待”代替人类坐到谈判桌前握手过招，还是退居“幕后”提供支撑？Vlad 明确指出，在涉及多方利益博弈的战略级交易中，Agent 绝大多数时候都扮演着幕后军师的角色，因为这种交易“90% 的工作都在于事前的极度充分准备”。最终在 ERP 或记录系统中看到的，可能只是简单的一行数字变动——原本 10 亿美元的采购额最终以 9 亿美元成交；但这看似轻松的 1 亿美元节约，背后凝结了 10 个人全职投入长达三个月的缜密推导与成本拆解。 然而，Agent 并非仅限于离线做案头准备，它正在走进真实的谈判现场。Vlad 描绘了一个极具实操价值的“谈判副驾”场景：当采购人员与供应商面对面谈判时，屏幕会实时弹出战术情报提示。如果供应商提出“近期原油价格指数上涨了 10%，因此零件单价必须跟涨 10%”，传统的采购人员很难在谈判桌上立即拆穿，但后台连接着工程物料清单（BOM）与大宗商品的 Agent 能在一秒内弹出精准提醒：“原油价格上涨 10% 属实，但该零部件中原油衍生材料的成本占比仅有 30%，供应商所主张的合理涨幅上限仅为 4% 而非 10%。”这种前置赋能彻底改变了现场博弈的力量天平。 ### [34:26](https://www.youtube.com/watch?v=OTQ-lFsq7zA&t=2066s) 超越通用大模型：线束编排与目标驱动微调的界限 随着 Harvey 和 Decagon 等垂直领域头部企业在垂直模型微调上持续发力，采购领域的模型底层架构也正在发生演变。Vlad 认为，来自各大科技巨头的通用基础大模型已经高度商品化，擅长处理基础通用任务，例如解析复杂的 PDF 文件、生成标准 Excel 表格等。在许多业务环节中，将基础模型配合优秀的外部控制线束（harness / 编排工程）就能完成全部工作；但对于某些极其专业的深度场景，即使堆砌最顶级的大模型与最复杂的编排框架，也只能达到 80% 的完成度，剩下的 20% 必须依赖领域微调才能跨越。 最典型的攻坚点在于成本工程师所做的“应付成本建模”（should-cost modeling）与跨企业价格基准测算（price benchmarking）。在理想状态下，采购员希望将一份零件工程图纸拖入系统，系统就能直接测算出其真实合理的物理制造成本。然而，相关的企业内部历史采购记录、工程折让条款与专属图纸数据均为非公开资产，通用基础模型在预训练阶段根本无法接触到这些专有数据。 此外，采购领域存在着大量无法用硬编码规则清晰描述的隐性“商业直觉”（gut feeling）。比如麦肯锡与波士顿咨询（BCG）就同一项目开出的报价可能相差十倍，内容看似相同，但经验老道的采购主管能够凭借直觉判断哪份报价真正合理；视频创作者制作一条视频应该付多少酬劳，同样难以通过明确的规则公式量化。通用语言模型擅长文本推导，但无法感知这种定价直觉。因此，行业正在探索类似 JEPA 等以目标/结果为导向（outcome-based）的架构，让模型基于海量专有业务规格输入，直接输出最优定价与成本估算，真正攻克应付成本建模的瓶颈。 - **通用任务的商品化**：基础语言模型完全胜任邮件解析、PDF/Excel 报表生成等流程层面的通用办公需求。 - **80% 的能力瓶颈**：单纯依靠通用基础模型结合工程 Harness 无法解决硬核工程核算与复杂定价博弈。 - **专有数据壁垒**：涉及企业真实采购流水、零部件图纸公差及商业协议的私有数据，公有基础大模型无法直接获取。 - **基于结果的端到端学习**：采购行业存在大量无法规则化的直觉经验，需要利用垂直微调与结果导向型模型实现精准的价格基准估算。 ### [37:58](https://www.youtube.com/watch?v=OTQ-lFsq7zA&t=2278s) 垂直 AI 企业的持久壁垒：从记录系统走向工作所有权 当被问及基础模型实验室正加速切入垂直行业时，什么样的垂直 AI 企业才能建立起持久的壁垒，Seema Amble 分享了 a16z 的投资视角。她认为，在行业早期试图精准预言某种“理论上的护城河”是非常不切实际的。纵观所有伟大的早期企业，它们在创立之初并不是按部就班地设计“先做前六步，到了第七步就能拥有护城河”的完美蓝图，而是脚踏实地专注于“赢取客户的深度信任，不断向他们销售更多价值并捕获更大的工作面”。 垂直 AI 公司的真正防御性与持久性，来源于“真正拥有端到端的工作执行流程”（owning the end-to-end work），让客户产生深度的业务依赖： - **旧时代系统（如传统 CRM / ERP）**：本质上只是被动的“记录系统”（system of record），充当各笔交易和日志的数据库，需要人类员工在系统外完成海量操作后再手动回填。 - **新一代 AI Agent 架构**：从一开始就全面接管了前期的销售准备、出海获客开发（outbound）、接听并处理潜在客户入站咨询（inbound）等大量具体业务。 - **防御性沉淀机制**：当 AI 系统直接承担了关键业务动作，客户企业对该软件的依赖程度就会产生质的飞跃。这种将业务执行、专有数据资产与客户日常运转深度锁定的状态，正是垂直 AI 公司最终形成网络效应与稳固护城河的关键前置信号。 ### [39:35](https://www.youtube.com/watch?v=OTQ-lFsq7zA&t=2375s) 自研还是外采：70% 的准确率并不等于 70% 的自动化 许多潜在客户常向 Lio 提出疑问：既然现在大模型如此易用，为什么企业不能直接调用基础模型，或者在现有的系统记录（System of Record）上套一层模型自己来做，而必须购买像 Lio 这样的第三方垂直产品？ Vladimir Keil 回顾了 Lio 三年前的起点。当时在采购界一炮而红的第一个用例非常具体：读取一份供应商报价单（Quote），由检索 Agent 提取结构化信息并直接写入 SAP。这个技术看起来不算极其高深但商业价值巨大的工具，当年由三四名工程师花费了数周乃至数月才打磨出来并推向市场。而今天，这已经变成了 Lio 面试工程候选人的标准试题，候选人通常只需 8 个小时就能从零搭出一个原型。这证明在基础模型成熟的当下，快速搭建一个 Demo 确实轻而易举。但企业采购部门若因此认为自己可以在两个月内搞定整套生产系统，就会落入巨大的陷阱： ```mermaid flowchart TD accTitle: 70% 原型与生产级交付的鸿沟 accDescr: 展示自研原型仅能达到70%准确率，导致人工需全量复核，甚至增加额外工作量 subgraph Prototype["内部自研原型 (8小时 ~ 2个月)"] A["调用基础模型 / 简易 RAG"] --> B["达到 ~70% 任务准确率"] end subgraph FailureMode["生产落地陷阱"] B --> C{"70% 性能 = 70% 自动化？"} C -- "否定" --> D["人工仍需复核 100% 的数据"] D --> E["错误处理与幻觉排查产生额外返工"] end style Prototype fill:#f9f9f9,stroke:#666,stroke-width:1px style FailureMode fill:#fff0f0,stroke:#d9534f,stroke-width:1px ``` 在严肃的企业业务中，“70% 的准确率或性能绝对不等于 70% 的自动化”。由于缺乏足够的置信度和确定性，员工必须对 100% 的字段和数据进行逐一核对。结果不仅没有省下时间，反而增加了查错、纠错的额外负担，甚至制造了比原本纯人工流程更多的工作量。 - **面试挑战**：Lio 如今将公司最初上线的核心产品作为面试编程考核，要求求职者在 8 小时内构建一个能将报价单录入 SAP 的 Agent。 - **自动化的虚假换算**：在复杂的企业流程中，30% 的错误率会导致整个流程重退回人工审核链条，自动化红利完全被边际审核成本抵消。 ### [42:05](https://www.youtube.com/watch?v=OTQ-lFsq7zA&t=2525s) “帕累托法则”下的最后 20% 与企业自研困境 Seema Amble 指出，基础模型与 ChatGPT 类产品的广泛普及对垂直 AI 创业公司是重大利好，因为市场教育让企业客户对 AI 工具建立了信任和兴奋感。然而，企业一旦真正动手搞“自研内部工具（DIY）”，就会重蹈传统软件时代的覆辙。 她分享了近期与一家财富 500 强高管团队的交流案例：该企业试图自行研发一套回款处理（Cash Collection）AI 工具。耗费了三到四个月后，项目陷入了僵局。开发人员虽然收集了大量的屏幕录屏、截图和交互日志，试图汇总至一个系统，但提取出的上下文质量极其糟糕，根本不足以支撑模型做精确判断。更致命的是企业系统的现实复杂性——公司内部本就运行着两套不同的 ERP 系统，近期又准备收购一家新公司引入第三套 ERP。此时企业面临着难以承受的维护成本：谁来编写并长期维护所有的字段映射（Mappings）？谁来做回归测试？谁来针对不同系统设计五花八门的一线异常处理（Exception Handling）机制？企业管理层迅速意识到，自研 AI 内部工具与以往痛苦的自建软件项目并无二致，不仅拖累了研发，而且偏离了公司自身的核心主业。 Vladimir 补充强调，企业虽然能轻易做到 80% 的效果，但决定能否进入生产环境（Production）的，恰恰是“最后的 20%”。根据帕累托法则（Pareto Principle），这最后的 20% 往往需要耗费 80% 的工程心血： - **工程外壳（Harness）**：包括深度的系统 API 集成、上下文记忆管理、状态机编排以及复杂的异常分支处理工作流。 - **专有垂直数据**：解决冷启动与行业特定术语映射问题。 - **核心能力权衡**：大企业必须拷问自己，是否有必要把“打造与维护一支 AI 采购数字化劳动力”变成自己的核心组织能力。 ### [44:46](https://www.youtube.com/watch?v=OTQ-lFsq7zA&t=2686s) 双边 Agent 协作：当买方与卖方都在同个平台上 针对 Elena Burger 提出的是否有供应商也在使用 Agent 的问题，Vladimir 证实了行业正在走向“供需两端皆有 Agent”的终局格局。有趣的是，过去在企业软件生态中，销售端（供应商侧）在技术采纳上通常领先于采购端，但如今在财富 500 强级别的工业供应链里，供应商端虽然会用一些轻量工具（例如会议记录工具 Granola），却并没有真正部署能够自主执行端到端工作的 AI Agent。 采购在传统观念中被视为枯燥、不受重视的“后台业务”（Unsexy），而销售则是充满光环的前台业务（Sexy）。但在工业制造（如汽车、航空航天主机厂）领域，采购部门面对供应商拥有极高的话语权。采购方不仅能规定零件的技术公差，还可以直接制定供应商必须使用的交互界面、数据标准乃至 RFQ（报价请求）的回复格式。这赋予了垂直平台巨大的战略借势机会：Lio 从采购端切入，凭借采购方的强制力要求下游供应商接入同一平台，从而顺理成章地向供应商端提供辅助他们自动处理回复的 Agent，最终实现“双边交易全流程掌控”。 ```mermaid flowchart LR accTitle: 工业供应链双边 Agent 协同网络 accDescr: 采购方凭借行业话语权推动供应商接入统一平台，实现两端 Agent 协同 subgraph BuyerSide["采购方 (核心企业 / 买方)"] B_Agent["Lio 采购 Agent"] B_Goal["核心诉求: 缩短上市时间 / 加速项目推进"] end subgraph Platform["统一协同基础设施 (Lio 平台)"] Coord["中立状态同步与上下文追踪"] Rules["标准化数据接口与 RFQ 协议"] end subgraph SupplierSide["供应方 (配套厂商 / 卖方)"] S_Agent["Lio 销售 / 响应 Agent"] S_Goal["核心诉求: 减少交易摩擦 / 快速完成签单"] end B_Agent -->|强话语权规则下发| Rules Rules --> S_Agent B_Agent <--> Coord Coord <--> S_Agent B_Goal -.->|在 5000 个执行子任务上诉求一致| S_Goal ``` Seema 补充道，这种多方在同一中立平台上协作的模式在法律科技（LegalTech）中同样成立。两家立场针锋相对的律师事务所，尽管代表着不同利益的客户，但在处理复杂交易或诉讼时，都极度渴望有一个平台能自动追踪：“哪一份是最新草案？”、“目前尚未达成一致的争议条款有哪些？”、“哪些核心要点已经敲定闭环？”以往这类跨机构协调需要消耗大量初级律师的人工精力，未来完全可以由统一平台上的 Agent 自主完成。 ### [47:39](https://www.youtube.com/watch?v=OTQ-lFsq7zA&t=2859s) 价格背后的 5,000 个任务与前置部署（Forward Deployed）优势 面对“当买卖双方都拥有谈判 Agent 时，系统是否会陷入纯零和博弈死锁”的疑虑，Vladimir 给出了深刻的业务拆解： - **利益分歧的表象**：表面上看，最终的“价格”确实是零和的，买方希望压价，卖方希望溢价。 - **5,000 个协同子任务的一致激励**：价格只是漫长流程中最后导出的单一结果，在此之前存在着多达数千个中间执行任务。在这些繁琐子任务上，买卖双方的核心利益不仅没有冲突，反而完全一致——销售希望尽可能消除流程阻力、少填繁琐表格并快速成交；买方为了让飞机按期下线或数据中心按期并网，极度渴望缩短项目周期，绝不想因为人工耗费六个月去肉眼对比供应商提案而延误工期。因此，双边部署 Agent 负责消除这 5,000 个环节中的协同摩擦，能为双方带来倍增的效率回报。 针对 Elena 提出的“传统 SaaS 忌讳过度为单一客户定制，而大模型时代是否让低成本高定制成为可能”这一议题，Seema 剖析了目前垂直 AI 软件交付的底层逻辑： - 当前行业普遍采用了大量“前置部署（Forward Deployed）”的重服务交付模式，团队中往往保留着大量人工支持。 - 其根本原因在于，理解第 $N$ 个客户的脏数据、遗留流程与业务上下文，其难度往往远大于后续服务第 $N+1$ 个客户时的复用经验。企业当前仍处于 AI 代理落地生产的非常早期阶段。 - 传统在位软件巨头（Incumbents）的组织架构高度标准化、研发与交付部门严重割裂，这种天生依赖通用标准化交付的体制，使得它们在应对当下需要深度现场定制、伴随客户数据清洗的 AI 部署浪潮时，反而处于明显的战略劣势。 ### [49:59](https://www.youtube.com/watch?v=OTQ-lFsq7zA&t=2999s) 从埃森哲式的定制到“自我淘汰”的工程师 关于 AI 产品的部署模式，传统的企业软件公司往往把实施团队当作事后追加的附属服务，实施过程中的反馈极少能反哺到核心产品中。AI 原生产品的优势在于能够随时间持续学习；在建立起严密的评估闭环（eval process）后，模型不仅能接手越来越复杂的任务，更能将部署本身的定制化过程逐步自动化。过往企业需要雇用埃森哲（Accenture）这样的咨询巨头来给 SAP 做冗长复杂的定制开发，而现在的软件正让客户能够通过自主调节软件内的控制杠杆，以极快的节奏实现定制。 ```mermaid flowchart TD accTitle: AI 原生部署模式的演进 accDescr: 对比传统咨询外包、前线工程师介入和最终完全由软件驱动的自动化闭环。 subgraph Legacy["传统模式"] A["采购系统 (如 SAP)"] --> B["外包咨询公司 (如 Accenture)"] B --> C["冗长的人工定制与交付"] end subgraph Modern["Lio 的前线部署与自进化闭环"] D["前线部署工程师 (FDE)"] --> E["深入客户场景，攻克定制盲区"] E --> F["将人工流程封装为自服务功能"] F --> G["FDE 的终极 KPI：实现自身岗位的自动化"] G --> H["完全软件驱动的端到端 Agent 闭环"] end ``` 这种产品理念直接映射在 Lio 的组织架构上。Lio 团队中有高达 85% 的成员都是工程师，哪怕部分岗位没有挂工程师头衔，其背景也绝大部分属于工程技术方向。Lio 坚决不做外包咨询公司。虽然在进入大企业客户时，不同企业的采购流程存在千差万别的微小细节（nuances），必须配备前线部署工程师（Forward Deployed Engineers, FDE）深入前线，但这些工程师的核心 KPI 极其明确且颠覆：他们要像 Google 内部的自动化工程实践一样，通过把定制逻辑沉淀为软件的“自助服务（self-service）”，最终在当前任务上“自动化掉他们自己的工作”（automate yourself）。一旦这个流程完全不需要人手，他们便继续转向下一个未被自动化的复杂任务。 ### [52:22](https://www.youtube.com/watch?v=OTQ-lFsq7zA&t=3142s) 纽约 Bots & Buyers 峰会与“人人痛恨”的采购现状 Lio 近期在纽约举办了第三届“Bots & Buyers 峰会”，聚集了上百位采购领域的资深高管，包括 CPO（首席采购官）、副总裁等高规格决策者。在交流中，Vlad 揭示了该行业长达近三十年的一个怪现状：打开 LinkedIn，采购领域的生态图谱（market landscape）里挤满了超过 500 到 1000 种各色软件工具，然而现实中几乎找不到任何一个发自内心热爱采购系统的人。无论是提需求的业务端、外部供应商，甚至是采购从业人员自己，都对现有的工作流程深恶痛绝。 这种极度痛苦的根源在于：过往 26、27 年里出现的软件仅仅是在已有流程的表面做局部增效，根本没有改变从业者实际干活的方式。一线采购人员的核心阵地依然陷在无休止的电子邮件、Microsoft Teams 消息、Excel 电子表格和 PowerPoint 汇报中，完全没有任何真正的 AI 深度赋能。Lio 在峰会及其实际办公室中所呈现的，是打破单点工具（比如“又一个酷炫的发票解析功能”）的局限，提供横跨非生产性采购（indirect）、生产性物料（direct）、物流运输与财务结算的端到端体验。为了让抽象的 AI Agent 变得直观可见，团队甚至在现场搭建了真实的沉浸式实体展台，让决策者能够亲身体验 Agent 在各个节点间的流转。继纽约之后，下一场峰会将在慕尼黑举办，报名规模已迅速飙升至 700 人。 - 采购软件现状：市面上存在数百上千款单点工具，但由于未能触及底层工作流，行业净推荐值（NPS）普遍垫底，甚至有老采购直言即使面对 Coupa 等头部记录系统也充满厌烦、抗拒沟通。 - Bots & Buyers 峰会规模：从纽约办公室附近的百人闭门高管聚会，快速扩展至慕尼黑站预计超 700 人的大型行业盛会。 ### [55:56](https://www.youtube.com/watch?v=OTQ-lFsq7zA&t=3356s) 万亿美元级商机：情绪、枯燥与巨额商业价值的三重奏 采购领域的独特魅力在于其深刻的反差感。许多人觉得企业级 B2B 软件缺乏吸引力，但正是这种被普遍视作“枯燥、小众（boring and niche）”的角落，蕴藏着巨大的创业红利。从业者对现有工具长期的挫败感使得采购成为一个“高度情绪化（highly emotional）”的领域；因为过去 20 年里除了偶有一套稍微顺眼的 UI 之外几乎没有任何真正的底层革命，一旦有能真正解决痛点的产品出现，便极易让客户感到震撼和惊艳。 更关键的是采购背后惊人的商业杠杆。采购直接关乎物理现实与整体宏观经济的运转——从数据中心、飞机、汽车到无人机的建造，无一不依赖严丝合缝的采购流程。而在企业的损益表（P&L）上，采购降本具有极高的乘数效应：在高度竞争的市场中，企业要想让利润率提升 1%，通常需要依靠前端多卖出 10% 的销售额；而通过系统优化在采购端硬性节约下 1% 的支出，其释放的净利润直接等同于拉动 10% 销售额所带来的成效。这种“高度情绪化、枯燥小众、以及牵动真金白银的巨大商业影响力”的三者合一，构筑了 Vlad 眼中规模达万亿美元级别的巨大商业机会。随着全场讨论在对实体经济与企业 P&L 的重塑展望中推向高潮，对谈也正式圆满收官。 ## Notable quotes > [00:58](https://www.youtube.com/watch?v=OTQ-lFsq7zA&t=58s) 初创公司在巨头获得创新之前能否拿到分发渠道，而巨头在初创公司获得分发之前能否完成创新 — Seema Amble > [21:00](https://www.youtube.com/watch?v=OTQ-lFsq7zA&t=1260s) 抱歉，该零件将推迟两周送达。 — Vladimir Keil > [16:49](https://www.youtube.com/watch?v=OTQ-lFsq7zA&t=1009s) 直接给大企业寄一张 4 万美元的账单，他们根本抽不出人手来谈判，大概率会直接付款，除非他们部署了 Lio 的 Agent 来主动拦截并反向压测 — Vladimir Keil > [32:36](https://www.youtube.com/watch?v=OTQ-lFsq7zA&t=1956s) 90% 的工作都在于事前的极度充分准备 — Vladimir Keil > [32:36](https://www.youtube.com/watch?v=OTQ-lFsq7zA&t=1956s) 原油价格上涨 10% 属实，但该零部件中原油衍生材料的成本占比仅有 30%，供应商所主张的合理涨幅上限仅为 4% 而非 10%。 — Vladimir Keil > [39:35](https://www.youtube.com/watch?v=OTQ-lFsq7zA&t=2375s) 70% 的准确率或性能绝对不等于 70% 的自动化 — Vladimir Keil > [28:03](https://www.youtube.com/watch?v=OTQ-lFsq7zA&t=1683s) 我们能否在内部调拨这批螺栓？ — Vladimir Keil > [49:59](https://www.youtube.com/watch?v=OTQ-lFsq7zA&t=2999s) 自动化掉他们自己的工作 — Vladimir Keil ## Mentioned a16z, Seema Amble, Elena Burger, Vladimir Keil, Lio, Alex Rampell, Claude, Anthropic, Salesforce, Agentforce, SAP, Oracle, OpenAI, ChatGPT, Boeing, Polymarket, Harvey, Decagon, JEPA, Granola, Accenture, Google, Coupa, Bots & Buyers 峰会, McKinsey, BCG ## Video description a16z’s Seema Amble and Elena Burger sit down with Lio co-founder and CEO Vladimir Keil to ask where AI-native startups have an advantage when incumbent software companies already own the customer, the data, and the system of record. Their answer comes down to the work that happens outside those systems. In procurement, a final price in an ERP can hide hundreds of emails, spreadsheets, supplier conversations, engineering analyses, and decisions across legal, finance, and operations. Vlad explains how Lio uses multi-agent systems to take on more of that end-to-end work, from sourcing and RFQs to negotiation, shipment tracking, and invoices. They also discuss how enterprises learn to trust agents with increasingly consequential decisions, why the last 20% of an internal AI build can require most of the effort, and what happens when both buyers and suppliers have agents working on their behalf. Timestamps: 00:00 - Intro 00:58 - Why AI startups still beat incumbents 04:19 - The hidden work behind an $8K line item 06:18 - Retrieval, process, policy, principal 09:27 - The incumbent's internal conflict 14:48 - What procurement actually looks like 21:13 - Procurement at Boeing-scale 28:03 - A bolt order, end-to-end 37:57 - What a durable vertical AI company looks like 44:46 - When both sides deploy agents Resources: Follow Vladimir Keil on X: https://x.com/askvladi?lang=en Follow Vladimir Keil on LinkedIn: https://www.linkedin.com/in/vladimir-keil/ Follow Seema Amble on X: https://x.com/seema_amble Learn more about Lio: https://www.lio.ai/ Seema Amble’s “Investing in Lio” article: https://a16z.com/announcement/investing-in-lio/ Stay Updated: If you enjoyed this episode, be sure to like, subscribe, and share with your friends! Find a16z on X: https://twitter.com/a16z Find a16z on LinkedIn: https://www.linkedin.com/company/a16z Listen to the a16z Show on Spotify: https://open.spotify.com/show/5bC65RDvs3oxnLyqqvkUYX Listen to the a16z Show on Apple Podcasts: https://podcasts.apple.com/us/podcast/a16z-podcast/id842818711 Follow our host: https://x.com/eriktorenberg Please note that the content here is for informational purposes only; should NOT be taken as legal, business, tax, or investment advice or be used to evaluate any investment or security; and is not directed at any investors or potential investors in any a16z fund. a16z and its affiliates may maintain investments in the companies discussed. For more details please see http://a16z.com/disclosures. [Source](https://www.youtube.com/watch?v=OTQ-lFsq7zA)

[0:58] Seema Amble: 初创公司在巨头获得创新之前能否拿到分发渠道，而巨头在初创公司获得分发之前能否完成创新 [21:00] Vladimir Keil: 抱歉，该零件将推迟两周送达。 [16:49] Vladimir Keil: 直接给大企业寄一张 4 万美元的账单，他们根本抽不出人手来谈判，大概率会直接付款，除非他们部署了 Lio 的 Agent 来主动拦截并反向压测 [32:36] Vladimir Keil: 90% 的工作都在于事前的极度充分准备 [32:36] Vladimir Keil: 原油价格上涨 10% 属实，但该零部件中原油衍生材料的成本占比仅有 30%，供应商所主张的合理涨幅上限仅为 4% 而非 10%。 [39:35] Vladimir Keil: 70% 的准确率或性能绝对不等于 70% 的自动化 [28:03] Vladimir Keil: 我们能否在内部调拨这批螺栓？ [49:59] Vladimir Keil: 自动化掉他们自己的工作

[0:58] Seema Amble: The legacy incumbent is limited to their system of record and that record that they have, and they're not completing the end-to-end job. [21:13] Vladimir Keil: We don't talk about laptops and pencils... that's easy, we solved this three years ago. [32:36] Vladimir Keil: 90% of the work is preparation. [32:36] Supplier: The index for crude oil went up 10%, so we must raise our unit prices by 10%

[0:00] If you want to build an aircraft, you [0:01] need to procure thousands of suppliers. [0:03] Someone sends a confirmation of like, [0:05] hey, sorry, like this part is going to [0:06] arrive 2 weeks later. And if they missed [0:08] this email, hundreds of millions of [0:09] damage. [0:10] >> Procurement historically may have been [0:12] more in a box and now it's touching [0:14] legal, it's touching finance. It's [0:15] touching a bunch of different software [0:16] systems and people. The opportunity for [0:18] the AI native startup is to say, we're [0:20] going to own that entire endto-end arc. [0:23] >> No company and no enterprise starts with [0:25] fully autonomous negotiation agents from [0:27] day one. Why? because they don't trust [0:29] us and they don't trust the technology [0:31] from day one. And by having this unit in [0:32] the loop approach, we have feeding our [0:34] agent with all the feedback and all the [0:35] learnings and then they suddenly trust [0:37] us for like 10k negotiations, 20k [0:39] negotiations, 100k negotiations. [0:41] >> When you think about what a durable [0:43] vertical AI company looks like, what are [0:45] the qualities that you look for? It's [0:47] really really hard to forecast your moat [0:50] going forward. If you look back at all [0:52] the best businesses at the early stages, [0:54] they were [0:58] Welcome back to the A16Z podcast. I'm [1:00] Elena Burgerer. And today I'm joined by [1:03] Sema Amble, a partner at A16Z, and Vlad [1:06] Kyle, a co-founder and CEO of LEO, which [1:10] builds AI agents for enterprise [1:12] procurement. Sema, you recently wrote a [1:15] piece called the incumbents are coming [1:17] and it asked a question facing almost [1:19] every AI application company. If an [1:22] established software vendor already has [1:24] the customer and the data and the model [1:27] can work across its tools, where does a [1:29] startup have an advantage? And um we [1:32] have Vlad here and Vlad can really help [1:35] us understand where a startup has an [1:37] advantage. And so I think a a good place [1:41] to start is the cases for and against [1:45] the incumbents. So if a company can [1:48] connect a capable AI agent to the [1:50] software it already uses, where does [1:53] another application actually uh have a [1:56] use case and an advantage there? [1:58] >> Yeah, let me back up and frame it up a [2:00] little bit. So historically we thought [2:01] about there's the incumbent and then [2:04] there's the startup and they're they're [2:06] fight they're you know it's a it's a [2:07] fight between distribution and [2:09] innovation which you know to to co to [2:11] take my partner Alex Rampel's phrase. Um [2:14] however now there's like this third [2:15] piece which you're you're you're [2:17] pointing at which is um the incumbent [2:21] can layer on um one of the models on top [2:24] or and and then there's a much more [2:26] formidable competitor in the market. And [2:29] so why do you need an AI native startup [2:32] if you've got cloud force which is [2:34] taking cloud plus Salesforce and putting [2:37] the two together and then you know [2:39] you've already got all your data and [2:40] your your customers uh sorry your [2:43] employees are all used to using the [2:44] product so why another product um I [2:48] absolutely still think there's obviously [2:49] still a case for uh the AI native [2:52] startup and it really centers around the [2:54] fact that um the legacy incumbent is [2:58] limited to their system of record and [3:00] that record that they have and they're [3:01] not completing the endto-end job. So let [3:04] me put that like more concretely in an [3:05] example. Um so say you're um a customer [3:09] and they customer calls and says they [3:12] got charged after they got cancelled. [3:14] that um resolving that cancellation [3:16] history isn't just the you know customer [3:19] going into the chat and saying hey um I [3:24] got uh I got overcharged and the [3:28] response there it has to hit billing it [3:30] has to look at all the chat history it [3:31] has to look at the contract that's not [3:33] one system of record that's the [3:35] knowledge around that customer and [3:38] everything it touched and that's [3:40] something that one system of record [3:41] wouldn't touch however the opport [3:44] opportunity for the AI native startup is [3:45] to say we're going to own that entire um [3:49] end to end arc that endto-end arc. So [3:51] that could be in legal. So owning [3:53] everything from uh brief all the way [3:56] through trial. Um well I can talk more [3:58] about procurement but it's it's really [4:00] this concept of owning the end to end [4:02] end to end work. [4:03] >> Yeah. Uh Vlad, do you want to talk about [4:05] just kind of where where that does show [4:08] up in procurement? Like where a existing [4:11] incumbent plus a model just is [4:14] insufficient and what have you seen just [4:16] with the companies that you work with? [4:19] >> Sure. So like when we think about [4:20] procurement, I like I would assume that [4:22] most people think about like prices, [4:24] right? So like what is the end price [4:25] that we negotiated on? And surprisingly [4:28] the the record looks always very very [4:32] simple and very easy. it's just 8K [4:34] that's the result for example and um I [4:37] mean it's the same for sales right so um [4:39] even if um if I come to Sema and tell [4:42] her like hey we we now partnered up with [4:44] another enterprise um look at the [4:46] signature here for like this contract it [4:48] looks like very easy but like Sema [4:50] doesn't see like all the work behind it [4:52] right so like there probably like 30 [4:54] stakeholder meetings happened 500 emails [4:57] 20 excel sheets and um that's the same [5:00] for the counterpart procurement right so [5:02] you see in ERP system, 8K for aluminum. [5:06] Um, but you don't see that maybe the um [5:10] the supplier did like a push back and [5:12] asked for like um 10K. Um you don't see [5:15] that like a cost engineer had run like 3 [5:18] weeks of Excel sheets and 3D modeling to [5:20] find out the prices of the um of the [5:23] part and everything else. Um so that's [5:25] what we see. like most of the work um in [5:28] procurement actually happens outside of [5:30] this um ERP or any system of record. [5:33] Yeah. [5:34] >> Yeah. Um, and I'm sure just in in the [5:37] workflow that you described, an agent [5:39] can do, you know, a huge number of [5:41] things and different kinds of agents can [5:43] do a large number of things, too. And I [5:45] actually I think that's a good bridge [5:47] into the next question, which is like, [5:49] you know, a year ago, I think a lot of [5:51] the incumbents were releasing chat bots [5:55] and that was kind of the extent of what [5:57] you would see. But Sema, in this piece [5:59] that you wrote, you you lay out four [6:01] different kinds of agents. retrieval [6:03] agents, which are kind of the the chat [6:05] bots that we're familiar with, process [6:07] agents, policy agents, and principal [6:09] agents. So, can you just walk us through [6:12] all of them and explain kind of what [6:15] changes in the kind of judgment that's [6:17] necessary across all of them? [6:18] >> Yeah. So, a year ago, um, I I I made [6:22] this meme, which was the slap on a [6:23] chatbot strategy, which is essentially [6:25] like all the incumbents effectively had [6:27] a a chatbot that sat on top of the [6:29] system record, which you could chat with [6:31] to retrieve information, maybe do some [6:33] light analytics. And that really um was [6:36] in that first bucket of what the [6:37] retrieval agent is. Um, and maybe let me [6:40] walk you through an example of what each [6:42] of like the retrieval, the process, the [6:44] policy, and the principle. Back to the [6:46] like customer um the the customer [6:49] support example just because it's very [6:50] it's easier to understand. So imagine if [6:53] you're a customer and there's a service [6:55] outage and you're calling in uh to say, [6:58] "Hey, I want to get compensated for this [7:00] service outage." the retrieval [7:02] assistant, which the incumbent may have, [7:04] is going to be able to pull up, yes, you [7:06] know, there this was what the contract [7:08] term said and um yes, there was an [7:11] outage and just verify that information. [7:13] It's pulling up information about the [7:15] customer. It's in the database and it's [7:17] just like sharing it back and and maybe [7:20] synthesizing. The second agent or the [7:22] second step in the agent um sequence is [7:25] the process agent. So that process agent [7:28] may be able to pull through an approval [7:30] of credit and say okay um based on the [7:34] our policy handbook um it was out from [7:37] these dates. Therefore we are say you [7:39] are entitled to um money and it can just [7:43] apply the bill and just do process. [7:45] There's no judgment involved. Then as [7:47] you keep going to the policy agent, so [7:50] the policy agent isn't just going to um [7:54] apply the process, but it's going to [7:55] say, "Okay, uh in this situation, it was [7:59] out for 20 minutes. That's enough to be [8:02] considered a significant outage." And [8:04] they're applying that judgment because [8:05] there isn't really like a strict [8:07] definition around it. And then the last [8:10] stage, which is the principal agent, [8:12] you're actually weighing um you know, [8:14] okay, should we offer more compensation [8:16] because that was a pretty terrible [8:18] outage and we want to reserve the [8:20] relationship and it's worth doing more [8:22] beyond even what a process or policies. [8:25] But the importance of these four things [8:26] is that most incumbents started out in [8:30] if at all in bucket one. Now, at least [8:32] they're marketing that they are moving [8:34] towards process and policy, meaning [8:36] they'll be able to apply more judgment. [8:38] And you know, if you look at what [8:40] they've they've launched, um these are [8:43] workflow agents that will help you get a [8:46] document signed or um input information [8:50] from a transcription or things like [8:51] that. They're very much still limited [8:55] to, I would say, retrieval and a little [8:57] bit of process. They've not gotten into [8:58] more judgment. And I can get into why [9:01] like there are all sorts of incentives [9:03] uh that are preventing them from that. [9:04] But the incumbents are trying and I [9:06] think what they're not able to do [9:07] they're like some of them are think okay [9:09] let me partner with open AI anthropic [9:12] one of the labs and try to build out um [9:16] take that model capability and [9:18] complement what they have um and [9:20] superpower it supercharge it I should [9:22] say. [9:22] >> Yeah. Well do you want to say why sort [9:25] of some of the incumbents are holding [9:26] back? [9:27] >> Yeah. Okay. So there I would say they're [9:29] holding back but they are held back. [9:31] Okay. [9:31] >> Um yeah I I'm sure they they want to be [9:34] full force but there's probably two [9:35] pieces. One they have this advantage of [9:38] distribution right they have the [9:39] customer trust which enables them to [9:42] then sell more products with the [9:44] customer. So take Salesforce when Agent [9:46] Force launched uh it was very easy for [9:49] customers to say yeah I'm going to sign [9:50] up for the Salesforce agent especially [9:52] if it was offered at almost no extra [9:55] cost. Um, and so they have this trust, [9:57] they have the distribution, it's often [9:59] pretty seamless to turn on the product. [10:02] The flip side is there's all these [10:04] internal incentive issues, right? Which [10:06] is if you start getting into more [10:08] complicated agents, it's um it's a it's [10:13] an internal conflict with an existing [10:15] product that's offering workflow versus [10:17] you're resolving the work. And those are [10:19] two products that are um like if you're [10:21] solving a customer support issue end to [10:23] end versus providing uh um a workflow [10:27] for a support agent, a human agent, [10:29] those are different buyers. How do you [10:32] um you know those two teams are in [10:34] conflict and then on top of that I think [10:36] from a sales perspective like what are [10:38] you selling to the customer? And then [10:40] oftentimes in the classic like incumbent [10:42] issue, right, is like there's two uh two [10:44] VPs, right? and they're different orgs [10:47] and they're selling different products [10:48] and like they will never be able to uh [10:51] figure out what the the right set of [10:53] incentives and the right person to sell [10:54] to. But anyway, so there's all these [10:55] sort of classic incumbent issues I think [10:57] that the incumbent also runs into. [10:59] >> Makes sense. Um Vlad, where across this [11:02] this kind of spectrum of retrieval [11:04] agent, process agent, policy agent, [11:07] principal agent, where where does LEO [11:09] sit? [11:10] >> Yeah. So like we span across like I [11:13] would say like all of those categories [11:14] and it really depends on the complexity [11:16] and the risk um our agents take right. [11:19] So there [11:20] >> sometimes we um can like we already have [11:23] use cases where we run fully [11:24] autonomously um sometimes you have the [11:27] human in the loop. It really depends on [11:28] like the budget approval how like how [11:30] how complex and how how risky this is. [11:32] Um but maybe like coming back to what [11:35] Sema said about trust for for for the [11:37] incumbents like you talked about like [11:39] internal trust. I think there's like [11:40] also an external trust thing, right? So [11:42] like how do you convince someone to go [11:45] through from hey just like an agent that [11:48] retrieves some information and maybe [11:49] runs processes to do to do like [11:51] something completely autonomously is [11:53] like obviously there's like a product [11:54] component to it but mainly is a people [11:57] component. So they need to trust you. Um [11:59] and like we as a startup scale up we [12:02] have to earn the trust. Incumbents [12:03] already have the trust but this also [12:04] means they like they can destroy the [12:07] trust if they ship something like too [12:09] early and the product maybe doesn't work [12:10] or it's bad or like works um decisions [12:14] um in in a bad way. So that's um that's [12:16] that's what we we can do. Um and then [12:19] interestingly or what was like really [12:21] surprising for us like those process [12:22] agents um actually became kind of like a [12:26] side quest um for us because it like [12:29] when we talk when we look at the invoice [12:31] process um or like invoice agents as a [12:33] process. So you we we retrieve some [12:35] informations um you match this across [12:37] like other documents and then you push [12:38] this back to um SAP or Oracle like a [12:41] very clear um clear process um and then [12:44] we figured out okay that's like 100% of [12:46] the software market right so that's [12:48] that's invoice software um that's how [12:51] you build it today but it's actually [12:53] only like 20% of the work of the job to [12:56] be done of the problem because 80% of [12:58] the problem is like um what if like what [13:02] does the invoice is fraudulent. What if [13:04] there's like a mismatch? Like what did [13:06] what is like if we don't take the happy [13:07] path? Um and so we like um very fast [13:12] shifted to to the next step of agents [13:14] like doing those doing those exception [13:16] um um exception handlings. Um and we [13:18] convinced the customers by um I think we [13:22] got like lucky being like always [13:24] slightly ahead of the curve. So we were [13:26] able to to pitch the next generation of [13:28] agents. So for example like when we [13:30] started out 3 years ago it was like just [13:33] retrieving a document like not [13:35] impressive at all today but like three [13:36] years ago this was in like crazy [13:39] impressive. Um so we pitched this to [13:42] customers we find out okay that's a real [13:44] problem that's a use real use case they [13:45] would pay for and then we were able to [13:47] ship this some weeks later. Um and in [13:50] the same way we are doing this now for [13:52] the next step for process agents and for [13:54] like fully autonomous agents and for [13:56] longrunning agents. um we are pitching [13:58] this to them, finding out it's a problem [13:59] and then we're able to ship this very [14:01] fast. [14:01] >> I think an interesting point on trust is [14:04] this internal versus external trust. The [14:06] other lens on that is yeah so you need [14:08] your customer to buy the procurement [14:10] software and trust to use it for their [14:12] internal processes. But one of the [14:14] really interesting things when we first [14:15] met Vlad was that they're doing they're [14:18] also doing the negotiation. So you have [14:21] to trust that the LEO agent is going to [14:24] then interface with a third party and uh [14:27] and there is that piece of trust. Um and [14:30] of course I think a lot of people feel [14:31] burned by the incumbents and like you [14:33] know they're pretty limited and haven't [14:35] been able to do what they've marketed [14:36] they've had in the past. So but putting [14:38] that aside like I don't know maybe Vlad [14:40] I'd love to hear a little bit how you [14:42] convinced the customers to to trust an [14:45] AI agent to now take on negotiations. [14:48] >> Yeah. And can when you do that, can you [14:50] also just like paint the picture of like [14:52] what's involved in procurement and who [14:54] who are your customers? Um and and what [14:58] kind of sort of legacy systems are they [15:00] used to? [15:02] >> Yeah. So like when we um when you think [15:05] about procurement, you like maybe just [15:06] think about like purchasing or like you [15:07] when you look at like B2C world like [15:09] just you buy something. Um but actually [15:11] it's like a very intense process process [15:15] which like runs the economy, right? And [15:17] it includes multiple stakeholders and a [15:20] lot of stakeholders and a lot of um um a [15:22] lot of departments um legal, cost [15:25] engineering, [15:27] um obviously procurement, finance um all [15:30] of them have to work have to work on [15:31] this um um on this decisions and [15:35] essentially like when we um [15:38] like I think I think the reason is like [15:39] how how we convince them is like what I [15:41] um already said said earlier on. So like [15:43] we we're like always a little bit ahead [15:47] of the curve. Okay. So we knew the [15:49] technology is coming. So we like even [15:51] before CHBT came out um like we we just [15:55] started a few weeks before like this [15:57] CHBT breakthrough. Um, so we already [15:59] heard all of the all of the problems. [16:01] Then we hear the the hype in the let's [16:04] say like tech bubble and we were able to [16:06] pitch this enterprises and we quickly [16:07] figured out okay like this is a this [16:09] could be like an interesting use case [16:10] like just chatbot applications or [16:12] retrieval agents um document processing [16:15] um and then we we are able to find out [16:18] the problem and then ship this quickly [16:19] to them. And then obviously like this is [16:21] like the people factor of like trusting. [16:23] So we we telling them something about [16:25] and then we're able to really ship [16:26] something in production. Um but then [16:28] there's also like this product [16:29] perspective where we have um a lot of [16:31] like evils in place, right? So like we [16:34] um the like no company and no enterprise [16:38] starts with fully autonomous negotiation [16:41] agents from day one. No one does that. [16:43] Um why? Because they don't trust us and [16:45] they don't trust the technology from day [16:47] one. Um so you have like a very easy [16:49] approach of like um having a like human [16:53] in the loop and this is like extremely [16:55] helpful for us because we see like we [16:58] have like let's say like the perfect [16:59] negotiation agent which is like overall [17:01] the perfect procurement negotiator but [17:03] we don't know exactly how a fortune 10 [17:07] enterprise like the specific fortune 10 [17:09] enterprise operates and by having like [17:11] this human in the loop approach we are [17:12] feeding our agent with all the feedback [17:14] and all the learnings um and then they [17:17] suddenly trust us for like 10k [17:19] negotiations, 20k negotiations, 100k [17:21] negotiations. Um, and then you also have [17:23] like other um other agents that are like [17:25] more longunning. So when we talk about [17:27] like multi-million dollar negotiations [17:28] where you analyze complex 3D models and [17:31] technical drawings um there we on [17:34] purpose have always experts in the loop, [17:36] right? So there's an agent running for [17:39] multiple hours and then we ask for [17:40] feedback of the cost like cost engineer [17:42] and then it does the next um the next [17:44] work um and so on. Maybe just to double [17:47] click on that, how do you what where do [17:49] you put the human in the loop on the [17:51] like negotiation side? Uh you mentioned [17:53] the cost engineer, but like if I were [17:55] you were going back and forth on a deal, [17:57] is it mostly around the like uh you know [18:00] data for you know something like cost [18:02] engineering or is there anything else [18:04] where you have humans in the loop there? [18:06] >> So again depends on the level of [18:08] negotiation right. So like we have to [18:10] distinguish between like um negotiations [18:13] where you just like a negotiation where [18:15] in u like where enterprises they never [18:18] did those because they didn't have the [18:19] capacity but by by by deploying agents [18:22] they can just um capture savings that [18:25] they were unaware of right so they they [18:27] like before agents or before Leo they [18:29] just didn't care about everything which [18:31] happened below 50k right so you can just [18:33] like this is maybe a hack for like other [18:36] startups you can just send um an [18:37] enterprise an invoice for 40k they will [18:39] probably not negotiate because they [18:41] don't have the capacity to do so. Um um [18:44] except they have leo agents then we are [18:45] going to negotiate um against you. Um [18:47] but other than that they are just like [18:49] just um just paying that. And obviously [18:51] there the risk of like like what is the [18:53] risk of like you don't you didn't ne [18:55] negotiate at all. So what is the risk [18:56] now of having a bad negotiation agent? [18:59] Nearly zero, right? So um maybe we miss [19:02] out on some negotiations but like it's [19:04] better than nothing. Um but still um in [19:08] like in we're talking about business [19:10] relationships and business relationships [19:11] are not always about the cost and the [19:14] money, right? So maybe you're not [19:16] spending a lot on the vendor. Um but [19:18] maybe you you're like you need this [19:21] business relationship, right? So good [19:23] example might be like podcast or [19:25] marketing services. Okay, that's like [19:27] probably like a a friction of a um of of [19:30] the spend, but you don't want to like [19:32] you have a clear business relationship [19:34] with with someone like setting up the [19:36] studio. Um and you don't want like like [19:39] a random some random people doing that [19:41] because they already know how how A6 [19:43] operates and how you want to record all [19:44] the stuff. Um so there we have human and [19:47] loop um approaches um where like [19:52] procurement people care about the [19:53] relationship. So they care about the [19:55] voice of tone and and how it works um [19:58] but mostly autonomously um and then we [20:00] have the other set of agents where we [20:01] always have um a human in the loop [20:04] approach and it's like um there like [20:06] multiple steps in like negotiating. It's [20:08] also like it's not only the price, [20:10] right? It's also like um how is the [20:12] contract design? So we talking about [20:14] legal, how is like um collection design? [20:18] So we we talk about finance obviously [20:20] like cost structure. So we're talking [20:21] about really like cost engineering. Then [20:23] we talk about commercials. Um that's [20:25] procurement. Um and those are not back [20:29] office people. Those are like highly [20:30] trained um people where they have very [20:34] specific knowledge of a very specific [20:36] process of a very specific company and [20:38] very specific industry. um and they feed [20:42] are those longrunning agents of LEO with [20:44] with those insights. [20:45] >> That's another example of how [20:47] procurement historically may have been [20:50] more in a box and now it's like okay [20:52] it's touching legal, it's touching [20:53] finance and it's touching a bunch of [20:55] different software systems and people [20:57] and both specialists and more [20:59] generalists. [21:00] >> Yeah. Um, can we can we map this on to a [21:03] specific uh c like not a specific [21:06] customer but a specific vertical like [21:08] I'm a drone manufacturer. I manufacture [21:11] humanoid robotics or something like [21:13] that. Like how many parts do I have to [21:17] you know order and procure? How many [21:19] factories am I touching? How many [21:21] suppliers am I touching? Just all of all [21:24] of those things. if you want to pick [21:25] maybe Vlad a vertical that um you know [21:30] is is just managing all of this [21:32] complexity with Leo and and just kind of [21:35] take us through what their experience [21:36] is. I think that would really help just [21:39] illustrate exactly um just everything [21:41] that you touch [21:42] >> again like when we like when we as Leo [21:44] when we talk about procurement of [21:46] purchases like we don't talk about like [21:49] laptops and pencils okay like we think [21:51] like that's solved also by Leo agents [21:53] but that's easy we solved this like [21:54] three years ago um we talk about like [21:56] when we like you want to build an [21:58] aircraft or robots or drones or even [22:00] like we now like doing a podcast about [22:03] um again like AI hype but even AI needs [22:05] to be built right so you data centers. [22:07] Um, and like building means procuring [22:11] like someone needs to like if you build [22:13] an aircraft, you need to procure [22:14] thousands of suppliers. You need to [22:16] build a factory to build this airplane. [22:19] Um, and like really small frictions can [22:24] have like a crazy impact, right? So like [22:26] there's a like if you're running a very [22:29] large project of like building a data [22:31] center, building an aircraft, if there's [22:32] like one specific part which arrives 2 [22:35] weeks later, this can have like a damage [22:38] of like hundreds of millions of dollars [22:41] um and postpone uh and postpone the [22:43] project. Um so like all of all of those [22:46] that's why like all of those decisions [22:48] have to be co coordinated. Um and one [22:50] part is like you you need to figure out [22:52] like what you need with what suppliers [22:54] you work. um what are like what is like [22:56] the best supplier to um to like to to [22:59] get this part. And then once you decided [23:01] all of the stuff, there's like all this [23:02] operational um back office um stuff [23:05] behind it which might sound like [23:07] unnecessary and boring but again like [23:09] operational means someone sends a [23:11] confirmation of like hey sorry like this [23:13] part is going to arrive 2 weeks later [23:15] and this is like one of 500 emails um in [23:19] the Outlook or Gmail um of a procurement [23:21] manager and if they missed this email [23:23] hundreds of millions of damage done and [23:26] this h like this happens regularly [23:28] because the only thing they store in [23:30] their um system of record is then just [23:33] the date, right? So like it will like [23:36] not like not this Wednesday, next [23:38] Wednesday. That's what you see in the [23:39] system, but you don't see like like [23:41] maybe that's okay, but maybe that's a [23:43] $100 million damage and someone has to [23:45] decide that. And that's that's also what [23:47] our agents are doing, right? They are [23:48] not only retrieving the information, [23:50] they are then making the decisions um or [23:53] like does this have an impact? What kind [23:55] of impact and how can we resolve this? [23:58] >> Yeah. um when when you are sort of so [24:02] deeply embedded in the physical world um [24:06] what what kinds of you know physical [24:08] world problems can you intervene with [24:11] like some things I would think are just [24:13] like unsolvable you know like let's say [24:16] you have a shipment coming in and a [24:18] bunch of stuff like falls off the ship [24:21] or the straight is closed or whatever [24:24] like they're there they're they're kind [24:25] they're things that like you can't do [24:28] and obvious Obviously there are things [24:29] that you can do. So, so where where can [24:32] you intervene and where does that really [24:34] make a difference? [24:35] >> But but that's like actually it's about [24:37] like probability, right? So obviously [24:39] like you can't like you can't change if [24:42] um if like some like if there's like a [24:45] damage on the ship like every every [24:47] example that you manage like you um you [24:49] can't um change that but you can if you [24:52] have like all of the context you can [24:54] predict that. [24:55] >> Mhm. um because you can predict like how [24:59] um how reliable is a supplier. Okay, so [25:02] they're like there are ways on like [25:04] protect the goods that you're shipping. [25:06] Um and if you have like all the context, [25:08] you have like one supplier where like [25:10] 20% of the goods are missing and then 1% [25:13] of the good is missing. Um and maybe [25:16] like this one with 20% is like 10x [25:18] cheaper. But for this use case, it's [25:20] fine for you to pay 10x the amount um [25:22] because you have like a higher [25:24] probability that this thing actually [25:25] arrives. Um so and this is the powerful [25:28] thing because you have like context not [25:30] only within one enterprise. Um so we [25:34] like we talked about like multiple [25:35] stakeholders but there's also context on [25:37] the outside world, right? So just the [25:40] agent should like have context of all [25:41] the news out there. um maybe even having [25:44] like context of like some bets like on [25:46] poly market I was like okay those [25:48] disruptions are going to happen um then [25:51] like um information about like on the [25:53] supplier side on the seller side on the [25:55] demand side and by combining all of [25:58] those contexts um I wouldn't say that [26:00] there is a limitation um in the long in [26:03] the long run obviously um like today we [26:07] have like different sets of like [26:08] probability but we can we can help [26:10] throughout the process and this is what [26:12] we are building um building at Leo right [26:14] so it's much bigger than just [26:16] procurement inter company it's more like [26:18] intra company and like how um how are [26:21] like businesses like how enterprises are [26:23] doing business with each other so like [26:24] buyer and supplier side [26:26] >> you describe LEO as a as a multi- aent [26:29] system so can you describe what the [26:34] different agents are doing [26:35] >> one level is that we um that Leo agents [26:38] spend across like all those four [26:40] categories that that Sema mentioned um [26:42] in her in her article and um it again [26:46] like depends on the um on the risk and [26:48] the complexity. Um so we use like all of [26:50] them. So like multiple agents um but the [26:53] other thing is that like in in order to [26:55] do a job end to end those agents need to [26:59] share information um with each other. [27:01] They need to do this in like a very [27:04] specific order. Um and when we talk [27:08] about like a multi- aent system this is [27:10] essentially what we are We are solving [27:12] like the task end to end and because the [27:15] also like human level task involves [27:17] eight people, eight stakeholders um and [27:20] maybe like three departments and five [27:22] different um software tools. Um we need [27:25] to cover like all of those to to do like [27:27] the job end to end and those agents need [27:29] to then communicate with each other and [27:31] only with a multi- aent system you can [27:32] do a job end to end. Um when we started [27:35] out we um like started off with like an [27:38] retrieval like more like a co-pilot [27:40] obviously like 3 years ago but then the [27:41] next step was like a single agent but [27:43] then we very quickly discovered okay [27:45] like that's like you can't solve an um [27:49] you can't solve in like negotiation even [27:52] um without having a contract um agent [27:55] without um maybe like having having an [27:57] agent looking at the news and everything [27:59] that I described before. Um so that's [28:01] what yeah that's what we define as a [28:02] multi- aent system. So if you have like [28:04] a bolt like an airline company needs to [28:06] procure a bolt for say you know Boeing [28:09] needs to procure a bolt what exactly is [28:12] that process for procuring the bolt and [28:15] like where does LEO step in on that [28:18] process? [28:19] >> Yeah so like this is like one of one of [28:22] the um one of the purchases where that [28:25] we can like run fully autonomously and [28:28] we can do this because of this multi- [28:29] aent system. So like first of all [28:31] someone has a demand right? So they need [28:33] to um somehow communicate it and even [28:36] like this part is extremely complicated. [28:39] Um so you need to call someone maybe you [28:41] like you like you open up your laptop [28:43] because you're like a construction [28:44] worker. You open up your laptop only [28:47] every second week. Um and now you are [28:49] required to like work with SAP um or any [28:52] other like um EIP system. So you can't [28:54] even like issue the demand. So this is [28:56] like how we make it very easy. So you [28:58] like you you you take a photo like you [29:00] you you upload a quote an Excel sheet um [29:02] and that's actually everything that you [29:05] that you should know about procurement [29:07] like no one cares outside of the [29:08] procurement department about like [29:09] categories GL accounts framework [29:11] contracts no one cares um we in the [29:14] procurement will care but no one outside [29:15] that cares and then our agents take off [29:17] and they're like they check the [29:18] inventory they find out okay they they [29:20] ask another plant okay can we like can [29:23] we source those um those bolts [29:25] internally no okay then I'm calling the [29:27] um am talking to the sourcing agent [29:29] finding out do we have internal [29:31] suppliers do we have external suppliers [29:33] then some like another agent has to [29:35] draft the RFQ send out the RFQ over um [29:38] over email then bunch of emails arrive [29:41] some of them are like completely [29:43] nonsense some of them are like just in [29:45] the email some of them are PDFs some of [29:46] them are Excel sheets we retrieve those [29:48] informations um then we do like the next [29:52] step oh maybe like based on our price [29:54] benchmarking there's an opportunity to [29:55] negotiate [29:56] Um and then then we have like agents [29:59] that essentially decide on the next [30:00] step. So negotiation could mean [30:03] strategic negotiation with a human in [30:04] the loop. This could mean autonomous [30:06] negotiation. This could mean auctions [30:08] and e auctions calling then the specific [30:10] agent doing the negotiation. Um and then [30:13] like doing this like end to end think [30:14] about like confirming the order, [30:16] shipment tracking, um invoices. Um and [30:20] we we are able to run this like fully [30:21] autonomously capture like all the [30:23] context. Um and then obviously the next [30:28] powerful thing is do this for like more [30:31] complex um parts where we talk about [30:34] like direct procurement um where we um [30:37] where we also operate. Yeah. [30:39] >> What what's uh direct procurement? [30:42] Yeah. So like um essentially like [30:44] everything I just described is um the [30:47] main goal is here automation, right? So [30:50] you can like run this process fully [30:51] autonomously and then throughout the [30:53] process you can like generate even more [30:54] savings, right? So it's not like so we [30:56] look at this like okay what is like this [30:59] like end to like job to be done? How [31:01] does it look like? So like what are they [31:03] doing like thousand times a day but [31:04] actually they want to do it like zero [31:06] times a day. We've like run like fully [31:08] autonomous agents, but there's like also [31:10] opportunities of like what are they [31:11] doing zero times a day, but if a [31:13] business would do this thousand times a [31:14] day, that was that would have a crazy [31:16] P&L impact. Autonomous negotiations on [31:20] um spend they never negotiated before. [31:22] Um so, so this is um and this is like in [31:26] the indirect procurement part like think [31:27] about um MRO parts, building a factory, [31:30] the bold example that we did. Um but [31:32] also laptops and pencils, marketing [31:34] services, someone who needs to build up [31:37] this podcast studio. Those are like all [31:39] indirect. And then we have like direct [31:40] parts. Um this is like when you build an [31:42] airplane. Those are like all the [31:44] suppliers that actually um that you [31:47] actually need to build the airplane um [31:50] or to build the drone or to build the [31:51] robot. Um and then we don't talk about [31:54] 50,000 suppliers. We talk about 100 [31:57] suppliers or 2,000 suppliers maximum. [32:00] Um, and those are like extremely [32:02] strategically important. Um, and you [32:04] have maybe on one supplier [32:07] like 1 billion of spend. So you don't [32:09] want to run an autonomous negotiation. [32:11] You want to run a negotiation which [32:12] which takes 3 months and where you're [32:14] like crazy prepared and where you have [32:16] engineers on your team analyzing, okay, [32:19] what's the indicy for aluminium? What's [32:21] the indicy for oil? Um, how did the [32:23] price change? Um, so you like really [32:26] take over like all of those drawings. [32:28] you check the quality um of this part [32:30] and this is what like where it gets like [32:33] really exciting deploying deploying [32:35] agents. [32:36] >> Yeah. And and for something like that [32:38] presumably like you'd have the expert [32:41] engineers and the other procurement [32:43] people kind of more as the front of [32:45] house and like the agent is more back of [32:47] house. Is that the idea or is the agent [32:49] like actually it's like you you sit [32:52] across the table and you're shaking [32:53] hands and it's like the robot instead of [32:55] the human who's like negotiating like um [32:58] is so yeah is it is it more back of [33:01] house or is it like still front of [33:02] house? It it's it's obviously more back [33:04] of house, right? Because you like [33:05] because you like [33:07] >> need like these complex multi-million [33:08] dollar negotiations and and that's [33:10] that's again like a beautiful example. [33:13] 90% of the work is preparation. [33:15] >> Like the end result that you see in your [33:16] system of record is like oh instead of [33:18] like 1 billion I paid $900 million. [33:21] Okay. But like there's like three months [33:23] of preparation and like 10 people [33:26] working full-time on that. Um and [33:28] obviously this is like happening in the [33:30] um in the back. Actually we have some [33:32] use cases where um it's also like h like [33:36] helping in real time. So think about um [33:39] let's assume we would now have a [33:40] negotiation and I have like a pro [33:42] perfect preparation the same as like [33:44] with like those um those those notes um [33:46] those notes that we are having here um [33:49] imagine like while we are negotiating um [33:52] I would have like real time insights on [33:54] my screen popping up um where you tell [33:56] me the indicy for oil changed 10%. So [34:00] like it's increased by 10%. So that's [34:02] why we need to increase the prices by [34:04] 10%. And I would have like an initial [34:06] like an an immediate popup with like [34:08] that's true like oil increased by 10% [34:11] but the product um has only 30% um of [34:15] oil contained. So like you you can you [34:17] shouldn't increase the price by 10% but [34:19] maybe only by 4%. Um so yeah there are [34:22] like exciting use cases also like in in [34:24] the real life um part. Yeah, we we know [34:27] that you know companies like Harvey and [34:30] Dakagon are really fine-tuning models [34:32] now. Um what what kind of underlying [34:36] models do you use and and how do you [34:38] approach things like fine-tuning or or [34:40] harnessing? [34:41] >> Yeah, so like we believe like you can so [34:44] like we use um multiple models from like [34:47] all um all providers and we really see [34:49] this as a like obviously like as a [34:50] commodity, right? So they like have um [34:54] really good like general business [34:56] purpose. So like reading creating a PDF [34:58] and like creating creating Excel sheet [35:00] like all all of this stuff. Um but we [35:02] also believe that like for some use [35:04] cases you um you you can get extremely [35:08] far with like combining the foundation [35:10] model with a with a hardness and you can [35:13] maybe reach like 100% of like the job to [35:15] be done. Um but there are also some use [35:18] cases where you can have like the best [35:21] foundation model, the best harness, um [35:24] whatever that means, but like um and the [35:26] best harness, but you still can get only [35:27] to 80%. Um and like I would say like [35:33] good examples for that is like for [35:34] example what um when we talk about [35:38] negotiations what like cost engineers [35:40] doing, right? So they're like analyzing [35:42] drawings and then they're like defining [35:44] okay what should this um this part [35:48] actually cost like and that's why it's [35:49] called like should cost modeling. Um and [35:52] there's um there is definitely like an [35:54] opportunity where we like thinking and [35:56] already started um um [36:00] fine-tuning the model to get then to um [36:03] to get them to to 100%. And this part [36:06] another example is like price [36:07] benchmarking um where think about like [36:11] the the like you would have a quote and [36:14] in a perfect world you would just drag [36:16] drag and drop the quote somewhere and [36:18] you would get the perfect price but it's [36:21] ex like and all those like all those [36:23] informations they're like not publicly [36:25] available right so there's all like [36:26] those are like all proprietary data [36:28] based on like one um one enterprise or [36:31] like across multiple enterprises so like [36:33] general purpose models can't train their [36:34] models on that. Um, so what we are like [36:39] what we are thinking about is like maybe [36:40] like not training just an LLM, but I [36:42] think like what we see now with models [36:44] like Jeff um or so popping up where you [36:47] have like an um you train it on like [36:51] text data but the out like the the [36:53] output is actually like an outcome um or [36:55] just like the perfect price. Um, and you [36:58] can't do this with harness because like [37:02] if you would give me like an um a quote [37:04] from BCG and a quote from McKenzie, they [37:06] they could do like the exact same work, [37:08] but this could be like a 10x different [37:10] price and I would have like no idea like [37:13] what is better. But if you give this to [37:15] a procurement manager, he would like [37:17] initially have a gut feeling or like [37:19] okay like this quote makes sense. Um or [37:22] like I think I think like good examples [37:24] again like content creation like always [37:26] like I don't know like how much I should [37:27] pay someone for creating a video. Um but [37:29] there's like a gut feeling behind it. If [37:31] I ask like another video creator of how [37:33] to um how to do that but if you ask them [37:35] to write down the rules they can't do [37:38] this because it's like just like gut [37:40] feeling and instinct. So, and that's [37:41] where we see a lot of opportunity [37:43] actually um training an agent but not [37:46] maybe like a classic LLM but more exact [37:48] like again like what what companies like [37:50] um or like models like like Jev are now [37:52] doing um on the outcome based so we have [37:54] like price benchmarking short cost [37:56] modeling. Yeah. [37:58] >> Um Sema uh we've talked a little bit [38:01] about how labs are really moving into [38:04] industry specific work or working with [38:06] incumbents to do this. um when you think [38:10] about what a durable vertical AI company [38:13] looks like, what are the qualities that [38:14] you look for? One is around, you know, [38:18] owning the endto-end work that we're [38:19] talking about, building up this data [38:21] asset and being able to do um something [38:25] that uh hasn't been done before. In many [38:27] cases, this is all said, I think when [38:30] you talk a lot about moes, it's really [38:33] really hard to forecast your moat going [38:36] forward. If you look back at all the [38:38] best businesses at the at the early [38:40] stages, they were they were just [38:42] thinking about, okay, I'm winning [38:44] customer trust. I'm selling more to them [38:46] and there's a lot of opportunity versus, [38:48] okay, I'm going to do these six steps [38:50] and then get to the seventh step and [38:51] then we'll have a moat. And so I think [38:54] we we talk a lot about defensibility and [38:55] durability. And I think part of that is [38:57] you're locking in the customer. They [39:00] there's more dependencies, they find it [39:02] valuable, and you're doing more of the [39:03] work. And here it's it's truly like [39:05] okay, you know, old CRM company was uh [39:08] log for all the deals. New sales AI [39:12] agent is actually owning a lot of the [39:14] sales prep process and the outbound [39:16] process fielding inbound and doing a [39:19] bunch of the work. um the comp the [39:22] overall customer is dependent on that [39:25] product and that's like a really [39:27] important signal of getting to the moat. [39:29] And everything we talk about in terms of [39:31] um in terms of stickiness and network [39:33] effects and all that is sort of [39:35] downstream of um of of that initial like [39:39] customer use and the value of the [39:41] product. [39:42] >> Yeah. Um Vlad, have have you had [39:45] conversations with customers or [39:47] potential customers who ask you, you [39:50] know, why should I buy your product? Why [39:52] can't I just, you know, plug into a [39:54] model and do this myself or like use [39:57] whatever existing system of record I [39:59] have plus a model? Um like like what [40:02] what do you tell them and and how do you [40:04] how do you convince them to to use Leo? [40:07] >> Yeah, 100%. And and that's a very fair [40:10] question, right? And like even if you [40:11] look like internally at LEO, so like the [40:15] the first use case 3 years ago which [40:17] like kind of like went viral in the [40:19] procurement world was like just like [40:21] having a quote and then getting this [40:23] information into SAP. So like very like [40:26] again like technically like but like [40:29] like tremendous business value. Um and [40:32] so you um you have like this retrieval [40:34] agent um getting like all of the [40:36] information putting it into SAP. We like [40:38] this was our like first product um and [40:41] we had an engineering team like [40:42] obviously small just like the three of [40:44] us or maybe like four people building it [40:46] um and then selling it but this is [40:48] nowadays a case study um if you if [40:51] you're applying to work at LEO so we [40:53] give this to people like to build this [40:55] and they have like 8 hours to do so. So [40:58] what I want to say by that is like a [41:00] product that we like one of our first [41:02] use cases um can now be somehow built by [41:05] engineers within 8 hours. Um so because [41:08] it's very easy to build stuff nowadays. [41:10] Um so obviously there's a question okay [41:12] well so someone can build this within [41:14] eight hours. Okay cool but then couldn't [41:16] like just procurement department [41:17] partners also just build everything in 2 [41:20] months. Um and the answer is like yes [41:23] you can build this in 8 hours and you [41:24] can build this but you will only reach [41:27] 70% let's say like the of the [41:30] performance. Um and the problem is 70% [41:34] of performance or accuracy or however [41:37] you measure it. It depends really on the [41:38] task doesn't mean 70% um automation, [41:42] right? So it's it's this can be mean [41:44] that you're like have 70% of the [41:46] performance but you still need to do [41:47] 100% 100% of the work um because 70% is [41:52] not not that much. So again like all of [41:54] the people have to um have to check the [41:57] data. Um, so maybe you even created like [42:00] more work um more work than um than you [42:04] did before. [42:05] >> Two things to to layer on to what Vlad [42:07] just said. One um overall it's good if [42:11] there's more adoption of the base models [42:14] or just like GPT products because it [42:17] means that people are also willing to [42:18] trust um vertical specific products as [42:22] well. So I think that increased uh [42:25] familiarity, comfort, excitement about [42:28] AI tools is generally just good for the [42:29] market. Um the second thing is I was [42:33] chatting with the management team of a [42:35] Fortune 500 company two or three weeks [42:36] ago and um one of the things they [42:39] mentioned was they had tried to build [42:40] out their own cash collection uh [42:43] product. This is a big enterprise [42:44] business and they like after I I don't [42:48] know three or four months of work at a [42:50] minimum they had found that there wasn't [42:53] enough context. There was poor quality [42:54] context. They had a lot of recordings, a [42:57] lot of screen grabs. Um they tried to [43:00] pull it all into one system, but they [43:02] there was no um there they wasn't good [43:05] enough. Uh and then there was this giant [43:06] question around okay like you know we've [43:09] got now two different ERPs and we were [43:12] about to acquire another one. Um who's [43:14] going to update all the mappings test [43:16] the you know test out okay does it work? [43:19] Um and then like we keep talking about [43:21] exception handling. You now need to map [43:23] that onto a totally different system a [43:25] different way of doing things. And I [43:27] think they quickly realized that the [43:28] internal build didn't make sense. And so [43:30] we keep hearing stories of this where [43:31] people are like okay I'm going to do the [43:32] internal build. and they're like, "Wait [43:33] a second, it's not different from what [43:34] DIY's ever been in the past, which [43:37] corporates have always tried." But, um, [43:40] I think enterprise companies generally [43:41] realize that like there there's their [43:43] core competency and then there's, um, [43:46] building internal tools and, um, they [43:49] should focus on the the first camp. [43:51] Yeah. Yeah. Makes sense. Um, [43:53] >> yeah. So, that's exactly what I what [43:54] I've meant with like obviously like they [43:56] can they can get to 80%, but those last [43:58] 20% really matter. um and you can only [44:01] get like they matter to get into [44:03] production. So that's why you need like [44:04] this harness, right? So you need all [44:06] those like integrations memory workflows [44:09] and sometimes you need vertical data um [44:12] um to do that and like like like the [44:14] parto princip so like those 20% can make [44:17] like 80% of the of the effort um or like [44:20] they are making 80% of the effort um so [44:25] like to all like those fortune 500 150 [44:28] companies so you can do this um but then [44:32] let's say like procurement and workforce [44:35] of AI procurement agents should then [44:37] become like one of your core [44:38] competencies. Um, and you should [44:41] evaluate whether this makes sense or not [44:42] for you um to to have this in as like [44:45] your core skill. [44:46] >> Um, Vlad, I'm curious if you're seeing [44:49] suppliers start to use agents or AI at [44:53] all and kind of like what happens when [44:56] both the buyers and the suppliers are [44:58] are fully AI enabled. [45:03] >> Yeah. So like we 100% believe that like [45:05] in the future there will be like agents [45:07] on both sides. Um and um this this makes [45:11] so much sense. Um but like surprisingly [45:13] what we what we see is like so like when [45:16] we look at the supplier side this also [45:17] kind of like equals the seller side [45:19] right. So what we see is like that the [45:21] the sales side was like always ahead of [45:23] the procurement side. Um but what we're [45:26] now seeing with those suppliers for this [45:27] Fortune 500 companies um this actually [45:30] is not true. So they are like maybe [45:32] advanced in like let's say um video [45:35] recordings and like using tools like [45:38] granola and all all on all the stuff but [45:40] like not like really um having agents [45:43] deployed they're automating the work um [45:46] and the cool thing is is like [45:47] procurement is like is unsexy right so [45:50] sales is sexy procurement is unsexy but [45:52] it's like one process and procurement is [45:54] the the counterpart but now the good [45:56] thing is in those industrial companies [45:59] for 5000 companies [46:01] Procurement has the bigger power to the [46:03] supplier because you as an like typical [46:07] use case like automotive supplier you [46:09] dictate to your suppliers what they [46:11] should use what the quality has to be um [46:15] how they have to how they have to answer [46:17] to an specific RFQ. Um so the [46:20] opportunity is now if we serve the [46:23] procurement department and then can and [46:24] they can dictate what a supplier should [46:27] use. um why are we not like not like why [46:31] aren't we like also pushing them to like [46:33] LEO agents um that are also helping them [46:36] automate the work um owning then both [46:38] sides of um of the transaction [46:41] >> and Elena I know we were talking earlier [46:43] today about you know how can you have [46:45] two parties on the same platform is how [46:48] does that work um I think it could even [46:50] extended to like legal work right and [46:53] and these are two very adversarial [46:55] parties right but like if you have two [46:57] law firms with clients with different [46:59] interests, but both benefit from [47:01] knowing, okay, here's the latest draft. [47:03] Here's where we are with the open [47:04] issues. Here are things that have been [47:06] agreed upon. And just even tracking that [47:08] that doesn't doesn't really exist right [47:10] now, right? That's all being human [47:11] that's being created by um humans. And [47:14] so being that coordination effort an [47:16] agent could be doing. [47:18] >> Yum. Yum. It's well, it's super cool to [47:20] think about how like, you know, both [47:23] sides are kind of maybe evolving in [47:26] tandem. one side might be going a little [47:28] bit faster as you're talking about Vlad. [47:30] Um, but over time potentially people are [47:33] just like on the same platform and [47:35] actually it's like it's way better [47:37] coordinated for for everyone [47:39] >> 100%. Because like like also like what [47:41] we mentioned in the beginning, right? So [47:42] like like the obvious like the obvious [47:45] question is like okay but like we we [47:47] also talked a lot about like negotiation [47:48] agents. What if like both parties have [47:50] negotiation agents and price is 100% a [47:53] point where it's like a zero something [47:56] or like they have like different [47:58] interests. Um but we what we also [48:00] discuss is that like price is like the [48:04] outcome of like 5,000 different like [48:06] other task that happened and on those [48:09] 5,000 other tasks they have the same [48:11] incentive. Um sales wants to have as [48:15] less like little friction as possible. [48:17] Um, buyers want to have a really fast [48:20] time to market, right? Again, like [48:21] coming back to building aircrafts, [48:23] building data centers. You want this [48:25] data center to be built as fast as [48:26] possible. You don't want to be build it [48:28] like 6 months later just because it [48:30] takes so much time um to analyze all of [48:33] the um responses from suppliers. Um, and [48:36] you want to make the sale also fast. So [48:38] like all of those five 5,000 other [48:40] tasks, the incentive is ex exactly the [48:42] same. And that's why you can deploy or [48:46] Leo can deploy agents also on both sides [48:49] um doing like automating the work um of [48:52] all those other tasks. Um this is the [48:54] beautiful thing. [48:55] >> I think that there used to be this logic [48:57] of like you shouldn't customize your [49:00] software too much to to one end user or [49:03] one end customer. But I think something [49:06] about LLMs and AI in general is like it [49:10] might increasingly be possible to [49:13] customize without slowing yourself down [49:16] as a business too much. Um, so I'm [49:19] curious if if that is something that [49:22] you're seeing Vlad or Sema um and and [49:25] kind of what what does that mean for for [49:27] end buyers of software? I think the [49:30] overall principle there's a lot of [49:33] forward deployed work happening right [49:34] now and part of that is because the [49:36] state of the customer data and [49:38] understanding [49:40] customer N um is a lot harder than [49:43] understanding N plus1 and so we're sort [49:44] of in the early stages of deployment [49:46] overall and that's why there's still a [49:47] lot of humans as part of this product um [49:50] and by the way that is something that is [49:51] harder for the incumbents to do because [49:54] they're also like not set up in in a way [49:57] to have even their the way their product [49:59] feedback cycle works where they have a [50:01] they have implementation teams but [50:03] that's very much an afterthought versus [50:05] something that feeds into the product [50:06] side. The beauty of AI is a it learns [50:08] over time. So that's when we're talking [50:09] about learning loops and you have the [50:12] right eval process, you can do more and [50:14] more complicated jobs over time and part [50:19] of that is automating the deployment [50:20] itself. And so you can and and I'd be [50:24] curious to hear how Vlad is doing it, [50:25] but a lot of our companies even are [50:28] doing that at you know at a rapid clip [50:30] where more of the customization a is [50:33] being handled in an automated way and b [50:36] the customer is able to turn the knobs [50:39] and levers around customization via [50:42] software um versus okay I needed to [50:45] bring in a the original it was like you [50:47] know you brought in Accenture to do your [50:48] SAP customization and now it's like [50:50] Okay, I've got a forward deployed team [50:52] that's going to build some, you know, [50:54] help build and spec it out. And [50:56] ultimately, it's going to be completely [50:57] software driven. [50:58] >> Yeah. Um I mean that's the reason why [51:00] like when you look at the or structure [51:02] of LEO, like 85% of the people are um [51:05] engineers. Um and even if you look at [51:08] like the people where you like they [51:10] don't have an engineering title, they [51:12] have like mostly like an engineering [51:14] background. And reason for that is um we [51:17] obviously like we don't want to be a [51:20] consulting company right so we make sure [51:21] that we have like overall the best um [51:24] agents in indirect and direct and [51:26] finance um and those part but then like [51:28] as you mentioned like there is a lot of [51:30] like forward deployed work um to do if [51:33] you go to enterprises because they have [51:34] like different nuances in their um in [51:36] their processes um but how we work is we [51:40] as you mentioned like building a product [51:42] in a way where it's like um where we [51:45] reducing this customization but also [51:46] where's a lot of like self-service. So [51:49] the job of our like FTDES and forward [51:52] deployed engineers um is on the one hand [51:56] making it like self-service but like in [51:58] internally is like automating their own [52:00] job right so like their KPI is like you [52:02] are seeing this happening like multiple [52:04] times like you like your literally your [52:05] job is like to automate yourself um and [52:08] then if you automate yourself you go to [52:09] the next task um I think it's like also [52:12] like an um approach that Google or so is [52:14] doing um so like but 100% agree with you [52:17] Um, and that's exactly what we are um [52:20] what we are building um and what how [52:22] we're doing it. [52:22] >> There was recently a a very big system [52:25] of record event and and we're not we're [52:28] not talking about Dreamforce, we're [52:29] talking about the the bots and buyers [52:32] summit that that Leo hosted in in New [52:35] York. Um and and just wanted to to kind [52:39] of hear, you know, stories from the [52:41] ground and and what what you're seeing [52:43] among buyers. What are people excited [52:46] about? what are people looking ahead [52:48] toward? What are people you know asking [52:51] you for? Um can you just kind of tell us [52:54] some stories from from that day and that [52:56] event? [52:56] >> So I mean it was the third time that we [52:58] are doing that we are doing this event. [53:00] Now we did it in um in New York just a [53:02] few blocks from our um from our office [53:04] here and over 100 procurement leaders [53:06] arrived right and what we like we made [53:08] sure that we like when we do when we do [53:10] such events that we only invite like [53:12] high caliber people right so like seale [53:15] co vice um vice president um and there [53:19] are like two things very like very [53:21] different um of of of how we do this or [53:23] like why why they're so amazed so the [53:25] first thing is um when we look at like [53:28] how procurement used to work in the last [53:30] 26 27 years um a lot of tools emerge [53:35] right so you can like they're like all [53:37] those technology landscapes and you can [53:38] find them on LinkedIn you will see like [53:39] there are like 500 procurement tools but [53:41] like if you like and this is also the [53:43] reason like but when you talk to [53:44] procurement people like very painful and [53:49] like I challenge someone like to find [53:52] someone who like loves to work with [53:53] procurement like no one does like you [53:55] can really state people hate working [53:56] with procurement and we're talking about [53:58] the requesters and we're talking about [53:59] the suppliers and even people in [54:01] procurement hate working on procurement. [54:03] Um so like what's going on if there are [54:06] like 1,000 tools? Um and the reason for [54:09] that is like the all of the tools. Um um [54:14] they just made the the process more [54:16] efficient. That's that's all. But they [54:19] never changed how though how those [54:21] people actually work. Um, and it's like [54:25] crazy to see that they work in emails [54:28] and in Microsoft Teams there's and in [54:29] Excel sheets and powerpoints. like their [54:31] main channel where they work on and it's [54:33] like zero like zero AI enabled um [54:37] obviously um um in in in this part and [54:40] the the other thing is like we um we [54:43] give them like a very cross department [54:45] perspective on procurement right so we [54:47] are not talking about like look at this [54:50] crazy invoice feature that we developed [54:53] but we more looking like like someone [54:55] needs something um and in the end you [54:57] have it on your table and this can be [54:59] like a cross indirect direct, logistics, [55:01] finance and you can see how we are doing [55:04] it here and we also putting it into like [55:05] more into like a physical world because [55:07] like AI agents that's very abstract um [55:10] so what we are doing is we building up [55:13] booths and we we even have this in our [55:15] offices um in um also in New York where [55:18] you can walk through the booths and [55:20] experience like um all of those agents [55:23] like really hands-on um and that's what [55:25] the what's what what the people love and [55:27] it's like um I mean the next event um [55:30] will be with around 700 people. Um so [55:33] you can imagine how how crazy this this [55:35] grows. [55:36] >> Procurement people gone wild. Yeah. [55:38] Yeah. It's going to be that one is in um [55:41] Munich. [55:42] >> Um yeah. So we doing them like in um in [55:46] Europe Munich and in New York um all the [55:49] time. Yeah. [55:50] >> Cool. If you're listening and in [55:52] procurement, you know where to go. Maybe [55:56] I guess one question uh one question for [55:58] me. What do you think it takes to get uh [56:02] to to get people excited about [56:04] procurement? Is it is it the agents? Is [56:06] it the people? Is it the time saved? [56:08] Like what or like something else like [56:10] you mentioned it like procurement is one [56:12] of these things that I I I remember you [56:14] know people aren't they don't like it [56:16] like it's like a universally kind of [56:18] disliked low NPS area. I remember [56:20] talking to a guy who was out of [56:21] procurement like I don't know seven or [56:22] eight years ago as I was looking at this [56:24] category and he was like h I hate [56:27] talking about this product like I I use [56:29] you know this legacy system of record. [56:31] I'm on Koopa and I don't I don't want to [56:33] buy anything else. I don't want to talk [56:34] about it. It's fine. It was like the [56:36] most disgruntled customer call I've ever [56:37] done out of like millions of them. Um [56:39] but I'm I'm curious uh yeah what it is [56:42] that you think you know really gets [56:44] people excited about this category. [56:47] >> Yeah. So I mean it's like it's like and [56:49] this is also why why why I like [56:51] procurement is like it's on one hand [56:54] like so like the reason like why we [56:55] started in procurement like it's not [56:57] essentially like what happened but like [56:59] how people react to it right so if you [57:00] talk to the to the people they're like [57:02] really frustrated so this means it's an [57:05] like highly emotional topic but it's [57:08] like like let's be honest like um B2B [57:10] Zars okay but it's like highly emotional [57:13] um so that's a good thing and then if [57:14] you combine this with um like something [57:17] which is boring and niche. Um this is [57:19] also an advantage because um again like [57:22] the it's um it's like also easy or easy [57:26] for us for to amaze those people, right? [57:28] Because like the the really last [57:29] revolution they have seen is like 20 [57:32] years ago um and then maybe a nicer user [57:35] interface 10 years ago um but nothing [57:38] else happened. Um, and so you have like [57:40] boring, highly emotional and then plus [57:44] crazy business impact, right? So like it [57:47] feels like it's unnecessary, but like I [57:49] I told you like some examples. So like [57:50] it has like obviously crazy P&L impact, [57:54] but it has impact on like the whole [57:57] economy, right? So we're like we are [57:59] talking about like how data centers are [58:01] built, how aircrafts are built, how cars [58:02] are built, um how drones are built. So [58:06] it's extremely important that um you [58:10] have a fixed procurement process um not [58:12] only to like make it happen and build [58:15] something um but then also when you talk [58:17] about like when we when you look at like [58:18] the competitive landscape so to get like [58:20] 1% margin increase um you need to make [58:24] 10% more revenue 10% more sales um so [58:28] like if you just manage to get like 1% [58:30] savings it's like equals like 10% of [58:33] sales um that you have to [58:35] um to get the same outcome in your P&L. [58:37] Um so like it's tremendously important [58:40] and like you combine all of those three [58:42] things and then you have like um a [58:44] trillion dollar business opportunity. [58:46] That's that's my opinion like for [58:48] procurement, but there are probably also [58:50] like other things that like emotional, [58:52] boring, and have a crazy business [58:53] impact. [58:54] >> Well, Vlad, thank you so much for [58:56] joining us. Uh this was a ton of fun.