Why Building an AI Agent Is Easier Than Deploying One

A16Z播客邀请了A16Z合伙人西玛·安贝尔（Seema Ambel）和Leo联合创始人兼首席执行官弗拉德·凯尔（Vlad Kyle），深入探讨了AI代理在企业采购中的变革潜力。讨论集中于AI原生初创公司在这种演变格局中相对于老牌企业的独特优势。 西玛·安贝尔将AI原生初创公司面临的挑战置于“老牌企业正在到来”的背景下。尽管老牌软件供应商可以在其现有系统上叠加AI模型，但他们的优势在于分销和客户信任，这通常受限于他们的“记录系统”。安贝尔认为，AI原生初创公司的机会在于掌握流程的整个“端到端弧线”。她通过一个客户服务示例说明了这一点：解决取消请求不仅仅是一个聊天机器人查询；它需要访问计费、聊天历史和合同——这些数据分散在多个系统中。这种从法律领域的“从简报到审判”或整个采购流程的综合方法，正是初创公司可以脱颖而出的地方。 安贝尔进一步将AI代理分为四种类型：检索型（用于获取信息的基本聊天机器人）、流程型（不带判断地应用规则，如批准信贷）、策略型（根据不那么严格的定义应用判断）和原则型（权衡更广泛的结果和关系）。她指出，由于内部冲突，如竞争的产品团队和销售激励，老牌企业主要在检索型和流程型领域运作，这阻碍了他们采用更具判断力和自主性的代理。 弗拉德·凯尔详细阐述了Leo在采购中这些代理类型的应用。他强调，传统采购远不止议价那么简单，它涉及众多利益相关者、会议、电子邮件和电子表格，这些都超出了主要的ERP系统。Leo的代理涵盖所有四种类型，其自主程度取决于任务的复杂性和风险。建立信任至关重要，尤其因为企业不会立即部署完全自主的谈判代理。Leo通过“人机协作”（human-in-the-loop）方法实现这一点，在扩大规模之前，通过较小、风险较低的谈判（例如，1万美元、2万美元、10万美元）向代理提供反馈。对于涉及复杂3D模型、价值数百万美元的谈判，人类专家仍需参与其中，提供见解。 凯尔生动地描绘了采购的实际运作，以建造飞机或数据中心为例。这涉及数千家供应商，单个部件的延迟可能导致数亿美元的损失。Leo的多代理系统介入，例如，通过捕获关于部件延迟的被遗漏的电子邮件确认，评估其影响，并推荐解决方案。虽然AI无法阻止诸如货物损坏等现实世界问题，但它可以通过结合内部背景信息和市场新闻甚至投机市场等外部信息来预测概率（供应商可靠性）并优化选择。 一个简单的螺栓的典型采购流程，由Leo的多代理系统完全自主处理，包括：绕过复杂的ERP系统进行简单的需求沟通（照片、报价）；代理检查库存和采购（内部/外部）；起草和发送询价单（RFQs）；解析各种电子邮件回复；基准测试价格；以及自主谈判、确认订单、跟踪发货和处理发票。 凯尔区分了间接采购和直接采购。间接采购（笔记本电脑、办公用品、服务）通常涉及因能力不足而从未谈判过的支出，这些是“每天零次”（zero times a day）的任务，如果自动化，将带来显著的节省。直接采购（飞机、机器人的战略部件）涉及高价值、长期谈判，AI代理主要支持“后台”（back-of-house）准备工作（分析图纸、市场价格），但也能在复杂的讨论中为人类谈判人员提供实时洞察。 关于AI原生初创公司如Leo为何能在老牌企业或DIY解决方案失败的地方取得成功，凯尔解释说，虽然一个基本的检索代理可以在八小时内构建完成，但它只能达到约70%的性能。这种不完全的自动化通常意味着人类仍然需要完成100%的工作，甚至可能增加工作量。生产就绪所需关键的“最后20%”，需要深度集成、记忆、工作流和垂直数据，这正是专业AI公司展现其价值的地方。 展望未来，凯尔认为代理将在买方和供应商双方运作。他指出，虽然销售在技术采用方面历来领先于采购，但工业部门的采购现在有能力向供应商规定技术使用，这为Leo这样的平台提供了自动化交易双方的机会。他强调，除了价格（可能是一场零和博弈）之外，交易中的其他5000项任务都有共同的激励（减少摩擦、加快上市时间），这使得多代理协作对所有相关方都有益。 Leo近期主办的“机器人与买家”（Bots and Buyers）峰会凸显了人们对传统采购工具的不满，这些工具仅仅提高了流程效率，但并未从根本上改变人们的工作方式（仍严重依赖电子邮件、Excel和Teams）。Leo对跨部门、端到端代理能力的现场演示让与会者感到兴奋。凯尔最后断言，采购尽管名声“不性感”，却是一个“万亿美元的商业机会”，因为它是一个高度情感化、小众且乏味的领域，对全球经济具有巨大且常被忽视的商业影响。

The A16Z Podcast, featuring Seema Ambel, a partner at A16Z, and Vlad Kyle, co-founder and CEO of Leo, delved into the transformative potential of AI agents in enterprise procurement. The discussion centered on the unique advantages AI-native startups possess over established incumbents in this evolving landscape. Seema Ambel framed the challenge for AI-native startups against the "incumbents are coming" backdrop. While established software vendors can layer AI models onto their existing systems, their strength lies in distribution and customer trust, often limited by their "system of record." Ambel argued that the AI-native startup's opportunity is to own the entire "end-to-end arc" of a process. She illustrated this with a customer service example: resolving a cancellation isn't just a chatbot query; it requires accessing billing, chat history, and contracts—data spread across multiple systems. This comprehensive approach, from "brief all the way through trial" in legal or entire procurement processes, is where startups can excel. Ambel further categorized AI agents into four types: Retrieval (basic chatbots for information), Process (applying rules without judgment, like approving credit), Policy (applying judgment based on less strict definitions), and Principle (weighing broader outcomes and relationships). She noted that incumbents largely operate in the Retrieval and Process realms due to internal conflicts, such as competing product teams and sales incentives, which prevent them from adopting more judgmental and autonomous agents. Vlad Kyle elaborated on Leo's application of these agent types in procurement. He highlighted that traditional procurement is far more complex than just negotiating prices, involving numerous stakeholders, meetings, emails, and spreadsheets that fall outside the main ERP system. Leo's agents span all four categories, with the level of autonomy depending on the task's complexity and risk. Building trust is paramount, especially since enterprises don't immediately deploy fully autonomous negotiation agents. Leo achieves this through a "human-in-the-loop" approach, feeding agents with feedback from smaller, lower-risk negotiations (e.g., 10K, then 20K, 100K) before scaling. For multi-million-dollar negotiations involving complex 3D models, human experts remain in the loop, providing insights. Kyle painted a vivid picture of procurement in action, using the example of building an aircraft or data center. Thousands of suppliers are involved, and a single part delay can cause hundreds of millions in damages. Leo's multi-agent system intervenes by, for instance, capturing missed email confirmations about delayed parts, assessing their impact, and recommending resolutions. While AI cannot prevent physical world problems like a damaged shipment, it can predict probabilities (supplier reliability) and optimize choices by combining internal context with external information like market news or even speculative markets. A typical procurement process for a simple bolt, fully autonomously handled by Leo's multi-agent system, involves: easy demand communication (photo, quote) bypassing complex ERPs; agents checking inventory and sourcing (internal/external); drafting and sending RFQs; parsing diverse email responses; benchmarking prices; and autonomously negotiating, confirming orders, tracking shipments, and handling invoices. Kyle differentiated between indirect and direct procurement. Indirect procurement (laptops, office supplies, services) often involves spend that was never negotiated due to lack of capacity, representing "zero times a day" tasks that, if automated, yield significant savings. Direct procurement (strategic parts for aircrafts, robots) involves high-value, long-term negotiations where AI agents primarily support "back-of-house" preparation (analyzing drawings, market prices) but can also offer real-time insights to human negotiators during complex discussions. Addressing why AI-native startups like Leo succeed where incumbents or DIY solutions falter, Kyle explained that while a basic retrieval agent can be built in eight hours, it only achieves about 70% performance. This incomplete automation often means that humans still need to perform 100% of the work, potentially increasing effort. The critical "last 20%" for production readiness, requiring deep integration, memory, workflows, and vertical data, is where specialized AI companies demonstrate their value. Looking to the future, Kyle believes agents will operate on both the buyer and supplier sides. He noted that while sales have historically been ahead of procurement in technology adoption, procurement in industrial sectors now has the leverage to dictate technology usage to suppliers, creating an opportunity for platforms like Leo to automate both sides of a transaction. He emphasized that beyond price, which can be a zero-sum game, 5,000 other tasks in a transaction share common incentives (reducing friction, faster time-to-market), making multi-agent collaboration beneficial for all parties. The recent "Bots and Buyers" summit hosted by Leo highlighted the frustration with traditional procurement tools that merely made processes efficient without fundamentally changing how people work (still heavily reliant on emails, Excel, Teams). Leo's hands-on demonstrations of cross-departmental, end-to-end agent capabilities excited attendees. Kyle concluded by asserting that procurement, despite its "unsexy" reputation, is a "trillion-dollar business opportunity" because it's a highly emotional, niche, and boring field with immense, often overlooked, business impact on the global economy.