Kavak's Playbook for Rebuilding a Company Around AI

以下是内容的中文翻译： Kavak人工智能负责人Ale Massa详细阐述了该公司如何彻底转型为一个“AI原生”组织，这一过程涉及将AI智能体深度嵌入到其运营的几乎每一个环节。在加入Kavak之前，Massa曾经营Opi Analytics，一家早期的机器学习公司，他现在认为这家公司“超前了时代十年”。Transformer模型和“ChatGPT时刻”的出现，为他雄心勃勃的愿景提供了实现所需的突破。 Kavak作为拉丁美洲的二手车交易平台，由于缺乏现有的基础设施，不得不垂直整合金融科技、物流和检测等服务。这种复杂性使其成为智能体驱动模型的理想选择。Massa的愿景是，将Kavak打造成一个仿佛置身于2035年、拥有“GPT-10级别智能”的公司。 Kavak战略的核心是“一客一智能体”架构。每天，公司会为特定客户实例化10万到20万个智能体，每个智能体都拥有自己的虚拟机。这些智能体能够记住多年的互动，制定长期策略以最大化客户生命周期价值，并致力于在Kavak的所有产品中保持客户满意度。目前，96%的客户交互和95%的交易都由智能体处理。 Massa强调了此次转型的三个关键决策： 1. **重塑整个公司：** Kavak没有简单地向员工提供AI工具，而是重建了其API接口和系统，以赋能智能体执行任务。 2. **构建“超人智能体”：** 这些智能体旨在关键指标（如转化率、生命周期价值和客户体验）上超越Kavak最优秀的人类员工，解决公司最具挑战性的问题。 3. **转变衡量成功的指标：** Kavak从衡量交易结果（汽车买卖）转向衡量关系型指标，专注于最大化其1000万客户的生命周期价值。 强大的评估系统是成功的核心。Massa表示，他们在构建评估系统上投入的时间、工程精力和计算资源与构建智能体本身一样多。这使他们能够充满信心地“极速前进”。评估系统衡量的是实际业务成果，如客户转化率和满意度，而非表面化的KPI。 这些智能体的影响是深远的。它们是“更优秀的销售员”，由于其无限的耐心、完美的记忆和即时学习能力，客户转化率比人类团队高出2.1倍。Kavak的智能体还处理复杂的金融服务，例如在不到三分钟内批准汽车贷款（而人工处理可能需要数月），并提供个性化利率。Massa甚至分享了一个案例：一个“AI首席执行官”管理着Kavak在库埃纳瓦卡市的业务，通过微观管理日常任务和协调实体工作人员，在六周内实现了1.5倍的利润增长。即使在物理世界中，智能体也充当技师的“搭档”（“El Maik”），提高了检测质量和速度，并将保修索赔减少了20-26%。 组织结构已被彻底重新定义。Kavak推出了一个“绝地学院”项目，旨在对从首席执行官到技师的每一位员工进行AI培训。该项目赋能员工构建智能体并与之协作。公司现在采用扁平化、高资深、高度赋能的团队结构，其中智能体可以充当“老板”，而人类则负责设计智能体。至关重要的是，Kavak避免了常见的“人工干预”陷阱——即智能体仅仅将问题上报给人工支持。相反，智能体通过调用专门的API接口寻求人工帮助，而这些人工干预则被用作数据，以持续训练和改进智能体，从而形成一个闭环反馈。 Massa对其他领导者和创始人的建议非常明确： * **拥抱自上而下的转型：** 领导层清晰、垂直的愿景至关重要，因为表面化地采用AI不会带来显著成果。 * **衡量AI的投资回报率：** 专注于那些能够直接衡量投资回报的“高价值AI应用/计算资源”（Tier 3 tokens），而不是无法量化的普遍使用。 * **拥抱“创造性破坏”：** 就像之前的工业革命（例如电力彻底改变了工厂设计）一样，AI将催生新的AI原生公司，颠覆那些不愿从根本上重塑其运营的现有企业。 * **着眼未来：** Massa认为，对于创始人而言，这是“人类历史上最激动人心的时刻”，他们可以前所未有地获得强大工具。他鼓励他们构想一个AI驱动的未来，并为那个现实构建解决方案。

Ale Massa, Head of AI at Kavak, detailed the company's radical transformation into an "AI-native" organization, a process that involved deeply embedding AI agents into nearly every aspect of its operations. Before Kavak, Massa ran Opi Analytics, an early machine learning company that he now views as being "10 years ahead of its time." The advent of Transformers and the "ChatGPT moment" provided the breakthrough needed to fully realize his ambitious vision. Kavak, a used car marketplace in Latin America, had to vertically integrate services like fintech, logistics, and inspections due to a lack of existing infrastructure. This complexity made it an ideal candidate for an agent-driven model. Massa's vision was to rebuild Kavak as if it were 2035 with "GPT-10 level intelligence." The core of Kavak's strategy is the "agent-per-customer" architecture. Every day, between 100,000 and 200,000 agents are instantiated for specific customers, each with its own virtual machine. These agents remember years of interaction, formulate long-term strategies to maximize customer lifetime value, and aim to keep customers happy across all Kavak's products. Currently, 96% of all customer interactions and 95% of transactions are handled by agents. Massa highlighted three key decisions for this transformation: 1. **Redesigning the entire company:** Instead of simply giving AI tools to staff, Kavak rebuilt its APIs and systems to enable agents to perform. 2. **Building "superhuman agents":** These agents were designed to outperform Kavak's best human employees across critical metrics like conversion, lifetime value, and customer experience, tackling the company's most challenging problems. 3. **Shifting success metrics:** Kavak moved from measuring transactional outcomes (cars bought/sold) to relational ones, focusing on maximizing the lifetime value of its 10 million customers. Central to this success are robust evaluation (eval) systems. Massa stated they spend as much time, engineering effort, and tokens on building evals as they do on building agents. This allows them to "move extremely fast" with confidence. Evals measure actual business results, such as customer conversion and satisfaction, rather than superficial KPIs. The impact of these agents is profound. They are "better sellers," converting customers 2.1 times more effectively than human teams, due to their infinite patience, perfect recall, and instant learning. Kavak's agents also handle complex financial services, approving car loans in under three minutes (compared to months) and offering personalized interest rates. Massa even shared an example of an "AI CEO" managing Kavak's Cuernavaca city operations, achieving 1.5 times profit growth in six weeks by micromanaging daily tasks and coordinating physical workers. Even in the physical world, agents act as "sidekicks" ("El Maik") for mechanics, improving inspection quality, speed, and reducing warranty claims by 20-26%. The organizational structure has been completely redefined. Kavak launched a "Jedi Academy" to train every employee, from the CEO to mechanics, in AI. This program empowers them to build and collaborate with agents. The company now features flat, highly senior, empowered teams, where agents can be "bosses" and humans design agents. Critically, Kavak avoids the common "human in the loop" pitfall of agents simply escalating issues to human support. Instead, agents call specific APIs for human help, and these human interventions are used as data to continuously train and improve the agents, ensuring a closed feedback loop. Massa’s advice for other leaders and founders is clear: * **Embrace top-down transformation:** A clear, vertical vision from leadership is crucial, as superficial adoption of AI won't yield significant results. * **Measure the ROI of AI:** Focus on "Tier 3 tokens" where the return on investment can be directly measured, rather than unquantifiable general usage. * **Embrace "creative destruction":** Like previous industrial revolutions (e.g., electricity transforming factory design), AI will lead to new, AI-native companies disrupting incumbents unwilling to fundamentally redesign their operations. * **Build for the future:** Massa believes this is the "most exciting time in human history" for founders, with unprecedented access to powerful tools. He encourages them to envision an AI-driven future and build solutions for that reality.