Kavak Rebuilds Operations Around AI Agents, Boosting Sales Conversions 2.1x

Kavak围绕AI智能体重构业务运营，销售转化率提升2.1倍 企业必须围绕AI智能体彻底重塑组织架构，而非仅将其作为简单的生产力工具，方能获得根本性的效率提升。 Kavak已从使用基础预测性AI工具转变为部署“每位客户专属智能体”架构，由自主运行的AI实例负责管理大额交易、即时贷款审批以及客户生命周期价值。通过重构公司API、将一半工程力量投入评估框架，并开展诸如使城市级利润提升50%的AI“CEO”等大胆运营测试，该公司在扁平化管理层级的同时超越了人工销售团队的业绩。Kavak的经验凸显了为何真正的AI转型需要打破传统工作流程，自上而下彻底重塑业务系统。 Kavak围绕AI智能体彻底重构了组织架构、内部API和系统，而非仅仅将LLM用作简单的生产力工具。 (0:00) 该公司的核心架构在独立的虚拟机中为每位客户实例化专属智能体，每日部署10万至20万个智能体。 (2:23) Kavak全公司96%的客户互动和95%的交易均由AI智能体完全自主处理。 (4:59) Kavak将约50%的全部工程带宽与算力预算直接投入到评估框架（evals）中。 (8:04) 专业销售智能体实现的转化率是顶尖人工销售团队的2.1倍，同时将整体NPS提升至三倍。 (10:49)

Kavak Chief Product & AI Officer Alejandro Maza Ayala details how the Latin American used-car marketplace restructured its core operations around autonomous AI agents, moving beyond incremental productivity tools to run end-to-end transactions across sales, lending, and physical servicing. Architectural Shift to Dedicated Customer Agents (0:00) Rather than deploying generic conversational interfaces, Kavak re-engineered its internal APIs and infrastructure to allow software agents to execute actions directly. The company transitioned away from rigid, multi-agent graphs—destroying two years of prior development after the arrival of advanced models like Claude 3.5 Opus—to establish an architecture where every incoming customer triggers a dedicated agent in an isolated virtual machine. Kavak now instantiates between 100,000 and 200,000 agents daily, executing tasks that range from minutes to several days before entering a sleep state. These individual agents maintain persistent memory across years of cross-channel interactions, accessing company APIs to optimize long-term customer lifetime value (LTV). According to Maza Ayala, 96% of all customer interactions and 95% of transactions are executed autonomously without human intervention, leaving humans to handle physical touchpoints such as key handovers. Activating just 1% of Kavak's 10-million customer registry through agentic relationship management drives hundreds of millions in revenue due to large vehicle transaction sizes. Automating Complex Sales and Sub-Three-Minute Lending (10:49) Kavak deliberately bypassed standard customer service bots to build sales agents capable of navigating vehicle trade-ins, financing options, insurance, and an inventory of 20,000 SKUs. By consolidating capabilities that previously required up to 15 human domain specialists into single agents that share daily operational learning loops, Kavak reports that its sales agents achieve 2.1 times the conversion rate of its top human representatives while tripling Net Promoter Scores. The company also automated its regulated financing operations. While regional auto loan approvals in emerging markets like Mexico historically average two to three months of manual review, Kavak's integrated credit models evaluate thin-file borrowers and underwrite loans in under three minutes. Because Kavak controls both the lending arm and vehicle assets, agents dynamically adjust loan terms, maximum limits, and risk pricing, even facilitating vehicle swaps for lower-cost models if a borrower encounters payment difficulty. Algorithmic Management in Physical Operations (7:19) To evaluate autonomous executive decision-making, Kavak deployed an AI agent to serve as the operational manager for its location in Cuernavaca, Mexico. Operating with granular oversight, the agent micromanaged local staff through daily schedules, performance tracking, and mandatory voice-note check-ins. Although it missed its target of doubling operating profits within a month, the system increased city-level profits by 50% (1.5x) over a six-week trial. In physical repair facilities employing around 800 mechanics in Mexico, Kavak integrated an AI assistant named 'El Mike' to assist with vehicle inspections and repair guidance. Rather than attempting to automate manual dexterity, the system provides real-time procedural instructions during inspections, which Kavak credits with reducing post-service warranty claims by roughly 26%. Evaluation Infrastructure and Internal Workforce Retraining (8:04) To sustain high deployment velocity without risking system failures, Kavak allocates approximately 50% of its total engineering bandwidth, compute budget, and token volume to evaluation harnesses ('evals'). Rejecting vanity metrics like call duration, the company tests agent iterations strictly on downstream business outcomes, such as completed conversions and voluntary repeat customer engagement. Operationally, Kavak flattened its middle management into senior teams that work alongside agents. The company instituted the 'Jedi Academy,' an internal six-week curriculum requiring staff from software engineers and mechanics up to the CEO to learn how to build and maintain production-ready agents. When agents encounter edge cases, they call human support APIs directly, ensuring that unresolved customer scenarios systematically feed back into the model's training data. Maza Ayala argues that enterprise AI adoption requires top-down organizational redesign rather than superficial tools or bottom-up hackathons. He advises founders to build systems designed around continuous linear improvements in underlying intelligence models, using deep structural reinvention to capture value as frontier capabilities expand.

Kavak 首席产品与 AI 官 Alejandro Maza Ayala 指出，企业若要真正释放人工智能的潜能，必须放弃将大语言模型作为边际辅助工具的表面做法，而是彻底推翻既有架构，转向以客户为中心、由自治智能体端到端驱动业务与运营决策的原生组织。 底层架构重构与单客户专属智能体 (28:34) Kavak 的根本性转变在于彻底重写底层系统接口与数据反馈回路，而非仅仅为员工提供 ChatGPT 或 Claude 等聊天界面。Alejandro 指出，在不改变底层系统的前提下引入外部模型无法带来任何运营效率提升。当 Claude 3.5 Opus 等新一代模型出现后，Kavak 果断废弃了此前投入两年、基于任务图谱的多智能体系统，重构为“每位进入平台的客户均由运行在独立虚拟机内的专属智能体提供服务”的全新架构。目前系统每天动态实例化 10 万至 20 万个智能体，单次任务运行时长从数分钟跨越至数天，并具备调用公司全部 API 的命令行接口与持久跨渠道记忆。 Alejandro 将这种转型比作 20 世纪初的工厂电气化革新：若仅把多层厂房内的蒸汽机替换为电动机而不重新设计流水线布局，生产力便无法实现实质飞跃。通过践行熊彼特式的“创造性破坏”，Kavak 将企业核心指标从传统的单笔交易量转向全生命周期价值（LTV）管理，依靠算法记忆动态推行长期维系策略，仅激活其千万级用户数据库的 1%，便能带来数亿美元的营收增量。 销售与金融全流程的超人化业务交付 (10:49) 在客单价高且决策链路复杂的二手车交易场景中，Kavak 绕开基础问答客服，直接打造具备全流程促成能力的专业销售智能体。传统模式下购车通常需要跨部门协调车辆选品、估值置换、信贷审批和车险等 15 位领域专家，而整合了 20,000 个 SKU 专业知识的“超级专家”智能体不仅具备无限耐心与全局上下文，还能在出现单点失误后于次日将教训同步至全量实例。实际运营数据显示，该销售智能体的成交转化率目前已达优秀人类销售团队的 2.1 倍（最初上线时领先 50%），同时使客户净推荐值（NPS）与整体满意度增至原先的 3 倍。 在受强监管的汽车金融与信贷风控领域，Kavak 依托自营二手车资产与垂直整合生态，将拉美传统新兴市场长达 2 至 3 个月的车贷审批周期大幅缩短至 3 分钟以内。智能体不仅能对缺乏传统征信记录的客户进行动态风险定价，还能在借款人还款困难时主动介入，通过置换低价车型降低月供等方式调整信贷结构。目前 Kavak 全平台已有 96% 的客户互动和 95% 的交易由 AI 智能体自主完成，人工团队主要负责钥匙交接等物理环节。 物理运营渗透与“AI CEO”管理试验 (7:19) 为了探索智能体在行政管理与复杂实体运营中的能力边界，Kavak 在墨西哥库埃纳瓦卡市开展了将 AI 作为区域运营“CEO”的受控沙盒试验。该智能体被赋予首月将城市营业利润翻倍的目标，依托高精度的运营数据分析建立预测模型，并通过下发每日任务规划与调取语音备忘录等方式微观管理现场一线员工。在部署后的 6 周内，尽管未达到 2 倍的激进目标，该 AI CEO 仍使该城市的净利润实现了 50% 的显著增长（达到原利润的 1.5 倍），并在车辆库存周转与金融渗透率方面展现出严密的管理改善。 在传统软件难以直接自动化的物理维修车间，Kavak 针对墨西哥约 800 名机修工推出了代号为“El Mike”的实时引导系统（灵感源于电影《料理鼠王》）。该系统作为智能助手在车辆检测与翻新过程中向工人提供即时技术指导与操作核对，促使售后保修索赔率下降了约 26%，证明了通过算法深度赋能一线人类技师比纯粹寻求物理替代更能快速提高整体交付质量。 评测基准先行与“绝地学院”组织扁平化 (8:04) 在模型落地与工程实践中，Kavak 恪守“唯有制动可靠才能放心踩油门”的安全准则，将高达 50% 的工程带宽、算力预算与代币消耗直接投向评估套件（evals）的构建。团队明确摒弃通话时长或交互频次等虚荣指标，严苛依据最终下游业务成果——即客户是否达成交易、核心价值是否交付以及用户是否持续自发互动——来构筑安全护栏，以此换取在生产环境中高速、激进部署最新模型的工程信心。 与底层系统重构相配套的是组织架构的深度变革。Kavak 设立了覆盖全员的“绝地学院”，要求从管理高层、软件工程师到机修工的所有员工在 6 周内掌握构建生产级智能体的技能，从而建立人机协同的工作文化。公司彻底去除了传统中层管理人员，转由扁平且经验丰富的高资历团队运作；当智能体在运行中受阻时，会自动调用“人类 API”寻求人工干预，形成闭环训练数据以供迭代，并对代币开销建立严格的三级 ROI 追踪框架。 Alejandro 认为当前创业者正处于人类历史上工具与智能极其平价的关键时期，主张不必过度揣摩指数级曲线，只需沿着当下的线性技术进步趋势提前布局系统架构。然而需要指出的是，Kavak 激进推行“单客户智能体”并取得超人绩效的前提，高度依赖其重资产垂直整合、自营信贷闭环以及对整车生命周期的深度控制；这种模式在资产较轻或数据链条分散的传统行业中能否同样有效复制，依然存在客观不确定性。

## Background 本期节目由风险投资机构 a16z 于 2026 年 8 月 13 日在 YouTube 平台发布。节目由 Angela Strange 与 Gabriel Vasquez 主持，邀请拉美二手车独角兽 Kavak 的首席产品与 AI 官 Alejandro Maza Ayala，深度拆解 Kavak 如何摒弃浅层的工具采纳，围绕 AI 智能体彻底重构企业系统、组织架构与业务流程，实现 96% 客户交互与 95% 交易均由智能体全权托管的实战经验。 ## TL;DR 面对大模型浪潮，Kavak 毅然推倒旧有系统，开创性地落地了“单客户单专属智能体”架构，每天动态实例化数十万个长程记忆 Agent 直接接管端到端交易，使其销售转化率达到人类团队的 2.1 倍、车贷审批缩短至 3 分钟以内，更通过“AI CEO”试点在单城市实现首月利润激增 50%。为了保障系统的高可靠性，Kavak 坚决贯彻评测（Evals）与研发 1:1 的工程投入法则，并通过内部“绝地学院”强制要求从高管到一线汽修工全员掌握 Agent 构建能力，消灭传统中层并围绕智能体全面重组协作链条。Alejandro 借用福特流水线彻底颠覆旧工厂的电力革命史指出，浅层套用 AI 只能带来微弱的边缘改善，唯有敢于自我破坏、围绕全新智能范式彻底重构自进化组织的企业，才能释放数倍的生产力跃迁。 ## Atmosphere & style 整场对话洋溢着极度硬核的务实精神与破釜沉舟般的颠覆决绝，节奏紧凑、信息密度极高，完全褪去了行业常见的概念包装与演示滤镜。Alejandro 的讲述坦率、尖锐且充满极客哲学，无论是直言“不感到恐惧说明没真正碰过 AI”，还是亲手销毁两年成熟盈利架构推倒重来的笃定，都透露出强大的战略魄力与终局自信。两位主持人与嘉宾在商业逻辑与技术落地的交叉口高频碰撞，现场氛围既像是一场严肃的战地复盘，又宛如一剂写给未来创业者的清醒剂，让人深刻感受到工业级重构正在真实发生的情绪震颤与时代力量。 ## Bulletin - [00:00](https://www.youtube.com/watch?v=n34CIw3gk1k&t=0s) 二手车独角兽 Kavak 重构为全智能体架构，目前 96% 的客户交互和 95% 的交易流程已由 AI 智能体端到端自主闭环处理，算力与 Token 支出已全面超越知识型员工的薪酬投入。 - [02:23](https://www.youtube.com/watch?v=n34CIw3gk1k&t=143s) Kavak 摒弃传统的 Multi-Agent 任务图谱，转向“单客户单智能体（Agent-per-Customer）”架构，每日动态实例化 10 万至 20 万个独立虚拟机，并赋予智能体最大化客户终身价值（LTV）的长程目标与底层系统 API 权限。 - [10:49](https://www.youtube.com/watch?v=n34CIw3gk1k&t=649s) 全能销售智能体的转化率达到人类团队的 2.1 倍，NPS 翻了三倍，并将传统银行耗时两个月的汽车信贷审批压缩至 3 分钟以内。 - [16:13](https://www.youtube.com/watch?v=n34CIw3gk1k&t=973s) 在墨西哥库埃纳瓦卡展开的“AI CEO”实验中，AI 代理接管了整座城市的车辆库存调度与每日执行计划，仅用 6 周时间便使该市业务利润提升 50%。 - [32:52](https://www.youtube.com/watch?v=n34CIw3gk1k&t=1972s) 企业 AI 落地面临类似 19 世纪“电气化”的重构考验：仅给员工配发 AI 插件等同于旧厂房换电动机，只能带来 6% 的微调；只有彻底推倒重构组织流程，才能获得 3 倍以上的生产力飞跃。 ## Key takeaways - [00:00](https://www.youtube.com/watch?v=n34CIw3gk1k&t=0s) Token 投资优先于人力资本：Kavak 将计算资源与 Token 消耗置于传统知识型团队之上，致力于打造在所有核心指标上全面超越人类的“超人智能体”。 - [28:08](https://www.youtube.com/watch?v=n34CIw3gk1k&t=1688s) 颠覆任务图谱范式：Kavak 毅然销毁了耗费两年搭建的 Multi-Agent 图架构，转向“单客户单专属智能体”架构，每日在独立微虚拟机（Micro-VM）中动态调度多达 20 万个实例。 - [04:59](https://www.youtube.com/watch?v=n34CIw3gk1k&t=299s) 从事务型向关系型企业蜕变：借助具备长程记忆的专属智能体，企业不再只是计算购销频次，而是以最大化 1000 万客户的终身价值（LTV）为战略核心，激活 1% 的存量即可撬动数亿美元营收。 - [07:35](https://www.youtube.com/watch?v=n34CIw3gk1k&t=455s) 评测开发与系统构建的 1:1 资源黄金法则：智能体如同超跑，评估系统（Evals）则是刹车；Kavak 在评估系统上投入的工程时间、算力与资金与构建智能体完全对等，并坚决拒绝通话时长等过程性虚荣指标。 - [10:22](https://www.youtube.com/watch?v=n34CIw3gk1k&t=622s) 超级全能销售智能体碾压人类专家：整合了选车、信贷、保险与置换等 15 项专业职能，凭借无限耐心与数据飞轮，AI 销售转化率达到人类团队的 2.1 倍，客户 NPS 提高至 3 倍。 - [14:23](https://www.youtube.com/watch?v=n34CIw3gk1k&t=863s) 垂直整合支撑 3 分钟极速车贷：通过掌控二手车资产与全流程数据，AI 能在 3 分钟内完成信贷风控，并通过无力偿还时允许置换低价车的兜底机制化解坏账风险。 - [16:13](https://www.youtube.com/watch?v=n34CIw3gk1k&t=973s) “AI CEO”管理实验：在墨西哥试点城市上线 AI 代理负责统揽库存周转与运营预测，以极度精细的数据追踪下发执行指令，6 周内推动当地利润提升 50%。 - [18:16](https://www.youtube.com/watch?v=n34CIw3gk1k&t=1096s) 人机协同的“料理鼠王”模式：AI 质检助手“El Mike”指导一线修车技师盲区排查与规范流程，使出厂车辆的售后质保返修率降低了约 26%。 - [21:02](https://www.youtube.com/watch?v=n34CIw3gk1k&t=1262s) 全员入读“绝地学院”：从首席执行官到一线蓝领机修工，全员接受为期 6 周的系统化 AI 培训并能够交付生产级智能体，构建人人与 AI 协同的激进文化。 - [22:46](https://www.youtube.com/watch?v=n34CIw3gk1k&t=1366s) 中层管理消解与组织重构：传统层级结构被废除，全员重新分化为三类角色——构建智能体、为智能体提供支持服务、以及深入物理世界处理实体业务。 - [24:01](https://www.youtube.com/watch?v=n34CIw3gk1k&t=1441s) 反转 Human-in-the-Loop：摒弃将失败案例粗暴甩给人工客服的单向流程，改由专属智能体主动调用人类 API 获取决策支持，将干预过程完整记录以反哺评估与微调闭环。 - [25:18](https://www.youtube.com/watch?v=n34CIw3gk1k&t=1518s) 转型必须自上而下推进：零散黑客松或自下而上的微小试水终将沦为“玩具”，AI 转型需要领导层具备极度清晰的顶层战略和如同军队推进般的决绝意志。 - [25:18](https://www.youtube.com/watch?v=n34CIw3gk1k&t=1518s) 三层 Token 价值衡量框架：盲目消耗算力毫无意义，企业应杜绝泛用型漫游（Tier 1），将核心资源集中在能端到端承接关键业务职责、可直接测算商业 ROI 的业务自主型智能体（Tier 3）。 - [28:08](https://www.youtube.com/watch?v=n34CIw3gk1k&t=1688s) 超越模型自我改进，构建“自进化组织”：商业价值归根结底是由组织创造的，唯有将公司本身打造成能无缝吸收迭代底层大模型红利的自进化组织，才能实现长期价值的指数级爆发。 - [32:52](https://www.youtube.com/watch?v=n34CIw3gk1k&t=1972s) 福特工厂与创造性破坏的历史警示：在旧系统上修补（将蒸汽机换成电动机）仅带来 6% 效率提升，唯有像福特一样打破旧厂房重塑流水线才能释放 300% 生产力，创业者应把握老牌巨头难以自我重构的结构性窗口。 - [34:45](https://www.youtube.com/watch?v=n34CIw3gk1k&t=2085s) 致创业者的底线思维：当今是构建新公司前所未有的黄金时代，创业者无需豪赌超指数爆发，只要依据确定性的线性技术演进“往深处做”，就足以打造出颠覆性的原生商业范式。 ## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=n34CIw3gk1k&t=0s) 开场：全力押注 AI 智能体与彻底重构企业 视频开篇剪辑了 Alejandro Maza Ayala（Kavak 的首席产品与 AI 官，昵称 Ale）的核心暴论：如今他在计算 Token 上的投资，已经超过了在知识型员工上的支出。Kavak 树立了一个极具野心的目标——打造“超越人类水平的超人智能体”（Superhuman Agents），这意味着在所有核心业务维度上，这些智能体的表现都必须超越公司有史以来雇佣过的最优秀人类员工。为了实现这一点，Kavak 毅然放弃了行业内常见的拼凑式 AI 工作流，开创性地采用了“单客户单专属智能体”（Agent-per-Customer）的架构，每天有 10 万到 20 万个专属智能体在各自独立的虚拟机中实例化运行。 面对各界对未来组织形态以及人类角色边缘化的焦虑，Alejandro 直言：“如果你在 AI 浪潮中还没有体会过恐惧，那说明你根本没有真正尝试过 AI。”为了带领全员跨越这种恐惧，Kavak 在内部创办了“绝地武士学院”（Jedi Academy），从 CEO 到一线修车汽修工，所有人都必须接受构建 AI 的系统化培训，并在短短六周内将前沿的智能体推向生产环境。对于正在探索技术路径的创始人，这场从底层重新发明公司的实践提供了极具颠覆性的参照样本。 - **Token 投资超越人力资本**：计算资源与 Token 消耗的优先级已凌驾于传统知识型团队之上。 - **全员培训与交付**：通过“绝地学院”，无论是高管还是汽修工，都能在 6 周内构建出生产级前沿智能体。 --- ### [01:03](https://www.youtube.com/watch?v=n34CIw3gk1k&t=63s) 从 2013 年机器学习探索到 ChatGPT 奇点时刻 主持人 Angela Strange 与 Gabriel Vasquez 欢迎 Alejandro 登场，并回顾了他在进入 Kavak 之前的创业背景。早在 2013 年深度学习变革前夕，Alejandro 就创立了 OPI Analytics。在那个 Transformer 架构尚未问世、行业普遍统称为“机器学习（Machine Learning）”的年代，他们就抱着超前十年的野心，坚信算法模型终有一天能够解决现实世界中一切复杂的商业决策难题。 当时 OPI Analytics 已经服务了包括《财富》500 强在内的诸多大型企业，为他们定制风险控制、物流配送、销量预测及精准营销算法。尽管业务扎实，但受限于早期算法的局限性，技术始终无法实现真正端到端的智能飞跃。直到 Transformer 普及以及 ChatGPT 带来的奇点时刻降临，Alejandro 意识到，彻底打破传统软件与组织架构局限的时代终于到来。带着这项技术洞察，他与 Kavak 联合创始人 Carlos Garcia Ottati 等人携手加入 Kavak，开启了一场激进的 AI 本生转型。 --- ### [02:23](https://www.youtube.com/watch?v=n34CIw3gk1k&t=143s) Kavak 的业务版图与“单客户单智能体”前沿架构 在深入探讨技术细节前，Alejandro 梳理了 Kavak 的核心业务生态。虽然 Kavak 最初是以二手车交易平台的模式起家——收购车辆、翻新整备、重新销售并提供配套金融服务，但由于拉美市场的基础设施严重匮乏，不存在成熟的信用评估体系、全国性物流网络或类似美国的 Carfax（车辆历史记录数据库），Kavak 只能被迫自建全套垂直生态，一网打尽了汽车金融公司、物流运输网络与车辆数据审计服务。 面对如此庞大的重资产与复杂流程，Kavak 团队向自己提出了一个终极设想：“如果我们身处 2035 年，手握 GPT-10 级别的超强智能，我们会如何重新构建 Kavak？”答案绝不是在现行架构上打补丁。Kavak 抛弃了业内主流的、按职能划分的专家多智能体系统（Multi-Agent System），转而采用“一客一代理”（Agent-per-Customer）的架构。当任何一个客户进入系统，后台便会为该客户专门在独立的虚拟机（VM）中实例化一个专属智能体。这个智能体拥有长程持久记忆，能清晰调阅客户过去数年间在官网的每一次点击、两年前的一通客服电话等所有交互数据，自主制定全局转化策略，以最大化客户终身价值（LTV）为硬性目标并驱动行动。 ```mermaid flowchart TD accTitle: Kavak 单客户单智能体系统架构 accDescr: 展示客户触发系统后实例化专属长程记忆智能体，并调用全流程底层业务系统的流程图。 subgraph User["客户入口"] C["客户触达 (买车/卖车/贷款)"] end subgraph AgentRuntime["专属智能体运行时 (VM)"] VM["独立虚拟机实例"] M["多年全触点长程记忆"] O["硬性目标: 最大化 LTV 与客户满意度"] VM --- M VM --- O end subgraph BackendSystems["垂直重构的业务中台"] B1["汽车金融与贷款审批"] B2["车辆整备与检测数据"] B3["物流运输调度"] B4["库存与动态定价"] end C -->|"动态实例化 (100k-200k/天)"| VM O -->|"自主调用与编排 API"| B1 O -->|"自主调用与编排 API"| B2 O -->|"自主调用与编排 API"| B3 O -->|"自主调用与编排 API"| B4 ``` - **全流程垂直自建**：涵盖二手车电商、Fintech 借贷、物流网络及整备工厂。 - **动态实例化机制**：每天动态创建 10 万到 20 万个智能体虚拟机容器，专属服务特定客户。 - **目标导向 vs 流程导向**：摒弃僵硬的预设工作流（Workflow），赋予智能体明确的长远目标（Hard Goals）和端到端自决策权。 --- ### [04:59](https://www.youtube.com/watch?v=n34CIw3gk1k&t=299s) 三大激进押注：重构企业系统、超人智能体与指标革命 在大多数企业只是浅尝辄止地进行“AI 试水”时，Kavak 经历了一整年推倒重来、甚至大规模裁员阵痛的彻底变革。Alejandro 剖析了许多企业在转型中陷入的误区：企业的第一本能往往只是“采纳 AI（Adopt AI）”，在不触动组织架构的前提下给员工分发 ChatGPT 或 Claude 账号，结果系统壁垒依然存在，运营效率未见改观，客户痛点也原封不动。对此，Kavak 确立了三项破局押注： 1. **彻底重构系统底座**：不搞渐进式接入，围绕智能体的交互需求重写公司绝大部分 API、底层数据架构与系统界面，让智能体可以直接接管并操作底层业务系统。 2. **打造超越人类的超人智能体**：AI 必须直接面对最棘手、最复杂的业务问题。通过将智能体投入真实业务环境产生反馈闭环与评测数据（Evals），反复微调训练，确保智能体在转化率、客户留存、用户体验等所有关键指标上全面碾压最优秀的人类员工。 3. **彻底颠覆企业核心考核指标**：Kavak 从前是一家以交易为中心的事务型公司（Transactional Company），考核的是买了多少辆车、卖了多少辆车、采购了多少副刹车片；如今转变为以关系为中心的关系型公司（Relational Company）。数据库中沉淀的 1000 万客户，大部分都有专属智能体在全生命周期内进行跟进和关系维护。 在二手车与大额个人信贷这种客单价极高的领域，客户交易的核心驱动力是信任，而建立信任的基础在于企业真正懂客户、能够长期维系关系。面对 1000 万高净值客户群体，哪怕依靠智能体仅多激活 1% 的存量客户，就能直接带来数亿美元的增量营收。 --- ### [07:35](https://www.youtube.com/watch?v=n34CIw3gk1k&t=455s) 96% 交互全自动化与评估驱动（Evals）的研发方法论 Gabriel 指出，许多公司在演示 Demo 阶段看起来很惊艳，但极少有人敢将 90% 以上的真实客户交互交给智能体托管。Alejandro 披露了 Kavak 当前惊人的落地数据：公司 **96% 的客户交互** 和 **95% 的交易流程** 已完全交由 AI 智能体闭环处理，整个生命周期中唯一的真人触点，仅剩客户线下取车时由工作人员现场递交车钥匙。 面对每天实例化出的 10 万到 20 万个智能体——它们有的只被唤醒执行 3 分钟的任务，有的连续运行 8 小时甚至 3 天，任务结束后自动设定“闹钟”进入休眠状态——保障系统在如此大尺度下稳定运行的秘诀在于**评估（Evals）**。Alejandro 形象地比喻道：“智能体就像一辆性能超凡的跑车，你只有配上了足够强大灵敏的刹车系统，才敢把油门踩到底。”很多企业之所以步履蹒跚，根源正是因为缺乏可靠的刹车（即评估机制）。 在资源调配上，Kavak 严格执行一条原则：**在评测系统（Evals）上投入的工程师时间、Token 算力和资金，与开发智能体本身的投入完全对等（1:1）**。在具体的考核评估上，Kavak 坚决拒绝用“通话时长”、“拨打频次”等表面化 KPI 自欺欺人，而是牢牢锚定核心业务成果：客户是否满意？贷款是否通过？买车或卖车的交易最终是否达成转化？ - **全流程接管比例**：96% 客户交互全自动化，95% 交易全自主闭环。 - **1:1 资源黄金法则**：评测开发与智能体开发耗费完全对等的工程时间与计算成本。 - **反虚荣指标**：抛弃通话时长等过程性 KPI，将最终商业转化与真实满意度作为评估的第一准则。 ### [10:22](https://www.youtube.com/watch?v=n34CIw3gk1k&t=622s) 销售智能体：超越人类专家的超级销售 许多企业在引入 AI 时往往态度保守，只敢将其放在后台或仅用于处理边缘任务，甚至认为高客单价、流程极其复杂的销售场景只能由人类负责。Alejandro 指出，Kavak 从未打算把 AI 局限在传统的“客服支持”角色上，而是从一开始就打造真正的销售智能体（Sales Agents）。在拉丁美洲买卖二手车是一个极其繁复的过程：用户需要在两万个库存（SKU）中挑选，还要处理贷款申请、购买保险与延保服务，并且往往伴随着旧车置换（Trade-in）的估值流程。在 2020 至 2021 年期间，Kavak 往往需要 15 个不同领域的专家团队分别对接客户；而现在的做法是打造一个在贷款、选车、估值、保险等 15 个细分领域都超越单项人类专家的“超级全能专家智能体”直接面对客户。 这套全能销售智能体带来了令人震惊的业务增长：上线后，客户净推荐值（NPS）和满意度直接翻了三倍；初期转化率就比人类销售团队高出 50%，随后更是达到了人类团队的 2.1 倍。Alejandro 强调，AI 销售之所以彻底碾压人类，是因为它们拥有近乎无限的耐心、掌握客户完整的历史交互数据、具备长期规划能力且永远不知疲倦。更关键的是飞轮效应——当某个智能体犯错并被修正后，次日不仅是它自己，全系统部署的 20 万个智能体都会同步学到教训，形成极其强大的自我迭代能力。 ```mermaid flowchart TD accTitle: Kavak 销售流程变革架构图 accDescr: 展示 Kavak 如何从过去 15 个分散的人类专家团队过渡到单一全能 AI 销售智能体的对比流程。 subgraph OLD["过去的人类专家流程 (2020-2021)"] direction TB C1["客户 (购车/置换)"] --> H1["车辆推荐专家"] C1 --> H2["汽车金融专家"] C1 --> H3["车险保障专家"] C1 --> H4["二手车估值专家"] H1 & H2 & H3 & H4 --> P1["分散耗时的购车方案"] end subgraph NEW["全新 AI 智能体架构"] direction TB C2["客户 (购车/置换)"] --> AG["超级全能 AI 销售智能体<br/>(无限耐心 / 掌握全量数据 / 零疲劳)"] AG --> S1["2 万 SKU 精准匹配"] AG --> S2["3 分钟极速贷款核保"] AG --> S3["个性化车险与质保"] AG --> S4["旧车即时估值"] S1 & S2 & S3 & S4 --> P2["全链路高转化交易 (2.1x 转化率)"] end ``` - **全流程覆盖**：智能体单点整合了选车（20,000 SKU）、金融分期、保险配置与旧车置换等 15 种专业职能。 - **业务转化暴增**：相比人类销售团队，AI 智能体的销售转化率飙升至 2.1 倍（初期即达 1.5 倍）。 - **满意度倍增**：将智能体直接推向一线客户后，NPS 与客户满意度得分增长至原来的 3 倍。 - **群体进化**：单点修复直接泛化至全网部署的 200,000 个智能体实例。 --- ### [14:23](https://www.youtube.com/watch?v=n34CIw3gk1k&t=863s) 3分钟金融核保与垂直整合闭环 Angela 提到了行业内的普遍质疑：大家或许勉强能接受 AI 做客户支持，但很难相信 AI 能够端到端搞定强监管、高风险的金融服务，尤其是为那些缺乏信用记录（Thin-file 或 No-file）的下沉客群完成风控定价与贷后服务。Alejandro 解释说，在墨西哥及其他新兴市场，传统银行审批一笔汽车贷款往往要耗费两个月甚至更长时间；而 Kavak 凭借对车辆资产和客户交易数据的全面掌控，直接将放款审批压缩到了 3 分钟以内。 Kavak 垂直整合（Vertical Integration）的商业模式在此发挥了巨大的护城河效应。如果借款客户后续出现还款困难，他们完全无需违约，而是可以直接将车辆退还给 Kavak 并置换为一辆更便宜的二手车，从而降低每月月供、化解流动性危机。这种端到端的闭环让 AI 能够在客户长达 3 至 4 个月的选车决策周期中深入了解对方的真实画像，并对贷款额度、利率水平和风险敞口进行超深度的动态个性化定制，不仅平衡了整体贷款资产组合的风险，还大幅拉高了留存率与转化率。 - **审批时效颠覆**：传统渠道需 2 个月以上的车贷审批，Kavak AI 仅需不到 3 分钟完成。 - **资产兜底机制**：借款人无力偿还时可通过换购低价车辆降月供，依靠垂直整合化解坏账风险。 - **长决策期运营**：在用户 3 至 4 个月的决策窗口期内，AI 持续根据动态风险和外部竞品报价调整利率与最大授信额度。 --- ### [16:13](https://www.youtube.com/watch?v=n34CIw3gk1k&t=973s) “AI CEO”实验：6周利润提升50% 当团队决定以 AI 为核心重塑公司时，Alejandro 和团队抛出了一个激进的问题：“AI 究竟能不能胜任领导岗位，甚至直接干 CEO 的活？”尽管外界普遍认为这种高阶决策在 2035 年左右才可能实现，Kavak 却决定立即动手实验。他们选择了墨西哥的库埃纳瓦卡（Cuernavaca）作为试验田，在自研的 Harness 框架内上线了一个 AI 代理作为该城市的“AI CEO”。该城市业务首月设定的目标是利润翻倍，虽然最终未能完全达成翻倍指标，但在运行仅 6 周后，该城市的实际利润就实现了 1.5 倍增长（提升 50%）。 这位“AI CEO”的工作模式被 Alejandro 戏称为“菲尔兹奖级别的天才下场进行极致微观管理（Micromanagement）”。它深入审查每一个数字、跟踪每一位客户、构建完美的销售预测，并严格管控日常执行。每天，AI CEO 会直接向库埃纳瓦卡的一线物理作业员工发送定制化工作计划，并要求员工通过语音备忘录（Voice notes）打卡汇报进度。在它的调度下，当地的库存周转率加快、车辆结构优化、金融渗透率提升，各项核心 KPI 均出现系统性好转。这打破了人们“管理层是 AI 最难取代的角色”的固有偏见。 - **试点区域成效**：库埃纳瓦卡市由 AI CEO 全权掌管运营，首月即实现利润提升 50%（1.5x）。 - **精细化协同方式**：AI CEO 统揽库存周转与金融渗透预测，每天下发具体执行指令，并要求一线员工发送语音便签汇报进度。 --- ### [18:16](https://www.youtube.com/watch?v=n34CIw3gk1k&t=1096s) “料理鼠王”副驾与绝地学院（Jedi Academy） 既然软件层面的管理和销售都可以被 AI 接管，那么哪些工作依然需要人类？Alejandro 指出，物理现实世界中依赖肌肉灵巧性（Dexterity）和感官经验的工作依然需要人类，比如 Kavak 在墨西哥聘请的约 800 名专业修车技师。但即便在这里，AI 同样没有缺席。Kavak 基于同一套底层架构为技师们打造了专属 AI 助理——“El Mike”。Alejandro 用皮克斯动画《料理鼠王》作比喻：“就像那只藏在厨师帽子里指挥做饭的老鼠一样，El Mike 就像一个副驾（Sidekick）。”它会实时指导技师如何检查车辆盲区、提供维保建议并规范操作流程。结果，车辆质检和维修速度大幅提升，更重要的是出厂质量显著改善，上线以来车辆的质保返修率下降了约 26%，客户满意度再度上升。 面对“人机协作”带来的组织形态巨变，Kavak 早在三年前就清醒地认识到，全员的工作性质都将被 AI 颠覆。为了防止员工被淘汰，Kavak 内部设立了硬核培训体系——“绝地学院（Jedi Academy）”。无论是公司最高层 CEO、AI 工程师，还是一线的汽车修理工，所有人都必须进入绝地学院学习如何构建和使用智能体。由于前沿 AI 技术迭代极为迅速，传统外部高校根本没有现成课程，Alejandro 亲自负责设计并不断刷新整个课程大纲，推动全员在实战中重塑自身的技能树。 - **料理鼠王式人机协作**：AI 质检助手“El Mike”深度辅助墨西哥约 800 名车辆检修技师。 - **质保指标显著改善**：在 El Mike 辅助下，车辆检验维修质量大幅提升，售后质保返修率下降约 26%。 - **全员绝地学院（Jedi Academy）**：由首席产品与 AI 官亲自设计与迭代，覆盖从 CEO 到蓝领机修工的全部员工，推动全员掌握 Agent 构建能力。 ### [21:02](https://www.youtube.com/watch?v=n34CIw3gk1k&t=1262s) 全员 AI 训练与“不可逆转”的组织变革 为了让公司全面转型，Kavak 展开了全员密集培训。经过仅 6 周的系统化训练，员工们便能够将达到业界前沿（SOTA）水平的 AI Agent 部署到生产环境。参与培训的不仅有软件工程师，还包括一线汽修工、财务人员等各类岗位。Alejandro Maza Ayala 坦言，这并不意味着所有人都会转型成为专职的 AI 工程师（尽管确实有一部分人做到了），但其核心价值在于让全员掌握了与这项新技术协同作战的技能。 在推动这场变革时，管理层的态度极其坚决，甚至带着破釜沉舟的气魄。他们明确告诉团队：“没有任何回头路可走，这就是 Kavak 未来的走向，这就是整个公司的形态。”无论是工程、财务还是产品，整个组织架构都将迎来颠覆性重构。管理层给员工提供了一个清晰的选择：要么接受培训、掌握在新现实中生存与竞争的新技能，要么如果不认同这种方向，就选择离开 Kavak。这种决绝反而极大地强化了团队文化，员工们迅速具备了构建 Agent 系统的能力。在如今的 Kavak，任何业务流程本质上都是人机协作系统——有时 Agent 是人类的上级和任务分配者，有时人类则是 Agent 的架构师与设计者。在技术飞速迭代的时代，唯一的生存法则就是每隔几个月甚至每个月持续更新自身的技能体系。 ### [22:46](https://www.youtube.com/watch?v=n34CIw3gk1k&t=1366s) 扁平化架构：人不再按层级管理，而是围绕 Agent 协作 当成千上万原本由人类处理的工作全面交由 Agent 接管后，传统中层管理的组织概念在 Kavak 彻底不复存在。取而代之的是高度扁平、高度资深且被充分赋权的敏捷团队。每个小团队都由工程、AI、业务运营等多元角色深度融合而成。在这样全新的组织蓝图下，全员基本上只分布在三种职责维度中： 1. **构建 Agent**（设计核心能力、工具接口与架构）； 2. **为 Agent 工作**（作为 Agent 无法闭环时的下游支持节点）； 3. **深入物理世界**（直接面对客户，执行无法完全数字化的实体业务，例如验车与交付）。 这种组织形态与 2022、2023 年前完全不可同日而语。过去没有任何人会为 Agent 编写技能，更不会接受由 Agent 发出的调度指令。如今，即便是一线面对客户与实体车辆的员工，也是在 Agent 的赋能与协同指导下开展工作。 ### [24:01](https://www.youtube.com/watch?v=n34CIw3gk1k&t=1441s) 重新定义 Human-in-the-Loop：从“工单丢失”到“闭环反哺” Alejandro 深入拆解了 Kavak 与业界绝大多数公司在 Human-in-the-Loop（人机协同闭环）设计上的根本差异。当下许多大型生产级 Agent 系统的通用做法非常粗暴：当 Agent 遇到瓶颈、任务失败或无法继续推进时，系统会直接把案例甩给二级人工客服（Tier 2 Support），随后 Agent 彻底撒手不管。Alejandro 尖锐地指出，这种模式存在致命缺陷：**它根本没有形成闭环，无法沉淀出高质量的纠错与决策数据来持续训练和提升 Agent**。 Kavak 彻底倒转了这一关系，构建了一套“以客户为中心”的反向调用机制： ```mermaid flowchart TD accTitle: Kavak Human-in-the-Loop Architecture accDescr: 客户专属 Agent 遇到边界情况时通过 API 请求人类支持并完成闭环学习的流程 subgraph AgentLoop["Agent 决策执行层"] A["客户专属 Agent (Agent-per-Customer)"] -->|"调用全业务 API / 工具"| B["自主执行交互与交易"] B -->|"遇到阻碍 / 边界场景"| C{"命中能力边界?"} end subgraph HumanAssistance["人类专家支持层"] C -->|"是：调用内部 Help API"| D["人类支持专家协助介入"] D -->|"返回决策逻辑与解决方案"| A end subgraph FeedbackLoop["数据飞轮层"] A -->|"全流程日志与闭环数据"| E["Eval 评测与微调训练"] E -->|"模型与能力持续进化"| A end ``` - **客户专属 Agent 拥有主导权**：平台上数以百万计的客户各自拥有一个专注其全生命周期的专属 Agent。该 Agent 拥有调用公司所有底层 API 与专业工具的最高权限。 - **人类扮演 API 服务**：当 Agent 遇到无法处理的复杂状况时，它会主动调用一个特殊的 API 发出求助信号——此时坐在屏幕另一端接单提供支持的正是人类员工。在微观层面，相当于人类团队在协同辅助 Agent 完成任务。 - **全闭环数据反哺**：通过这种方式，Agent 从头到尾没有离开上下文，人类的干预被完整记录并直接转化为提升 Agent 评估基准与后续迭代的核心数据。 ### [25:18](https://www.youtube.com/watch?v=n34CIw3gk1k&t=1518s) 企业落地 AI 的核心：自上而下的意志与三层 Token 价值法则 针对许多大型机构高管频繁前来请教“模型能力明明足够，为何企业内部推行落地极其艰难”的心理与组织瓶颈，Alejandro 给出了两条极为关键的破局建议： 第一，**AI 转型必须是绝对自上而下（Top-Down）的战略推进**。自下而上（Bottom-Up）的零散尝试是死胡同。很多企业热衷于举办内部黑客松、让各个部门天马行空地提交零碎场景并给予微小资助，最终几乎无一例外沦为“玩具”。普通员工缺乏全盘商业战略的嗅觉，更无法形成一致的品味与方向。企业领导层必须拥有极度清晰的 3 至 5 年终局视野，并以类似军队打仗般的纵向执行力带领团队冲锋。“一支军队不可能让士兵在战场上各自凭空构想战术、为所欲为；处于转型期的企业需要的是绝对清晰的顶层战略。” 第二，**必须建立严密的 Eval 评估体系，严格衡量 Token 的真实 ROI**。盲目炫耀内部 Token 消耗量毫无意义，重点在于“Token 的质量”。Alejandro 提出了评估企业 AI 投入的三层 Token 价值框架（Token Tiers）： | 分层 | 类别 | 评估方式与商业价值 | |:--- |:--- |:--- | | **Tier 3（最高价值）** | **业务自主型 Agent** | 能够端到端承接公司核心组织职责（如直接成交、审核贷款），**可以直接精确核算出消耗每一枚 Token 所换取的实际业务 ROI**。Kavak 目前绝大多数 Token 支出均处于此层级。 | | **Tier 2（中等价值）** | **可间接评估价值的生产力系统** | 例如代码库中的 AI 辅助编程，能通过代码差异度、合并率与发布上线间接评估 Token 价值。 | | **Tier 1（低价值/浪费）** | **泛用型工具漫游** | 绝大多数企业所处的阶段。员工只是在日常零散使用 ChatGPT、Copilot 生成文本，管理层对 Token 究竟产生了什么具体效益一无所知。 | ### [28:08](https://www.youtube.com/watch?v=n34CIw3gk1k&t=1688s) 亲手毁掉两年心血：从 Multi-Agent 图架构转向“单客户单虚拟机” 在系统架构演进上，Alejandro 分享了一次极其冒险却决定性的战略决断。业界以往的普遍范式是沿着业务工作流（Workflow）搭建 Agent，构建复杂的 Multi-Agent 图系统（Graphs）或函数树——每个微型 Agent 分别负责贷款、车辆检索或推荐等细分动作。到 2025 年底，Kavak 依靠这套多 Agent 编排系统实现了盈利和惊人增长，每天有数万个此类 Agent 在规模化运转业务。 然而，当大语言模型能力跃升至 Claude 3.5 Sonnet / Opus 4.5 级别时，Alejandro 敏锐地意识到旧范式已经彻底过时了。高度结构化、层层嵌套的 Agent 图谱网络，反而成了更高级通用智能的枷锁与束缚。于是，团队做出了一个惊人的决定：**亲手彻底销毁过去花费两年搭建、运行极其成功且带来巨大盈利的 Multi-Agent 图系统，推倒重来**。 新架构摒弃了“按任务拆分”的逻辑，确立了**“每个客户配备一个专有虚拟机（Micro-VM）与独立 Agent”**的新范式： - **单客户专属环境**：每天系统动态实例化数十万个专属 Agent 虚拟机； - **全权限命令行与工具流**：Agent 配备持久化记忆、全量 Eval 监控以及全套系统 CLI 权限，可任意调度公司的底层 API； - **长期目标驱动**：Agent 的指令不是完成单一步骤，而是锚定宏观业务指标，例如“最大化该客户的全生命周期价值（LTV）”。 Alejandro 强调，当前 AI 界痴迷于算法和模型本身的递归式自我改进（Recursive Self-Improvement, RSI），但这忽视了更大的商业现实：在人类过去四千年的文明史中，经济价值始终是由“组织”而非孤立的“个体”创造的。因此，真正需要进入自进化循环的，是**能够交付更大经济价值的“自进化组织”（The Self-Improving Organization）**。当企业构建出一套能够自动吸收迭代模型新能力、持续自我调优与业务闭环的组织架构时，企业价值才能真正迎来与底层模型智能同步的指数级爆发。 ### [31:16](https://www.youtube.com/watch?v=n34CIw3gk1k&t=1876s) 创造性破坏与福特工厂：为什么浅层采纳 AI 远远不够 加布里埃尔·巴斯克斯（Gabriel Vasquez）提出，鉴于传统企业在深度采纳 AI 面临重重阻力，当前最大的商业机会或许属于那些从第一天起就以全新范式运转、进而颠覆既有市场的新创公司。亚历杭德罗·马萨·阿亚拉（Alejandro Maza Ayala）对此深表赞同，并引入了经济学家约瑟夫·熊彼特（Joseph Schumpeter）著名的“创造性破坏”（Creative Destruction）理论。他指出，技术创新对经济体系的重塑，从来不是靠老牌巨头温和采纳新技术来完成的，而是既有企业固步自封，掌握新技术的颠覆者彻底摧毁老牌公司。这种颠覆虽然在短期内会造成局部的经济价值破损，但在长远尺度上对全社会有利，因为更高效、更具生产力的新型企业能够为大众提供品质远超以往的产品与服务。 这一历史规律为当下的创业者带来了千载难逢的窗口期。对于如今任何一家大型成熟企业或上市公司的 CEO 来说，彻底转向 AI 是一项近乎不可能完成的任务——没有人有魄力站出来宣布：“我要把公司过去 40 年积累的一切全部推倒重建，押上全部身家重构成一家 AI 原生公司。”这种巨头转型的天然阻力，给新成立的创业公司留下了充足的空间：它们完全可以围绕 AI 的核心优势从零构建商业模式，直接接管市场。 为了向团队解释这一点，亚历杭德罗经常分享一个电力普及的历史典故：福特装配线流水线所依赖的核心电气技术，早在 1879 年至 1881 年托马斯·爱迪生（Thomas Edison）发明高效发电机并商业化供电时就已经成熟。理论上，福特工厂完全可以在 40 年前诞生。然而当时工业界采纳电力的方式极其表面化：工厂主们保留了原有的四层高楼、垂直传动轴和复杂的皮带滑轮系统，仅仅是将底层的燃煤蒸汽机换成了一台大型电动机。这种“修修补补”的改造虽然带来了收益，但生产效率提升仅有微不足道的 6%。 真正的生产力革命，直到有人彻底砸碎旧工厂才得以发生——企业家们放弃了纽约市中心昂贵的多层厂房，搬迁至新泽西或康涅狄格州的平坦开阔地带，建起单层厂房，将整个制造流程完全围绕小型分布式发电机和电力网络重新设计。正是这一根本性的架构重组，带来了高达 3 倍的生产力飞跃，直接奠定了美国 20 世纪工业繁荣的基础。从个人计算机的普及到今天的 AI 浪潮，历史正在原样重演：多数企业只想停留在浅层套用，这最终只能带来 6% 到 10% 的边缘改善；只有敢于颠覆整个组织架构的重塑者，才能斩获 10 倍乃至更高维度的跃迁。这就是在工业级尺度上再度上演的“创新者的窘境”。 ```mermaid flowchart TD accTitle: 浅层采纳与深度架构重组的效率对比 accDescr: 对比传统企业浅层替换动力与原生企业全面重构组织所带来的生产力差距 subgraph Superficial["浅层套用模式 (提升 6% ~ 10%)"] A1["保留旧组织架构与流程<br/>(如多层工厂/传统职能层级)"] --> B1["单一组件技术替换<br/>(蒸汽机换电动机 / 员工配发 AI 辅助工具)"] B1 --> C1["边缘效率微增<br/>维持旧系统摩擦力"] end subgraph Deep["深度重塑模式 (带来 3x ~ 10x 跃迁)"] A2["彻底拆除旧式流程<br/>(平地重构单层厂房 / 废弃传统层级)"] --> B2["围绕核心技术原生构建<br/>(分布式发电机流水线 / 全智能体驱动闭环)"] B2 --> C2["生产力几何级爆发<br/>重写行业成本结构"] end ``` - **创新的破坏性规律**：经济史上每一次重大工业革命，生产力红利都是由以新技术为核心的新公司兑现的，而非依赖传统巨头的自我改良。 - **效率数字的残酷反差**：将电动机简单替换进燃煤旧厂房仅带来约 6% 的效率提升；而福特式全面重构工厂布局后，释放了超过 300%（3x）的生产力跃升。 - **当下的 AI 误区**：企业如果只将大语言模型当成现有岗位的“辅助插件”，其效果等同于在 19 世纪的旧厂房里装电机，注定错失 10 倍级的范式红利。 --- ### [34:45](https://www.youtube.com/watch?v=n34CIw3gk1k&t=2085s) 给创业者的建议：人类历史上最令人振奋的时代 访谈尾声，加布里埃尔总结道，亚历杭德罗的复盘为所有未来的创业者提供了无可辩驳的行动号召——当下正是躬身入局、创造新公司的黄金时代。结合自身创立公司并成功将 Kavak 重构为全智能体驱动平台的实战经历，亚历杭德罗向所有初次创业者及未来建立者分享了核心心得。 “我们正生活在整个人类历史上最令人振奋的时代。”亚历杭德罗情绪饱满地强调，这也是有史以来成为一名创始人最美妙的节点。这是人类文明第一次实现让世界上任何一个普通人，只需每月花费 20 美元甚至近乎免费，就能直接调用全球最顶尖的智能与生产力工具。制造与构建工具的民主化程度，达到了前所未有的顶峰。 面对这片尚未开垦的蓝海，亚历杭德罗给创业者的核心建议是：放手去干，并且一定要“往深处做”（go for it deep）。创业者无需对未来的技术做任何过于激进的指数级假设，哪怕仅仅假设 AI 按照最平稳的线性轨迹持续演进，并坚定地围绕这个确定的趋势去构想未来世界的模样、提前布局构建原生产品。以此为出发点，创业者必将孕育出不可思议的商业构想，并为整个世界创造出无可估量的巨大价值。 - **工具门槛彻底瓦解**：全球顶尖的机器智能已被普惠化，每月只需 20 美元即可拥有世界上最强大的生产工具。 - **构建策略的底线原则**：无需过度押注无法预测的超指数爆发，只要依据“线性提升”的稳定预期深度构筑产品逻辑，便足以支撑颠覆性业务的诞生。 ## Notable quotes > [00:00](https://www.youtube.com/watch?v=n34CIw3gk1k&t=0s) 如果你在 AI 浪潮中还没有体会过恐惧，那说明你根本没有真正尝试过 AI。 — Alejandro Maza Ayala > [02:23](https://www.youtube.com/watch?v=n34CIw3gk1k&t=143s) 如果我们身处 2035 年，手握 GPT-10 级别的超强智能，我们会如何重新构建 Kavak？ — Alejandro Maza Ayala > [07:35](https://www.youtube.com/watch?v=n34CIw3gk1k&t=455s) 智能体就像一辆性能超凡的跑车，你只有配上了足够强大灵敏的刹车系统，才敢把油门踩到底。 — Alejandro Maza Ayala > [16:13](https://www.youtube.com/watch?v=n34CIw3gk1k&t=973s) 菲尔兹奖级别的天才下场进行极致微观管理。 — Alejandro Maza Ayala > [18:16](https://www.youtube.com/watch?v=n34CIw3gk1k&t=1096s) 就像那只藏在厨师帽子里指挥做饭的老鼠一样，El Mike 就像一个副驾。 — Alejandro Maza Ayala > [21:02](https://www.youtube.com/watch?v=n34CIw3gk1k&t=1262s) 没有任何回头路可走，这就是 Kavak 未来的走向，这就是整个公司的形态。 — Alejandro Maza Ayala > [25:18](https://www.youtube.com/watch?v=n34CIw3gk1k&t=1518s) 一支军队不可能让士兵在战场上各自凭空构想战术、为所欲为；处于转型期的企业需要的是绝对清晰的顶层战略。 — Alejandro Maza Ayala > [31:16](https://www.youtube.com/watch?v=n34CIw3gk1k&t=1876s) 我要把公司过去 40 年积累的一切全部推倒重建，押上全部身家重构成一家 AI 原生公司。 — Alejandro Maza Ayala > [34:45](https://www.youtube.com/watch?v=n34CIw3gk1k&t=2085s) 我们正生活在整个人类历史上最令人振奋的时代。 — Alejandro Maza Ayala ## Mentioned Kavak, Alejandro Maza Ayala, Angela Strange, Gabriel Vasquez, Carlos Garcia Ottati, a16z, OPI Analytics, Carfax, ChatGPT, Claude, Claude 3.5 Sonnet, Claude Opus 4.5, El Mike, 绝地学院 (Jedi Academy), 约瑟夫·熊彼特 (Joseph Schumpeter), 托马斯·爱迪生 (Thomas Edison), 福特汽车 (Ford) ## Video description Angela Strange and Gabriel Vasquez are joined by Alejandro Maza Ayala, Chief Product & AI Officer at Kavak, to unpack how the Latin American used-car marketplace rebuilt itself around AI agents, with 96% of customer interactions and 95% of transactions now handled by agents. Alejandro explains why Kavak decided that simply giving employees AI tools wasn't enough, and instead redesigned the company's systems, teams, and customer experience around agents. They discuss why Kavak spends as much engineering effort on evals as it does building agents, how its AI sellers outperform its human teams, and an experiment where an AI "CEO" increased profits in one city by 50% in its first month. The conversation also explores what happens to organizational structure when agents do most of the work, why Kavak trains everyone from executives to mechanics to build with AI, and Alejandro's argument that companies looking for incremental AI adoption may be missing the larger opportunity: redesigning the organization itself. Timestamps: 00:00 - Intro 01:03 - Machine Learning Before Transformers 02:23 - What Kavak Does & the Agent-Per-Customer Architecture 04:59 - Three Bets: Redesign the Company, Build Superhuman Agents, Change the Metrics 10:49 - Agents That Sell: 2.1x Better Conversion Than Humans 14:23 - Car Loans Approved in Three Minutes 16:13 - The AI CEO Experiment: 1.5x Profits in Six Weeks 20:13 - The Jedi Academy: Training Mechanics to Ship Agents 28:44 - Destroying Two Years of Work: From Multi-Agent Graphs to One Agent Per Customer 32:52 - Creative Destruction & Ford's Factory: Why Adoption Isn't Enough 34:45 - Advice for Founders: The Most Exciting Time in Human History Resources: Follow Alejandro Maza Ayala on X: https://x.com/alehandromz Follow Angela Strange on X: https://x.com/astrange Follow Gabriel Vasquez on X: https://x.com/GEVS94 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=n34CIw3gk1k)

[0:00] Alejandro Maza Ayala: 如果你在 AI 浪潮中还没有体会过恐惧，那说明你根本没有真正尝试过 AI。 [2:23] Alejandro Maza Ayala: 如果我们身处 2035 年，手握 GPT-10 级别的超强智能，我们会如何重新构建 Kavak？ [7:35] Alejandro Maza Ayala: 智能体就像一辆性能超凡的跑车，你只有配上了足够强大灵敏的刹车系统，才敢把油门踩到底。 [16:13] Alejandro Maza Ayala: 菲尔兹奖级别的天才下场进行极致微观管理。 [18:16] Alejandro Maza Ayala: 就像那只藏在厨师帽子里指挥做饭的老鼠一样，El Mike 就像一个副驾。 [21:02] Alejandro Maza Ayala: 没有任何回头路可走，这就是 Kavak 未来的走向，这就是整个公司的形态。 [25:18] Alejandro Maza Ayala: 一支军队不可能让士兵在战场上各自凭空构想战术、为所欲为；处于转型期的企业需要的是绝对清晰的顶层战略。 [31:16] Alejandro Maza Ayala: 我要把公司过去 40 年积累的一切全部推倒重建，押上全部身家重构成一家 AI 原生公司。 [34:45] Alejandro Maza Ayala: 我们正生活在整个人类历史上最令人振奋的时代。

[4:59] Alejandro Maza Ayala: giving staff ChatGPT or Claude without altering underlying systems yields zero operational efficiency. [8:04] Alejandro Maza Ayala: drivers can only hit the gas pedal if they trust their brakes. [34:45] Alejandro Maza Ayala: the most exciting time in human history [34:45] Alejandro Maza Ayala: go for it deep

[0:00] I'm investing more [0:02] in tokens than in knowledge workers. We [0:04] could build superhuman agents. This [0:07] means that by every dimension that [0:09] matters, our [music] agents would [0:10] outperform the best human if we had ever [0:12] hired. [0:13] >> ambitious companies listening to this [0:14] will decide to follow suit, which is you [0:16] decided to build an agent per customer. [0:20] >> Yes. Every day between 100 and 200,000 [0:23] [music] [0:24] agents get instantiated specifically for [0:27] this customer with its own virtual [0:28] machine. [0:29] >> There's a lot of people worried about [0:30] how the organizations of the future are [0:33] going to look like and the role that [0:35] humans are going to play. [0:36] >> If you haven't faced fear before, you [0:37] haven't felt it, then you haven't tried [0:40] AI. We launched a program [music] [0:42] inside Kavak that's called the Jedi [0:44] Academy. From the CEO to like AI [0:46] engineers to mechanics, we train [0:48] everyone and after [music] 6 weeks, they [0:50] launch state-of-the-art agents to [0:53] production. [0:53] >> What advice do you have to future [0:55] founders or first-time founders that [0:57] might be listening? [0:57] >> What works right now is [1:03] >> Welcome back to the ACC podcast. Today [1:05] we have Ali Massa, the head of AI at [1:06] Kavak. We're going to discuss today the [1:09] transformation that Ali led within Kavak [1:11] to turn it into an AI-native company. [1:13] Thank you, Ali, for being with us today. [1:15] >> Thanks for having me. [1:16] >> Before starting at Kavak, [1:18] you were running a company called Opy [1:20] Analytics. [1:21] >> That's right. [1:22] >> And you were very much into AI before [1:25] ChatGPT. You want to tell us a little [1:27] bit about that journey? [1:28] >> Yes, yes, of course. Well, we called it [1:30] machine learning back then. It was a [1:31] different family of of algorithms and [1:33] and we founded a company [1:35] with this very like [1:37] like ambitious vision there that that [1:40] new machine learning models would be so [1:42] powerful that they could solve any [1:45] complex problem. This was [1:46] pre-transformers, right? This was like [1:48] 2013. [1:50] So, we started building the company that [1:51] way and [1:53] I think we were like [1:54] 10 years ahead of time. But, we built a [1:57] great company. We served like 14 500 [1:59] companies [2:00] uh around like risk algorithms, [2:02] logistics, forecasting, marketing, um [2:06] but like really the the the power of [2:11] what transformers and then like the [2:13] ChatGPT moment [2:15] uh when it arrived make things like very [2:17] clearly that that we could now build a [2:21] whole new [2:23] uh [2:24] company and and [2:25] and way of of of building companies. [2:29] Um and we joined Kavak through and [2:32] Carlos to build it. [2:33] >> Amazing. All right, so we're going to [2:34] spend the bulk of this podcast talking [2:36] about exactly how you've identified [2:39] Kavak, but maybe just to start, what [2:41] does Kavak do and what is your role [2:42] there? [2:43] >> Kavak uh started out as a used case as a [2:46] used car marketplace. So, we buy cars, [2:49] we refurbish them, and then we sell them [2:52] and finance them. But to do that, we [2:54] also had to build a fintech and a [2:56] logistics company and a Carfax and like [3:02] basically all the infrastructure for [3:04] this to work didn't exist in in LatAm. [3:06] So, we had to build everything [3:08] vertically so we could so serve our [3:09] customers the right way. [3:11] >> I want to sort of start with the framing [3:12] of what the architecture looks like. So, [3:14] a consumer comes in and says, "I want to [3:16] sell my car." Like how many agents do [3:18] they touch? Like what's the harness look [3:20] like? Like ground us in how you design [3:22] this. [3:22] >> Right. So, so so we bet the company in [3:25] transforming [3:26] to a company run by agents. The [3:28] questions we ask ourselves is, "How [3:30] would we build Kavak in 2035 with Fable [3:33] 10 or or or GPT 10 level intelligence?" [3:38] And actually that company looks very [3:39] different that than what we had built or [3:41] what we had back then. [3:44] So, when a customer comes in right now, [3:46] um agent will get spawned specifically [3:50] for this customer with its own virtual [3:51] machine. It'll remember years of [3:55] interaction of this customers with with [3:56] Cabal, what they visited in the web page [3:59] or call they had 2 years ago. [4:01] Remember everything like in its memory, [4:04] come up with a strategy and set a [4:07] long-term goal to maximize the lifetime [4:09] value of this customer and do whatever [4:12] it takes to make the customer happy and [4:14] convert them into like all our different [4:18] products [4:19] like across time. And [4:22] this is a completely new and [4:24] groundbreaking architecture at scale, I [4:26] think, because [4:27] like people are still building [4:29] multi-agent system with with experts and [4:31] and we realized too bad [4:33] that long-running agents with hard [4:36] goals, not just workflows, [4:38] could could maximize our our customers [4:41] satisfaction and obviously their their [4:43] lifetime value. [4:44] >> Awesome. Okay, so we're going to jump to [4:46] the nuances that but maybe versus many [4:49] companies that say, "Hey, we want to be [4:51] agentic." And they try some workflows. [4:54] >> Yes. [4:54] >> You guys took the just rip like we had [4:57] to make this work. You had to downsize [4:59] dramatically. It didn't work for a year. [5:02] >> Right. [5:03] >> So, do you want to talk through [5:04] obviously you had to [5:05] tune a lot of things to make that work. [5:07] Like describe the harness at that time [5:09] and like what models you were using and [5:11] sort of specifically. Yeah. [5:12] >> So so there there were like three main [5:15] decisions that that we had to make. The [5:17] first and this is what where I think [5:19] many companies are stuck right now is [5:21] the first instinct is, "Okay, let's [5:24] adopt AI." And you you basically leave [5:26] your structure as it is and just give [5:28] ChatGPT or Claude to to your team and [5:31] then [5:32] there's no efficiencies. Your customers [5:34] have the same problems and and nothing [5:35] happens, right? And so so you need to [5:38] redesign your whole company [5:40] around the agents and around the future [5:42] capabilities. And this means really like [5:45] rebuilding most of your APIs, rebuilding [5:48] your system so the [5:49] agents can use them to to perform. [5:53] Then, you need to start generating the [5:55] data and the feedback loops to fine-tune [5:57] these agents. The only way to really [5:58] make them work is if you teach them. [6:01] Like, how do you teach them? You you put [6:02] them out in the open, you you put them [6:05] in front of customers, you get that [6:06] data, you get those evals, and then you [6:08] train your your [6:11] your agents. And this is the second bet [6:12] that we made that that we could build [6:16] superhuman agents. This means that by [6:19] every dimension that matters, like [6:21] conversion, [6:22] lifetime value, uh customer experience, [6:24] our agents would outperform the best [6:27] human we had we had ever hired. And we [6:29] put them in front of the hardest [6:30] problems. [6:31] And and finally, you you start to change [6:35] how you measure the success of the [6:36] company. Cazoo was uh transactional [6:39] company. We used to measure how many [6:41] cars we bought, how many cars we sold, [6:43] how many brakes we we needed to to break [6:45] pads we needed to buy. [6:47] Uh and we moved to a relational company [6:50] where now I have 10 million customers in [6:52] my database, and I have agents assigned [6:55] to most of them [6:56] with the task of maximizing their [6:58] lifetime value. Now, [7:00] we're selling cars and and and personal [7:03] loans and very high-ticket items. So, [7:05] just activating 1% of this customer [7:09] base, it's like hundreds of millions of [7:12] of dollars uh if we do it the right way. [7:15] So, so it's a bet that made sense [7:17] for us because of our industry, because [7:19] of the ticket, and because at the end of [7:21] the day, [7:23] customers need to build trust with with [7:25] a company because they're buying a used [7:27] car. [7:28] And the way to build trust is to to know [7:30] them and to and to plan and and nurture [7:33] a long-term relationship. [7:35] >> I like I just wanted to double-click on [7:36] something, you know, evals over agent [7:39] demos. Um you probably get pitched a lot [7:42] of, you know agents and you know [7:44] it's never been easier to build things [7:46] like before but um one of the questions [7:48] is like how do you how do you guys go [7:50] about evaluating this because not [7:51] everybody tests them across 90% of the [7:54] customer interactions to see if they're [7:55] really working and you know you guys I [7:58] believe it's a about 98% of the [8:00] interactions or something like [8:02] >> Yes. [8:02] >> that are now handled by agents. [8:04] >> Yes, totally. So so so to give you a [8:05] sense of the scale [8:07] um 90 like 96% of all interactions uh [8:12] are handled by agents. [8:14] So so no humans there. [8:16] Um like 95% of all transactions are [8:19] completely handled by by agents. [8:21] Obviously you meet a human when you pick [8:23] up your car like there's someone [8:24] physically there to give you the keys [8:26] but the rest of the of the of the [8:28] experience of the journey is handled by [8:30] by an agent. [8:31] Every day [8:33] between 100,000 [8:36] agents get instantiated in a day. They [8:38] wake up. They work sometimes for 3 [8:41] minutes, sometimes for 8 hours, [8:43] sometimes for 3 days and they like set [8:46] an alarm clock for their next task and [8:48] they go back to sleep. So the scale of [8:50] this [8:51] is just is just amazing and and it's [8:53] working. Now, how do you get this to [8:55] work at scale? [8:58] And the answer you mentioned it is is [9:00] Evals. Like [9:03] I like to move extremely fast [9:05] but in order to move fast you need to [9:07] have brakes, right? Imagine a car [9:09] uh [9:10] you'll hit on the gas just if you have [9:12] the right brakes. [9:14] And AI is super powerful and I've seen [9:16] many companies get this wrong because [9:17] they try to go slow [9:20] because they they don't have the right [9:21] brakes. So so I thought about it the [9:23] other way around like how fast can we [9:25] go? [9:26] Well, it depends on the quality of our [9:28] Evals. So a good rule of thumb here is [9:31] we spend [9:32] about the same amount of time [9:36] engineer time [9:38] tokens and and and money on building the [9:41] evals, the building the agents. And this [9:43] is how you get better and better and [9:45] better. Not not letting evals as an [9:47] afterthought. So, what do we measure? [9:50] First and foremost, like the the the [9:52] resource for the business. Like, if my [9:54] customer is happy, [9:55] they'll buy a car, they'll they'll get [9:57] their loan approved, uh they'll sell a [10:00] car to us. And and that's the like first [10:03] check. Like, did it convert? [10:05] And that's where most things break. [10:07] Like, I I see companies [10:10] like measuring number of calls or [10:13] minutes during the call or or some like [10:16] superficial KPIs that give you some [10:18] information, but that doesn't really [10:20] work. Like, the important thing is [10:22] did this customer convert? Is it [10:24] bringing value to the customer? And is [10:26] the customer happy to reengage with us [10:29] after a while? And once you get those [10:31] evals connected, then it's just [10:33] optimizing the right agent and the tech [10:35] architecture and giving the agent skills [10:38] to to scale this and and cater to [10:40] millions of customers. [10:42] >> It's really really amazing. And you [10:44] know, related to this is like, okay, so [10:46] you create the right evals, you know [10:47] it's working. [10:49] You know, some people some companies [10:51] still feel a little bit risk averse in [10:53] putting them in front of of of the [10:55] customers and being able to perform [10:58] >> Mhm. [10:59] >> the highest leverage task, which in your [11:01] case would be selling. Do your agents [11:03] really sell to customers? [11:04] >> Yes. So, so we never built customer [11:07] support or customer service agents. We [11:09] we we built like sales agents. It's [11:12] extremely hard to sell a car in in Latin [11:15] America. So, [11:17] imagine someone wanting to buy a car, [11:19] they can choose like among like 20,000 [11:21] SKUs. Then they need to pick like [11:24] financing and go through the financing [11:26] process, insurance, and and coverage. [11:30] And then they're probably trading in [11:32] their car. So, so we need to quote that [11:34] car. So, [11:35] it's a process that if someone does it [11:37] or or the way Kavak did it back in in in [11:39] 2020 2021, was you need to be extremely [11:42] good at 15 different things and have 15 [11:46] different experts in 15 different teams. [11:48] And usually the person would go and [11:50] speak with the expert in financing, the [11:52] expert in car car advisory, the expert [11:55] in buying, the expert in insurance, and [11:57] they'll build a package and buy a car. [12:01] That's extremely hard to do. [12:03] But, like [12:05] the first thing we did was, okay, can we [12:07] get an agent to be better than the [12:09] expert in each of these things? And then [12:13] put it together and have like a mega [12:16] expert that's an expert in insurance, [12:18] financing, etc. And that's who we put in [12:21] front of the customer. So, the [12:22] experience for the customer is amazing. [12:24] We tripled uh NPS and customer [12:26] satisfaction scored by putting the the [12:29] agent in front of the of the customer. [12:32] And it first it converted like 50% more [12:36] than our human team, and now it's [12:37] converting over that like [12:40] 2.1 [12:42] uh x [12:43] more. So, it's a completely different [12:45] company. [12:45] >> agents are better sellers. [12:47] >> Totally better. And and [12:49] and you get this, right? Because they're [12:50] experts and they're infinitely patient [12:52] and they know all your history and they [12:56] they they can plan for the long term and [12:58] they never get tired. So, [13:00] and if they make a mistake, they learn [13:02] it and the next day, not just them, but [13:05] the other 200,000 agents will have [13:08] learned from that mistake. So, that's [13:09] the feedback loop that we engaged and [13:11] and that's showing in the growth and [13:14] results and satisfaction of our [13:15] customers. [13:16] >> Yeah. [13:17] One of the well, one of two of the very [13:19] cool things I think about Kavak is I [13:20] think I think the world has gotten [13:22] comfortable with AI can do customer [13:24] service. It's still very hard to do [13:26] well. But, you know, as Gabe said, [13:28] there's still a view that well, [13:30] customers aren't going to want to buy [13:31] expensive things from AI. And you are [13:34] proving them wrong. [13:35] >> Yes. [13:35] >> The next layer on that is well, you're [13:37] not actually going to be able to do [13:39] regulated financial services end-to-end [13:42] with AI. [13:43] >> But if you walk through what you're [13:44] doing, you are underwriting a thin or no [13:46] file customer. [13:48] >> Yes. [13:48] >> Pricing them correctly. [13:49] >> Yes. [13:50] >> Doing servicing. So so maybe talk [13:52] through how did you how did you write [13:54] the evals to get comfortable with that? [13:57] And then versus I don't know, going to a [13:59] bank branch or or even a fintech. Sort [14:02] of how how is that experience? [14:03] >> Yes. [14:04] >> That much better. [14:05] >> the first financial product that that we [14:07] launched was a car loan. And usually in [14:09] in Mexico and in in some emerging [14:11] markets, it'll get like [14:13] two months or or or or more to get a car [14:17] loan approved. [14:18] >> Mhm. [14:18] >> Um we usually approve it in under three [14:21] minutes. [14:22] Uh [14:23] which is like pretty cool because we [14:26] have all this data around the customer [14:28] and the car. [14:30] And if the customer can't pay for the [14:32] car anymore, [14:33] they'll just return it to us and we can [14:35] give them a cheaper car. And and then [14:36] they're pay an uh uh [14:39] a smaller amount each month and they [14:40] like get out of the water, which is [14:42] amazing about the the vertical [14:43] integration of of the business. [14:45] But then like [14:47] when we started launching other [14:48] financial products, we we realized that [14:52] this is a very important decision for [14:53] the customer, right? Like like they [14:55] usually take three to four months to [14:57] make their make up their mind and buying [15:00] a car and and and and getting a loan or [15:04] getting a personal loan, like a large [15:05] personal loan that that we also [15:08] um do. So [15:11] if you get to know your customer [15:12] throughout this process and make the the [15:14] the process easy for them, then just [15:17] your conversion and retention metrics [15:20] start going through the roof. It's not [15:22] just the transaction, it's understanding [15:25] each customer personally [15:26] and get them to to convert when they're [15:30] ready [15:31] with a very deep personalization of the [15:33] interest rate, the [15:35] the risk, the max amount of the loan [15:38] in a way that makes sense for the [15:40] portfolio as a whole, obviously, but [15:42] that's optimized to the risk level and [15:45] and [15:46] probably the other offers that the [15:48] customer is getting. [15:50] >> And then maybe give us just to be [15:52] you know, Evals are always a very hot [15:53] topic. You kind of led with that. What [15:55] is like what is an example of maybe a a [15:58] hard to design area for Evals or one [16:00] where you had to spend extra amount of [16:02] time? What's just given the fact that [16:03] like there's real money PII at risk. [16:06] Yeah. [16:07] >> So [16:08] when we decided to to to to redesign the [16:11] company around the AI [16:13] you ask [16:14] the question, okay, is is AI going to be [16:16] able [16:18] to do this job, like even the CEO job or [16:21] or jobs where the leadership is? [16:23] And the answer honestly is probably yes, [16:25] like in 2035 with a rate of improvement, [16:28] it will be able to do so. We said, okay, [16:29] let's let's try it now. [16:31] Let's try and build an AI CEO. [16:35] So [16:36] we carved out a city in Mexico, it's [16:38] it's Cuernavaca, and we put like an [16:41] agent in one of our harnesses [16:44] as a CEO and it starts learning and it [16:47] starts making decisions and evaluating [16:49] on those decisions. And it's only been [16:51] running for for 6 weeks now. [16:53] The goal of the first month was to [16:55] double the the the profits of of [16:58] Cuernavaca. It didn't reach it, but it [17:00] was 1.5x, [17:02] like 50% more profits. Just like [17:05] managing the the city, which is it's [17:09] crazy, right? It's it's amazing and and [17:11] it's it's a CEO, like people were [17:14] like that was the last job AI was [17:16] supposed to stick, and no, it isn't, [17:17] really. And [17:19] how did this happen? And and [17:22] it's like [17:23] uh very smart person, like like Fields [17:26] Medal level smart, like going into every [17:29] single number, every single customer, [17:31] making the perfect forecast, and going [17:34] to micromanage every single things that [17:37] needs to be executed every day [17:40] to reach a plan. So, he'll literally [17:42] send messages to all the physical [17:44] workers in in Cuernavaca with their [17:46] plans for the day and ask them to send [17:48] voice notes back [17:50] to to know their their progress. [17:53] So, customer satisfaction grew, [17:56] uh we got a better inventory, we rotated [17:58] better, better financing penetration, [18:00] like every KPI started to to improve. [18:04] Um so, it's super cool, it's super [18:05] exciting. Now, [18:09] what are the jobs where where we think [18:12] uh [18:13] we're still like training and hiring [18:14] humans. [18:16] Those are related to the physical world. [18:18] So, when we talk about mechanics, Kavak [18:20] has around I think in Mexico around 800 [18:23] uh mechanics. There's lots of dexterity [18:26] and and senses that's super hard to [18:29] substitute. So, there we also build this [18:31] agents with exact same harness that's [18:33] scaling. [18:34] And the the mechanics have the psychic. [18:38] Um I was telling you guys earlier, it's [18:39] like [18:40] the the movie Ratatouille, like the the [18:44] mouse that's actually a chef [18:46] collaborating with a with a human. [18:49] It's kind of like that. So, it's a [18:50] psychic, we call it El Mike, and it [18:53] tells them how to inspect a car and [18:55] gives them tips and and shows them the [18:58] way to to do it, and [19:01] the quality of inspections, again, went [19:03] through the roof. We're inspecting [19:05] faster, we're repairing faster, it's [19:07] cheaper, but most importantly, we're [19:09] delivering higher quality cars. [19:11] Um [19:13] warranties came down around like 20 26% [19:17] since we since we launched and customer [19:20] satisfaction again [19:21] went up. So, [19:24] it's about this. Like how would you [19:25] design your organization from scratch [19:29] with with [19:31] with abundant super intelligence that's [19:34] that's cheap and just go build it. [19:37] >> Now, like this this is a good segue to [19:39] key topic right now in Silicon Valley [19:41] where you know, there's a lot of people [19:43] worried about how the organizations of [19:45] the future are going to look like and [19:48] and the the role that humans are going [19:50] to play. [19:51] >> Yes. And I think you touched a little [19:52] bit on that. So, we'd love to hear yeah, [19:55] like how you guys are thinking about [19:56] that and yes. [19:58] >> And [19:59] the organizations. [20:00] >> Yeah. [20:01] Totally. So, um [20:03] we we took that question very seriously [20:05] 3 years ago and the truth is that [20:08] everyone's [20:09] job will change. [20:11] So, [20:14] and [20:15] what we were doing a couple of years ago [20:17] will probably be be be performed better [20:20] by an AI agent. Right? So, what does [20:22] this mean? [20:24] We need to train everyone. So, so we [20:27] launched a program inside Kavak that's [20:29] called the Jedi Academy where anyone [20:32] from Kavak like from the CEO [20:34] >> We love this. [20:35] >> to Yeah, and it tells them like from the [20:37] CEO to like AI engineers to mechanics [20:41] like go into the academy. It's super [20:44] hard. Like I I I've [20:46] I led [20:47] I can you [20:48] >> You got that myself. [20:49] >> You designed the program. [20:50] >> I designed the program. [20:51] >> And but constantly [20:52] >> Constantly because you [20:53] you need to be upgrading the the program [20:55] because everything's changing so fast. [20:57] And there there's [20:58] like you can't send these people like [21:00] outside to Stanford to to learn this [21:02] because like it's new stuff, right? So, [21:06] we train everyone and after 6 weeks [21:08] they launch state-of-the-art [21:11] um agents AI agents to production. [21:16] And it's [21:17] mechanics and and finance guys and [21:19] engineers, like everyone can do it. [21:22] And what this generated is maybe this [21:24] person won't become an AI engineer, some [21:27] of them have, [21:29] but they they know how to collaborate [21:31] with this new technology, right? So, the [21:33] way we looked about it was guys, there's [21:35] no way back. Like this is the way Kavak [21:37] is going. This is the way the company [21:39] will look like. These are the changes [21:41] for the engineering team, the finance [21:44] team, the product team. Like this is [21:47] what's going to change. [21:49] You have the choice to like train and [21:52] and get the skills to perform in this [21:54] new reality in this new world, um or [21:58] maybe leave Kavak if this is not for [22:00] you, but this is the way we're going. [22:02] And [22:03] like we're great. Like like we we we [22:04] strengthened the culture, and we were [22:06] super excited. Um [22:09] people really know how to build these [22:11] agentic systems. And then, if you look [22:13] at Kavak now, any process is really a [22:16] collaboration of agents and humans, and [22:19] sometimes like agents are the bosses or [22:21] of of humans, and sometimes humans are [22:24] designing the agents, but [22:26] like we we we managed to really build [22:28] this and and change this. [22:30] Um [22:31] And it's through this [22:34] idea that we need to be learning every [22:36] day, and things will continue to change. [22:39] And the only way to to continue being [22:41] relevant is to [22:43] upgrade your skills uh every month or [22:45] every couple months. [22:46] >> But you do have or did have, you know, [22:49] thousands of people. Now agents do most [22:53] things. [22:53] >> Yes. [22:53] >> So, like what is the org structure of [22:56] Kavak? Like does the middle management [22:58] concept even exist anymore? Like what [22:59] does your org look like? [23:00] >> Right. So, [23:02] the way it looks like now is very flat [23:04] teams, very senior teams, super [23:05] empowered. [23:07] If you look at a team, you'll have [23:09] engineering, AI, like operations, like [23:13] everything. And they're either [23:15] building the agents, [23:17] working for the agents, [23:19] or being in the physical world in front [23:21] of the customer. Like most of our [23:23] organization looks like that. So, so [23:25] it's really built [23:27] around [23:28] um [23:29] a around the idea of how organizations [23:31] will look like in the future and the [23:32] round AI and really harnessing this this [23:35] new technology. [23:36] Obviously, this required lots of [23:38] retraining because in 2023 [23:40] or 2022, no one was building agents, no [23:43] one was helping agents or or taking [23:45] orders from from agents. And the way you [23:48] cater to the physical world or the [23:49] customers was [23:51] in a different way than if an agent's [23:53] telling you what to do or helping you [23:55] make your job uh better. [23:56] >> Yeah. [23:57] >> Um so, it's a completely different [23:59] structure than than we had just 2 years [24:01] ago. [24:01] >> Yeah. Explain um we talked about this [24:03] before what working for the agents look [24:05] like. Like I think the way you described [24:06] it was the agentic system and then [24:08] sometimes when it fails, it's like, "Oh, [24:09] let's kick out to kind of a human [24:11] queue." But then that's loss. And so, [24:13] how have you brought that together? [24:14] >> So, so [24:15] like the the way we see human in the [24:17] loops and and and most of these agentic [24:19] systems in production right now, like [24:20] large-scale agentic systems, usually if [24:23] an agent hits a wall or can't perform [24:25] anymore, [24:26] it'll [24:27] like send uh [24:29] this case or this customer to a tier two [24:32] support and forget about it. [24:34] That doesn't really work because you [24:36] don't close the loops, so you don't [24:38] generate the data to train the agent to [24:40] do this better. [24:41] What works right now is we have an agent [24:43] that's obsessed with each of the [24:45] customers, like millions of this. [24:47] They have access to every single API, [24:50] every single skill, like and we have [24:51] agents building those humans building [24:53] those skills for them. [24:55] And then if an agent hits a wall or [24:57] cancels something, it'll call this API [25:01] saying I need help. [25:02] And on the other side, it's not an agent [25:05] or software, it's a human [25:07] helping them out. But if you map this [25:09] out in an org chart, it's really human [25:12] teams that have an agent I'm getting [25:14] better results. It's super clear, like [25:17] it makes sense. [25:18] >> That's actually a perfect segue. And I [25:19] know you get lots of of leaders at [25:21] larger institutions inbounding to you. [25:24] So so maybe this will save you many [25:25] phone calls. But I think [25:26] rationally many leaders of companies [25:30] intuitively understand this. It is very [25:32] hard still to deploy AI through the [25:36] organization. Like the models are good [25:38] enough. You know that. It's a it's an [25:39] org problem. It's a psychology problem. [25:41] Like what advice do you have or what [25:42] have you seen? [25:44] >> I think it's two things. [25:45] Uh the first is [25:47] it has to be top-down [25:49] because of this. Like [25:50] if you just get adoption, it won't go [25:53] anywhere because it's it's hard to [25:54] generate this taste or a strategy for [25:58] people to to bottom-up decide what to [26:00] build or what not and and come up with [26:02] something that works for the company. So [26:04] the transformation has to be top-down [26:06] and leaders need to adopt and leaders [26:07] have to have a very clear [26:10] plan on what to build. I've seen so many [26:13] companies that's just like, oh, like [26:15] we're doing a hackathon. People are [26:17] coming up with use cases. We're [26:19] sponsoring some of these use cases. That [26:20] doesn't work. It's [26:22] like [26:23] be very clear on what the company will [26:26] look like in three or five years. [26:29] And then start building that. And be [26:30] like very vertical in in guiding your [26:33] troops towards that. Like an army [26:35] doesn't really work if everyone comes up [26:37] with ideas on the strategy and tactics [26:39] and goes to the battlefield and [26:41] like does whatever they want. Like you [26:42] need a very clear strategy. [26:44] And that's what we need now. It's a like [26:46] transformation stage. [26:48] The second one is you need to measure [26:50] what what really matters. And it's [26:52] evals, but it's also the right evals. [26:54] So, [26:55] I see a lot of companies spending now [26:57] huge amounts and they say, "Okay, I got [27:00] adoption. I'm just spending like [27:02] hundreds of millions of dollars in in [27:04] tokens now. What about that? Like, [27:06] there's quality in the tokens." So, I [27:08] have a framework here that's also [27:10] useful. Like, level tier three tokens, [27:12] the most valuable, [27:14] are these agents where you can [27:18] get the ROI of each specific token. And [27:21] I can do that now. That's great news for [27:23] me. [27:23] Because I'm growing and because I know [27:26] the ROI of each token, because it goes [27:28] to agents that are [27:30] performing the job of the organization, [27:32] right? These are the best tokens. [27:34] Tier two tokens are [27:36] things that you can measure indirectly. [27:38] Do I see devs in the in the code base [27:41] and I can evaluate the value of these [27:42] tokens at least indirectly and then push [27:45] those to production. [27:46] Tier one, when most companies are, is [27:49] people are just just saying, "Click code [27:51] or ChatGPT or code work or whatever. [27:53] What happened with those?" I have no [27:55] idea. [27:56] So, it's not just about adoption. It's [27:58] really about having a very clear vision [28:01] and then measuring that each token you [28:03] spend is bringing you those benefits and [28:05] just iterate, iterate, iterate from from [28:08] there. [28:08] >> And then one of you we touched on this a [28:10] little bit, but I think it's worth a [28:11] dive as [28:13] maybe the the most ambitious companies [28:14] listening to this will decide to follow [28:16] suit, which is you decided to build an [28:19] agent per customer [28:21] >> Yes. [28:21] >> versus per task. And then discovered [28:23] along the way that each one of those [28:25] agents needs its own kind of micro [28:27] virtual machine. So, maybe kind of walk [28:30] us through [28:32] >> Yes. [28:32] >> those decisions that architecture. [28:33] >> And and [28:35] I think the the [28:36] we're seeing these results now, but it [28:38] was a really risky bet because [28:40] people usually go from workflows. [28:42] >> Mhm. [28:44] >> And like, if I could advise everyone, [28:46] don't build agent workflows to graphs or [28:49] or functions or or objectives. And we [28:52] built that that these are multi-agent [28:54] systems that can perform [28:56] a whole function for a complex goals. [28:58] Like the ones I told you that to sell a [29:00] car you need to do financing, [29:02] purchasing, like recommendations, etc. [29:05] And we had thousands, like tens of [29:07] thousands of these agents working at [29:09] scale running the business back in [29:10] December. [29:11] But then [29:13] Opus 4.5 came out. [29:15] And I realized like this isn't the right [29:17] paradigm anymore. [29:19] Like the the intelligence now doesn't [29:22] need like the graph and the multi-agent [29:25] lattice work and harness because it will [29:28] constrain [29:29] this level of intelligence. [29:31] So we decided to like [29:34] destroy everything we had been building [29:36] for for two years that was working, that [29:38] brought us to profitability, that [29:39] brought us amazing growth. [29:41] And start over with a harness that we [29:44] thought would be robust and scalable and [29:47] and [29:48] leverage [29:50] recursive self-improvement. And or new [29:52] models, more intelligent models coming [29:53] out every month. So the way this looks [29:56] like it's it's a virtual machine [29:58] with an agent with access to memory and [30:01] evals and a CLI where they can access [30:04] every tool and every API in my company [30:07] and the long-term goal. [30:09] And I instantiate hundreds of thousands [30:12] of these each day with long-term goals. [30:15] Like maximizing the the lifetime value. [30:17] >> Yeah. The self-improving organization. [30:19] >> Exactly. The self-improving [30:20] organization. And I think people are [30:22] super obsessed with with RSI now [30:25] and this will improve the the [30:28] the models, but if you look at it this [30:31] way, [30:32] economic value [30:34] in humanity for the past 4,000 years [30:38] has been delivered by organizations, not [30:41] by individuals. So, what you want [30:43] to self-improve and to engage in that [30:45] loop is the organization that can [30:47] deliver more economic value. Right? So, [30:49] that's the loop that I think companies [30:52] will start to to focus on because if you [30:54] get that loop working and it's an [30:56] organization that is really [30:58] self-improving [30:59] and harnessing the the newer models and [31:02] the better intelligence that we're [31:03] getting every couple of days now, [31:05] uh [31:06] then like you hit the exponential not [31:08] just in intelligence, [31:10] but in the value uh that you can [31:12] generate as a company. So, so that's [31:14] really exciting. That's what we're [31:15] working on. [31:16] >> You mentioned that because of all the [31:18] challenges on adopting AI, you saw the [31:21] biggest opportunity on net new companies [31:23] being formed [31:25] working on this new way and then [31:26] disrupting markets. Like you want to [31:28] talk a little bit about that? [31:29] >> Yes, um [31:31] there's this concept in economics about [31:33] creative destruction from from from [31:35] Joseph [31:36] Schumpeter. [31:38] And what it says is that the the the way [31:40] innovation hits the economy isn't by [31:43] companies [31:45] adopting the the the new technology, [31:47] but by companies remaining the way they [31:49] they were and incumbents with the new [31:52] technology [31:53] destroying [31:55] the old companies. So, this destroys [31:57] value in the short term in the economy, [31:59] but in the long term it's better for [32:01] everyone because of these new [32:03] more efficient, more effective uh [32:05] companies will provide better products [32:06] and services for the for the economy as [32:08] a whole. [32:09] And this has happened in the past like [32:11] industrial revolutions and this has [32:12] always happened. And [32:14] this is a great opportunity for [32:15] entrepreneurs and and people today [32:18] because [32:20] it's hard to adopt [32:22] AI deeply. [32:24] Uh it's really hard for a CEO today, [32:27] especially of a large company or public [32:28] company, to go and say, "Hey, like I'm [32:31] betting everything on AI. [32:33] Uh the company has to look this way. [32:36] I'll [32:36] like destroy and rebuild everything I've [32:39] been building for the past 40 years [32:42] to become an AI native company. Like how [32:45] many CEOs will will will do that in a [32:47] company at scale? So, while they they [32:50] adopt [32:52] new companies can be formed that are [32:54] built around the the the strengths of of [32:57] AI [32:58] and and take over and and and and bring [33:00] new products and services to to to the [33:02] masses. And and this has happened [33:05] before. Like this happened with [33:07] electricity. This is the story I always [33:08] tell tell my team. [33:10] The the the technologies for Ford's [33:13] production line were developed in 1879 [33:17] and 1881. Edison started commercializing [33:21] electricity in in New York and then [33:23] London and he invented a dynamo that was [33:26] extremely efficient. So, you could have [33:28] built [33:30] Ford's factory [33:32] 40 years before Ford. The technology was [33:35] there. Everything was there. But the way [33:37] people adopted electricity and and [33:39] Ford's dynamo was okay. I'm going to [33:41] leave my factory like four floors, [33:43] shafts, and belts. I just change my coal [33:47] engine for an electric engine. And this [33:49] will bring you benefits. Yes, but like [33:51] 6% efficiency. [33:53] What needed to be done was like to [33:55] destroy that factory, [33:57] build it in a flat surface, not in the [34:00] center of New York, but in Connecticut [34:02] or New Jersey. [34:04] And redesign your coal factory around [34:08] small dynamos and and electricity. And [34:10] then you get like the 3x uh improvement [34:14] in productivity that that like powered [34:17] the US during the 20th century. And the [34:20] same happened again with a computer. And [34:22] the same is happening again today. [34:23] People want to adopt it, but they're not [34:25] willing to redesign the whole company. [34:28] And they just adopt it superficially. [34:30] And in the end that'll give you a 6% or [34:33] a 10% improvement, not a 10x [34:36] improvement. And it's like the [34:38] innovator's dilemma [34:40] at an industrial [34:42] uh scale again. [34:45] >> I think you've just made an amazing case [34:47] for any future founders out there that [34:49] it's time to build. [34:51] And and maybe a great a great place to [34:53] to end this, you know, you've built and [34:55] scaled your own company, you've now [34:56] turned Kavak fully agentic. Like, what [34:59] advice do you have to future founders or [35:02] first-time founders that might be [35:03] listening? [35:04] >> So, this is the most exciting time [35:06] in human history. I I believe that. Like [35:08] like we're living in the most exciting [35:10] time in human history. And it's the most [35:13] exciting time to to be a founder because [35:16] it's the first time that anyone [35:19] has access [35:21] to the most powerful tools and [35:24] intelligence in the world. [35:27] Like for almost for free or for $20 a [35:29] month. So, so literally there [35:31] the the democratization of the tools for [35:33] people to build [35:35] has never been uh this way in human [35:37] history. And there's so much [35:40] problems to be solved and a new reality [35:42] to be built around this this new [35:45] paradigm. So, [35:47] say like just go for it, but go for it [35:49] deep. Like imagine [35:51] what the future around AI will look [35:53] like. [35:55] It's just a uh it's not even an [35:56] exponential. Just just map [35:59] a trend that's linear. If things keeps [36:01] getting like AI keeps getting better at [36:03] a linear scale, [36:04] and just build for that. And and you'll [36:07] come up with with wonderful ideas that [36:09] will like bring a lot of value to the [36:11] world. [36:13] >> Amazing. And I thank you for joining us. [36:15] >> Thank you. Thanks for having me. [36:30] >> Mhm.