Lassie Deploys Autonomous AI Agents to Eliminate Paperwork at Small Businesses

Lassie部署自主AI智能体，消除小微企业文书工作 AI驱动的劳动力自动化将通过直接取代人工行政任务，而非仅仅存储数字记录，来拓展小微企业的服务能力。 本次对话探讨了AI智能体相较于传统企业软件的根本性转变——从单纯的文档存储转向执行实际的人工作业。在直接深入真实医疗诊所现场后，Lassie打造了能够处理高达98%复杂行政任务的自主智能体，为不堪重负的诊所经营者消除了巨大的运营瓶颈。联合创始人详细介绍了其技术架构、消费级易用性的上手指引流程，以及将AI驱动的后台自动化从牙科诊所推广至全球各类小微企业的多步规划。他们还探讨了如何应对付款方与医疗服务提供方争议中零边际成本的转变、数字支付领域的监管利好，以及打造一个高敏捷度、面向2026年的原生组织。 传统企业软件仅仅是将纸质记录数字化，并没有执行企业运营所需的实际人工工作。 (0:00) 由于严重的招工短缺，小微企业主每月往往要在手工行政文书工作上耗费高达200小时。 (1:12) 直接深入客户诊所现场让Lassie创始人发现，行政阻碍是从业人员产生职业倦怠的主要原因。 (3:46) AI流程自动化依赖三个架构层：上下文（数据/历史）、工具（行动执行）与智能（AI推理）。 (5:39) 与充当数字存储的传统SaaS不同，AI将软件转变为能够做出决策并执行业务任务的数字劳动力。 (7:23)

Alex Rampell, Steijn Pelle, and Frédéric Renken argue that enterprise software is shifting from passive systems that merely store digital records to autonomous AI agents that directly execute administrative labor. In small businesses constrained by acute staffing shortages, automating operational workflows unlocks suppressed supply and allows practitioners to focus on their core services. Shifting Software from Storage to Autonomous Execution (7:23) Alex Rampell traces enterprise software back to early systems like Sabre, PeopleSoft, and QuickBooks, arguing that traditional software merely replaced physical paper filing cabinets with digital equivalents. Instead of eliminating labor, digitization required organizations to hire staff to manage software inputs and IT departments to maintain the infrastructure. AI labor fundamentally alters this dynamic by allowing software to independently edit records, make decisions, and complete operational tasks. To translate this shift into practice, Lassie decomposes administrative automation into context (historical patient and office data), tools (APIs and software interfaces), and intelligence (reasoning models). According to Frédéric Renken, early product versions relied on human-in-the-loop manual reasoning to handle complex edge cases until foundation models matured enough to take over operational decision-making. Administrative Bottlenecks in Main Street Healthcare (14:35) Small healthcare practices face severe labor bottlenecks rather than demand constraints. In the United States, approximately 160,000 dental practices spend an estimated $200,000 each year on administrative overhead. Because qualified office managers are increasingly difficult to recruit and retain, practice owners report spending up to 200 hours per month—often working until midnight—reconciling ledgers, checking bank accounts, and manually submitting claims. Pelle and Rampell contend that eliminating this administrative burden directly expands provider capacity. Rather than eroding profit margins through increased competition, reducing operational overhead allows doctors to double patient throughput in localized markets that are currently underserved due to labor shortages. Engineering 98% Autonomy for Regulated Billing (17:24) Unlike enterprise environments where administrative workers oversee software tools, small businesses lack excess personnel to monitor AI interfaces. Lassie therefore requires a 95% automation threshold before launching new features, ultimately targeting 98% autonomous execution to free owners from routine manual interventions. Automating tasks like ledger reconciliation saves practices between 10 and 20 hours per week. This level of autonomy is particularly impactful in claims processing, where Rampell notes the marginal cost of arguing drops to near zero. While human managers struggle to track 50 to 100 core dental codes alongside variable insurer guidelines, AI agents can immediately match procedure narratives with required X-rays. This counters payer friction strategies, such as claim denials and physical check issuances, by enforcing regulatory compliance without human operational drag. Defensibility and Distribution in Fragmented Markets (22:08) In the small business sector, the primary competitive incumbent is not legacy software but manual human labor or outsourced billing agencies. Defensibility requires solving what the founders call a technical 'schlep': building custom read-and-write integrations across fragmented, legacy systems and establishing a unified data ontology for claims and payments that bridges disparate industry definitions. Reaching overworked, non-technical owners presents a distinct distribution hurdle that traditional B2B sales tools cannot address. Lassie targets practices by analyzing signals like practice management software signatures and local job postings on platforms like Indeed. To avoid rapid churn, onboarding mirrors consumer applications: automated flows configure bank links and payer settings using basic tax metadata, bypassing the roughly 50 hours of manual setup normally required to transition practices from paper checks to electronic funds transfers. Lassie outlines a multi-stage roadmap starting with the $1 billion addressable US dental market before expanding into adjacent outpatient specialties and general small businesses. However, significant technical hurdles remain: Frédéric Renken notes that frontier models still lack innate knowledge of unindexed, real-world business workflows and payer standard operating procedures, requiring software providers to continuously train agents on proprietary ERP records and human operator feedback.

a16z合伙人与Lassie联合创始人提出，传统企业软件仅将纸质档案数字化而未替代人工，而新一代AI代理能直接执行中小企业的后台运营劳动，从而在劳动力短缺的实体商业中释放巨大的服务供给潜能。 软件演进：从数字档案柜转向自主劳动 (7:23) a16z合伙人Alex Rampell将企业软件的原点追溯至IBM与美国航空（American Airlines）合作开发的Sabre系统。他指出，从Sabre替代纸质索引卡，到PeopleSoft和Workday替代人事文件夹，再到QuickBooks替代纸质账本，半个世纪以来的传统SaaS本质上只是数字化的档案柜。这些工具不仅未能大幅缩减企业人员编制，反而增加了IT维护与网络安全部门的开支，人类员工仍需在屏幕前手动输入数据并推进流程。 AI劳动的出现打破了这一范式，使软件能够自主理解上下文、做出决策并执行诸如催收电话或账目核对等实际任务。Alex Rampell以餐饮软件Toast通过捆绑支付手续费（约2%抽成）拓展市场规模为例，认为AI直接提供劳动力，其商业变现将直接切入传统软件无法触及的人工预算领域。这一机制能够挽救因关键行政人员离职而被迫提前出售业务的实体雇主，从根本上解决由人力匮乏引发的市场失灵。 三层架构与高精度自动化解决诊所瓶颈 (17:24) 全美约有16万家牙科诊所，单家诊所平均每年需承担约20万美元的行政管理支出。Lassie联合创始人Steijn Pelle指出，即便是在Yelp上广受好评或毕业于哈佛的牙医，也常因招不到熟练的前台行政人员，被迫每月加班多达200小时手动处理保险报销与账单。由于中小企业缺乏冗余员工来监督软件运行，AI代理必须具备极高的自主性，并在准确性上近乎免检，才能让医生按时下班回归家庭生活。 联合创始人Frédéric Renken阐释了Lassie的技术演进框架，将其拆解为上下文（Context，历史病历与运营数据）、工具（Tools，软件API交互界面）以及智能（Intelligence，模型决策）。自2020年起步时，创始团队长期肉身驻扎在牙科诊所内充当“人在回路”的手动推理节点，以此排查长尾边缘情况并编写自动化代码。随着基础大模型推理能力的迭代，团队在保持数据底座不变的前提下替换智能层，确立了新产品上线必须达到95%以上劳动自动化、核心流程达到98%全自主运行的工程标准，每周代为核对约200份患者账册与保险户头。 政策红利与零边际成本的理赔博弈 (49:18) 目前约70%的美国小企业仍依赖纸质支票进行结算，而保险机构往往倾向于利用邮寄纸质支票或默认拒赔来拖延资金流出。近期出台的联邦监管法规成为关键转折点，该法案强制要求健康险与牙科保险机构在医疗机构提出申请时，必须提供电子资金转账（EFT）及标准化数字汇款格式。然而，单个诊所自主完成所有保险公司的系统设置通常需要耗费超过50小时，Lassie借此研发了仅需诊所税号（TIN）即可一键迁移至全数字支付的底层引擎。 Alex Rampell指出，AI技术的引入使得“商业争辩的边际成本归零”。在传统模式下，前台员工难以熟练记忆50至100种常见的核心诊疗账单代码和各家保险公司的严苛附带要求；而AI代理能够以近乎零的边际成本，自动匹配治疗叙述、提取所需X光片并按规定合规发起索赔与反驳。由于保险公司在雇主年度续约时面临维持优质医疗网络的压力，标准化的自主理赔代理能够极大提高赔付通过率并加速资金回笼。 消费级分发机制与小微企业自动化愿景 (33:35) 针对中小企业客户分散且缺乏技术背景的特性，Lassie摒弃了依赖高端商务宴请或LinkedIn与Apollo等传统B2B销售工具的渠道策略。团队通过抓取本地诊所在Indeed等招聘平台上发布的行政岗位等替代性意向信号，建立起精准的目标诊所数据库；同时借鉴Superhuman与Robinhood的消费级设计准则，构建了类似Stripe的高度自助化入驻流程，要求在48小时内让繁忙的业主直观看到核心价值，以防客户迅速流失。 Steijn Pelle公布了公司的三阶段总体规划：第一阶段聚焦全美16万家牙科诊所，填补其每年20万美元的人力缺口，撬动10亿美元年度经常性收入（ARR）的细分市场；第二阶段横向拓展至门诊胃肠科等流程相似但缺乏数字服务的医疗专科；第三阶段将经过训练的代理抽象化为通用商业组件，全面推广至水暖工、烘焙坊等各类全球小微商业。创始团队认为，当前本地实体服务主要受制于供给瓶颈而非需求不足，自动化后台杂务能够释放专业人员的生产力，推动服务容量成倍增长。 Lassie创始人坦言，公共互联网数据并未索引真实商业的标准作业程序（SOP）与隐性账单工作流，这构成了基础大模型落地的核心技术障碍。公司当前依赖ERP历史数据挖掘与前台员工交互反馈系统来训练代理，其长远商业构想建立在AI劳动力能显著扩充本地实体供给的假设之上，未来仍需检验通用化代理在跨行业复杂场景中的鲁棒性。

## Background 该内容由知名风险投资机构 a16z 于 2026 年 8 月 3 日发布在 YouTube 平台。视频中，合伙人 Alex Rampell 和 Olivia Moore 与 Lassie 联合创始人 Steijn Pelle 及 Frédéric Renken 展开对话，探讨如何将 AI 智能体引入牙科诊所等长期被忽视的实体小微经济体，通过自动化繁琐的账单和理赔流程，让医疗从业者从行政负担中解放出来并实现业务的自主运转。 ## TL;DR Lassie 创始团队分享了他们离开 Robinhood 和 Superhuman 后扎进牙医诊所一线打杂的经历，揭示了小企业主普遍深陷招工荒、每月被迫耗费数百小时处理报销文书的深层痛点。Alex Rampell 阐述了企业软件正在从单纯存储数据的“数字文件柜”跨越到直接承担劳动的“AI 代理”范式转变，Lassie 借此颠覆传统 SaaS 模式直接切入客户的人力薪资预算。通过构建读写双向集成和极简的免触入驻流程，系统实现了 98% 的端到端自主化运行，并在巨头完全缺位的小微医疗垂直蓝海中建立起非典型竞争壁垒。团队不仅锁定了全美 16 万家牙科诊所对应的高达十亿美元级 ARR 市场，更勾勒出未来将自主运行网络横向扩展至全球各行各业小微企业的长期蓝图。 ## Atmosphere & style 整场对话在硅谷前沿技术洞察与美国内陆腹地真实商业生态的鲜明反差中展开，兼具深刻的理论思辨与极接地气的行业纪实感。嘉宾与主持人在探讨 TiVo 困境、供需弹性等严肃商业命题之余，频繁穿插着“心力交瘁退休的牙医老友”、“亲手拆封十万美元纸质支票”等鲜活生动的细节与风趣调侃。交流避开了浮夸的 AI 概念炒作，代之以对一线痛点与底层脏活的敬畏，整体洋溢着一种通过技术解放人类琐务、让手艺人重拾核心业务热情的理性乐观主义。 ## Bulletin - [12:17](https://www.youtube.com/watch?v=Fpg_8aiBxx4&t=737s) 小微医疗机构深陷严重的行政人手荒：全美约有 16 万家牙科诊所，每家诊所平均每年在纯行政运营上的支出高达 20 万美元，单月文书处理常超 200 小时。 - [15:35](https://www.youtube.com/watch?v=Fpg_8aiBxx4&t=935s) Lassie 颠覆传统 SaaS 商业模式，直接切入客户损益表（P&L）中的人工薪酬预算，首个 AI Agent 仅接管每月约 30 小时的工作量即实现五位数美元年费变现。 - [17:29](https://www.youtube.com/watch?v=Fpg_8aiBxx4&t=1049s) 通用大模型缺乏具体的商业工作流 SOP，Lassie 通过肉身驻场接管业务、打通底层 ERP 历史记录逆向解析流程，将首个理赔 Agent 的自动化率推升至 98%。 - [53:18](https://www.youtube.com/watch?v=Fpg_8aiBxx4&t=3198s) 联邦政府强制推行直接存款（Direct Deposit）及电子明细单据法规，正强力推动仍有约 70% 依赖纸质支票结算的小微医疗机构完成底层数字化，为 AI 接管创造了政策顺风。 ## Key takeaways - [07:23](https://www.youtube.com/watch?v=Fpg_8aiBxx4&t=443s) 过去数十年软件的本质只是静态的“数字化文件柜”，虽然提升了检索效率但脏活累活仍需人工完成，生成式 AI 的核心突破是软件开始真正承担劳动做功。 - [08:44](https://www.youtube.com/watch?v=Fpg_8aiBxx4&t=524s) 市场空间的同心圆模型显示，最内层是万亿美元级的纯软件存储市场，中层是绑定交易分润的 Fintech 市场，最外层则是体量呈数量级爆发的人工薪酬预算市场。 - [12:17](https://www.youtube.com/watch?v=Fpg_8aiBxx4&t=737s) 小微企业面临的核心危机并非“AI 抢走工作”，而是现实中极度缺人且招不到合适文员，繁重的报销和账单文书严重摧毁了专业手艺人的核心生产力。 - [17:29](https://www.youtube.com/watch?v=Fpg_8aiBxx4&t=1049s) 中小企业（SMB）没有专门的 IT 或运营人员来学习复杂系统，软件若无法实现端到端高自治，就只会变成深夜把医生拉回电脑前的又一套繁琐工具。 - [17:29](https://www.youtube.com/watch?v=Fpg_8aiBxx4&t=1049s) Lassie 采取务实的自动化推进策略：单模块自动化率达到 95% 以上即行交付投产，依靠后续极少数复杂拒赔案件的人工介入反馈，驱动长尾自动化率攀升至 98%。 - [20:00](https://www.youtube.com/watch?v=Fpg_8aiBxx4&t=1200s) 为了具备极高毛利和快速交付能力，Lassie 坚决拒绝以廉价外包充当后台的“假 AI”重资产路线，坚持纯软件工程路径。 - [22:08](https://www.youtube.com/watch?v=Fpg_8aiBxx4&t=1328s) “TiVo 困境”揭示了缺乏底层分发管道的纯功能型软件极易被巨头挤压，初创企业必须优先死磕枯燥的底层数据与结算管道，牢牢握住客户关系。 - [24:27](https://www.youtube.com/watch?v=Fpg_8aiBxx4&t=1467s) AI 辅助编程大幅抹平了大公司平庸工程师的研发劣势，巨头抄袭浅层功能的速度加快，但小微垂直领域不存在类似 Workday 或微软的敏捷巨头，为 AI 原生代理留出了蓝海。 - [26:11](https://www.youtube.com/watch?v=Fpg_8aiBxx4&t=1571s) 在缺乏成熟软件覆盖的 Main Street 场景中，初创公司的真正对手不是传统 SaaS，而是离职的前台员工“Betty”或低效的外部账单代理公司。 - [29:13](https://www.youtube.com/watch?v=Fpg_8aiBxx4&t=1753s) 面向日程极度碎片化的小微企业主，必须打造如 Stripe 或 Robinhood 般的自服务免触入驻体验，在数分钟内打通银行、EHR 与保险门户。 - [33:35](https://www.youtube.com/watch?v=Fpg_8aiBxx4&t=2015s) Lassie 宏伟蓝图第一步是专注吃透全美 16 万家牙医市场，单垂直领域即可支撑超 10 亿美元 ARR 软件空间，进而抽象出“读写核心记录系统 + 客户预约支付”的底层共性拓展至全球各行各业。 - [45:24](https://www.youtube.com/watch?v=Fpg_8aiBxx4&t=2724s) 通用大模型并不知道真实世界的工作流 SOP，真实业务诀窍往往只存在于老练前台的脑海中，从 ERP 历史数据逆向推导业务规则构成了关键专有数据壁垒。 - [49:19](https://www.youtube.com/watch?v=Fpg_8aiBxx4&t=2959s) 牙科报销存在严苛明细规则，AI Agent 能够将患者 X 光影像与数十种计费代码精准对齐并递交，有效终结了保险公司利用人工失误制造摩擦与推诿拒赔的博弈。 - [55:59](https://www.youtube.com/watch?v=Fpg_8aiBxx4&t=3359s) 主街小微商户缺乏 LinkedIn 等传统 B2B 档案，Lassie 转向全新 GTM 打法，通过全美诊所全景图谱映射并捕捉 Indeed 等求职网站的前台招聘缺口意图信号实施精准触达。 ## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=Fpg_8aiBxx4&t=0s) 序幕：AI 的炒作断层与“做工作”的软件革命 在节目的片头剪辑中，嘉宾与主持人们抛出了一系列犀利观点。Alex Rampell 指出，AI 目前的现状是“在硅谷被过度炒作，但在像艾奥瓦州这样的美国内陆腹地却被严重低估”。传统的软件演进实际上只是把纸质档案迁移到绿屏计算机再到云端数据库，但背后的脏活累活依然要靠人力来干。Lassie 联合创始人 Steijn Pelle 回忆起他在 Yelp 上排名第一的牙医诊所看到的震撼场景：一位顶级医生每个月居然要花 200 个小时在琐碎的案头文书上。模型虽然经过海量数据训练，但一开始根本不知道该如何完成这些实际业务，Lassie 团队最初不得不充当系统里的“人在回路”（Human-in-the-loop），靠亲自动手帮诊所干活来把自己的痛点一步步自动化。 Alex Rampell 随后点出初创企业对抗巨头的经典规律：“每一家创业公司与巨头之间的战争，本质上都取决于创业公司能否在巨头获得创新之前拿到分发渠道。”Lassie 的切入方式则是直接告诉小微企业主，他们已经打造出了能够接管数十小时劳动量的 AI Agent，诊所往往极为迅速地接受，因为小微企业主并非担心 AI 会抢走工作，而是现实中根本招不到愿意干这些枯燥文书的人。正如客户 Dr. Quan 所说，Lassie 的意义不是取代人，而是让医护人员免于分身乏术、身兼数职。 --- ### [01:12](https://www.youtube.com/watch?v=Fpg_8aiBxx4&t=72s) 创立契机：从 Robinhood 到牙医诊所的后勤案头 访谈正式开始，Olivia Moore 询问 Steijn Pelle，他是如何从金融科技独角兽 Robinhood 的优渥岗位离职，跑到牙医诊所帮医生手工处理账单和付款的。Steijn 笑着打趣道，他从没想过自己的“美国梦”竟然会变成这个样子。他大约 6 年前从阿姆斯特丹搬到硅谷，骨子里一心想寻找一个真正难啃的硬核问题（hard problem）来创业。他在他的牙医 Dr. Quan 那里看牙，一年去两次。在一次常规检查后（Alex 开玩笑追问当时是不是还躺在牙科椅上、嘴里塞着钻头无法说“不”，Steijn 澄清当时检查完并没蛀牙），医生得知他正在到处寻找值得解决的痛点，便随口问他：“你想不想看看我是怎么经营这家诊所的？” Steijn 欣然答应并跟着 Dr. Quan 走到诊所后方办公室，眼前的景象让他终生难忘：Dr. Quan 明明是 Yelp 上评分第一的顶尖名医，却每个月要耗费 200 个小时应付繁琐的行政文书。由于招不到合适的人来负责患者账单和理赔提交，医生看完病人后还得留下来亲自手动填表报销。Steijn 感到难以置信，因为上世纪 70 年代他母亲在医院工作时，干的就是拿着一袋袋现金去银行、手工核对款项的活计，他本以为半个世纪过去这早该是被软件彻底解决的领域，没想到现实依然如此落后。为了验证这是否仅仅是 Dr. Quan 的个例，Steijn 又去到宾夕法尼亚州斯克兰顿（Scranton）的一家胃肠病专科诊所调研，发现那里的医生也在完全靠手工完成所有的报销与财务流程。这让他们确信，全美成千上万的小微医疗机构都深陷这一泥潭，背后蕴藏着巨大的痛点与机遇。 --- ### [03:46](https://www.youtube.com/watch?v=Fpg_8aiBxx4&t=226s) 肉身入局：在客户办公室“驻场”打杂的信任建立 对于初创团队而言，在还没成型产品之前就直接要求接管客户的核心财务甚至进入诊所内部，显得极其反常。Steijn 坦言，这听起来确实很荒谬——一个在 Robinhood 做增长和推荐项目的工程师，加上一个在 Superhuman 做产品设计的同伴（Frédéric Renken），跑到医疗诊所跟医生说“能不能把财务和账单交给我们来做”。但令他们意外的是，医生们几乎毫无保留地向他们大吐苦水，滔滔不绝地聊上几个小时。 这种毫不设防的开放态度是一个强烈的市场信号：行政负担绝不是“可有可无的维生素”，而是让诊所老板夜不能寐、甚至心灰意冷到想要关门转行的核心痛苦（painkiller）。医生们进入这个行业是为了行医救人，而不是当全天候的账房先生。因此，医生们不仅没有以“缺乏财务经验”或“HIPAA 隐私合规”为由将他们拒之门外，Dr. Quan 反而直接让他们朝九晚五坐在后勤办公室上手干活；宾夕法尼亚州的 Dr. Shah 甚至像对待自家子侄一样把系统所有最高权限开放给他们，手把手教他们各种报销规则。正是这种毫无保留的接纳，坚定了团队解决这一严重脱节问题的决心。 --- ### [05:39](https://www.youtube.com/watch?v=Fpg_8aiBxx4&t=339s) 自动化进化论：上下文层、工具集与智力层的解耦 谈及技术架构的演变，Frédéric Renken 回顾了公司自 2020 年创立以来的技术探索路径。当时底层的 AI 模型还非常原始，尤其是复杂的推理模型（reasoning models）根本尚未面世。但团队从第一天起就对“将业务置于自动驾驶（autopilot）状态”有着极深的执念。Frédéric 剖析，要将任何现实岗位实现自动化，核心都可以拆解为三大要素： ```mermaid flowchart TD accTitle: 三层自动化架构图 accDescr: 展示上下文层、工具层与智力层如何协同工作以取代人工完成复杂流程 subgraph ContextLayer["1. 上下文层 (Context Layer)"] A["历史就诊记录与患者档案"] B["过往保险理赔与账单数据"] end subgraph ToolLayer["2. 工具层 (Tool Layer)"] C["外部保险门户与 EHR 对接接口"] D["报销表单生成与操作执行工具"] end subgraph IntelligenceLayer["3. 智力层 (Intelligence Layer)"] E["早期基础规则 / 人工参与"] F["现代大语言模型与推理能力"] E -.->|随着大模型演进平滑替换| F end ContextLayer --> ToolLayer ToolLayer --> IntelligenceLayer IntelligenceLayer --> G["全自动跑通诊所行政闭环"] ``` 在成立初期，由于智力层能力孱弱，对于最基础的自动化任务并不需要复杂的推理，他们便将精力全部扑在最难啃的基础设施上：构建深度整合的**上下文层**（抓取诊所海量历史数据、患者档案）以及**工具层**（用于模拟和执行具体业务动作的工具集）。当近年来以大语言模型为代表的高级推理能力爆发时，Lassie 顺理成章地迎来了巨大的技术顺风：由于他们早已建好了现成的上下文和工具底座，只需无缝换上更聪明的“智力引擎”，整个产品便能以惊人的速度随之升级变聪明。他们从未将 Lassie 定位为给医生使用的繁琐“新软件”，而是从始至终旨在把繁重的具体工作彻底自动化。 - **自动化的核心三支柱**：掌握业务全貌的历史数据与患者记录（上下文层）、真正能执行业务动作的操作接口（工具层）、以及负责决策和判断的智能内核（智力层）。 - **技术迭代优势**：通过提前占领数据上下文与执行工具，底层模型能力越强，Lassie 的替代效率就越呈指数级放大。 --- ### [07:23](https://www.youtube.com/watch?v=Fpg_8aiBxx4&t=443s) 软件的演进本质：从抽屉数据库到“替人干活”的 AI Olivia 指出，随着模型能力跃进，业界正加速目睹“软件真正承接人类劳动”这一趋势，而 Alex Rampell 正是这一论点的著名提出者。Alex 随即分享了他关于“软件起源”的经典洞察。 在 Alex 看来，商业软件在过去几十年里的本质演变极其单调：它无非就是把办公室里的“铁皮文件柜”（filing cabinet）搬进数据库而已。这一进程的零时刻（$t=0$）始于 IBM 与美国航空（American Airlines）联合开发的 Sabre 系统（这也是 Sabre 名字中带两个“A”的渊源）。在 Sabre 出现之前，航空公司的所有机票预订全部靠纸质卡片分类存放在实体文件柜里；Sabre 的出现让记录数字化，第一次能在绿屏终端上快速修改订票，免去了手工用橡皮擦改纸质单据的麻烦。 此后，整个企业软件行业都在不断复刻这个逻辑： - 人力资源（HR）的文件柜被数字化，演变成了 PeopleSoft； - 法律合同与文献的文件柜被数字化，演化成了 LexisNexis； - 财务与记账的文件柜被数字化，演变为了 QuickBooks 以及后来的 NetSuite。 Alex 深刻地总结道，这些传统软件固然提升了检索和修改的效率，但本质上只是把纸质媒介变成了屏幕上的输入框，**具体的活依然得由坐在电脑前面的人一点一滴敲出来**。从把纸质记录塞进绿屏电脑到搬上云端，人类在实际工作效率上的改善并没有想象中那么彻底，而真正意义上由工具替人干活的变革，才刚刚随着 AI 智能体的登场拉开序幕。 ### [08:44](https://www.youtube.com/watch?v=Fpg_8aiBxx4&t=524s) 从“数据存储”到“替人干活”：软件效率范式的根本转变 Alex Rampell 剖析了传统软件行业几十年来存在的“效率假象”。回顾像 PeopleSoft 或 Workday 这样的企业级人力资源管理系统，它们并没有真正让人力资源部门变得更高效——在 1950 年和 2000 年，同等规模的企业所雇佣的 HR 人数几乎没有差别。过去，企业需要靠 Stein 或 Frédéric 这样的人肉保镖在走廊里守着放有纸质 HR 档案的物理文件柜；而今天，企业只是换成了养着一个庞大的 IT 部门和首席信息安全官（CISO），防范黑客入侵云端文件柜。软件在很长一段时间里仅仅充当了“更昂贵的愚蠢存储工具”。 如今生成式 AI 的突破让软件跨越到了行动层：它不再只是静态的文件柜，而是能对柜子里的数据主动做聪明的事情。如果是 HR 软件，它应该直接执行背景调查、办理入职手续、向新员工解释福利政策；如果是会计财务，当牙医在 QuickBooks 里看到一堆逾期未付的发票时，系统不应该只是呈现这些账单，而是应该直接拿起电话帮诊所催收。Alex 强调，“实际工作本身的体量，比用于存储这些工作信息的数据体量要大上好几个数量级”。直到近期模型性能显著突破，软件才真正具备了执行具体业务任务的能力，而不再停留在简单的“下一个词预测”（next-word prediction）上。 ```mermaid flowchart TD accTitle: 软件演进的三层同心圆模型 accDescr: 展示从纯软件工具到金融科技绑定，再到AI替代人工劳动的市场空间爆发式增长。 subgraph S1["传统纯软件阶段"] A["静态文件柜 / 数据记录<br/>(只存储不干活，如 PeopleSoft)"] end subgraph S2["Fintech 扩展阶段"] B["捆绑支付与金融交易<br/>(如 Toast 通过抽取支付流水变现)"] end subgraph S3["AI 代理时代"] C["直接承担劳动工时<br/>(自主完成催收、对账、报销)"] end A -->|"打包金融抽成"| B B -->|"直接切入人工预算"| C ``` - **市场规模的同心圆**：最内层是纯软件市场（约 1 万亿美元规模）；中层是通过捆绑支付交易扩展的金融科技市场（如同 Visa 或餐饮系统 Toast 的商业模式）；最外层的巨大同心圆则是被 AI 代理吞食的人力劳动预算市场，其体量呈数量级爆发。 - **Toast 的经典案例**：Toast 所需的基础技术在 1985 年搭配 IBM PC 和 MS-DOS 就能实现，但当年一家年营收 500 万美元的餐厅绝不可能掏出 10 万美元购买点单排班软件；而 Toast 将支付处理打包进来，依靠约 2% 的交易提成，间接实现了 10 万美元级别的变现。 --- ### [12:17](https://www.youtube.com/watch?v=Fpg_8aiBxx4&t=737s) 牙医老友的退休遗憾与劳动力短缺的真相 Alex 分享了一个极具代表性的身边真实故事。他的第一任牙医 Ronald Sloop 博士是他父母多年前搬到佛罗里达州后结识的第一位朋友。最近 Sloop 医生彻底退休并将诊所转让给了年轻合伙人，搬到南加州过上了打高尔夫的退休生活。然而，他退休的核心导火索并非厌倦了看牙，而是负责全套账目和前台杂务的“关键员工”离职后，他根本招不到合适的人来顶替，最终身心俱疲选择放弃。当 Sloop 医生看到 Lassie 的发布视频后激动地致电表示：“如果当年有这个系统，我就根本不会退休了。” 大众普遍担忧“AI 是否会抢走工作”，但现实世界中尤其是在中小微经济体中，企业面临的真实危机是根本招不到人。这在经济学上属于一种供给与需求均衡点右侧的“市场失灵”——比如全美每一家牙医诊所如果都能拥有一个说荷兰语的接待员来服务偶尔光顾的母语患者，那体验固然很好，但诊所不可能为此每年花费 4 万欧元全职雇佣一个人。然而，如果这项服务的边际边际成本接近于零或者只要 1 美元，这种潜在的巨大市场需求就会被瞬间激活。 - **小企业主的真实痛点**：拥有哈佛等名校学历的高级牙医，下班后不得不亲自在诊所前台坐到半夜去核对报销单据，琐碎的文书工作彻底摧毁了他们对行医本身的职业热情。 - **全美牙科诊所概况**：仅在美国就有约 16 万家牙科诊所，每家诊所平均每年在纯行政运营上的支出高达 20 万美元。 --- ### [15:35](https://www.youtube.com/watch?v=Fpg_8aiBxx4&t=935s) 切入人工薪资预算：Lassie 的无人工在环设计 Steijn 与 Frédéric 指出，由于中小企业客户每天都在经历缺人的极端痛苦，他们对真正好用的 AI 代理几乎毫无抵触心理，同行推荐后往往迅速采购部署。在商业变现上，Lassie 颠覆了传统 SaaS 的定价逻辑：它不再去挤压每月几十、上百美元的软件订阅费，而是直接切入客户损益表（P&L）中的“人力薪酬预算”（labor budget）。目前 Lassie 推出的第一款 AI Agent 每月仅承担约 30 个小时的行政工作，诊所就极其爽快地支付了高达五位数（万美元级）的年费，而一家诊所每月待处理的人工杂务总计超过 200 小时，商业延展空间巨大。 这种业务属性也对技术架构提出了极为苛刻的挑战。很多所谓的 AI 软件本质上依然是“副驾驶”辅助工具，底层依赖人类最终把关（Human-in-the-loop）或依靠工程师手动部署。但 Lassie 的定位是“接管诊所业务运行”，客户根本没有多余的前台人员来扮演校验员。为了能自主处理涉及真金白银的保险理赔对账和患者账单催缴，团队耗费数年时间，将系统打造成一个不依赖人工干预、自主运行且极致追求准确率的端到端自动化平台。 - **商业收费逻辑**：从软件成本转向劳动报酬，以远低于人类全职员工的成本提供更高质的服务，即使只替代几十小时的工作量，也能轻松拿下五位数的企业合同。 - **产品高容错门槛**：直接与患者沟通收款并核销医保，对正确性有着近乎绝对的要求，促使系统必须朝完全无需人类干预的真正自治系统（Autonomous Systems）方向构建。 ### [17:29](https://www.youtube.com/watch?v=Fpg_8aiBxx4&t=1049s) 从人肉替班到98%自动化：Lassie的构建哲学与长尾学习 在 Lassie 的官方发布视频中，一位牙医客户感慨自己终于能准时下班，去执教孩子们的足球队并观看比赛。正如主持人 Olivia Moore 所指出的，Lassie 的核心价值并非“用 AI 取代人类”，而是“将人们从分身乏术的杂务中解脱出来”（"freeing them from wearing so many hats"）。然而，从渴望减轻负担，到真正放心让 AI 接管诊所的核心收款流程，并实现依靠牙医之间口耳相传爆发式增长，这中间存在着巨大的鸿沟。 面对如何将系统推进到 98% 自动化率的提问，联合创始人 Frédéric Renken 透露，唯一的诀窍就是团队亲自扎进诊所一线去“做苦力”。在创业初期，Frédéric 和 Steijn 充当了诊所的“人在回路”（Human-in-the-loop）——他们全面接管诊所的所有繁重流程，手动帮客户做账、跑医保、核对账目，进而在亲身体验中把自己遭遇的痛点用代码逐一自动化。这种与大企业（Enterprise）截然不同的打法，根源在于中小企业（SMB）根本没有人手去专门学习或操作复杂的工具；如果不实现高度自治，软件无论多精美，最终也只会在深夜把忙碌了一天的医生重新拉回电脑前。 因此，Lassie 确立了敏捷而务实的自动化推进策略：团队不会死磕某个单一功能直到 100% 完美才推向市场，而是在单项流程达到 95% 以上的高自动化水准后，便果断将软件交付诊所，让剩余极少数边界个案（Edge Cases）交由诊所前台处理，自己则迅速抽身去开发下一个自动化模块。系统会在日常运转中从诊所员工对这些罕见个案的处理中持续学习，逐步覆盖那漫长的长尾场景。 ```mermaid flowchart TD accTitle: Lassie自动化演进路径 accDescr: 展示Lassie如何从人工接管流程演进至98%自主运行并借助长尾反馈持续优化的闭环 subgraph Phase1["第一阶段：人肉替班与痛点发掘"] A["团队驻扎诊所一线"] --> B["全面人工接管繁琐流程"] B --> C["把自身痛苦转化为自动化代码"] end subgraph Phase2["第二阶段：高阈值交付"] C --> D["单模块自动化率达 95%+"] D --> E["交付诊所自主运行<br/>（释放前台精力）"] end subgraph Phase3["第三阶段：长尾闭环迭代"] E --> F["遇到极罕见复杂拒赔/个案"] F --> G["诊所员工介入示范处理"] G --> H["AI 吸收反馈提升模型能力"] H --> I["综合自动化率攀升至 98%+"] end ``` - **中小企业的工具困境**：不同于大企业配有专门的 IT 或运营岗位，中小企业没有专职人员来操作系统；工具如果不能直接做功，最终就是医生在深夜牺牲休息时间亲自上阵。 - **冷启动的人在回路**：创始人先通过人工代工（Do things that don't scale）彻底吃透牙科理赔全流程，将自己遇到的机械重复操作编写为自动化程序。 - **95% 的发布门槛**：不追求单个产品线从 0 到 100% 的绝对完美才发布，达到 95% 以上便立即投产，将精力投入横向扩展覆盖诊所的其他劳动力缺口。 --- ### [20:00](https://www.youtube.com/watch?v=Fpg_8aiBxx4&t=1200s) 告别200笔账单梦魇：软件做功与消费级产品标准的结合 Steijn Pelle 在讨论中开起玩笑，提到了此前节目中聊过的退休牙医“斯卢普医生”（Dr. Sloop）。大家调侃说，如果靠 Lassie 能把这类因繁重行政负担而心灰意冷的退休医生重新激活，“把斯卢普医生请出山重操旧业”简直可以作为 Lassie 讲给投资人的 B 轮融资绝妙故事，甚至能一举缓解全美上亿人的看牙荒。 调侃归调侃，Steijn 严肃指出了诊所老板当下的日常噩梦：一家中小型诊所每周通常要接待约 200 名患者，这意味着前台员工每周必须在牙科管理系统、各大保险公司门户网站以及银行对账单之间，来来回回手动更新和核对 200 次账目。哪怕自动化软件每周依然留下极少数棘手账单需要人工干预，只要将原本 200 次的折磨减少为屈指可数的几次，就能直接为诊所每周省下 10 到 20 个小时的纯体力消耗。Steijn 以自身连续三次创业的经历作类比：今天的科技初创团队之所以能用极小的人员规模跑出极高的效率，正是因为各类 SaaS 工具已经解决了大部分基础流程，即便它们还没达到 100% 的完全自治。 为了兼顾软件的无缝扩展与极高稳定性，Lassie 坚定拒绝了以廉价外包后台充当“假 AI”的模式，坚持纯软件路径。与此同时，创始团队的消费级产品背景（Frédéric 曾任职于邮件应用 Superhuman，Steijn 曾任职于 Robinhood）赋予了产品近乎苛刻的交付门槛。目前，其首个 AI Agent 的自动化率已稳定达到 98%，而全美成千上万名合作诊所的员工在日常处理剩下的 2% 复杂上诉（Appeals）时，正在不断向系统提供高质量的人工标注与反馈，推着自动化率向更高的极限迈进。 - **每周 200 笔账单的减负效应**：从前台每周手动对账 200 次骤降至个位数，单周直接释放 10 至 20 个工作小时，业务运转流畅度发生质变。 - **纯软件扩展路线**：坚决不走依赖廉价外包充当人工后台的重资产模式，确保产品具备极高的边际毛利和秒级部署交付能力。 - **消费级团队标准**：团队继承了来自 Superhuman 与 Robinhood 的工匠精神，严防半成品上线，依托一线医护人员的自然操作反哺最难的拒赔申诉逻辑。 --- ### [22:08](https://www.youtube.com/watch?v=Fpg_8aiBxx4&t=1328s) 巨头与初创的博弈：从TiVo困境到拥抱底层管道 在谈及产品落地策略时，Olivia 指出 Lassie 选择直接打通并集成诊所现有的管理系统（Practice Management Systems, PMS），而非强迫医生更换整套核心软件。她借此向 a16z 普通合伙人 Alex Rampell 提问：在 AI 时代，初创企业如何在现有巨头反应过来并展开创新之前抢先建立分发优势？ Alex Rampell 回顾了自己创办 TrialPay 期间的深刻领悟，并将其形象地归纳为“TiVo 困境”（The TiVo Problem）。当年 TiVo 与 ReplayTV 共同发明了划时代的数字视频录像机（DVR），实现了让用户暂停正在直播的电视节目的神奇体验。然而，这虽然是一项惊艳的发明，却注定是一家悲剧的公司。因为没有自身分发管道的纯功能型硬件公司，最终通常只剩三种糟糕的结局： | 破局路径 | 巨头博弈机制 | 最终商业后果 | |:--- |:--- |:--- | | **折价出售给分发垄断者** | 卖给 Comcast 或时代华纳有线（Time Warner Cable），由于收购后其他竞争有线电视网络会集体封杀该硬件，收购方会强行压低对价（Alex 称之为“控制权折价”而非“控制权溢价”）。 | 团队虽能退出，但估值被严重压缩。 | | **被行业巨头劣质抄袭** | 巨头拥有现成的庞大用户基数，会在数年后自行开发一个功能粗糙但免费/内置的版本。 | 初创公司因缺乏直达客户的管道，被巨头直接用平庸产品挤死。 | | **屈辱的技术许可授权** | 将底层技术授权给手握所有客户资源的巨头。 | 巨头凭借垄断分发拿走绝大部分经济利益，初创公司沦为技术苦力。 | Alex 坦言，这一惨痛逻辑促使他意识到，绝大多数创业公司真正该做的是那些“枯燥乏味的事情”（"do the boring thing"）——先去做最底层的脏活累活管道，把客户关系在最上游死死抓在自己手里，然后再去搭建上层那些光鲜亮丽的增值功能。Alex 甚至提到，在创办 TrialPay 四年后他猛然发现，团队真正应该做的是像 Stripe 刚起步时那样（当时 Stripe 只有 5 个人）去死磕底层极为同质化、枯燥却至关重要的支付处理管道；因为一旦掌控了底层的交易通路，上层高毛利的定制化支付解决方案才能拥有不可撼动的护城河。 --- ### [24:27](https://www.youtube.com/watch?v=Fpg_8aiBxx4&t=1467s) AI时代的非典型竞争：为什么牙科软件不是Workday Alex 进一步深入分析了 AI 时代传统竞争法则的“变与不变”。在传统的 SaaS 领域，初创公司往往面临残酷的绞杀。例如，有人提出了一个绝妙的创意：为企业员工入职流程做一套无缝背景调查软件，并集成到人力资源巨头 Workday 中。但这属于显而易见的常规功能延伸，拥有全部企业客户分发权的 Workday 随时可以自己抄一个并将其内置。在 AI 时代，大公司工程师效率低下的问题甚至被 AI 编程助手抹平了一大半（"AI kind of makes a bad engineer into a pretty good engineer"），导致大公司抄袭并集成小功能的速度变得更快。 Alex 还列举了早年由 Idealab 孵化并获得顶级 VC 投资的 X1 公司的经典案例：微软 Outlook 原生邮件搜索体验极差，X1 开发了体验极佳的外挂搜索工具风靡一时，但这种寄生于巨头生态的单点功能最终逃不过微软自身修补功能后的归零或被贱卖（最终被雅虎收购）。 然而，Alex 强调，AI 正在创造前所未有的“非既有品类”（Non-categories），这正是 Lassie 这类专注于“实际替人干活的 AI 软件”（AI that does the work）的巨大机会所在。牙科及广大医疗小微商家的后台管理系统虽然盘踞着传统软件，但那些陈旧系统的背后根本不存在一个如同 Workday 或微软那样敏捷、充满技术野心、能够随时下场做功的科技巨擘。这使得初创公司无需在传统 SaaS 的功能点战役中与大厂贴身肉搏，而是能够利用 AI 智能体跨越系统鸿沟，开辟出一片全新的自动化服务战场。 - **Workday 式的集成陷阱**：在成熟企业级软件生态中，单纯依附于巨头的单点插件极易被巨头自研吞并，因为巨头控制了客户账户且具备产品叠加能力。 - **AI 抹平开发劣势**：大型科技公司过去研发缓慢的部分原因是人员臃肿与平庸工程，但 AI 辅助工具提升了大公司平庸工程师的代码产出效率，加速了巨头吞噬轻度功能型应用的能力。 - **“替人做功”的非竞争蓝海**：在垂直 SMB 领域，现有传统记录系统（Systems of Record）并非现代科技型巨头，AI 智能体直接从事复杂的真实劳动，开拓了巨头完全缺位的全新商业品类。 ### [26:11](https://www.youtube.com/watch?v=Fpg_8aiBxx4&t=1571s) 现有巨头、初创公司与“ Betty ”式的竞争对手 Alex Rampell 剖析了创业公司与传统巨头（Incumbent）之间的经典博弈动力学：是创业公司先掌握分发渠道，还是现有巨头先搞定技术创新？在 AI 浪潮下，现有巨头确实能比以往更快地集成 AI 功能，公开市场甚至已经把传统软件贴上“软件已死、毫无价值”的标签，转而疯狂追捧纯 AI 概念；但 Alex 犀利地指出，AI 的底层本质依然是软件，“软件即 AI，AI 即软件”，任何不想被淘汰的传统软件巨头（如 Workday、NetSuite）都在拼了命地加码 AI 功能。然而更关键的变革在于：有极大量的小微企业场景，历史上**根本不存在任何现有软件巨头**，因为那里的任务自始至终全靠纯人力完成。 ```mermaid flowchart TD accTitle: AI时代小微企业劳动力替代结构 accDescr: 展示诊所等小微企业中AI替代传统人力Betty及其背后的数据与接口壁垒 subgraph Problem["传统诊所痛点"] A["传统劳动力: Betty / 外包账单机构"] -->|离职 / 成本高昂| B["诊所医生身兼数职: 治疗/采买/对账"] end subgraph Solution["Lassie AI 劳动力层"] C["读写双向集成: 各类异构记录系统"] --> D["统一业务本体与数据模型"] D --> E["Lassie 业务执行智能体 (AI Betty)"] end Problem -.->|AI替代劳动力| Solution ``` Steijn Pelle 和 Alex 调侃道，在这类场景里，真正的竞争对手根本不是某家软件巨头，而是“两周前刚刚辞职的诊所前台 Betty”，或者是远在海外的外包账单代理。由于此前没有软件渗透，这片蓝海对创业公司极具吸引力，同时也筑起了坚实的护城河：创业团队必须躬身入局去做大量脏活累活（Schlep）。系统不仅要在各家底层系统之间建立读写双向集成，还必须在全生态中构建统一的“本体论”（Ontology），统一保险理赔与患者付款在不同系统里截然不同的定义，并在此基础上开发出能在复杂流程中替人干活的 Agent。 - 公开市场认知偏差：将软件视为“归零”的旧事物，将 AI 捧为神话，忽视了二者在工程本质上的一体性。 - 真正要竞争的对象：不是传统 SaaS，而是诊所里离职的员工 Betty、医生的私人助理或低效的外包账单公司。 - 隐形防御壁垒：不同诊所系统对于保单、报销的定义千差万别，必须花费数年搭建标准化数据模型与底层接口体系。 --- ### [29:13](https://www.youtube.com/watch?v=Fpg_8aiBxx4&t=1753s) 敲开爱荷华州诊所的大门：极度忙碌诊所的 GTM 与自服务入驻 小微企业软件最大的痛点往往不在于医生抗拒 AI，而在于**根本联系不上医生**。以虚构的典型牙医“ Dr. Sloop ”为例：他晚上 7 点刚想下班，急诊电话就来了，只能返回诊所诊治患者；好不容易赶回家陪孩子吃完晚饭，坐到电脑前，却发现因为 Betty 离职了，自己还得亲自在网上填单订购医疗耗材。面对这些完全不懂技术、日程被撕裂的普通店主，如何实现分发并在他们有限的注意力窗口内完成系统配置，是决定商业成败的核心生死线。 Alex Rampell 形象地总结道：“AI 在硅谷被高估了，但在爱荷华州却被严重低估了。”要攻克爱荷华州这种下沉市场，绝不能指望给每家诊所派一个专属实施人员上门部署。Lassie 创始团队借助此前在 Robinhood 和 Superhuman 等消费级产品的操盘经验，将入驻时间压缩到极短的窗口期——如同消费软件在 48 小时内必须展现核心价值一样，B端产品若不能即插即用，就会被立刻扫地出门。 ```mermaid flowchart LR accTitle: Lassie 自动化免触入驻流程 accDescr: 展示从医生确认使用到系统后台自动打通数据的极简流程 subgraph Onboarding["极简前端操作 (类 Stripe / Rippling 体验)"] A["医生一键确认使用"] --> B["连接诊所银行账户"] B --> C["授权核心系统与保险门户"] C --> D["后台自动核对执业医生信息"] end subgraph Background["后台自动化处理"] D --> E["自动生成数据接口与本体映射"] E --> F["AI Agent 自治接管数字账单与理赔"] end ``` - 消费级无感入驻体验：Lassie 将入驻流程打造成类似 Stripe 结账或 Rippling 入职的极简界面，数分钟内连通银行、记录系统（System of Record）以及所有保险商门户。 - 借鉴 Robinhood 的金融级 KYC 理念：正如 Robinhood 终结了用户跑去嘉信理财（Charles Schwab）线下网点开户的历史一样，Lassie 把复杂的企业合规与医保打通在后台全自动化处理。 - 目标：将原本极度依赖人工配置的系统实施转变为“完全免触碰的自服务（Self-serve）模式”，彻底扫清技术落地的最后一公里障碍。 --- ### [33:35](https://www.youtube.com/watch?v=Fpg_8aiBxx4&t=2015s) 终极蓝图：从牙科诊所到自主运转的万千小微企业 当被问及长期业务版图以及如何对目标市场进行优先级排序时，Steijn Pelle 揭晓了 Lassie 的“宏伟蓝图”（Master Plan）：终极愿景是**让所有小微企业都能实现自主运转（Every small business should run itself）**，把所有耗费精力的繁杂事务全部交由 AI Agent 处理。未来的商业网络形态很可能是：小微企业的业务 Agent，与消费者的个人 Agent 对话，并直接与保险公司等上下游机构的 Agent 协同处理事务。 但在迈向这个终局的过程中，团队有着非常严格的阶段性规划：第一步是专注吃透全美 16 万家牙科诊所。这些诊所每家因招不到合适行政人员而造成的劳动力缺口高达 20 万美元，单单这一垂直细分领域，就能支撑起一个年经常性收入（ARR）超过 10 亿美元的软件市场。在此之后，Lassie 才会将这种高完成度、消费级易用性的 Agent 方案，横向复制到其他未被充分服务、市场规模巨大且面临同样行政泥潭的医疗诊所类型中。 - 宏观战略目标：以自治 Agent 接管小微企业的后台脏活，构建小微企业 Agent 与上下游 Agent 全自动对接的网络。 - 牙科垂直市场基本面：仅美国本土即有 160,000 家牙科诊所，每家平均存在约 $200,000 的劳动力缺口，对应超 10 亿美元 ARR 的潜在市场空间。 - 横向扩张节奏：坚决不盲目过早铺摊子，先在牙科行业把准确率与无缝入驻打磨至极致，随后再进入物理治疗等同构的医疗小微业态。 ### [03:50:05](https://www.youtube.com/watch?v=Fpg_8aiBxx4&t=13805s) 终极蓝图：从牙科诊所到全球所有小企业 Steijn Pelle 描绘了 Lassie 的长期愿景与宏伟蓝图。在入门体验上，Lassie 力求做到像 Coinbase 或 Stripe 一样简单丝滑。在扩张路径上，团队首先会纵深拓展到类似的医生诊所场景，但终极目标远不止于此。 ```mermaid flowchart LR accTitle: Lassie 的跨行业横向拓展逻辑 accDescr: 展示 AI Agent 如何从垂直的牙科核心能力抽象并扩展到全行业小企业 subgraph Core["底层业务共性"] SOR["记录系统 (System of Record)<br/>读写数据"] CUST["客户 / 患者交互<br/>预约排期与支付交易"] end subgraph Vertical["垂直领域突破"] DENT["牙科诊所实践<br/>深度验证 200 小时替代"] CLINIC["其他专科 / 全科医生诊所<br/>复用医疗流转逻辑"] end subgraph Expansion["全行业小企业"] SMB["全球小企业<br/>爱荷华、肯塔基到欧洲本土"] end Core --> Vertical Vertical --> SMB ``` 从抽象层面审视，所有小企业都遵循相同的底层规律：它们都有一个需要频繁读写的记录系统（System of Record）；都拥有自己的终端客户（诊所面对的是患者，本质上都是围绕排期与支付开展交互与交易）。因此，如果在牙科和医疗诊所验证成功，Lassie 将具备全球最顶尖的 AI Agent 构建能力，去服务那些“脚踏实地的普通人（salt of the earth people）”——不仅是美国爱荷华州或肯塔基州帕迪尤卡（Paducah）的小商户，未来还会走向 Steijn 的故乡阿姆斯特丹以及 Frédéric Renken 的故乡德国汉堡。不过，团队依然保持着在 Superhuman 和 Robinhood 积累的克制：像激光一样高度聚焦于先把单点做到极致，再全面铺开。 - **小企业共性抽象**：底层均为核心记录系统（System of Record）的读写，以及围绕客户进行的预约排期与支付结算。 - **全球化蓝图**：从美国本地诊所逐步渗透，未来规划覆盖欧洲本土（阿姆斯特丹、汉堡）乃至全球的各行各业小商户。 --- ### [03:56:11](https://www.youtube.com/watch?v=Fpg_8aiBxx4&t=14171s) 2026 年创办 AI 原生企业：传承、迭代与彻底颠覆 当 Olivia Moore 问及从 Coinbase、Robinhood、Superhuman 等明星独角兽带出来的经验与在 2026 年创办 AI 企业需要做出的调整时，Frédéric 分享了两个维度的思考。首先是沿用至今的准则：对理想客户画像（ICP）的极端挑剔。如果让错误的诊所上线，由于 AI 无法完全自动化其流程，团队就会陷入既无法兑现承诺、又不得不面临痛苦退订流程的泥潭；同时，入门流程被精心设计得宛如故事剧本或电影镜头，每个时间节点都有严格量化的交付检查点。 然而，产品设计哲学发生了范式转移。以往做软件是提供“功能（Functionality）”，给用户一个趁手的工具去操作；而现在衡量一切的标准是“自动化了多少真实劳动力（Labor）”。Frédéric 以“患者账单（Patient Billing）”举例： | 维度 | 传统 SaaS 思维 | AI Agent 原生思维（Lassie） | |:--- |:--- |:--- | | **产品目标** | 提供发送账单、在线收款的 UI 工具 | 诊所端完全脱手，零人工耗时 | | **实际耗时痛点** | 忽略核对金额与后续患者沟通成本 | 识别出大量时间消耗在解释账单与保险覆盖上 | | **交付形态** | 用户登录平台手动点选、核算、催收 | Agent 全程核对金额、答疑解惑、自动完成收发 | 传统软件以为给诊所一个发账单按钮就解决了问题，但实际上大量时间浪费在后续琐碎的纠纷中——患者会打电话质问“为什么我欠这么多钱？保险不是报销了吗？”。Lassie 的开发逻辑是：一旦交付了该模块，诊所前台在该事项上的耗时就必须彻底归零。 --- ### [03:59:27](https://www.youtube.com/watch?v=Fpg_8aiBxx4&t=14367s) 招聘标准与极速交付：用奥运运动员的标准打造高密度团队 谈及团队招聘时，Steijn 指出虽然产品找到了巨大的市场契合点（PMF），但在选人上他们依然持有反直觉的传统观念：底层能力模型并没有改变。公司不强求应聘者拥有牙科行业背景（大家笑称只要按时一年看两次牙医就行），而是寻找那些处于前 5% 的工程或销售顶尖人才——拥有极其陡峭成长曲线（Steep Slope）与强烈野心的驱动型人才。 Steijn 借自身经历做了一个生动的类比：他曾是一名田径运动员，差点走上职业道路，身边的队友许多都参加了奥运会。奥运选拔的法则从未改变，运动员每天必须严苛训练两次、挑战体能与心智的极限；初创公司对高强度的追求同样如此。唯一的全新变量是候选人的“AI 融入度（AI-pilledness）”： - 程序员是否笃信当前时代的编码方式将被彻底颠覆； - 财务或业务人员是否意识到部门的搭建逻辑已经与过去截然不同。 Lassie 的内部目标是构建一个 2026 年版本的组织架构：通过全员、全流程嵌入 AI，实现发版交付速度是同行的两倍，推进节奏是普通团队的四倍。团队正把整个创业当成一个庞大的飞轮实验：以最快速度拿下全美牙医市场、帮他们砍掉 200 小时案头工作，随后迅速跨行业复制。 - **组织速度目标**：实现常规团队 2 倍的代码交付量与 4 倍的推进速度。 - **核心实验命题**：验证 AI 原生团队如何以极高人效在全美乃至多垂直领域实现超大规模渗透。 --- ### [04:02:00](https://www.youtube.com/watch?v=Fpg_8aiBxx4&t=14520s) 供给瓶颈与护城河悖论：小企业的市场规模上限在哪里？ Alex Rampell 抛出了一个深刻的“护城河悖论”：如果 AI 让任何人都能够随手变出一个像“Betty”这样能干的虚拟雇员，创办和运营小企业的门槛被大幅拉低，这是否会导致市场严重过剩，甚至出现传奇棒球手约吉·贝拉（Yogi Berra）所说的名言——“那里太拥挤了，以至于再也没人去了（It's so crowded nobody goes here anymore）”？小企业原先依靠“堆砌人力交付服务”所建立的壁垒，是否会因此被削弱而引发利润侵蚀？ ```mermaid flowchart TD accTitle: 小企业 AI 化后的供需模型与护城河逻辑 accDescr: 对比传统观点对过度竞争的担忧与当前真实医疗及本地服务供给严重短缺的现状 subgraph Skeptic["怀疑论观点 (护城河侵蚀)"] A["创建门槛大幅降低"] --> B["店铺与服务泛滥"] B --> C["利润空间被严重压榨"] end subgraph Reality["现实约束 (供给严重受限)"] D["真实需求巨大"] --> E["优质供给极度稀缺<br/>(好牙医、好水管工、家庭医生)"] E --> F["AI 释放产能，满足未被服务的市场"] end Skeptic -.->|假设存在需求天花板| Reality ``` Steijn 对此给出了明确的反驳：这种悲观假设的前提是“市场需求存在天花板”。但现实情况恰恰相反，无论是在牙科诊所还是寻找一个优秀的水管工，目前的真实需求至少是现有供给能力的两倍。在医疗领域尤甚，许多人根本没有得到应有的牙科保健服务。 Steijn 回忆了自己在荷兰长大的经历，欧洲的初级保健体系与美国截然不同。在美国，居民很难在本地社区找到一位真正熟悉自己、了解家庭病史且长期提供关怀的优质家庭医生。AI 的普及不仅不会导致所谓同质化竞争带来的恶性踩踏，反而是打破优质医疗与专业服务供给天花板的历史性机遇，让被繁杂行政琐事绑架的专业人员得以释放生产力，真正填补巨大的市场缺口。 ### [44:11](https://www.youtube.com/watch?v=Fpg_8aiBxx4&t=2651s) 让手艺人专注本业：消除繁琐行政带来的产能释放 Steijn Pelle 指出，全美目前约有 16 万名牙医，而如果利用技术将他们从繁重的文书工作中解放出来，要么能让全美享受相当于 50 万名牙医的服务产能，要么能让现有的牙医接诊量翻倍。这种逻辑在管道工、烘焙师等各类小企业中完全相通——小商户的愿望其实很简单，比如“烤出更多馅饼”，但往往受制于有限的劳动力与琐碎的杂务。Lassie 的核心愿景是让人们能拥有双倍的时间专注于自己的“手艺”（craft）。当 Lassie 最初公开亮相并宣布业务方向时，引发了外界极大的共鸣，大家普遍认为这是一个技术向善的乐观范例。在公众对 AI 的未来充满焦虑的背景下，“没有人会反对帮小企业主清理行政杂务，让他们去安心烤馅饼、做美甲、洗牙或补牙”。 - **行业产能对比**：全美目前约有 16 万名牙医，自动化后有望释放出相当于 50 万名牙医的接诊与服务能力。 - **核心诉求**：消除占用手艺人时间的低效“琐事”（busy work），让各行各业的专业人士重新回归核心服务本身。 --- ### [45:24](https://www.youtube.com/watch?v=Fpg_8aiBxx4&t=2724s) 大模型的盲区：隐性工作流与小企业专有知识壁垒 Alex Rampell 提出，Lassie 始于 2020 年，跨越了 2022 年底 ChatGPT 发布前后的分水岭。如今随着推理模型（Reasoning Models）的爆发，实现“完全由软件替代人工”的技术阻碍究竟还有哪些？那些处于“最后 10%”却决定了产品完整度的硬骨头又是什么？Frédéric Renken 对此给出了反直觉的洞察：通用大模型虽然在海量互联网数据上完成预训练且参数庞大，但它们“根本不知道如何做具体的实际工作”。真实世界的工作流从未被系统性地编码进公开互联网中，比如向不同保险支付方提交理赔索赔的具体标准作业程序（SOP），往往只存在于老练的诊所前台经理的大脑中。 ```mermaid flowchart TD accTitle: Lassie Agent Knowledge Acquisition accDescr: 展现大模型通用数据与诊所专有隐性数据结合的工作流 subgraph LLM_Limitation ["大模型能力边界"] A["公开互联网预训练数据"] --> B["缺乏垂直 SOP 与具体业务流程"] end subgraph Lassie_Data ["Lassie 专有壁垒"] C["历史 ERP / 诊所系统数据"] --> E["推导实际工作流规范"] D["诊所前台人类经验与偏好反馈"] --> E end subgraph Agent_Execution ["AI 代理智能执行"] B & E --> F["具备业务语境的智能代理"] F --> G["精准生成理赔叙述与账单"] end ``` Lassie 的巨大壁垒正是在于能够接入客户的 ERP 历史数据，并从中反向推导出这些未被文字化的业务流程。Frédéric 坦言，团队在引入最新推理模型时曾以为“它们懂得全天下的数据，理应天生就会这些操作”，结果发现模型对绝大多数现实业务的复杂细节一无所知。相比庞大而笨重的通用模型，他更期待那些学习速度更快、所需训练数据更少的小模型，希望未来“智能可以无处不在”，甚至像“皮克斯标志性的跳跳台灯那样拥有独立性格”。而在产品与交互设计层面，Lassie 花费了大量精力研究如何高效收集一线员工的操作习惯与偏好，将人类隐性知识转化为让 AI 代理不断进化的结构化输入。 - **模型认知断层**：海量数据训练的大模型缺乏真实商户内部的工作流 SOP，许多业务诀窍仅作为隐性知识留存在员工经验中。 - **数据护城河**：通过提取诊所管理系统（ERP）中的历史往来记录，逆向解析并沉淀出保险索赔等深水区流程。 - **产品与交互重心**：设计新型交互方式，低摩擦地捕捉一线操作员的偏好反馈，用于持续精调 AI Agent。 --- ### [49:19](https://www.youtube.com/watch?v=Fpg_8aiBxx4&t=2959s) 零边际博弈成本：AI 代理对决保险公司与实体支票的现实 Alex Rampell 提出了一个深刻的行业博弈命题：**“当争辩的边际成本归零时，世界会发生什么？”** 他以自己的健康保险公司 Cigna（信诺保险）为例，对方至今坚持只邮寄纸质支票，根本不开通线上转账渠道，其深层逻辑就是制造人为摩擦，赌客户会弄丢支票或忘记承兑；保险公司习惯性地“拒赔、拒赔、再拒赔”，而投保方也可能在索赔时夸大损失（如声称“被烧毁的房子里有一幅毕加索名画”）。如今双方都拥有了成本近乎为零的 AI 工具，可以进行无限期的拉锯与申诉。 Steijn Pelle 从医疗报销的底层规则回应了这一现象。牙科报销实际上处于极其严苛的规则监管之下，例如进行牙冠修复（Crown）手术时，只要诊所按照 Cigna 内部规章提供特定角度的 X 光片以及符合规范的临床病情叙述（Narrative），保险公司就必须按约赔付。但人类员工（如诊所老员工 Mildred 或 Betty）面对多达 50 到 100 种常用计费代码和复杂的附带条款时极易出错或遗漏；而 Lassie 的 AI 代理则将规章手册、病患 X 光影像和治疗方案完整打通，能够分毫不差地提交索赔，迫使保险公司必须照章履约付款。 与外界预想不同的是，保险公司与诊所之间并非纯粹的零和死敌。Steijn 解释道，Cigna 每年必须向雇主客户（例如 Andreessen Horowitz 这样的机构）争取续约，如果旗下网络无法覆盖 Alex 或 Olivia 认可的高水平本地牙医（如 Dr. Lu 或 Dr. Quan），企业就会转投其他竞争对手，因此保险公司有充足的动力维护好优质诊所的关系。更荒诞的现实在于物理层面的脱节：全美大量诊所至今完全依靠纸质支票结算。Steijn 曾亲自坐在 Dr. Quan 诊所的吧台椅上，当邮递员送来成堆的信件时，他亲手拆开一封封信封，核对并向银行存入高达 10 万美元的纸质支票，然后逐一核对支票附带的明细发票。这种不仅需要“把办公室的文件柜彻底数字化”，还要手工处理海量物理票据的沉重现实，正是 AI 代理改造传统小商业的最前线。 - **博弈规则化**：牙科常见计费代码约 50 至 100 个，AI 能够严格匹配保险公司的赔付要求与临床影像，将非标的“扯皮”转化为标准化的合规递交。 - **商业反制平衡**：保险公司受制于雇主端对医疗网络覆盖质量的要求，必须保证核心医生留在网络内，无法单方面无理拒赔。 - **纸质支票现实**：美国医疗小微诊所仍重度依赖物理邮件结算，单次桌面上堆积的未承兑纸质支票金额可达 10 万美元，亟待全流程数字化与自动化核销。 ### [53:18](https://www.youtube.com/watch?v=Fpg_8aiBxx4&t=3198s) 监管顺风与纸质流程的数字化护城河 几年前牙科诊所的现状是真正的“物理文件柜时代”——文件柜里实实在在地塞满了支票与逐项列出的纸质账单，诊所全靠这套纸质流程勉强维持日常运转。然而，联邦政府介入并明确表态：必须终止纸质支票操作。监管新规强制要求该行业必须向医生提供直接存款（Direct Deposit）选项，且针对每一份项目明细账单，都必须生成对应的数字化文件格式。Lassie 恰好借上了这股监管政策的东风。据统计，目前仍有高达 70% 的小型企业依然依赖纸质支票结算，整个行业正因联邦监管转折点而经历一场规模庞大的数字化变革。 ```mermaid flowchart TD accTitle: 医疗小企业支付数字化流程 accDescr: 展示从传统的纸质支票与账单，通过监管与Lassie自动化引擎，转换为AI可处理的数字化流程 subgraph Legacy["传统纸质状态"] A["物理文件柜"] --> B["纸质支票 & 账单"] B --> C["员工手动核对纸质单据"] end subgraph Transition["监管推进与阻碍"] D["联邦法规强制要求直接存款"] E["繁琐对接: 耗时50小时对接各大保险公司"] F["隐私顾虑: 诊所不愿对员工开放银行账户"] end subgraph LassieEngine["Lassie 自动化解决方案"] G["收集基础信息: 税号及营业资质"] H["后台全自动转换电子支付对接"] I["输出结构化数字数据并由 AI Agent 自动化处理"] end Legacy -.-> Transition Transition --> LassieEngine ``` 即使大模型在五年前就已经问世，如果底层的支付和单据仍然停留在纸质媒介上，AI 也根本无从下手。诊所医生之所以迟迟没有一键切换到数字化支付，背后有极其现实的阻碍：以一位普通的医生为例，要对接 Cigna、Delta Dental 等各大主流保险公司，可能需要耗费整整 50 个小时逐个梳理流程；更关键的是，一旦转为直接存款，款项将直接汇入诊所的核心银行账户，而医生绝不希望前台或普通员工看到账上的房租支出、合伙人分红等高度敏感的财务隐私。针对这些痛点，Lassie 打造了一套底层转换引擎，诊所只需提供税号（Tax ID）等基本营业信息，后台就能自动完成向电子支付的合规转换。这种将物理文件柜里的数据提炼为结构化数字格式的能力，构成了将尖端 AI 引入实体街区（Main Street）时极具深度的产品壁垒。 - 牙科等小型医疗机构中，大约有 70% 的账单和支付结算至今仍以纸质形式流转。 - 联邦监管法规强制推行直接存款选项，成为推动行业底层数字化的关键政策顺风。 - 传统流程依赖纸质不仅仅是技术滞后，如果单纯把纸质账单扫描成 PDF，员工往往觉得在纸上勾画比在屏幕上看 PDF 更顺手，缺乏数字化的内生动力；只有当数字化配合端到端的工作自动化时，转型的价值才真正释放。 --- ### [55:59](https://www.youtube.com/watch?v=Fpg_8aiBxx4&t=3359s) 触达主街小企业：完全不同于大企业的 GTM 新打法 面对分散在全美各地、成千上万家极度渴求自动化产品的小诊所，如何触达并规模化获客成了一个全新的商业课题。目前 AI 领域的常规打法多集中在大型企业（Enterprise）：通过几顿高规格的牛排晚宴建立高管关系，签下一份大单便能入账千万美元级 ARR。然而，Lassie 面对的是数十万家分布在普通社区的小微企业，这套大企业销售逻辑完全失效。 ```mermaid flowchart LR accTitle: 企业级销售与主街小企业获客模式对比 accDescr: 对比传统企业软件获客方式与Lassie面向小微企业的全新GTM打法 subgraph EnterpriseGTM["传统企业级软件销售"] E1["Clay / Apollo 爬取线索"] --> E2["LinkedIn 锁定HR/采购决策人"] E2 --> E3["高规格商务宴请 / 漫长销售周期"] E3 --> E4["签下单笔千万级大合同"] end subgraph MainStreetGTM["Lassie 小微企业新型获客"] M1["全美诊所全景图谱映射"] --> M2["求职网站招聘意向信号捕捉"] M3["针对性痛点信息精准触达"] M4["规模化赋能千家万户小诊所"] end ``` 小微企业主的画像完全不同：普通的牙科医生不仅不在 Clay、Apollo 等主流 B2B 销售线索数据库中，甚至连 LinkedIn 账号都没有。Lassie 正在摸索一套全新的 GTM（Go-To-Market）打法：全方位测绘全美所有牙医诊所的分布图景，挖掘诊所所有者身份、所用管理系统，并通过外部意图信号（例如诊所在 Indeed 等招聘网站上发布急聘前台或账务人员的招聘信息，表明其正遭遇严重的行政人手荒）进行精准识别。随后，团队通过高度契合其痛点、能穿透信息噪音的信息与其取得联系。这种将 AI Agent 大规模推向全美数以万计实体小企业的获客模型，是一片极具商业潜力且尚未被行业充分开发的全新蓝海。在节目的尾声，主持人鼓励渴望投身这一代际创新机遇的人才关注 Lassie 团队的招聘动态（lassie.ai），共同打造赋能实体小企业的自动化未来。 - 传统企业级销售依靠数据库富化与高管直聘，而面向小企业需要建立专属的覆盖地图与另类意图信号追踪机制。 - Indeed 等招聘网站上的行政、账务职位空缺，成为了判断诊所是否面临人力短缺、亟需 Lassie 介入的高价值意向信号。 ## Notable quotes > [00:00](https://www.youtube.com/watch?v=Fpg_8aiBxx4&t=0s) 每一家创业公司与巨头之间的战争，本质上都取决于创业公司能否在巨头获得创新之前拿到分发渠道。 — Alex Rampell > [08:44](https://www.youtube.com/watch?v=Fpg_8aiBxx4&t=524s) 实际工作本身的体量，比用于存储这些工作信息的数据体量要大上好几个数量级。 — Alex Rampell > [12:17](https://www.youtube.com/watch?v=Fpg_8aiBxx4&t=737s) 如果当年有这个系统，我就根本不会退休了。 — Dr. Ronald Sloop > [17:29](https://www.youtube.com/watch?v=Fpg_8aiBxx4&t=1049s) 将人们从分身乏术的杂务中解脱出来。 — Olivia Moore > [26:11](https://www.youtube.com/watch?v=Fpg_8aiBxx4&t=1571s) 软件即 AI，AI 即软件。 — Alex Rampell > [29:13](https://www.youtube.com/watch?v=Fpg_8aiBxx4&t=1753s) AI 在硅谷被高估了，但在爱荷华州却被严重低估了。 — Alex Rampell > [33:35](https://www.youtube.com/watch?v=Fpg_8aiBxx4&t=2015s) 让所有小微企业都能实现自主运转（Every small business should run itself）。 — Steijn Pelle > [42:00](https://www.youtube.com/watch?v=Fpg_8aiBxx4&t=2520s) 那里太拥挤了，以至于再也没人去了（It's so crowded nobody goes here anymore）。 — Alex Rampell > [44:11](https://www.youtube.com/watch?v=Fpg_8aiBxx4&t=2651s) 没有人会反对帮小企业主清理行政杂务，让他们去安心烤馅饼、做美甲、洗牙或补牙。 — Steijn Pelle > [45:24](https://www.youtube.com/watch?v=Fpg_8aiBxx4&t=2724s) 通用大模型虽然在海量互联网数据上完成预训练且参数庞大，但它们根本不知道如何做具体的实际工作。 — Frédéric Renken > [49:19](https://www.youtube.com/watch?v=Fpg_8aiBxx4&t=2959s) 当争辩的边际成本归零时，世界会发生什么？ — Alex Rampell ## Mentioned Alex Rampell, Olivia Moore, Steijn Pelle, Frédéric Renken, Dr. Quan, Dr. Shah, Dr. Ronald Sloop, Yogi Berra, a16z, Lassie, Robinhood, Superhuman, Coinbase, Rippling, Stripe, TrialPay, IBM, American Airlines, Sabre, PeopleSoft, Workday, LexisNexis, QuickBooks, NetSuite, Toast, TiVo, ReplayTV, Comcast, Time Warner Cable, Idealab, X1, Microsoft Outlook, Yahoo, Cigna, Delta Dental, Clay, Apollo, Indeed, LinkedIn ## Video description Alex Rampell and Olivia Moore speak with Lassie cofounders Steijn Pelle and Frédéric Renken about bringing AI to one of the most overlooked parts of the economy: small businesses. Inspired by time spent working inside dental practices, Pelle and Renken set out to automate the administrative work that keeps healthcare providers away from patients. They discuss how AI agents are changing billing, insurance claims, patient payments, and other operational workflows, allowing practices to spend less time on paperwork and more time delivering care. The conversation explores AI agents, software that performs work rather than simply storing information, onboarding AI into real-world businesses, and why healthcare administration offers one of the biggest opportunities for automation. Along the way, they discuss product design, go-to-market strategy, and what it takes to build AI systems that operate reliably in complex business environments. Timestamps: 00:00 - Intro 01:12 - Story Behind Lassie 03:46 - Embedding in the Customer's Office Before Building a Product 05:39 - How the Product Build Has Changed with Better Models 07:23 - Software Never Did the Work but AI Finally Does 17:24 - 98% Automation: How Lassie Got Agents to Actually Run a Practice 22:08 - Startup vs Incumbent in the AI Era 33:35 - The Master Plan: From Dentists to Every Small Business 39:27 - What It Takes to Hire & Build When You're Selling to Main Street 55:59 - How Do You Reach Hundreds of Thousands of Small Businesses? Resources: Follow Steijn Pelle on X: https://x.com/steijnpelle Follow Frédéric Renken on X: https://x.com/fredericrenken Follow Alex Rampell on X: https://x.com/arampell Follow Olivia Moore on X: https://x.com/omooretweets 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=Fpg_8aiBxx4)

[0:00] Alex Rampell: 每一家创业公司与巨头之间的战争，本质上都取决于创业公司能否在巨头获得创新之前拿到分发渠道。 [8:44] Alex Rampell: 实际工作本身的体量，比用于存储这些工作信息的数据体量要大上好几个数量级。 [12:17] Dr. Ronald Sloop: 如果当年有这个系统，我就根本不会退休了。 [17:29] Olivia Moore: 将人们从分身乏术的杂务中解脱出来。 [26:11] Alex Rampell: 软件即 AI，AI 即软件。 [29:13] Alex Rampell: AI 在硅谷被高估了，但在爱荷华州却被严重低估了。 [33:35] Steijn Pelle: 让所有小微企业都能实现自主运转（Every small business should run itself）。 [42:00] Alex Rampell: 那里太拥挤了，以至于再也没人去了（It's so crowded nobody goes here anymore）。 [44:11] Steijn Pelle: 没有人会反对帮小企业主清理行政杂务，让他们去安心烤馅饼、做美甲、洗牙或补牙。 [45:24] Frédéric Renken: 通用大模型虽然在海量互联网数据上完成预训练且参数庞大，但它们根本不知道如何做具体的实际工作。 [49:19] Alex Rampell: 当争辩的边际成本归零时，世界会发生什么？

[33:35] Steijn Pelle: Every small business should run itself [5:39] Frédéric Renken: humans in the loop [5:39] Frédéric Renken: putting the business on autopilot [49:18] Alex Rampell: the marginal cost of arguing goes to zero [45:24] Alex Rampell: BC/AD divide [42:00] Alex Rampell: It's so crowded nobody goes there anymore [52:32] Steijn Pelle: the file cabinet was literally a file cabinet [44:11] Steijn Pelle: twice as much time for the craft

[0:00] AI is overhyped in Silicon Valley but [0:02] underhyped in Iowa. I would actually [0:03] argue software just kind of took things [0:05] that were stored in paper format and [0:07] then they made them available first on [0:09] prem via green screen computers but [0:12] people still had to do the work. [0:13] >> I never forgot what I saw the number one [0:15] great adopter on Yelp spending 200 hours [0:17] a month on paperwork. [0:18] >> The models are trained on so much data [0:21] and they're so large and yet they [0:23] actually don't really know how to do any [0:25] of this work. Initially we were actually [0:26] the humans in the loop. We kind of [0:28] automated away our own problems. [0:29] >> The battle between every startup and [0:31] incumbent comes down to whe the startup [0:32] gets the distribution before the [0:33] incumbent gets the innovation. [0:34] >> We come by and we say, "Hey, we actually [0:36] have built this agent that can provide [0:38] you already with tens of hours of labor. [0:40] They then adopt it like very quick. They [0:42] see us as someone they break in to like [0:44] actually run the practice [music] for [0:45] them." [0:45] >> There was a great quote from Dr. Quan [0:47] about you guys, which is that Lassie [0:49] isn't replacing humans, but like freeing [0:51] them from wearing so many hats. [0:53] >> It's not like, oh, AI is going to take [0:55] the jobs. In many cases, you can't find [0:57] somebody. [0:57] >> How are you prioritizing what you build, [1:00] who you sell to? Is there of the world [1:02] where Lassie for dentists [music] makes [1:05] lassie for physical therapists better? [1:07] >> The end goal here is that [1:12] >> so welcome and thank you for joining us. [1:15] >> Thank you for inviting us. Excited to be [1:16] here. [1:17] >> Maybe we'll start with the basics. So, [1:18] Stein, this whole company started with a [1:20] conversation with you and your own [1:22] dentist, Dr. Quan. What did he tell you [1:25] that made you decide to quit your tech [1:27] job at Robin Hood and go process [1:29] payments for him by hand? [1:30] >> Yeah, I did not uh know that my American [1:32] dream would look like this. Um it was uh [1:35] it was interesting like I I um was at [1:38] Robin Hood at the time. Um and uh I came [1:40] to this country to start a company. So [1:42] after 6 years um I moved from Amsterdam [1:44] to Silicon Valley like I was looking for [1:46] a heart problem to solve. And then my [1:48] doctor Dr. Quan uh I was a patient [1:50] there. So I saw him twice a year as you [1:52] do with a dentist, knew that I was [1:54] looking for a heart problem and he said, [1:56] "Do you want to see how I run my [1:58] business?" And I said, "Absolutely." Um, [2:00] he walked me to the back and I never [2:01] forgot what I saw there. Just like a [2:03] small business owner that like is the [2:05] number one rated doctor on Yelp spending [2:08] 200 hours a month on paperwork and busy [2:10] work. So submitting claims by hand and [2:12] he had to stick around himself because [2:14] he couldn't find people to like build [2:16] the patients. So I'm like, "Wow, it's [2:18] fascinating. This is like a couple of [2:20] years ago." So I thought that this was a [2:21] solved problem cuz in the 70s my mom [2:23] worked in hospital and that's what she [2:25] did. She brought backs of cash to the [2:27] bank and then processed payments by [2:29] hand. Uh but it wasn't a solved problem [2:31] here. So that's uh why this piqued my [2:33] interest [2:33] >> though I have to ask when he gave you [2:35] this offer were you like was it like [2:37] that kind of reclined? Could could he [2:40] actually process your answer as yes [2:41] versus no if your mouth was open and [2:43] drills were in your mouth or how did [2:44] that go down? [2:45] >> Well uh up until now I don't have [2:47] cavities so you know there was no [2:48] drilling happening yet. Uh so you know [2:50] knock on wood [2:51] >> examination. [2:51] >> Exactly. Exactly. Yeah. Um um no like he [2:56] he uh took me aside like after that. [2:59] >> Okay. [2:59] >> Um because he he knew that I was like [3:01] looking for this like heart problem. I [3:03] was roaming around. Um and um then like [3:06] uh like after like that appointment he [3:07] he showed me kind of like what was going [3:09] on. Uh and then I thought maybe it's [3:12] him, right? But that didn't really uh [3:14] make sense to me because he was very [3:15] like well rated. He used all these [3:17] modern technologies. So then um we [3:19] started like talking to other doctors [3:21] like because maybe this Dr. Guan was [3:22] just like an anomaly but then you know I [3:25] also like worked for a gastronologist in [3:27] Scranton Pennsylvania and we saw the [3:29] same there. I'm like wait you're doing [3:31] this all by hand. Um so then we figured [3:33] out wait there are like hundreds of [3:34] thousands of these small businesses that [3:37] literally like do this all by hand like [3:39] that would be quite like a fascinating [3:41] problem to solve. We knew it was going [3:42] to be a hard problem but uh we were kind [3:44] of like looking for that. Yeah, the [3:47] Lassie story is so unique to me because [3:49] you both spent months if not years [3:51] before kind of fully releasing the [3:53] product like literally in the office of [3:55] the customers. [3:56] >> Yeah. [3:57] >> Um, how did you convince them to let you [3:59] in and to kind of look through the the [4:02] heart of the business and get into the [4:03] financials? [4:04] >> Yeah, it's a little weird, right? It's [4:05] like, hello, I work at Robin Hood on [4:07] growth on the referral program. Uh, can [4:10] I get a job like here? And by the way, [4:12] Frederick worked at Superhuman on [4:13] product. can we do the billing for you [4:15] and take over the finances? Um I think [4:18] that was the first sign that we were on [4:19] to something big. Um because to our [4:22] surprise all these doctors when we asked [4:23] them so we we first talk to all of them [4:25] like Dr. Guam because everybody likes to [4:27] talk about their problems. Um and uh [4:30] when all these doctors started talking [4:31] to us for hours we knew that okay this [4:33] is a real problem they have. This is not [4:35] some vitamin that you know maybe it's [4:38] nice to solve that for them. Uh but this [4:40] is this is something that keeps them up [4:41] at night. it it makes them almost quit [4:43] their job and say you know I got into [4:46] this industry because of my passion and [4:47] craft in this case like because I want [4:49] to take care of patients. Um so that was [4:52] the first time that these people were [4:54] not no chance that you can come work for [4:56] me because you know like you don't have [4:58] any experience running the finances uh [5:01] what about like HIPPA and security [5:03] reasons that you have access to all this [5:05] information. So I think the first sign [5:06] to us that these people said yes Dr. [5:09] Quan said, "Just come and sit here night [5:11] five. You can do the job." Um Dr. Sha [5:14] was in Scranton, Pennsylvania. Uh he sat [5:17] us down behind the desk and said, you [5:19] know, uh you can have access to anything [5:21] you need to have access to or teacher [5:23] like my son. Um so that was the first [5:25] sign that like this was very broken uh [5:27] and not a solved like problem. Uh and [5:30] then I think that triggered like our [5:31] intuition for okay we might be on to [5:34] something because this is a real pain [5:36] that uh people are desperately looking [5:38] for a solution. [5:39] >> Yeah. And from a technical perspective. [5:42] So Frederick you the business started in [5:45] 2020 and so much was different then in [5:47] terms of what was even possible to build [5:50] like how has your product building [5:52] process changed over time. How is what [5:54] you thought possible then different from [5:56] what you think is possible now? I would [5:58] say when we started the business, we [6:01] were always obsessed with automating and [6:04] and putting putting the business on [6:05] autopilot. So that hasn't really [6:07] changed. Um back then the models weren't [6:10] that good though. Uh especially uh [6:12] especially reasoning models or didn't [6:14] really exist in that form. Um but if you [6:17] think about what it takes to automate [6:18] any job, you're really looking at um [6:21] getting context on the work, which you [6:24] know in the case of a doctor office is [6:26] basically you need to have access to all [6:29] the historical data, the patient [6:30] records, uh that sort of stuff. Uh and [6:32] then you need uh tools to do the work. [6:34] Um you know, this is true if you're a [6:36] human in the office or if or if you're [6:38] an agent. Um and so we started building [6:41] uh building the context layer and [6:42] building the tools and it and it's just [6:44] that the intelligence layer wasn't that [6:46] intelligent. Uh but for the first uh for [6:49] the first job it wasn't that necessary [6:51] like the the most basic kind of uh you [6:54] know automation didn't require that much [6:56] reasoning. Um but then as the models got [6:58] really good we we had this kind of huge [7:00] tailwind because we could uh we had all [7:02] all this context already built uh all [7:04] the tools already built and we could [7:05] kind of you know as as the models got [7:07] better just replace our intelligence uh [7:09] and and the product would just get [7:11] smarter over time. Um so in that way we [7:13] we got a little lucky but I think the [7:15] you know the the core vision hasn't [7:17] really changed at all. It was always [7:19] about automating the work and not uh not [7:21] building tools for for them that they [7:22] would have to use. I feel like as the [7:24] models have improved, we are more and [7:26] more seeing software do the job of [7:28] labor, which Alex, I would say you were [7:30] the first to argue and famously argue [7:32] that that would be the case. Curious [7:34] like how you think about that when you [7:36] look at companies and maybe how it [7:37] played into like the thesis around [7:39] Lassie. [7:40] >> Yeah. Well, so I've given this whole [7:42] presentation on the origin of software [7:44] was basically take a filing cabinet and [7:46] put it in it to a database [7:48] >> and kind of pick the uh the the time [7:50] equals zero moment for that with this [7:52] company the Saber Systems because [7:55] airlines would just keep reservations [7:57] and filing cabinets. Saber Systems was a [7:59] joint project between IBM and American [8:01] Airlines. That's why Saber is spelled [8:02] with two A's, Saber. Um but then this [8:05] kind of took wind everywhere else. So [8:07] like there are HR filing cabinets and [8:09] that became something like Peopleoft. [8:11] There are legal filing cabinets and that [8:14] became you know all of these Lexus Nexus [8:15] products. There are accounting filing [8:18] cabinets and that became Quickbooks and [8:19] that became Netswuite. So every that was [8:22] the origin of software. So software just [8:24] kind of took things that were stored in [8:25] paper format and then they made them [8:27] available first on prem via green screen [8:30] computers because that was a lot more [8:32] efficient to book an airline ticket and [8:33] change it if you didn't have to use an [8:34] eraser anymore. uh starting with Saber [8:38] and um but people still had to do the [8:40] work. So I I would actually argue that [8:42] the world didn't get that much more [8:44] efficient with software because all that [8:47] software did was like take HR like did [8:50] Peopleoft and then workday make HR [8:53] efficient make make HR departments more [8:55] efficient like I don't think so because [8:56] the same number of people worked in HR [8:59] for the exact same size company in like [9:01] 1950 as probably 2000 and instead of [9:04] using filing cabinets that are you know [9:06] that are guarded by Stein and Frederick [9:08] you know making sure that nobody breaks [9:09] into the HR files. Now you have an IT [9:12] department and a CISO to make sure [9:14] nobody hacks into the IT files or the [9:16] you know the HR filing cabinet. So [9:17] nothing really got more efficient. I I'm [9:19] somewhat exaggerating for effect here. [9:21] But what you can now do with software is [9:25] it can do it can edit the filing [9:27] cabinet, right? It's no longer just the [9:28] dumb storage. It's actually like the [9:31] smart implementation of changes against [9:33] those things. So like if it's HR, let's [9:36] do a background check, right? Right? Or [9:37] let's do an onboarding or let's explain [9:39] the benefits to this person. If it's [9:41] accounting, what do you do with the [9:43] financial statements? Like imagine I'm a [9:45] dentist and I see I have all these [9:46] overdue invoices and I can look up look [9:48] up in QuickBooks. What do I do? Well, I [9:50] might want to call and say please pay [9:51] me. Like that's what the filing cabinet [9:53] should be doing and not just giving you [9:54] the information. So it just turns out [9:57] that the work is orders of magnitude [10:00] bigger than the storage of information [10:02] that the work is done on. Um so that [10:05] that's really the been the thesis and [10:07] that you know you need the technology to [10:09] catch up so it can actually do it. [10:10] >> Mhm. [10:11] >> Um because in 2023 [10:14] it's like you know next word prediction [10:16] wasn't really good at like going and uh [10:18] which is basically what AI is. Uh that [10:21] was not good enough to go say I'm going [10:22] to go run my practice or you know do [10:25] background checks. I know I'm going to [10:26] have statistical, you know, inference [10:28] play out and that's how I'm going to do [10:30] a background check and make sure that [10:31] Frederick didn't commit any crimes [10:33] before I go hire him for my company. [10:34] Like, no. But now things have gotten [10:37] good enough and that just massively [10:38] expands the market size. And if you [10:40] think about fintech, fintech massively [10:43] expanded the size of many non-financial [10:46] markets because now you could you could [10:48] bundle in financial products with [10:51] non-financial products. And what I mean [10:52] by that is like my favorite example of [10:54] this is Toast. And I know you and I have [10:56] talked about this a bunch, [10:57] >> but like Toast could have existed in [11:00] 1985. [11:02] >> Like you know, everybody had an IBM PC. [11:04] They worked pretty well. Microsoft DOS [11:06] worked pretty well. Why didn't why [11:07] wasn't there a restaurant software [11:09] company in 1985? Well, number one, it [11:11] was too hard to use. But then number two [11:13] is you have this cact issue because [11:16] could you get a big restaurant that [11:18] grosses $5 million a year to spend [11:21] $100,000 on an MS DOS software product [11:24] for keeping reservations and paying [11:26] weight staff and you know having a [11:27] little menu that showed up for the cook [11:29] so they can you know make your hamburger [11:31] more quickly or something. Um nobody [11:33] would pay $100,000 for that. But if you [11:35] bundle in payment processing, you're [11:38] effectively charging $100,000 for that, [11:40] right? Because maybe you get a 2% vig. [11:42] So fintech made the market much much [11:44] bigger for software because of this [11:47] bundling effect. Um, and that pales in [11:51] comparison to now software doing the job [11:53] of labor because it's like, yeah, [11:55] fintech made it a little bit bigger, but [11:56] now instead of just being a dumb pipe [11:59] for data or a dumb storage of data and [12:01] instead of just like charging [12:02] incrementally more by bundling in, you [12:03] know, financial processing, now we can [12:05] do work and we can charge for work and [12:08] we can charge for work at a way that is [12:10] cheaper than humans, better than humans. [12:12] But I think both of those sell the [12:13] opportunity short because in many cases [12:15] you can't even find a human. actually [12:17] the the best and funnest story of the [12:19] the lassie um introduction or like the [12:22] our announcement that we made together [12:24] with you or your your announcement your [12:25] your amazing video. [12:26] >> It was a great collaboration. [12:28] >> Um so um my my first dentist hopefully [12:31] he's listening to this podcast. His name [12:33] is Ronald Sloop. Um it was my parents [12:36] like first friend when they moved to [12:37] Florida. I'm from Florida. Um he retired [12:40] as a dentist. [12:41] >> Was he a Dutch uh guy? He sounds very [12:43] Dutch though. [12:43] >> No, you know, Ashkanazi Jew from, you [12:45] know, somewhere in Poland, Ukraine, [12:47] whatever my family's from, too. So, you [12:49] know, uh, but, you know, he's probably [12:52] 75, 80 years old right now. But part of [12:54] why he retired was he lost his like, you [12:57] know, key woman that did the books and [12:59] everything else. He's like, I can't deal [13:00] with this anymore. I quit. Yeah. [13:01] >> And then he sold his practice to his [13:03] like junior practitioner and now he's [13:05] out of the dentistry business. So, he [13:06] saw this announcement. He's like, oh my [13:07] god, this is amazing. And I'm not [13:09] talking this up because your like my dad [13:10] called me about this because he saw the [13:12] prosperity. Wow. I talked to Dr. Sloop [13:14] about this. And he said that if this had [13:16] been around, he wouldn't have retired. [13:18] >> Like this is why he's now like doing [13:20] nothing with his life. U just like you [13:22] know playing golf or something in [13:23] Southern California. He moved there from [13:25] Florida apparently. Um because it's just [13:26] too hard to hire the person. So it's not [13:28] like oh AI is going to take the jobs. In [13:30] many cases you can't find somebody. This [13:32] is the part that people don't realize or [13:35] you can't find somebody but there's like [13:37] a imagine that there is something that [13:40] every human on earth would pay a dollar [13:41] for but the cost of manufacturing that [13:44] thing is $100. You just have a market [13:46] failure and I kind of call this [13:47] everything to the right of the supply [13:48] demand equilibrium point on like an econ [13:50] 101 graph. So, it's like, wow, everybody [13:53] would, you know, have somebody like, why [13:55] isn't there a a Dutch receptionist at [13:58] every dentist in America? [14:00] >> Because, you know, there might be a guy [14:01] that only speaks Dutch that shows up. [14:03] Yeah. Like, why not hire somebody who [14:05] speaks Dutch? Well, because there's only [14:06] a one in aundred chance that a Stein [14:08] that only speaks Dutch shows up at their [14:10] office. You're going to have to pay that [14:12] person 40,000 euro. Like, it just [14:14] doesn't make sense. But if it were free [14:16] or if it cost a dollar, then every [14:19] dental receptionist would have a Dutch [14:21] counterpart, right? You know, stuff like [14:23] that. So, it's just it's anyway, so that [14:25] that's where the the market just expands [14:27] massively once you throw in labor [14:30] because it's like you have this tiny [14:31] tiny market for software, which by the [14:33] way is not that tiny. It's like a [14:34] trillion dollars. [14:35] >> Um, concentrically around that, you have [14:37] like you know big on financial [14:39] transactions that's even bigger. That's [14:40] why Visa has a very big market capital. [14:42] >> White can exist or businesses. But then [14:45] you go the the the concentric circle [14:47] around that is just like you know it [14:49] really is orders of magnitude bigger. [14:51] >> Yeah. [14:52] >> And we also see that so like the um the [14:56] there about like 160,000 dental [14:58] practices in the US alone and like they [15:00] spend roughly $200,000 a year on like [15:03] administrative costs. And then the [15:05] interesting part is that because we [15:06] serve hundreds um already like they [15:08] can't find people. So like what Alex [15:10] like shared is we come across that [15:11] literally every day that like it is the [15:13] doctor themselves with their Harvard [15:15] degree that sits there till like [15:16] midnight uh and it makes them like uh [15:19] not like their job anymore. Um so I [15:21] think that's that's very interesting [15:23] because like these small business owners [15:25] they just want to mainly spend time on [15:26] their patients. Um they don't really [15:28] want to spend time on uh the [15:30] administration part let alone like [15:31] working with Betty and then in this case [15:33] it's Wilder they can't find Betty. Um, [15:35] so we combined I think a lot of people [15:37] are also surprised that but isn't there [15:39] a lot of skepticism? It's like no these [15:41] people are in real pain and they are to [15:43] Alex's point about to quit or just like [15:46] they hate their job uh at least this [15:48] part of the job. Um so we come by and we [15:50] say hey we actually have built this [15:52] agent that can provide you already with [15:54] tens of hours of labor um and ask your [15:57] friends or the people in your study club [15:59] like if this is real and then uh or they [16:01] Google it and they see that this is [16:02] real. [16:03] >> [snorts] [16:03] >> um they then adopt it like very quick. [16:05] So it's it's super interesting to see. [16:06] And then uh indeed like on the why [16:09] that's an interesting business because [16:10] um uh this is indeed it comes out of the [16:13] P&L uh on the uh on the on the labor [16:16] budget. Um so like we are already [16:18] charging five figures for kind of like [16:20] this first agent that only does 30 hours [16:22] uh of labor a month and there's 200 [16:24] hours of labors to be done for Dr. Loop. [16:27] Um and and that's like really [16:29] interesting like to see that like they [16:30] see us as um someone they break in to [16:34] like actually run the practice for them [16:36] which is also on the other hand [16:37] complicated to do because if all of a [16:39] sudden the requirement for software [16:41] becomes hey this is not a tool that I [16:43] give Dr. Sloop and then Dr. the loop is [16:45] still on the line. In fact, one could [16:47] argue a lot of AI companies are still [16:48] like that. There is a human in the loop [16:50] that ultimately the software engineers [16:51] decide what's get deployed. Um like we [16:54] can't do that. So we needed to build an [16:56] agent. That's why it took us years that [16:59] like airs on the side of correctness [17:00] because like if you take over a job uh [17:03] reconciling all the insurance payments, [17:05] interacting with the patient to kind of [17:07] like bill um it needs to like work. Um [17:10] so that was like technically like hard [17:12] to do. Uh, which also makes this super [17:13] interesting from a technology [17:14] perspective because you all of a sudden [17:16] need to build autonomous systems that [17:18] run on its own and don't have a human in [17:20] the loop. It runs the business for Dr. [17:22] Sloop, which makes it technically super [17:24] interesting. [17:24] >> Yeah, there was a a great quote from Dr. [17:27] Quan about you guys, which is that [17:29] Lassie isn't replacing humans, but like [17:31] freeing them from wearing so many hats. [17:33] And [17:34] >> your launch video had a clip of him [17:36] talking about how he can actually coach [17:38] his kids soccer teams now and go to his [17:40] their games, which is amazing. [17:42] I there's a gap between wanting that and [17:45] being willing to adopt AI and actually [17:48] you know having it run payments in a [17:50] practice and in fact doing it so well [17:52] that most of your growth is word of [17:53] mouth so it's dentists recommending it [17:55] to other dentists. [17:56] >> They do. [17:57] >> How did you approach the technical build [17:59] process for that? What was it like [18:01] getting the product to I think you guys [18:02] are at 98% automation. Like walk us [18:06] through kind of that that journey. Yeah, [18:08] I think um a big part of it was us [18:11] actually spending the time in offices [18:12] doing the work ourselves. Um I don't [18:15] think we could have built uh built a [18:17] product that works as well as it does if [18:19] we if we didn't know how to do the job. [18:21] >> Um I think another part we already kind [18:24] of talked about it, but I think a huge [18:25] difference uh between SMBs in general [18:28] and and enterprises is that in SMBs [18:30] there's nobody to to use the tools. Like [18:32] you can you can build a tool, but uh [18:34] there's nobody sitting there that that's [18:35] going to use it. And so uh [18:38] >> loop at night. [18:38] >> Yeah. [18:39] >> Yeah. [laughter] [18:40] Needs to go into the tool. [18:41] >> We uh from the very beginning we we [18:43] focused on um you know initially we were [18:46] actually the the humans in the loop. So [18:48] we we kind of took over all of the work [18:50] and we were like you know we'll we'll [18:52] just do this work for you and uh we and [18:53] we kind of automated away our own [18:55] problems. Um, and then at at some point, [18:58] you know, we we got to uh a high enough [19:01] level of automation that that we felt uh [19:03] we felt comfortable handing it back back [19:05] over to the the remainder back over to [19:07] the office. And I think now um you know, [19:10] we learn obviously when uh when we can't [19:12] uh do something for for some reason, [19:14] which is pretty rare, but but say say we [19:16] we don't know how to do a certain case, [19:17] we learn from what what the staff tells [19:19] us. Uh and we also think about I think [19:22] um you know like we we want to get to [19:25] like a sufficient level of automation [19:26] across any product to uh before we sell [19:29] it. So you know for us I think that's [19:30] like 95 plus say but not necessarily 100 [19:34] like I don't think we're going to wait [19:36] until we get into uh to 100 with one [19:38] product and then do the next one. Um I [19:41] think the uh we really think about the [19:43] business more like as a whole like how [19:45] much of the business can we uh uh can we [19:47] automate and how much of the labor can [19:48] we can we do with uh with software. Um [19:50] and as as soon as we can take over a job [19:52] we take it over and then we move over to [19:54] the next one. Um and then over time [19:55] we'll just learn uh learn the long tale [19:57] of cases. [19:58] >> Yeah. [19:59] >> For a business it also doesn't matter [20:01] that much right like if um let's make [20:03] Dr. Sloop famous in this podcast. Let's [20:05] >> he's going to learn. [20:06] >> Yeah. If Dr. loop um said he can become [20:09] a customer. So maybe we should talk [20:11] about like someone [laughter] else [20:12] that's still active. [20:12] >> We could reactivate him. [20:14] >> Yeah, we he might he might come out. [20:16] That's how he sounded. [20:17] >> Yeah. Yeah. [laughter] Dang. That's the [20:19] series B story. We got Dr. Mac out of [20:21] sto out of [20:22] >> We had a dental shortage and now we [20:23] don't [laughter] have come out at [20:25] retirement. [20:26] >> We got 100 million more people in the [20:27] states that get good uh dental care. um [20:30] the um it's fine for like a business [20:33] owner if there is a tiny sliver left of [20:36] claims that need to be touched every [20:37] week, right? So like the way to see this [20:39] because they live in a nightmare world [20:40] where you have like to update 200 [20:41] ledgers a week because that's how much [20:44] patience you see and then you have to [20:45] like go to an insurance portal update [20:47] like a system of record check against [20:48] the bank account 200 times a week. If [20:50] you instead need to do that like a [20:52] handful of times rather than 200 times, [20:54] >> it saves 10 20 hours. um if there are a [20:57] handful of uh claims that you need to [20:58] file yourself, but the majority kind of [21:00] like is on autopilot makes it [21:02] tremendously more easy to run a [21:03] business. Um I often compare it with [21:05] like how we are served as tech [21:07] companies. [21:08] Um it's the third company I'm building [21:10] and it is a lot easier and with a lot [21:12] smaller team than we used to do kind of [21:14] like 10 15 years ago and it is because [21:17] there's great tools for us that we can [21:19] use. um uh are all these tools [21:21] completely like running our finances [21:23] autonomously yet or our payroll or HR? [21:25] No. But they do save a tremendous like [21:27] amount of time. So I think that's how we [21:29] approach like this like built as well [21:31] that we didn't want a human in the loop [21:33] because we want software that skills and [21:35] can be implemented quick. Um but it's [21:38] fine if it does in this case for the [21:39] first agent 98% of the work and then [21:41] there's a sliver left. And then the [21:43] interesting thing like we have thousands [21:44] of staffers that are basically giving us [21:46] input on how to make that appeal that [21:48] the agent currently cannot do. Uh we [21:50] come from the consumer world, right? [21:51] Like Frederick worked at Superhuman. Um [21:53] I worked at Robin Hood. So we have a [21:56] very high bar for shipping stuff. Um so [21:58] we don't release it before it actually [22:00] like okay this is this is good and it [22:02] can be hands off and it like works and [22:03] then these like staffers help us to kind [22:05] of like um get it to like a even higher [22:08] percentage. [22:08] >> Yeah. From an implementation and [22:11] onboarding perspective, you guys [22:13] integrate with existing practice [22:15] management systems for the most part [22:17] versus kind of making them switch a [22:19] bunch of software to adopt Lassie. And [22:22] Alex, you have written and talked a lot [22:24] about kind of startups getting [22:25] distribution before an incumbent can [22:27] innovate. Curious your thoughts on like [22:29] that in the AI era and then would also [22:32] love to hear from you guys how you [22:33] thought about which path to take there. [22:36] Yeah, I mean I had this epiphany when I [22:38] was building my company which is holy [22:40] crap like there really aren't if you [22:42] build something I call this the to [22:44] problem and to famously to and replay TV [22:47] both invented the digital video recorder [22:49] so you could pause online television [22:51] >> which is an amazing innovation but a [22:53] terrible company because you really have [22:54] very few outcomes that are good. you [22:57] either end up selling to one of the the [23:00] big guys like a Comcast or Time Warner [23:02] Cable, but they're not going to pay you [23:03] that much partially because if if [23:06] Comcast bought you um you you start Teo, [23:09] Comcast buys you. Well, all of the comp [23:12] all of the competitors to Comcast, [23:14] they're like, "Well, we're going to not [23:15] allow this to work." So, like uh you you [23:18] have what what I call a control discount [23:19] versus a control premium. So, that's [23:21] option one. Option two is they copy what [23:24] you've done many years later. are much [23:26] crappilier if that's a word because they [23:29] have all the customers. Um or you know [23:32] may maybe number three you do a [23:33] licensing deal with them [23:35] >> and uh they take all the economics [23:37] because they have all the customers. So [23:38] that hence you know my my recognition [23:41] was like the thing that a lot of [23:43] startups should do is they should do the [23:44] boring thing like they should build like [23:46] the raw pipes the raw like just own the [23:49] customer and then you get to build the [23:51] fun feature on top. Um which I still [23:53] stand by. I mean I like the vast and and [23:55] this is why like maybe four years into [23:57] trial pay I realized what I should build [23:59] is this thing called stripe and this was [24:01] not like revisionist history because [24:02] stripe had five people it's like wow we [24:04] should we should do boring payment [24:05] processing which is a commodity business [24:07] because if we do that then we own the [24:10] customer chronologically you get them [24:12] first it's a very very boring thing but [24:14] then we had this other thing which in my [24:16] case was offerbased payments which was [24:18] very lucrative but you can only do that [24:20] if you control the pipe just in the same [24:22] way that you can only build the digital [24:23] video recorder if you have digital video [24:25] to record. Um, so how does this change [24:27] with AI? It changes with AI because in [24:30] many cases these are non-categories like [24:32] there is no incumbent. So for most [24:35] categories like imagine that I say I [24:36] have a great idea. I'm going to do [24:38] background checks for new employees as [24:40] part of onboarding and I'm going to [24:42] integrate with workday. Workday being [24:44] the b the biggest HR information system. [24:46] That's a great idea. Um however um it's [24:49] such a great idea that it's a very [24:51] obviously great idea that this thing [24:52] called workday might copy and they own [24:54] all the customers and that's where it's [24:55] like you know battle between startup and [24:57] incumbent it's like the incumbent might [24:59] win there because their ability to add [25:01] things and I feel like that is actually [25:03] magnified in the AI era because you can [25:05] have like why are big companies not good [25:08] at replicating small companies? There [25:09] are lots of different reasons, but the [25:11] ones that one of the reasons is they [25:13] hire very bad engineers and they have [25:14] lots of process, but now AI kind of [25:17] makes a bad engineer into like a pretty [25:19] good engineer, you know, kind of. Um, so [25:23] that excuse kind of goes away a little [25:25] bit. But this is the cool thing about a [25:26] lot of the AI software companies or the [25:28] AI that does the work. Like who is the [25:31] giant ass incumbent of dental software? [25:34] You and I know the answer on this, but [25:35] it's not it's not the same thing as it's [25:37] like here's workday. It's a tech [25:39] company. They already have software and [25:41] they can add something to I I remember [25:43] actually this is a cool story. There was [25:45] a company I think it was called X1. [25:47] Microsoft Outlook had really bad search. [25:49] >> And this company that was funded by [25:51] Ideal Lab and all these VCs, you know [25:53] what they did? It was like search for [25:55] your Outlook email, which was so good, [25:58] [clears throat] but it's like, you know [25:59] who I think is going to do this [26:00] eventually? I think Micro and again like [26:02] that company unfortunately went to zero [26:03] or actually Yahoo bought that long time [26:05] ago. So I kind of think the same rules [26:08] apply. You know, the battle between [26:10] every startup and incumbent comes down [26:11] to whether the startup gets the [26:12] distribution before the incumbent gets [26:13] the innovation. But with many changes, [26:16] one change is the incumbent can get the [26:18] innovation much more quickly. Um [26:20] >> uh but the other is that there are a lot [26:22] of categories where there never was an [26:25] incumbent software company because the [26:27] only job to be done was like actual [26:30] human labor. And that's really exciting [26:34] because now you don't have to worry [26:36] about like oh shoot these guys are going [26:37] to come in and eat my lunch like who [26:40] right like who does like there are a lot [26:42] of industries that just don't have an [26:43] incumbent software solution for the [26:46] industries that do have an incumbent [26:47] software solution yeah the risk is very [26:49] high that uh they will start releasing [26:51] AI features and it's a really [26:53] interesting way that the market is [26:54] playing out right now because if you [26:56] look at the public markets the public [26:58] markets are saying in many cases like oh [27:00] you're a software company software is [27:02] dead, software socks, software is zero. [27:04] Oh, you're an AI. It's the opposite of [27:06] what VCs are saying. It's like, oh my [27:07] god, you do AI stuff, but AI is [27:10] software. Software is AI. Like the two [27:12] are the same. Um, but like I don't think [27:14] everybody's come to that realization [27:15] yet. So like if you're doing, you know, [27:17] workday but AI, if you're doing, you [27:19] know, Netswuite, but AI, it's like, you [27:21] know, the AI, the pure play AI thing is [27:24] software at its core. and the the Pure [27:26] Play software thing. It's like they're [27:28] they're smoking crack or not showing up [27:30] to work if they're not working on [27:32] implementing AI features because that's [27:33] what their customers are demanding of [27:34] them. [27:36] >> Yeah. And we see exactly that is that um [27:38] there isn't an incumbent [27:40] >> that kind of like does this job or can [27:42] do this job like quickly. [27:44] >> Well, the incumbent was named Betty and [27:45] she quit two weeks ago. That's the [27:47] incumbent. [27:48] >> Yeah. Or or a billing agency. [27:49] >> It's Dr. Sloop's old assistant. that's [27:51] the incumbent [27:52] >> or or version of that that like is [27:54] somewhere overseas or in the states. Uh [27:57] so that's indeed exactly like what [27:58] you're competing with and that's what [28:00] intrigued us so much about these small [28:02] businesses because like there is no [28:05] major player there. [28:07] >> Um and then if you show up with software [28:09] that they have never seen before which [28:10] is now possible they will adopt it and [28:13] you can grow. Um which is uh very [28:15] defensible. how uh maybe we get to talk [28:18] about it later. The the schle you have [28:20] to do to like uh do the actual labor and [28:23] talk to all these like systems and you [28:25] need to build an ontology to make sure [28:26] that everybody in this whole ecosystem [28:28] is on the same page about an insurance [28:30] claim and the patient payment because [28:31] all these different systems have like a [28:33] slightly different definition of that. [28:35] Uh which makes it also then harder to [28:37] build because there is no incumbent. Um [28:39] you need to kind of like stitch like a [28:41] lot of things together but makes it [28:42] extra defensible. Um because like what [28:45] we had to do was like okay first figure [28:47] out kind of like all these read and [28:49] write integrations between all these [28:51] systems that Betty the the AI version [28:53] need access to uh then you need to [28:55] figure out like okay what is the data [28:57] model that can be used across like all [29:00] these like systems and then on top of [29:01] that you need to build agents that uh [29:04] yeah you can't really build without [29:05] actually doing the work. Um, so there's [29:07] like years of work that you need to do [29:09] to kind of like get that like going [29:11] makes it very defensible. Like the the [29:13] go to market side as well, right? Is [29:15] this like you need to knock on millions [29:16] of doors and say uh and the interesting [29:19] thing we just talked about that is not [29:20] necessarily that they're skeptical about [29:22] AI because Dr. Sloop is like, "Oh, I [29:24] wish this was there." Is how do you get [29:26] hold of Dr. Sloop? Right? Because Dr. [29:28] Sloop um is not done at 7 p.m. There's [29:30] an emergency patient that calls then he [29:32] goes back into the office to treat that [29:34] patient. uh dinner with the kids and [29:37] then opens the computer and then oops [29:39] the supplies need to be ordered because [29:41] Betty left so I need to do this myself [29:43] right now. So like for us but for any [29:45] company selling to SMBs the interesting [29:47] puzzle here that's why this is also uh [29:50] super interesting go to market work [29:52] because like how do you adopt and spread [29:54] AI in small businesses is a super [29:55] interesting puzzle to solve. Um but [29:57] that's that's the crux there that is [29:59] like how do you get bit of busy [30:01] nontechnical owners [30:02] >> that adoption. Well, I imagine the other [30:04] part of the question for you, I'd love [30:06] to hear your thoughts on this. Um, or [30:08] I'm sure our audience would love to hear [30:09] our thought thoughts on this, but this [30:11] is not like you download the AI app from [30:13] the app store and then you're all done, [30:16] >> right? Like how do you actually do the [30:18] onboarding? [30:18] >> Yeah. [30:19] >> Um, and how much of that can you [30:21] automate? Because that's part of what [30:23] makes a business that is selling these [30:25] things work or not work, [30:26] >> right? Because if you have to, you know, [30:28] send your own Betty [30:30] >> to every single office in the country. [30:33] Yeah, [30:35] >> it would be amazing if it's just like [30:37] download the lassie app [30:41] like these these systems and processes [30:43] are very manual [30:44] >> and um I kind of think like AI is [30:47] overhyped in Silicon Valley but [30:49] underhyped in Iowa [30:50] >> and like there are a lot of people in [30:52] Iowa like how do you solve that dist [30:54] that like glass mile distribution [30:56] problem for Lassie? [30:58] >> Yeah. Yeah. I think uh it's a super [31:00] interesting problem. assume you can [31:01] knock on all these doors and get them to [31:03] try it, right? Then like how do you get [31:05] this adopted? Because the same argument [31:07] still stands. They're very busy. They [31:09] are not technical. So like good luck [31:11] setting up AI in their like business. [31:13] And I think that's also why um um our [31:16] consumer backgrounds come in handy here [31:19] because at Robin Hood or Superhum, you [31:21] get the time window 48 hours. If the [31:23] thing doesn't work and you provide core [31:26] product value, um you're out, right? And [31:28] these doctors are very much the same. [31:30] because not only are they hard to reach [31:32] and hard to work with, but if it doesn't [31:34] work in a couple of like months from [31:35] now, like you're out of the door. Um, it [31:37] really needs to be plugged in and then [31:39] do the job. So, I think a lot of the [31:42] work went into not only building this [31:44] agent, but also how do you set people up [31:46] on that agent such that almost all the [31:48] friction is gone. Um, and and that was [31:51] like a lot of work and it's it's already [31:53] to a point that it's self-s serve [31:55] almost. Um so there are a few more [31:57] things like left but it's literally that [31:59] the doctor in Iowa let's not use loop [32:01] again um like says yes I want this um [32:04] they then uh go to almost like a stripe [32:07] like checkout or a repling like [32:09] onboarding like flow where they hook up [32:10] the bank account of the practice in the [32:12] product they link the system of record [32:14] they link all the insurance portals that [32:15] claims come in from um it confirms [32:18] business information like are these the [32:20] doctors that actually like work in your [32:22] practice um and then under the hood kind [32:24] of like configures like things. There's [32:26] like one or two pieces left, but I think [32:28] we're like months out until you [32:30] basically like have an agent that you [32:32] can almost like set up like self-s [32:33] serve. Um, so I think that was a big [32:35] part or is a big part of bringing this [32:37] technology to people in Iowa. Um, is [32:40] that can you almost build a [32:41] consumer-like uh on boarding like flow [32:44] where a lot of the complexity is [32:45] abstracted away and kind of like happens [32:47] under the hood like the [32:50] maybe not many people I think I'm very [32:52] impressed by Robin. I'm a little biased [32:54] because I work there. But like in order [32:56] to set up an account for a user without [32:58] a human in the loop, which is what Robin [32:59] Hood pioneered because back in the days [33:01] there was Charles CH swap and then hello [33:02] I want to open an account. Before that [33:04] you had to go into the office. Uh but [33:07] one of the many things that Robin Hood [33:08] pioneered is that you could you can [33:10] almost self-s serve your way onto an [33:11] account and KYC is getting done your [33:13] bank is getting linked. Uh and a lot of [33:15] product work went under the hood there [33:17] to make that happen. I think we ran into [33:19] a lot of these like similar like like [33:20] situations. is like how do you connect [33:22] all these insurance portals um and kick [33:25] off like the the uh the process such [33:27] that the digital claims are coming in [33:29] like reliably how do you reliably link a [33:30] bank account all these system of records [33:32] so I think a big part of this was [33:34] figuring that piece out yeah [33:35] >> yeah I guess to that point like there's [33:37] more you can build and are building for [33:39] dental practices then there's all these [33:42] other types of healthcare practices that [33:44] could use lassie and then there's like [33:46] the broader universe of small businesses [33:48] that could use you guys [33:50] >> how are you prior Prioritizing what you [33:52] build, who you sell to. Is there, you [33:55] think, a world where Lassie for dentists [33:57] makes lassie for physical therapists [34:00] better? [34:01] >> The master plan. [34:02] >> Yes. [34:03] >> Are we at that part of the episode? [34:05] >> Yes. [34:06] >> Um, yeah. I think three steps like the [34:10] end goal here is that uh every small [34:12] business should run itself, right? And [34:14] the busy work is done by uh agents. uh [34:17] we want to build agent for the business [34:19] that then will interface with the [34:21] personal agent of a consu consumer [34:22] highly likely that then will interface [34:24] with an agent at the insurance company [34:26] or other parties that the business needs [34:28] to interface and interact like with. Uh [34:30] but step one is to Alexis's point like [34:33] there are 160,000 dental practices in [34:35] the US alone $200,000 in labor that Dr. [34:38] Sloop and others can't find. So like um [34:41] serving that market first. You're [34:43] looking at a $1 billion in like a [34:44] recurring revenue as a market um so [34:47] we're that's like step one and then uh [34:50] likely like we will pick another doctor [34:52] office type um that like is not well [34:54] served and has a big temp um and and [34:57] needs consumer-l like product right [34:59] because what we discussed a big part of [35:00] not this you get the AI to work 95% [35:03] accurateish it needs to really work and [35:05] the onboarding needs to be as simple as [35:07] onboarding on Coinbase. um or stripe. Um [35:10] so like it will likely be another like [35:12] doctor office type like um and then [35:15] [sighs] uh the last uh part there is I [35:18] think uh we've then trained AI agents uh [35:20] to like run the small business and all [35:22] small businesses at an abstract level [35:24] like have a system of record they need [35:26] to read and write into. Um they all have [35:28] customers in doctor offices they happen [35:29] to be patients but it's interacting and [35:31] transacting around payments. You need to [35:33] book appointments. So I think the end [35:35] goal is if we served all the the doctor [35:37] offices that we help all the small [35:39] businesses across the world because [35:40] we're just the best in you know building [35:42] AI agents that uh salt of the earth [35:45] people or people in Iowa and Paduka [35:47] Kentucky and hopefully in Amsterdam like [35:49] down the line where I'm from um and [35:51] Germany Hamburg where Frederick is from [35:53] can can start like using as well. Um so [35:55] that's the that's the end goal but I [35:57] think again similar to superhuman and [35:59] Robin Hood we work that like laser focus [36:01] on getting one thing really really [36:02] right. Um and then scale it like from [36:05] there. Yeah. [36:06] >> But the end goal is to help them all. [36:09] >> Amazing. I love that as a master plan. [36:11] It's a good one. A big one. Um you both [36:14] have been part of scaling many important [36:17] companies in the past. Robin Hood and [36:19] Coinbase and Superhum among others. [36:23] And building a company in 2026 is like a [36:26] whole brave new world. It's so different [36:27] than ever before. What are like the [36:30] biggest things you've carried over? You [36:31] mentioned some of them already. And then [36:33] what have you kind of had to unlearn or [36:35] what do you think is new to being [36:37] founders right now? [36:40] >> There's a bunch of stuff that uh we're [36:42] applying now that I I learned at [36:43] Superhum. Um I think for one uh focusing [36:47] on uh the right ICP and being really [36:49] strict about who you onboard uh to [36:51] basically guarantee that uh they're [36:52] going to have a great experience. Um, I [36:54] think in our case it's uh particularly [36:56] important because if we onboard the [36:58] wrong practice and say we can't actually [37:00] automate that much of their work, then [37:02] now we're kind of stuck with this [37:03] customer that uh, you know, we we [37:05] claimed we're going to automate a bunch [37:07] of labor for them. We can't do it. Are [37:08] we going to do it? Are we going to [37:09] offboard them? Uh, it's it's [37:11] particularly painful, maybe more painful [37:12] than uh than in a kind of old world [37:15] product. Um, so there's that. There's [37:17] also uh we talked a bunch about the [37:19] onboarding already, but uh we we've [37:21] really obsessed about uh getting to uh [37:24] the core product value really quickly. [37:26] Um and almost like our onboarding is is [37:29] a little bit like a uh you know, it's [37:31] kind of like a story or a playbook or [37:33] like a movie. We we have like set uh set [37:36] points and checkpoints that we want to [37:37] reach in certain time frames and we make [37:39] uh we make sure it happens every time. [37:41] Um and uh you know and we measure that [37:44] of course. Um I think one thing that's [37:46] that's really different in particular [37:47] about product building uh from uh from [37:50] back then to now is that uh if you think [37:53] about the products that you built uh in [37:55] the past I think it was much more about [37:57] kind of functionality or like the [37:59] ability for the user to do something. Um [38:02] and now uh we think really only about uh [38:06] like what what kind of labor can we [38:08] automate and what uh or like work can we [38:10] do and where can we save time. So if [38:12] you're thinking about say uh like [38:13] patient billing as an example um I think [38:16] previously you would have built uh you [38:18] know the ability to send a statement and [38:19] the ability to receive a patient [38:21] payment. Um but if you're looking at [38:23] where does the actual work of patient [38:26] billing go today uh it's really like [38:30] figuring out is the money or is the is [38:32] the statement that we're going to send [38:33] the patient the correct amount and or [38:36] like once the statement is sent um you [38:38] know the patient calls and asks about [38:39] like why do I owe this amount? uh do I [38:41] really have to pay this? I thought this [38:42] would be covered. Uh so if you're just [38:44] looking at like where does the time go, [38:45] it actually goes in like oftentimes the [38:48] customer communication or or some other [38:50] kind of more like you know fuzzy fuzzy [38:52] part of the work. Uh and and when we're [38:54] thinking about like shipping say a [38:56] product like that, we're really thinking [38:57] about okay once we you know once we [38:59] deliver patient billing the office [39:01] should not have to spend any more time [39:03] on patient billing which is super [39:04] different to giving them a tool that [39:06] they then need to use uh which doesn't [39:08] really save them a lot of time. And then [39:10] neither of you I think were dental [39:12] experts before you started the company. [39:14] Although [39:15] >> that's that's safe to say. [39:16] >> You go twice a year. That's pretty good. [39:18] I think for the average [39:19] >> you're supposed to, right? [39:20] >> Yeah, of course. [39:21] >> I'm just doing my um [laughter] [39:23] >> not a saving duty to say. [39:25] >> Um when you're Well, maybe tell us like [39:28] how big is the team now? When you're [39:30] hiring, are you looking for expertise in [39:32] dental? What what are the kind of [39:34] characteristics of of team members that [39:36] you want to hire? [39:37] >> Yeah. uh which is now mainly on our [39:39] mind, right? Because like we found there [39:41] a really great product market fit in a [39:43] large market uh where you can go after [39:45] the labor um that they can find. [39:47] [snorts] [39:47] >> Um we we we currently have two takes on [39:50] that like one not much has changed [39:52] contrarian maybe take here. You still [39:55] need people with steep slope um that [39:57] like are very ambitious and and driven [40:00] uh and have skills that are just like [40:02] top five percentile either in [40:03] engineering or in selling, right? Um and [40:06] I think that remained the same getting [40:08] hold of doctor um slope um is just like [40:11] the same exercises before you could do [40:13] it in a more AI like native way. Uh but [40:15] I think the skills are the same and [40:16] you're uh so I think that has not [40:18] changed. Maybe the only thing is that [40:20] the AI pilness of a person. Uh I think [40:23] that we see a pretty clear a clear [40:25] division between uh across the board [40:27] like um do you believe that you know the [40:30] way you code will change completely as a [40:32] result of building a company in this [40:33] era? Um building out a finance [40:35] department uh do you think that's going [40:38] to be completely different uh than [40:40] before? So in that way we we interview [40:42] specifically for for that uh because we [40:44] want to build a 2026 version of a of a [40:47] of a big organization right uh where [40:50] like we ship twice as much than others. [40:52] We move like four times as like fast [40:54] because uh we should not only like have [40:56] AI adopted in these businesses but the [40:58] big puzzle for us is how do you build a [41:00] team um that kind of like um also [41:03] incorporates AI in all these like [41:04] functions. So I think that yeah on one [41:07] hand nothing has changed um because yeah [41:09] the bar is still the bar right the I [41:11] used to be track and field runner um [41:13] almost became a professional like track [41:15] and field runner took a different path [41:17] in life um but my friends went on to the [41:19] Olympics uh and um yeah Olympic like uh [41:23] running is still the same right you need [41:25] to train twice a day you like like push [41:27] it to the edge and that's not given like [41:30] to everyone mentally and physically so [41:32] like that I think has not changed in [41:33] company building I think what you can do [41:35] and the output you can generate is just [41:37] like four or five times but that's the [41:39] big experiment that we're doing together [41:40] right it's fun to see okay how quick can [41:43] we get to all the dentists in America [41:45] and build a really good product um such [41:48] that 200 hours gone like how quick can [41:50] we then bring it to another vertical and [41:52] then help all small businesses um I [41:54] think that uh that's that is the big [41:57] interesting experiment that we are going [41:58] to do over the coming years [42:00] >> if it's so much easier if everybody can [42:03] hire Betty [42:05] right? Just materialize a Betty. [42:07] >> Yeah. [42:08] >> Um how does that change? I mean like it [42:11] could actually work out where like small [42:13] businesses it's much easier to start one [42:15] and run one but then actually [42:16] paradoxically it's much harder to be one [42:19] >> because you do have if you think about [42:21] moes in the AI era in general [42:24] >> we often talk about it with respect to [42:26] software companies. [42:27] >> Y [42:27] >> um so it's so easy to go replicate XYZ [42:30] software. I did it on Replet or I did it [42:32] on lovable or I did it on Claude. Like [42:34] you hear this right, you know, left and [42:36] right all the time. [42:37] >> Um much much harder to say I'm going to [42:40] go replicate Dr. Sloop's practice. [42:42] >> Yeah. [42:42] >> But one of the things that makes a small [42:44] business somewhat defensible is actually [42:46] it is an accumulation of people that are [42:48] required to deliver the end product. [42:50] >> Yeah. [42:51] >> And it's like uh if you know who Yogi [42:53] Barra is, Yeah. you know, famous Yankees [42:55] baseball player that said all these [42:57] things that make no sense. [42:58] >> And um but like [42:59] >> quoted often though quoted often, right? [43:01] And my one of my favorite ones, it's so [43:03] crowded nobody goes here anymore. [43:04] >> Yeah. [43:05] >> Yeah. It doesn't make sense. Um but I [43:08] guess my question is how do you think [43:10] small business changes if it's easier to [43:14] run a small business and start one? [43:16] Because theoretically that could erode [43:18] the margin of a small business such that [43:20] it's so crowded nobody goes here [43:22] anymore. M um and like this one moat [43:24] that exists of just materializing people [43:27] to deliver a product, it's now so much [43:30] easier, but therefore it's actually [43:31] harder. [43:32] >> Yeah, I think our our current take on [43:34] that is that this assumes there's a cap [43:37] on kind of like demand and if you just [43:39] look at the dental practice, but the [43:40] same applies to try to find a good [43:42] plumber. um there's just like twice as [43:45] much demand that currently can be [43:47] supplied and I think this is a great [43:49] opportunity for you know everybody to [43:50] have great dental care and go twice a [43:52] year um and I think the same can be said [43:55] about primary care doctors. I grew up in [43:57] the Netherlands. I had a very different [43:58] primary care experience than like most [44:00] people in America here. Um because it's [44:03] just really hard to find the good [44:04] primary care doctor that is in your [44:06] community, knows you, your family, um [44:08] and takes takes care of you. So I think [44:11] this we see as an opportunity to kind of [44:13] like create like now there are 160,000 [44:16] dentists or hundreds of thousands of [44:17] dentists. What if uh America has half a [44:20] million dentists um or can see kind of [44:23] like uh twice as much patients maybe the [44:25] same amount of dentists? Plumbing like [44:27] is the same and I think there's a lot of [44:29] these like small businesses if they they [44:32] wish they could you know bake more pies [44:34] um but they're constrained on labor [44:35] basically. And I think what this unlocks [44:37] is that people have twice as much time [44:39] for the craft and that is an exciting [44:41] future. Um because then all of a sudden [44:43] yeah you're going to see a better world. [44:45] I think that's that's currently our view [44:46] which is super exciting um about this [44:49] technology. I think that's also to your [44:51] point like when we launched and we [44:53] finally told the world hey this is what [44:54] we've been up to. Uh that was the main [44:57] uh piece of like um a lot of people [44:59] talked about that it's like wow an [45:00] optimistic example um of like how this [45:04] new technology like can be used. uh [45:05] because there's a lot of like what's [45:07] going to happen to the world. Um [45:09] [snorts] but yeah, nobody really can be [45:10] against cleaning up busy work for small [45:12] business owners that should be baking [45:14] pies or polishing nails or cleaning [45:16] teeth or drilling. Um uh depending on [45:20] like who you are, of course, but uh I [45:22] think that's so interesting about about [45:24] this. Yeah. [45:24] >> Yeah. So you started the business in [45:26] 2020 and that was like arguably pre-AII [45:29] or pre what we think of as being this [45:32] generative AI revolution uh which I I [45:34] would call like the the BC80 divide of [45:36] like November of 2023 when chat GBT [45:39] launched right so um if you or I guess [45:42] 2022 sorry uh November of 2022 when when [45:45] chat GBT launched public what is what [45:49] are the kind of remaining so you know [45:52] then you had reasoning models you have [45:54] all of these things that have kind of [45:55] built on top of the original revolution [45:58] of you know four years ago call it like [46:00] what what are the hardest problems to [46:02] solve like what is it when we talk about [46:03] like software that does the job of labor [46:05] what cannot be done right now what do [46:07] you feel like we still need more [46:09] technical advance to get there and [46:12] sometimes it's like a 9010 thing where [46:15] it's like you know the last 10% is [46:17] really really hard but you can't be a [46:20] feature complete solution until you've [46:21] done that so I'm just kind [46:22] [clears throat] of curious like from a [46:23] technical lens [46:24] What are the things where you would say [46:26] okay and it's not Alassie specific [46:29] question it's just kind of more the [46:30] technology and what it enables at large [46:33] where does work still need to be done [46:35] >> where it's just not quite good enough um [46:38] and then what do you think the curve of [46:39] that looks like when so it's not you [46:41] know it's like a question of AGI for [46:42] small business like you know what do you [46:44] need and where are we on that curve if [46:47] you had to estimate [46:49] >> I think one thing that's interesting is [46:50] that the models are trained on so much [46:53] data and they're they're so large and [46:55] yet they actually don't really know how [46:57] to do any of this work. Like they don't [46:59] have the uh workflows encoded in any [47:02] way. Um so for example, we're working on [47:05] a product now where we we have to like [47:08] collect all of these like basically SOPs [47:11] on like and documents about like how are [47:13] you supposed to bill insurance claims to [47:15] certain payers and all this kind of [47:17] stuff. uh which uh to some extent humans [47:20] would do the same but there's also a big [47:22] amount of um just like human knowledge [47:25] that is encoded in say these office [47:27] managers and they just like know how to [47:29] do this work uh that's weirdly not that [47:31] accessible on the internet um [47:34] >> I think we have a big advantage there [47:36] because we have um all of this like [47:38] historical data out of their ERPs that [47:40] we can look at and kind of infer you [47:42] know some of these workflows from um but [47:45] that's something we notice a lot I think [47:46] actually when we [47:48] started using some of the, you know, [47:49] later reasoning models, uh, we kind of [47:51] assumed like, oh, they probably just [47:52] know how to do this work because like [47:54] why would they not, right? Like they're [47:55] trained they're trained on on all of [47:56] this data. Um, but it it it turns out [47:58] that they they don't know all the [48:01] intricacies of most of these workflows. [48:02] I think um, yeah, this is like less [48:05] specific to Lassie, but I'm I'm [48:07] personally kind of excited about uh [48:10] smaller models that uh, you know, learn [48:12] faster and can learn on less data. I [48:14] think uh that will be uh that will be [48:16] really cool to see because I you know I [48:18] think over time um we'll have [48:21] intelligence kind of I think [48:22] disseminated everywhere uh and uh you [48:25] know like to the point where like you [48:26] know like the the Pixar lamp that has [48:28] its own personality like why why not I I [48:30] I feel like I I want my intelligence to [48:32] be to be everywhere. Um and it feels [48:34] like we're super far from that. Um so [48:37] I'm you know curious about curious about [48:38] that. I don't I don't know if it'll [48:40] benefit us necessarily that that much. [48:42] Uh but there there is something about um [48:45] you know the models not being able to [48:47] learn very uh very very fast. Uh and uh [48:51] you know as it relates to us like we we [48:53] put a lot of product work into how do [48:56] you actually get input from uh you know [48:59] from people uh about you know their [49:01] preferences and like uh how they've done [49:03] done the work uh in a way that we can [49:06] then kind of like useful input in a way [49:08] that we can then use to actually make uh [49:10] make our our agents smarter. Um and I [49:13] think from a you know like product UX [49:15] perspective I think there's a bunch of [49:16] interesting stuff to to solve there as [49:18] And maybe one kind of final question [49:19] between technical and non-technical is [49:22] uh I I've been thinking about this a lot [49:23] how how the world changes when the [49:26] marginal cost of arguing goes to zero, [49:28] >> right? So I was thinking about this [49:31] because Sigma, who I have for my health [49:34] insurance, will only send paper checks. [49:36] I was like, "Oh, I must have missed the [49:38] whole like online enrollment." Nope. [49:39] Nope. They don't have one. Why don't [49:41] they have one? Well, they're kind of [49:43] hoping that like you might lose the [49:44] check, you might not deposit it. It's [49:46] just like this kind of like intentional [49:48] delay. And it's the same thing for like [49:50] a lot of insurance. It's like, you know, [49:52] they're it's like deny, deny, deny. They [49:54] want to deny. Um, and like on the other [49:56] side, it's like, you know, pretend that [49:57] when my house burned down, I had a [49:59] Picasso in there. Like both parties are [50:00] trying to cheat each other. This is not [50:02] new to the insurance agency or the [50:03] insurance like industry and both sides, [50:05] right? [50:06] >> But now that everybody has this like [50:08] superpowered thing that costs [50:09] effectively nothing to go argue in [50:12] perpetuity, it's like I can argue with [50:13] you and then you can argue with me. [50:15] Yeah, [50:16] >> but like how does that change the [50:18] business dynamic of things like [50:20] insurance payments and collections? I [50:22] mean, I'm sure you thought about this a [50:23] lot [50:24] >> because in many cases the counterparty [50:25] that the dentist is dealing with is the [50:28] insurance carrier. [50:29] >> Yeah. Right. [50:29] >> A big part of the accounts receivables [50:31] or or revenue comes from insurance [50:33] companies, [50:34] >> right? [50:34] >> Um yeah, I think it's quite interesting [50:38] because um it's pretty like well [50:40] regulated. So [50:43] um if a dentist does a crown like and [50:46] you provide them with the correct [50:48] narrative, Frederick was already talking [50:49] about it. There is just a document at [50:51] Sikna that specifies if you give me you [50:54] know this X-ray and you give me this [50:55] narrative then we will cover that. Um [50:57] but for a human it's quite hard to like [50:59] a to follow those like rules because [51:02] there are like take dentistry there like [51:04] 50 60 maybe 100 common billing codes. Um [51:08] and it's quite a lot to like remember [51:09] for Mildrat um or Betty the staffer. Um [51:14] and in this case kind of like our agent [51:16] like has like all the like uh [51:18] documentation [51:20] um it has access to the X-ray that is [51:22] the relevant like one it it knows how to [51:24] build like a treatment plan um and we'll [51:27] then submit that with the insurance [51:28] company and the insurance company like [51:30] just needs to follow the rules. Right. [51:32] Right. [51:32] >> Um and then they will like bait that [51:33] out. Uh the other thing because indeed [51:35] we we are very deep in this industry. Um [51:38] that is interesting dynamic um it took [51:40] me a while to understand this but um [51:43] Sigma also has an incentive um uh to [51:46] have good doctors in network that are [51:48] happy with them um because every year [51:50] they need to go um uh to Andre and say [51:55] uh do you want to renew your dental plan [51:56] with us and then reason is going to look [51:58] at like are there good dentists like in [52:00] the area that are covered that Alex and [52:01] Olivia can go to um and if the answer is [52:04] no then because there's competition in [52:06] this market they will go to like you [52:08] know another insurance company um so [52:10] like there is actually an incentive to [52:11] keep Dr. loop and quan like in network [52:14] um because like otherwise they will not [52:16] renew the plan like with their [52:17] employers. So like I think that it is a [52:19] lot more uh for us it's quite [52:20] interesting because why I also think [52:22] these agents are so well equipped to do [52:24] this work because there's very clear [52:26] documentation to Frederick's point uh [52:28] and we can talk about this for hours [52:30] maybe another time but like we are also [52:32] literally digitizing the file cabinet [52:34] like um uh that's it's really [52:36] interesting that a it's really hard to [52:38] find these like files they're there uh [52:40] and once we have them we know how kind [52:42] of like to do this uh but all these [52:44] payments you talked about a lot of [52:46] doctors across America still get paid on [52:48] paper, too. So, like, um, they get [52:50] $100,000 in checks like deposited on [52:52] their desk. Um, and I did this, right? [52:54] So, for Dr. Quan, I literally knock [52:56] knock knock, open the door, I sit on my [52:58] bar stool, and there's the mailman in [52:59] this case that gives me a stack, and I [53:02] kid you not, I open all these envelopes, [53:04] deposit $100,000, like cash uh, in the [53:07] bank account of the doctor, and then I [53:08] have to go to work on all these itemized [53:10] invoices that are attached to the [53:11] checks. And now the interesting thing, [53:13] so even if we had the models, um, you [53:16] couldn't really do this up until a [53:18] couple of years ago because that was the [53:20] status quo. So the file cabinet was [53:22] literally the file cabinet. In the file [53:23] cabinet like were the checks and the [53:25] itemized invoices that you need to do [53:27] kind of like part of the job to keep the [53:29] doctor office like running. And then the [53:31] federal government stepped in and they [53:32] said this has to stop, right? You cannot [53:34] do paper checks anymore for much longer. [53:37] So then they mandated this industry to [53:40] uh switch the direct deposits offer that [53:42] as an option to a doctor. So if a doctor [53:43] says I want to flip the switch um you [53:46] need to do that the same as these [53:48] itemized invoices you need to create a [53:49] digital file format like for that. Um [53:52] and we are also writing that like till [53:54] win. So you will see that like a lot of [53:55] these small businesses I think the stats [53:57] are 70% are still paid uh on paper and [54:00] it's going through this massive like [54:02] digitization revolution right now [54:03] because federal inflection point that's [54:06] regulatory. Um, so I think that combined [54:08] with these models being so good makes [54:10] this like super interesting because even [54:12] if you have the models or had the models [54:14] 5 years ago like yeah the payments are [54:16] still on paper and we are digitizing [54:18] that in the meantime which is also what [54:19] we talked about earlier like it is not [54:21] one click and then kaboom because [54:23] otherwise the whole country would be on [54:24] digital payments right like why why [54:26] would kind of like Dr. Sloop not do this [54:29] well Dr. loop needs to also then spend [54:32] 50 hours figuring out with Sigma and [54:34] Delta and all the other insurance [54:36] companies, how do I flip the switch? It [54:38] then comes into the bank account. Do you [54:40] want the staffer to have access to that [54:41] bank account? Likely not because the [54:43] rent payments aren't there like the [54:45] other data that you don't want your [54:47] staff to see. Um, [snorts] so that's why [54:48] the status quo is the way it is. And [54:50] then we built this like massive engine [54:52] that basically hey give us the business [54:54] info that we need the tax ID number and [54:57] some other nonsense and then we will go [54:59] do that conversion to digital payments [55:01] uh which is also a hard product problem [55:02] to solve but we have kind of like also [55:04] productized like that. So yeah it's [55:06] super interesting to think about what [55:08] other industries like still have that [55:10] because that's I think even a more [55:12] interesting mode where it's like okay [55:13] you can apply these models bring them to [55:15] main street is super hard. Um and then [55:18] in addition to that [55:20] um how do you convert kind of like data [55:22] that you need that's currently living in [55:23] a file cabinet uh to like digital file [55:26] format such that you can actually [55:27] automate the work. Yeah. To your uh [55:30] earlier point actually about digital [55:31] file cabinets not being that much more [55:33] efficient. Part of the reason that uh [55:36] all of these payments are still on paper [55:37] is that the staff basically prefers [55:39] doing the work by hand on paper. Right? [55:41] Because if you just digitize everything [55:43] now you have a PDF instead of like a [55:44] paper in front of you, right? like you [55:46] actually need, you know, some tools to [55:48] to use that uh use the PDF and it's it's [55:51] preferable to just kind of, you know, [55:52] work off of a sheet. Uh and so like [55:55] there wasn't really much of an incentive [55:56] to to digitize. Um now that's obviously [55:58] different. [55:59] >> So obviously there are a lot of dental [56:00] practices that want or even desperately [56:03] need products like Lassie, but they are [56:06] distributed and they're all over the [56:07] country and there are a lot of them to [56:09] reach. How do you think about reaching [56:12] them? like how do you bring agents to [56:14] kind of these mainstream American [56:15] businesses? [56:17] >> Yeah, this is a very different playbook [56:19] than where currently I think the cutting [56:21] edge is. It's like you have these models [56:22] good enough and apply them in [56:24] enterprises and you do like a few stake [56:26] dinners and then you sign a contract and [56:29] then you have like 10 million in AR [56:32] booked, right? Um we literally need to [56:34] go find like thousands, tens of [56:37] thousands, hundreds of thousands of [56:38] small businesses. Uh so it's a super [56:41] interesting problem to solve because [56:42] we've built this agent that's really [56:44] good. Um and what we're doing right now [56:46] is is literally mapping out like where [56:48] are all these dentists in this case. [56:50] Then after the the next small business [56:52] type in the country, who's the owner? Um [56:55] what systems are they on? Like are there [56:57] any intent signals that we can find, [56:58] right? Like they're looking for a job [57:00] because they they say it on Indeed. Uh [57:03] and then get in touch with these like [57:04] people uh with a message that resonates [57:06] with them. Um and that cuts through the [57:08] noise. I think that's especially with [57:10] our um um customer type is very [57:13] different, right? Because if you were to [57:15] sell to like me, you like find me, you [57:18] know, in clay and you enrich it with [57:20] some Apollo data and then you look me up [57:22] on LinkedIn, then you know, okay, this [57:23] is the guy that's going to buy my HR [57:25] system. We can't really do that because [57:27] Dr. Sloop is not in that database. He's [57:30] often not on LinkedIn. Um so like this [57:32] is a completely different playbook that [57:33] we're developing here. Um, and I think [57:35] that's that's another very compelling [57:38] kind of like thing that um we're [57:39] figuring out basically what is the go to [57:42] market playbook look like to adopt uh AI [57:45] and like at many many businesses. Um so [57:48] yeah that's it's quite an exciting [57:50] opportunity and untapped. [57:51] >> Thank you both so much for for coming to [57:53] chat with us today. This was awesome. We [57:55] are very very excited for the future of [57:57] Lassie. Uh, anyone who's listening who [57:59] might be interested in in working with [58:02] the Lassie team to build something [58:04] generational here, um, check it out at [58:06] lassie.ai. And you guys are very [58:09] actively hiring from what I understand. [58:11] >> Oh, yeah. [58:12] >> Amazing. Great. Well, thank you guys [58:14] again. [58:15] >> Thank you so much. Thanks for hosting [58:16] us. [58:18] [music]