What Happens When Al Agents Can Spend Your Money?

当AI智能体可以替你花钱时会发生什么？ Plaid前首席技术官Jean-Denis Greze解释了为何将Town从季节性税务软件转型为AI个人助手，因为他意识到日常知识工作的瓶颈集中在电子邮件、日历和重复的客户沟通上。Greze与Alex Rampell认为，消费级用户体验优于合规优先的企业软件，且随着基础模型实验室向上拓展技术栈，保持模型中立对保护应用层初创公司至关重要。他们强调，AI的广泛普及受阻于职场中善于动手尝试者的匮乏，而非模型能力不足，这要求软件主动挖掘工作流，而不是等待用户输入提示词。最后，他们分析了自主智能体代为花费人类资本将如何瓦解以广告为支撑的互联网模式、供给聚合壁垒，以及依赖交易摩擦的收入机制（如信用卡未兑现收益）。 Town从第一年收入不足100万美元的税务软件，转型为专注于电子邮件、日历和日常业务工作流的个人AI助手，以摆脱年度报税周期的低频和季节性陷阱。 (3:00) AI的主流普及受限于职场中能够构建多步骤自动化流程的非技术型“动手尝试者”的稀缺，这要求软件主动发现并建议工作流，而不是只提供空白的提示词输入框。 (45:26) 随着前沿AI实验室在应用层直接展开竞争，初创公司和中型企业需要保持模型中立并拥有专有数据，以规避供应商绑定风险以及基础模型性能的不断变动。 (35:00) 智能体网络流量的兴起冲击了传统的广告支撑型互联网经济，不仅威胁到高利润的广告引擎，还通过自动兑换优惠券和基础商品自动补货，瓦解了依赖未兑现损失（breakage）的商业模式。 (52:32)

Alex Rampell and Jean-Denis Greze argue that the true bottleneck in artificial intelligence is not raw model intelligence, but embedding neutral, autonomous agents into everyday human workflows. As software shifts from discrete applications to autonomous agents executing tasks, traditional business models reliant on advertising, supply aggregation, and consumer breakage will face structural obsolescence. From Seasonal Tax Automation to Daily Workflow Habits (3:10) Greze’s initial venture applied large language models to automate small-business tax returns, reaching nearly $1 million in revenue in its first year and tracking toward $2.5 million to $3 million. However, Greze realized tax software suffered from a fatal structural flaw: an annual purchase and usage cycle that prevented daily active usage and organic word-of-mouth growth. In a market where incumbents could easily adopt foundation models before an entrant achieved distribution scale, the narrow seasonal sales window created an insurmountable go-to-market barrier. Recognizing this limitation, Town pivoted to email and calendar workflows. Rather than forcing users into new interfaces, Greze targeted external-facing roles such as sales, business development, and recruiting, where daily work already takes place across inboxes and scheduling tools. By focusing on mundane, high-frequency tasks—such as meeting preparation, routine client inquiries, and drafting authentic replies—Town secured immediate executive adoption and high retention, validating daily habit formation over annual compliance. Model Neutrality and Consumer-First Enterprise Design (13:16) Drawing on lessons from Dropbox and the mobile transition, Greze rejects the traditional enterprise approach of prioritizing compliance checklists over user ergonomics. He argues that building a compelling consumer-grade product and hardening enterprise governance post-adoption is feasible, whereas retrofitting an intuitive user experience onto compliance-first software is nearly impossible. Town adopts an entry wedge via email and calendar to establish high-utility personal leverage before requesting access to sensitive enterprise systems of record. To counter platform risk from foundation model labs like OpenAI and Anthropic—which increasingly build downstream applications—Town emphasizes model neutrality and customer context portability. Greze notes that different foundation models lead across distinct modalities like coding, reasoning, and voice. By routing tasks dynamically across frontier and open-weight models, Town insulates enterprise clients from vendor lock-in, ensures proprietary operating data is never used to train public weights, and avoids expensive, human-dependent forward-deployed engineering contracts. The Economics of Cheap Tokens and Diffusion Bottlenecks (23:25) Rampell and Greze view autonomous agents through Jevons Paradox: as the cost of model tokens declines toward near-zero, tasks historically considered uneconomic for human executive assistants become standard automated routines. Greze characterizes Town as a "token allocation machine" that surfaces latent high-return opportunities without user prompting. For example, an agent can audit fluctuating produce price sheets for independent restaurants across thousands of items for roughly a dollar a month, creating significant savings if a single quarterly overcharge is detected. Despite these economic benefits, AI diffusion across the broader economy remains stalled by what Greze calls the "tinkerer bottleneck." While tech hubs are saturated with software engineers who write custom automations, non-technical businesses rarely configure workflows inside blank chat interfaces. As a result, users evaluate artificial intelligence based on outdated static impressions from past trials rather than the steep trajectory of current model capabilities, making proactive, concrete product suggestions essential for mainstream adoption. Agentic Commerce, Broken Ad Covenants, and Breakage Collapse (52:32) The rise of agentic web traffic disrupts the historic economic covenant of the ad-supported internet, where platforms provided free services in exchange for human eyeballs. Because autonomous agents scrape data and complete purchases without viewing ads, platforms heavily dependent on retail media advertising face conflicting incentives between serving human sponsors and fulfilling neutral agent queries. Greze advocates for consumer-funded, neutral agents that remain uncorrupted by marketing distortions, arguing that merchants who embrace headless agent checkout will outmaneuver ad-dependent walled gardens. Furthermore, autonomous execution undermines traditional price discrimination and business models built on breakage—unredeemed credits or forgotten coupons. Historically, high-income consumers with more money than time forgave unused card perks or discounts due to high cognitive effort. Rampell points out that when AI agents automatically redeem every promotional credit and track price cuts, corporate breakage models like premium credit card perks become financially unsustainable, shifting consumer relationships into an adversarial posture and eroding the historical value of human-facing supply aggregators. As software agents evolve to handle any action a human can perform at a computer, value shifts from the individual application layer to autonomous execution. However, an unresolved question remains whether ad-dependent commerce giants will successfully wall their gardens to preserve ad revenues, or whether open, neutral agents will force the broader economy into programmatic, headless transactions.

Alex Rampell 与 Town 创始人 Jean-Denis Greze 指出，AI Agent 的核心瓶颈并非模型智力不足，而是能否无缝嵌入知识工作者的日常工作流；随着 Agent 获得代为执行与消费能力，基于注意力稀缺的传统互联网广告与价格歧视商业模式将面临结构性重构。 低频税务转向高频工作流的战略抉择 (0:00) Jean-Denis Greze 在离开 Plaid 创立 Town 时，最初尝试利用大语言模型（LLM）解析非结构化文档以自动化中小企业（SMB）报税，首年营收虽接近 100 万美元且次年有望达到 250 万至 300 万美元，但在经过首个报税季后便决定彻底转型。他与 Rampell 认为，税务软件具有致命的结构性缺陷：每年仅有一到两次狭窄的销售窗口，缺乏日常活跃用户（DAU），口碑传播几乎为零，且面对坐拥渠道优势的传统软件巨头，初创公司赖以生存的迭代速度优势会被年度周期彻底消耗。 团队随后转向高频的邮件与日历场景，这一灵感源于税务客户对日常会议准备、跨会议上下文记忆及重复邮件起草的强烈诉求。随着底层模型（如 Opus 4.5）在代码代理层面的突破，Town 将重心放在外部沟通密集的岗位（如销售、商务拓展与招聘），在封闭测试期间实现了每日 1% 至 1.5% 的产品迭代。这一转型避开了传统 IT 采购在初期对合规清单的严苛限制，复制了 Dropbox 和早年 iPhone 依靠卓越消费级体验自下而上驱动企业普及的路径。 模型中立架构与企业级上下文主权 面对 OpenAI 与 Anthropic 等基础模型提供商向下挤压应用层的局面，Greze 提出了“模型中立”的防御策略。对于拥有 500 至 1,000 名员工的中端市场企业而言，单一锁定某家大模型存在极高的平台风险，因为不同模型在代码、语音合成、视频或幻灯片生成等细分模态上各具领先优势。Town 并不自研专有基座模型，而是作为动态编排层，在各家前沿模型与开源权重之间择优调度，以平衡效能与推理成本。 与此同时，Town 将企业积累的机构上下文与运行记忆视为客户的核心资产，反对大模型厂商通过高昂的人工驻场工程（FDE）模式实施生态锁定。Greze 强调，平台必须保证客户数据的可移植性，确保企业如需迁移至竞争产品能在一天内完成，并严格杜绝使用客户私有数据训练公共模型。为在保护隐私的同时实现深度定制，Town 拒绝人工查阅用户收件箱，而是通过聚合指标识别高频用户并配合自动化访谈，先沉淀出针对特定场景的种子程序库，再逐步过渡到动态自生成的定制工作流。 以极低代币成本重构内部协作与日常决策 (40:00) Rampell 与 Greze 指出，硅基助理的普及遵循杰文斯悖论（Jevons Paradox）：当边际推理成本降至几分钱时，人们对自动化任务的需求会呈千百倍爆发。传统的内部协作往往受制于“人类是低速路由器”的物理瓶颈，例如在 Slack 频道公开求助会引发大范围注意力打断，而在 Town 的团队产品中，代表各个员工的 Agent 可在毫秒级内完成权限审查与内部人脉检索，实现去中心化的瞬间协同。 这种低成本代币分配机制还催生了以往在经济上不可行的高价值微观决策。例如在餐馆日常运营中，店主每个月都会收到长达数千行物料的比价单，但极少有人力定期核查；AI Agent 能以每月约 1 美元的算力成本自动监控进货单价，即便前 11 个月均无异常，只要在第 12 个月抓出一次 20% 至 30% 的违规涨价，就能为小企业挽回成百上千美元的损失。这也印证了软件价值的转移：核心效用正从单一静态软件（App）上升为具备完整计算机操作能力的 Agent 系统。 Agent 经济学与供给聚合模式的瓦解 (52:32) 当 Agent 开始具备浏览、比价乃至代付能力时，其与互联网现有平台的利益关系迅速分化为对抗与协同两派。Rampell 指出，以展示广告为核心变现支柱的平台（如每年广告营收约 700 亿至 800 亿美元的 Amazon）倾向于采取对抗姿态，因为 Agent 抓取信息直接跳过了广告曝光，打破了“免费服务换取用户眼球”的传统互联网契约；相反，广告包袱较轻的零售商（如 Walmart）则有望通过对 Agent 开放原生 API 和结算接口获得增量流量。 更深远的影响在于传统“损耗模型”（Breakage Model）与价格歧视机制的瓦解。以往高端信用卡（如 Amex Platinum 或 Chase Sapphire Reserve）的丰厚权益，高度依赖于“有钱无闲”的高净值用户忘记兑现各类消费返现与优惠券；一旦个人 Agent 自动追踪并 100% 兑领每一项福利，发卡机构的商业闭环将被迫颠覆。Rampell 总结认为，互联网时代靠组织海量商品目录而获得超额估值的“供给聚合商”正在失去护城河，因为当人类稀缺的注意力能够被机器完全代劳时，单纯的供给聚合已不再具备以往的控制力。 当前主流经济体与硅谷之间存在显著的 AI 认知温差，瓶颈在于缺乏将通用技术转化为日常工作流的操作者；但随着计算机控制能力在近两三个月内从 30% 跃升至 90% 可靠性，Agent 终将从被动的对话框演变为主动调动企业外部工具与资金结算的执行中枢。

## Background 本期内容由 Alex Rampell 与 Jean-Denis Greze 于 2026 年 10 月 19 日在 RSS 播客平台发布。页面简介指出，对谈聚焦于为何 AI 尚未像改变编程那样迅速变革日常工作，以及 Town 创始人兼 CEO Jean-Denis Greze 为何将公司从中小企业报税软件果断转型为围绕邮件与日历构建的个人 AI 助理。两人深入探讨了消费级体验在企业级 AI 中的胜出逻辑、模型中立性与数据所有权的重要性，以及当 AI Agent 具备自主花钱和执行能力时商业模式面临的重构。 ## TL;DR 节目详述了前 Plaid CTO Jean-Denis Greze 将 Town 从低频、强周期的中小企业报税业务，战略转型为高频、自下而上渗透的邮件与日历 AI 助理的全过程。面对前沿大模型实验室向下挤压应用层的竞争，Jean-Denis 阐述了 Town 依靠模型中立路由、绝对保证客户上下文所有权以及避免定制化工程（FDE）模式实现规模化 PLG 增长的核心壁垒。Alex Rampell 则从“杰文斯悖论”、计算机控制（Computer Use）及电商比价案例切入，剖析了 Agent 流量对传统互联网广告契约的冲击。双方一致认为，随着人类注意力瓶颈被 AI 彻底打通，传统依靠信息不对称与用户精力限制维系的商业沉淀模型（Breakage Model）与供给聚合护城河正在加速瓦解。 ## Atmosphere & style 整场对话呈现出硅谷一线顶级投资人与硬核二次创业者之间高度同频、直率且极具技术洞见的思维交锋。没有浮夸的公关辞令，Jean-Denis 以极为坦诚的复盘开场，毫不避讳地解剖转型前的焦虑、沉没成本与战略死线，随后在探讨底层架构与商业哲学时展现出从容笃定的工程师理性。Alex 的追问犀利敏锐，善用生动的商业历史类比与亲身经验抛出尖锐命题，激发出深层次的战略推演。整体氛围既有深夜喝咖啡般的真实与亲切，又充斥着高密度、甚至略带颠覆性紧迫感的未来学思辨。 ## Bulletin - `00:00` 前 Plaid 首席技术官 Jean-Denis Greze 放弃具备数百万美元收入预期的中小企业税务软件，果断转型打造基于邮件和日历的个人 AI 智能体产品 Town，并在 Ben Horowitz 推荐后迅速获得 a16z 领投。 - `20:00` 大模型处理单次长尾任务的成本已降至分钱级，使得对所有日常会议深度背景调研、针对小微餐饮每月比对供货商万行食材报价等长尾自动化成为可能，重现了廉价硅基算力驱动的“杰文斯悖论”。 - `35:00` 针对基础模型实验室（如 OpenAI、Anthropic）向下渗透侵蚀应用层的挑战，Town 主张“模型中立”、保障企业对上下文记忆层的绝对所有权并恪守不拿用户数据进行模型训练的底线。 - `01:00:00` AI 代理全面自动化执行正在终结传统商业中的“未兑现沉淀模型”（Breakage Model），使得商家依赖用户精力不足而设立的价格歧视机制（如信用卡返现、优惠券）彻底失效。 ## Key takeaways - `05:00` 低频与强周期性业务（如一年一度的中小企业报税）无法形成 DAU 习惯与口碑自传播，极易在时钟走完前被具备庞大分发生态的传统巨头直接以同等 AI 技术吞噬。 - `13:17` 企业级软件从合规与 IT 限制出发几乎无法反向打磨出世界级体验，而坚持以极致消费级体验为基座、后续再补齐安全合规的 PLG 路径才是主流胜道。 - `20:00` AI 个人助理正经历“杰文斯悖论”：当单次助理服务的交付成本趋近于零且不涉及真人翻看收件箱的隐私尴尬时，市场对智能化执行的需求将暴增千百倍。 - `26:40` 软件的价值单元正从具体的“独立 App”跃迁为“对整台计算机的接管能力”，随着计算机使用能力（Computer Use）成功率的大幅上升，传统应用层的既有壁垒正被迅速抽象消解。 - `30:00` 大模型实验室推行的前线部署工程师（FDE）人工驻场定制模式极难规模化，Town 坚信依靠 AI 自动调研与自适应生成 80-20 定制化工作流才是可持续之道。 - `35:00` 中立应用层在面对基础大模型供应商下场竞争时的核心壁垒，在于充当动态调度最优模型的全局路由器，并确保企业随时能够平滑带走其核心上下文资产。 - `40:00` 未来企业内部协作的瓶颈在于人类是“低速路由器”，Town 通过推动员工 AI Agent 之间的安全通信直接对齐信息，大幅削减低效的“人在回路”。 - `45:00` 当前 AI 扩散的真正瓶颈不在于智能不足，而在于各行各业普遍极度缺乏有闲暇和技能去鼓捣工具的“修补者”，因此产品必须从提示词框走向开箱即用的主动交付。 - `52:41` 个人智能体必须保持绝对商业中立，唯有以用户的直接付费而非外部广告赞助作为唯一收入来源，Agent 才能在复杂的商业交易中完全代表用户自身的意志。 - `01:00:00` 当 AI 代替人类精力自动穷尽一切比价、省钱与权益兑现时，原本建立在人类精力有限与注意力稀缺基础上的“供给聚合”与广告商业模式将面临根本性重塑。 ## Detailed walkthrough ### `00:00` 从 Plaid 到税务软件，再到 Ben Horowitz 的惊喜推荐 对创业者而言，对世界产生真正的积极影响是驱动一切的核心动力。在离开金融科技公司 Plaid 并卸任 CTO 职位后，Jean-Denis Greze 与 a16z 普通合伙人 Alex Rampell 曾经在门洛帕克或帕洛阿尔托喝咖啡。当时 Jean-Denis 透露自己正准备做一款面向中小企业（SMB）的税务自动化工具。Alex 当时心里直犯嘀咕，心想“做中小企业报税？这也太难啃了”，但表面上为了维持良好关系，只能微笑着表达鼓励。 转折发生在几个月后，a16z 联合创始人 Ben Horowitz 突然给 Alex 发了一条热情洋溢的信息。Alex 回忆道：“我收到本·霍洛维茨发来的消息说：‘我发现了一个极其惊艳的工具，名叫 Town。它彻底改变了我的生活，简直太好用了。’”（“I get pinged by Ben Horowitz about, like, I found this amazing tool. It's called Town. It changed my life. It's so great.”）Alex 第一反应以为这是一个消费级应用，正好属于自己基金的投资范畴，结果一查发现创始人正是 Jean-Denis。Alex 顿时陷入困惑：“他不是在做中小企业税务吗？这根本说不通。”于是立刻发短信询问 Jean-Denis，两人火速重启沟通，仅一周后 a16z 便与 Town 正式确立了投资合作关系。 ```mermaid flowchart TD accTitle: 创始人转型历程与投资契机 accDescr: 展现 Jean-Denis 从 Plaid 离职到探索中小企业报税，再到通过 Ben Horowitz 引荐获得 a16z 投资的流程。 subgraph Phase1["离开 Plaid 与初步探索"] A["Jean-Denis 卸任 Plaid CTO"] --> B["休整陪家人数月"] B --> C["探索 B2B 市场与 AI 结合点"] C --> D["锁定中小企业报税自动化"] end subgraph Phase2["转折与二次联手"] D --> E["Alex 听闻初创想法（内心担忧其难度）"] E --> F["Town 原型赢得 Ben Horowitz 极高赞誉"] F --> G["Ben 引荐给 Alex，一周内敲定合作"] end ``` - Jean-Denis 在创立 Town 之前曾长期担任 Plaid 首席技术官，Alex Rampell 当时主导了 a16z 对 Plaid 的投资。 - 决定再次创业后，Jean-Denis 特意留出数月陪伴家人，因为他深知一旦重新成为创始人，将很长一段时间无法再享受真正的假期。 --- ### `03:00` 为什么中小企业报税在产品上可行，但在商业上受限 Jean-Denis 解释了当初选择税务赛道的逻辑。大语言模型（LLM）天生极度擅长处理文本与非结构化信息，而商业报税恰恰涉及海量的法律条款、租约文件、财务报表等规范化材料。与个人报税不同，中小企业税务流程繁琐冗长，从工程技术角度看，完全可以通过 AI 实现全流程自动化。经过严格的技术尽职调查，团队判定在两到三年内，把凌乱的企业财务凭证转化为最终合规报税单的技术路径是完全成立的，于是倾注整整一年时间构建该产品。 然而，产品跑通并不等于商业模式成立。在经历了第一年报税季后，业务虽然产生了接近 100 万美元的收入，并且第二年有望按节奏增至 250 万到 300 万美元，但这仅仅是“AI 时代之前的一家好生意”，根本无法支撑起风险投资所期望的超高增长曲线。真正致命的症结在于获客渠道与分发：“即便我对几年内报税流程全面自动化深信不疑，但如何高效触达数以百万计的中小企业并向他们兜售这款产品？这里的问号比我们预想的要多得多。”技术可以攻关，但进入市场（GTM）的阻力却难以逾越。 - 第一年收入指标：首年营收略低于 100 万美元，次年预期原本在 250 万至 300 万美元之间。 - 技术验证与落地的反差：LLM 处理结构化财务文档与文本法则游刃有余，但企业端销售链路极长、获客成本高昂。 --- ### `05:00` 周期性生意的死穴：缺乏 DAU 属性与口碑传播窗口 任何初创企业对抗既有行业巨头（Incumbents）的唯一先机就是速度与差异化的产品体验。然而巨头往往掌握着压倒性的分发渠道和客户存量。如果新入局者不能凭借强大的口碑在极短时间内抢占显著的市场份额，在背后“滴答作响的时钟”走完之前，巨头就会运用相同的 AI 技术改造自己的传统系统，瞬间抹平初创公司的技术红利。Jean-Denis 直言：“对于税务软件来说，我们感觉根本无法在竞争对手和老牌巨头赶超之前做到数亿美元的营收规模，因为你一年只能卖一次税。”（“And it felt for tax, like, we wouldn't be able to get to... hundreds of millions of revenue before our competitors, the incumbents could, because you can only sell tax once a year.”） Alex 补充指出，报税工具在本质上永远无法成为一个“每日活跃用户”（DAU）产品，这让口碑效应变得支离破碎。一般人在搞定复杂的年度报税后，几乎不会在日常生活中向朋友吹嘘自己的税务软件有多好；即便真有人夸赞，听者当时往往也不处在寻找新会计师或报税软件的时间节点上，转头便忘到了九霄云外。Jean-Denis 联想到自己一位做教育科技（EdTech）的好友，该好友也常常抱怨学校和学区每年只在秋末和冬季新年前后采购两次软件，销售窗口极其狭窄，极大地限制了业务的超级扩张（hyperscale）。 | 商业维度 | 中小企业报税业务 (SMB Tax) | 理想的风险投资级 AI 应用 (Town 方向) | |:--- |:--- |:--- | | **使用频次** | 低频（一年仅交互 1 次，无法成为 DAU 工具） | 高频（日常工作流、日历、会议、日常沟通） | | **销售窗口** | 极度狭窄（错过报税季前后便无法成单） | 全年持续销售、持续转化 | | **口碑自传播** | 极其微弱（用户报税后极少日常讨论） | 极强（高效处理工作后自然分享与协作扩散） | | **巨头防御性** | 极高（老牌税务机构拥有垄断级分发渠道） | 依赖灵活性与对跨工作流中立整合的深度体验 | --- ### `07:15` 痛苦而决绝的转型：直面团队沉没成本与时间窗口 在首个报税季结束后，Jean-Denis 和联合创始人整整一周都在一边喝着无数杯咖啡散步，一边严肃反思创业的本质与人生的意义。继续做税务固然能维持一家小有盈余的公司，但这违背了他们“在世界上产生真正大规模影响力”的初心。同时，团队招揽的顶级人才在硅谷都有着极高的机会成本，如果创始人自己都不再相信公司能成就一家百亿级巨头，就无法坦然面对每天在办公室并肩作战的员工，更无法在面对手握大厂丰厚现金 Offer 的候选人时给出诚挚的承诺。 直面“放弃”在情感上极为痛苦，就像是在承认自己失败了，但理性思考后会明白，真正的朋友不会因为你转型而离开，只要给予离开的团队成员（例如原本专注税务业务的同事）丰厚的补偿与体面的安置，这其实并不是一个难以跨越的心理门槛。促使他们迅速做出转型裁决的关键因素，恰恰是报税业务的季节性死线： “因为税务是一年一度的，我认为在那个节点做决定，是全年中唯一可行的时刻。”（“One last little bit is because tax is annual, I think... Making the decision at that point in time, it was the only time we could make the decision.”） 在 10 月 15 日报税截止后的一个月时间里，门还虚掩着；一旦迈入 1 月，公司就要重新与各类合作伙伴签约并对客户许下报税承诺，届时将彻底被锁死在下一个闭环中。正是这种不可逆的强时间压力，逼迫创始团队砍断退路，果断启动业务蜕变。 - 创始人的道德责任：如果无法坚信能带来巨大的商业与社会成果，就对不起放弃高薪机会加入早期团队的核心员工。 - 季节性窗口的逼迫效应：非周期性业务容易陷入“再试一个月看数据”的无限拖延陷阱，而 10 月 15 日之后的约一个月窗口期是团队在“继续报税”与“彻底转身”之间做抉择的最后死线。 ### `10:00` 从税务软件到邮件智能体：Town 的关键转型与 PMF 在面临资金窗口期即将关闭的压力下，团队意识到必须立即做出决断，否则一年后资金链将无法支撑转型。保留了价值不菲的优质域名 Town.com 后，团队认真探讨了大约四个方向。最终促成新产品诞生的核心疑问非常纯粹：“为什么从来没有人打造一款基于电子邮件和日历的优秀 Agent？” 这个洞察直接脱胎于他们此前做税务软件时的痛点：用户固然喜欢税务自动化，但更常抱怨的是把大量时间消耗在与客户安排会议上；开会时常常遗忘会议目的；面对客户重复的税务问题，用户非常希望能有一款工具在开会前自动预先起草好回答。彼时恰逢 Opus 4.5 推出，AI 在代码等领域的自主执行能力突飞猛进。这种底层模型能力的跃升，加上直击日常工作痛点的洞见，让 Town 迅速迎来了强劲的 Product-Market Fit（产品市场契合点）。 团队并没有花时间去推导繁琐的 TAM（总可寻址市场）模型或定义获胜标准，而是直接顺应用户的极度渴望全力冲刺。他们逐渐意识到，几乎所有商业运作——尤其是销售、商务拓展、招聘等对外岗位——全都是依托日历和邮件来驱动的，其潜在市场本就遍布全球。将大语言模型（LLM）接入邮箱看似简单，但要交付深度的产品体验却极其复杂。从去年 12 月到今年 4 月，产品以每天 1% 到 1.5% 的速度迭代，最终促使团队决定结束封测、全面推向市场。 Jean-Denis Greze 回顾取得 PMF 的状态时说道： > “It's a magical thing. Once you have it, you kind of want to hang on for dear life and keep pulling at the string forever and ever.” ```mermaid flowchart TD accTitle: Town Pivot and PMF Engine accDescr: 展现从税务痛点到基于邮件/日历 Agent 的转型逻辑与 PMF 爆发路径 subgraph Origin ["转型起点与痛点"] A["原有税务软件客户反馈"] --> B["日程繁重/忘会前背景/重复答疑"] C["时间窗口紧迫<br/>资本耗尽前必须转型"] --> D["确立新方向：<br/>基于邮件与日历的 Agent"] end subgraph Catalyst ["模型升级与 PMF"] E["底层能力突破<br/>(Opus 4.5 / 代码 Agent 进化)"] --> F["AI 生成更逼真且具自主性"] B --> F D --> F F --> G["强劲的 PMF 爆发<br/>每日产品体验增长 1-1.5%"] end subgraph Expansion ["市场认知深化"] G --> H["对外核心岗位全依赖邮件/日历"] H --> I["超越简单的 LLM 接口封装<br/>深挖工作流复杂度壁垒"] end ``` --- ### `13:17` 消费级体验与企业级合规：自下而上的产品哲学 对于应该做 PLG（产品驱动增长）还是直接面向大企业走采购路线，Alex Rampell 指出，若买单者不是终端使用者（例如为了迎合 CISO 的合规需求而忽视一线员工的体验），往往会做出一款糟糕的产品。 Jean-Denis Greze 与其联合创始人在 Dropbox 的任职经历深刻影响了这一决策。基于过往经验，他们坚信从极致的产品体验起步、后续再去补足企业级合规与 IT 配置是可行的；但反过来，如果从合规和 IT 限制出发作为产品基座，几乎不可能反向打磨出世界级的使用体验。在早期，专注把产品本身做到极致所带来的 Alpha（超额收益）远远大于过早陷入企业管理功能的泥潭。 未来的大企业 CIO 和 CEO 们都在个人生活中体验过顶级的消费级 AI，他们必然会要求企业内部工具也具备同等的流畅度与智能化。因此，谁能把企业级产品做出消费级软件般的丝滑体验，谁就能主导未来的企业级市场。随着产品开始切入大公司，核心命题正从单点验证转向如何体系化满足全行业的数据安全与隐私标准。 Jean-Denis Greze 总结了他们在产品定位上的核心哲学： > “It is almost impossible to go from an experience that starts from the compliance IT enterprise requirements and then figures out how to make the product experience incredible if you take that as the baseline.” --- ### `16:03` iPhone 时刻的重演：从消费级震撼到企业级渗透的现实路径 Alex Rampell 回忆起 2007 年初代 iPhone 发布时的情景：当时微软 CEO Steve Ballmer 嘲讽没人会花 800 美元买一部手机，但 BlackBerry 却迅速走向消亡。究其原因，往往是企业 CEO 们在假期给孩子拍照时感受到了 iPhone 的神奇之处，随后直接向 IT 部门下达死命令要求适配。如今许多企业高管在个人场景体会到了 AI 的魔力，然而在工作场景中却由于权限和安全顾虑难以铺开。 Jean-Denis 赞同该类比，并以 Good Technology（曾为 iPhone 提供合规但体验极其丑陋的企业邮件客户端）为例指出，Town 绝不能成为第二个“Good”——光有企业合规却缺乏优质体验终究会被淘汰。AI 个人助手在全社会的普及目前尚处于极早期（全球仅数百万用户在深度使用），大多数人仍需要 6 到 18 个月才会迎来自己的“iPhone 震撼时刻”。 AI Agent 的有效性严重依赖于访问企业数据与内部 API 的能力；缺乏上下文，Agent 便毫无智能可言。Town 采取的渗透策略不是一步到位去撬动企业核心的底层记录系统（Systems of Record），而是先从连接少数员工的邮件和日历切入。日历与邮件蕴含着大量对外业务的交互洞察，Agent 仅凭这一入口即可完成协助会议准备、按照企业设计风格微调演示文稿（Decks）等高价值动作，从而自下而上地为企业员工展现真正的 AI 生产力。 对于 Agent 智能与底层数据权限的关系，Jean-Denis Greze 明确指出： > “If you have access to those things in a secure, privacy-preserving way, then they can do lots of things for you. But if they don't, they're only really as smart as the data they have the access to.” ### `20:00` 赢得信任、自下而上渗透与助理领域的“杰文斯悖论” 谈到企业软件的落地路径时，Jean-Denis Greze 指出，AI 工具必须先用触手可及的“魔法感”让用户尝到甜头。比如，只要系统能根据刚刚收到的 PDF 附件，把最新的数字准确更新进去，用户就能立刻感受到价值。但接下来的关键问题是：“我们是否已经赢得了向你提出连接 CRM 的权利？我们是否赢得了让你主动去找 IT 团队、申请打通所有其他工具的权利？”这个过程是随着人类信任的建立而逐步演进的。向大企业渗透需要兼顾硬币的另一面：必须提供完整的审计日志（audit logs）、确保受特权保护的信息不被随意喂给 Agent 的安全姿态，以及让 CFO 清楚掌握整个团队在 Town 上的花销和 Token 使用效率的成本控制机制。Jean-Denis Greze 总结道： > “向财富 100 强公司销售必须具备某种程度的严谨性，所有这些都需要更多人类时间去适应。” 而在中小企业（SMB）与中端市场（Mid-market），这些繁文缛节大幅减少，采用速度也快得多。因此 Town 当前的策略非常明确：在产品体验与自主式增长（Self-serve growth）上主打 SMB 和中端市场，同时在安全、合规与财务功能上针对大型企业发力，从而通吃整个市场。 ```mermaid flowchart LR accTitle: Town's Dual-Track Market Expansion accDescr: SMB and mid-market drive self-serve adoption while enterprise requires security and compliance. subgraph SMB["中小企业与中端市场"] A["即时价值体验"] --> B["自服务增长 (PLG)"] B --> C["高频日常工作流渗透"] end subgraph Enterprise["大型企业市场"] D["审计日志与合规"] --> E["CFO 成本与 Token 控制"] E --> F["IT 权限与工具打通"] end SMB -.->|"赢得用户信任"| Enterprise ``` Alex Rampell 随后引入了“杰文斯悖论”（Jevons Paradox）来审视个人助理市场。在历史上，当詹姆斯·瓦特改进了蒸汽机使其效率大幅提升后，人们原本以为煤炭需求会因单次消耗减少而下降，但实际上各行各业因成本骤降而广泛使用蒸汽机，煤炭需求反而暴增了上千倍。在人类助理领域，这种逻辑同样成立，甚至带来了一种“隐私层面的杰文斯悖论”。Alex Rampell 举了自己的亲身经历：他配有全职行政助理（EA），平时很信任对方也处得很好，但依然不想让真人助理翻看自己所有的私人邮件。前不久他在伦敦出差，总觉得自己有一场晚宴要参加，但日历上空空如也，他便发短信向自己的 Town 智能助理“Snap”（他给自己的 Townie 起的名字）核对，Townie 立刻回复确认他确实有这个晚宴安排。原来发件人漏抄送了人类助理，而 Alex 自己也没有及时回复邮件，导致日程未被排入日历，但不知疲倦、持续阅读他所有邮件的 AI 却敏锐地捕捉到了这件事。Alex Rampell 指出，由于隐私顾虑和高昂的人工成本，全职助理的市场表面上看很小，可一旦这项能力能通过“硅基芯片”（silicon）廉价且安全地交付，助理市场的潜在需求将被瞬间放大千百倍。 - **产品增长战略**：以自服务模式进攻 SMB 和中端市场，以安全合规驱动大企业销售。 - **隐私与助理边界**：用户不愿让真人翻看全部收件箱，但对硅基 Agent 毫无保留，创造了全新的持续监控与提醒场景。 - **杰文斯悖论映射**：单次处理成本趋近于零时，人类对“助理服务”的需求将爆发式增长，原本不值当委托的小事都将变成常态化任务。 --- ### `22:45` Token 分配机与长尾价值挖掘：从会议准备到餐厅比价 面对 Alex Rampell 提出的“将产品定义为助理是否画地为牢”的疑问，Jean-Denis Greze 提出了自己对 Town 核心本质的理解。他认为 Town 是一个“Token 分配机”（Token allocation machine），其核心使命是与人类协同工作，不断为用户发掘能够部署 AI 的全新场景，让用户在工作中表现得更出色。系统向人类呈现应用 Token 的选项，并由人类决定是否需要这一结果。随着 Token 成本持续断崖式下跌，曾经由于人工昂贵或初期 Token 成本高而不可行的场景纷纷变得极具吸引力。 Jean-Denis Greze 用两个生动的例子阐述了这一点： 1. **全量会议行前简报**：即使是有专职助理的高管，助理也不可能为每半小时一场的会议去撰写完美的行前简报（prep notes），因为这太浪费人类时间了；但如今大语言模型（LLM）只需花费大约一分钱，就能深入调研与会人员背景、检索他们过去一周在社交平台 X 上的言论，并提炼出高度相关的背景信息。 2. **餐馆批发价监控**：美国的独立餐馆老板每月或每季度都会收到批发商发来的 Excel 表格或 PDF，里面罗列着成千上万种食材原料的最新单价，价格瞬息万变。餐馆老板雇不起全职财务人员去逐项比对是否有食材暴涨了 20% 到 30%，这种低频但繁琐的分析在过去往往被直接放弃。 ```mermaid flowchart TD accTitle: Town's Long-Tail Value Discovery Engine accDescr: Town automates low-frequency, high-value tasks cheaply and alerts users only when actionable. A["非结构化输入: PDF/Excel/收件箱"] --> B["Town 定期自动化比对分析"] B -->|"每月成本约 1 美元"| C{"是否存在异常波动?"} C -->|"11 个月无异常"| D["静默不打扰"] C -->|"1 个月发现异常"| E["主动邮件告警: 建议更换供货商或下架菜品"] E --> F["为小企业挽回成百上千美元损失"] ``` 在 Town 的体系中，用户只需要花五分钟输入一段提示词（prompt），就能设置一个每月定时自动运行的比价机器人。即使一年中有 11 个月它什么问题都没发现，每月一美元左右的运行成本也是完全值得的；而在剩下的那个月里，它一旦找出供货商的隐性涨价，就能为餐馆省下数百乃至数千美元。Jean-Denis Greze 强调，餐馆老板并不缺乏商业头脑，只是过去受限于工具和人力，根本不会想到还可以这样操作。Town 的真正价值，正是主动向用户揭示这些以前被忽略的自动化机会。 - **行前调研平民化**：LLM 凭借一分钱的极低成本，使得对每场会议、每位与会者进行长尾背景调查成为可能。 - **隐形财务审计**：将成千上万行批发商食材报价表自动化对齐，一年只要抓住一次不合理涨价便能收回百倍成本。 - **主动式交互**：不是等用户发出指令，而是产品主动询问用户“是否需要我持续监控批发商的价格变动”，极大降低了技术使用门槛。 --- ### `26:40` 横向平台与纵向垂直的抉择：软件价值单元的范式跃迁 Alex Rampell 顺着餐厅的例子进一步推演：如果直接向餐馆老板推销“行政助理”，对方绝无兴趣；但如果像 Geico 汽车保险的经典广告语（“想要在车险上省 15% 吗？”）那样问他们“想不想在结球生菜（iceberg lettuce）采购上每年省下 500 美元，而我只收你 5 美元？”，所有人都会毫不犹豫地答应。但这随之引发了企业战略上的经典抉择：到底是做垂直方案（Vertical），还是做横向平台（Horizontal）？做纯垂直方案很容易让公司被定型为“那个帮人省生菜钱的工具”；而 Town 的选择是先尽可能做横向，从成千上万家企业和各类用户的实际使用中学习，让市场自然涌现出具体的痛点，之后再顺势将特定场景垂直化。 对于横向与垂直的辩证关系，以及当前 OpenAI、Anthropic 等前沿大模型实验室的步步紧逼，Jean-Denis Greze 给出了深刻的技术洞察。他断言，今天绝大多数 AI 产品在底层能力上最终都会高度同质化——在办公场景下，主流的 Agent 外壳（harness）能做的事情大同小异。硅谷长期以来依赖的“打造领先市场的独特产品功能”这一核心护城河，未来将越来越难以维持。对此，Jean-Denis Greze 做出了关于软件行业本质变革的核心判断： > “我认为软件的价值单元已经从 App 变成了计算机本身；但其真正含义是，它已经从一个 App 变成了人类所能做的一切事情。” 因为人类几乎可以用电脑完成任何任务，当 AI Agent 具备了像人一样操作计算机的能力时，传统的应用层（app layer）正在被彻底抽象剥离，整个应用层的既有商业价值也将面临剧烈缩水。 - **营销心智差异**：从“卖虚无缥缈的 AI 助理概念”转向“卖可量化的确定性成本节省”。 - **演进路径设计**：先通过横向平台吸纳海量真实边界案例（如生菜比价），再依托沉淀的数据和工作流实现精准的垂直化变现。 - **价值单元转移**：软件交互的核心从具体的独立 App 转向对整台计算机通用操作能力的接管，传统单一功能软件的护城河正在快速消解。 ### `30:00` 构建定制化 AI 软件：从人类经验沉淀到自动化规模化 Jean-Denis Greze 阐述了 Town 打造成功企业的底层逻辑：随着 AI 技术持续迭代，它能够越来越精准地为每一位用户量身定制产品。如果一位电工注册了 Town，系统向其推荐的自动化工作流会极其精准地匹配电工的日常痛点；而如果是风险投资家（VC）或行政助理（EA），推荐的内容也会与他们的特定行业和职能严密贴合。然而眼下基础模型的能力尚未达到“仅凭一句指令就能生成深层推荐”的水平——如果直接询问大模型“给投资人 Alex 推荐 Town 的自动化流程”，往往只能得到浮于表面的建议。 为了解决这一问题，同时遵守企业级软件严禁直接窥探用户数据的安全红线，Town 采用了一种由浅入深的“飞轮式”构建路径：通过宏观使用指标筛选出平台上的超级用户（Power Users），直接由团队与他们深入访谈。比如在访谈了 10 位餐厅经营者后，提炼出他们共有的 5 类核心业务场景（如批发商供应商管理），并在内部将其打造成预置的自动化原型（在 Town 内部被称为 routines）。新用户注册时，系统通过识别其收件箱的特征判断其“似乎是一位餐厅老板”，便会主动向其推送这些精心打磨的模板方案。 ```mermaid flowchart TD accTitle: Town 的用户定制化工作流飞轮 accDescr: 展示 Town 如何从人工访谈提炼常规工作流，逐步演进为 AI 自主访谈与大模型个性化生成的全过程。 subgraph Phase1["第一阶段：人工提炼与冷启动"] A["筛选平台超级用户 (如餐厅老板)"] --> B["深入人工访谈<br/>发现共性痛点"] B --> C["内部手工打造标准化原型<br/>(自动化 Routine 模板)"] C --> D["匹配新用户画像<br/>主动推送精准推荐"] end subgraph Phase2["第二阶段：AI 加速与飞轮运转"] D --> E["AI 驱动的自动化调研访谈<br/>(发放 Amazon 礼品卡)"] E --> F[" Routine 模板库从 5 个扩充至 15+"] end subgraph Phase3["第三阶段：模型自主生成与复合角色适配"] F --> G["大模型能力提升<br/>减少对预设库的依赖"] G --> H["80-20 定制化软件：<br/>兼顾餐厅运营、个人财务、<br/>短租托管与子女夏令营日程"] end ``` 这种模式在后续演进中被大幅加速。Town 不再依赖纯人工节奏，而是通过 AI 录制访谈、赠送亚马逊礼品卡等方式全自动拓展用户调研，使预设工作流库从 5 个快速扩充到 15 个甚至更多。随着基础大模型的能力逐步跃升，系统对预设库的依赖程度随之降低，最终能够实现“80-20 定制化”——让每个人都拥有完全针对自身复合身份的量身定制软件。因为现实中的人从来不是单一维度的职业标签：一个开餐厅的人，可能自己兼任财务主管，同时名下还有一套用于出租的度假民宿（Vacation Rental），家里还有两个上学的孩子需要安排暑假活动。用户期望其 AI 助手能够无缝贯通这些全生命周期的任务。 相比之下，Jean-Denis 认为大模型实验室（AI Labs）目前普遍推行的“前线部署工程师”（Forward Deployed Engineer, FDE）模式存在本质缺陷。实验室派驻人类工程师深入客户业务现场、使用自身模型去手工编写定制软件，在当下对愿意每年砸下 500 万美元给 OpenAI 的大型企业或许行得通，但它极难规模化。 > “I think if you're an enterprise that's willing to spend $5 million a year paying OpenAI for FDs, it looks really good right now. But I think... Over time, as the models get better, the FDE approach actually gets much worse. And our approach, which is much more around automation and using AI to build the custom software, will scale much better.” —— Jean-Denis Greze Town 的核心战略是坚决走产品主导增长（PLG）路线，优先切入中小企业（SMB）及崇尚技术的前沿中型企业，放任竞争对手在大企业手工定制的红海中消耗战力。 --- ### `35:00` 基础设施与应用层的竞合：模型中立、上下文所有权与防训练壁垒 Alex Rampell 抛出了一个让所有应用层创业者如鲠在喉的难题：OpenAI、Anthropic、Google 等上游模型供应商正在加速向下渗透，陆续推出了集成 Slack、LinkedIn 等工具的连接器。这在历史上是罕见的异常生态——20 世纪 90 年代，思科（Cisco）和 Akamai 构筑了互联网的基础设施，但思科绝不会跳下场去做线上书店（Amazon）、拍卖网站（eBay）或门户目录（Yahoo），它只满足于售卖交换机和负载均衡设备。而如今，模型供应商在享受算力代工利润的同时，正在亲自将触手伸向应用层，成为客户的直接竞争对手。 针对 Town 如何在面对“供应商下场抢饭碗”时立足，Jean-Denis 给出了三个维度的战略防线： | 竞争维度 | 模型实验室（AI Labs / FDE 路线） | Town 的中立应用层路线 | |:--- |:--- |:--- | | **模型供应绑定** | 强制绑定自有模型生态，一旦技术落后客户即被锁定 | **模型中立（Model Neutrality）**：根据语音、视频、文本和代码等场景动态路由最优模型 | | **数据与上下文归属** | 借助 FDE 人工定制系统抬高客户的迁移与退出壁垒 | **用户绝对拥有上下文（Context Layer）**：承诺支持客户在一日内平滑迁移至替代平台 | | **商业机密与训练风险** | 存在客户业务运行模式被隐性吸收进未来模型权重的顾虑 | **坚决不参与模型训练**：企业运营智慧绝不外溢给潜在竞品 | 首先是**模型中立性（Model Neutrality）**。当面向拥有 500 至 1,000 名员工的中型市场企业销售时，Jean-Denis 会直接发问：有谁敢断言未来 3 到 5 年哪家大模型供应商必定能持续领先？事实上没有任何一家企业能给出确切答案。如果企业在一棵树上吊死、全面采购某一家实验室的应用，一旦该实验室落后或借机大幅提价，企业就会瞬间陷入被动与劣势。例如两年前大量企业选定 OpenAI，但随后 Anthropic 的 Claude 在代码能力上持续领跑整整 12 个月，许多企业被迫同时采购两者以维持竞争力。Town 的价值在于承担了全局路由的角色：在通用智能、语音、视频、幻灯片生成等细分能力上，在最前沿闭源模型与开源权重（Open Weight）之间为企业动态选用当下最具性价比的最优解。 其次是**记忆与上下文层的所有权（Ownership of Context Layer）**。Jean-Denis 强调，在数字时代之前，企业的资产是实体资产和人力资产；而在 AI 时代，企业最宝贵的核心资产是 AI 掌握的业务运转全量记忆与上下文。实验室通过派遣 FDE 建立定制系统，实质上是在加深系统绑定、提高企业的切换成本；而 Town 的立足之本，恰恰是反其道而行之，让企业毫无保留地彻底掌控自己的数据资产： > “And I think you should own that entire bucket. And I think you should own it so much that if there was ever a better solution in town, switching to that other solution... would take, you know, a day. That's what I'm selling. I'm selling you want to own that asset.” —— Jean-Denis Greze 虽然承诺让客户“随时能用一天时间迁移离开”意味着 Town 必须直面极低的切换壁垒风险，但 Jean-Denis 坚信这正是面对基础模型巨头时最有力的卖点——靠最好的产品体验留住人，而不是靠数据绑架客户。 最后是**数据防训练红线**。即便大企业能与实验室签订不使用其数据微调的协议，但企业运转的逻辑框架一旦在交互中渗入模型权重并被转售给行业对手，实质上就等于将企业的核心竞争壁垒拱手让人。Town 将“绝不把用户数据投入训练”视作不可动摇的商业底线，以此构建与模型供应商在人机关系哲学上的根本分野。 ### `40:00` 告别“低速路由”：让 AI 介导的企业内部协作 Jean-Denis Greze 详细拆解了 Town 团队版的核心设计哲学——打造一个由 AI 驱动的内部 CRM 体系。在该模式下，企业员工不再需要像过去一样在 Slack 或 Microsoft Teams 等通讯软件里大海捞针般地发消息询问（例如：“有人认识某潜在客户财务团队的人吗？”），然后再苦等同事们逐一阅读、回忆并在一天后回复。在 Town 的架构中，用户的 AI 智能体可以直接与公司内其他员工的智能体进行安全通信，瞬间锁定谁与目标企业存在联系。Jean-Denis 强调，人类本质上是极其低效的信息路由器，要打破这种瓶颈，就必须重新设计隐私与权限体系，让企业敢于授权 AI 来协调内部数据的流动。 谈及未来的工作形态，Jean-Denis 认为整个行业的核心课题在于让组织学会适应机器的高效节奏，而不是让 AI 受制于人类的迟缓流程。尽管目前许多人对数据安全心存顾虑，但伴随效率成倍提升带来的巨大商业收益，未来人机协同中“人在回路（Human-in-the-loop）”的比重必然会大幅削减。Jean-Denis Greze 明确指出：“从竞争和未来的角度来看，你必须思考未来的样子，而我认为未来的人在回路会比今天少得多。我们的职责就是让企业适应这种环境，因为由此带来的经济收益远远超过他们当前感受到的风险。” ```mermaid flowchart LR accTitle: 人类手动路由与智能体直接路由对比 accDescr: 展示传统方式通过人工询问和阅读与 Town 方式通过 Agent 间安全协同获取人脉信息的差异 subgraph 传统人类低速路由 A["员工发起询问"] -->|"Slack / Teams 发帖"| B["全员逐一阅读与回忆"] B -->|"数小时或数天后"| C["得到零散反馈"] end subgraph Town 智能体高速路由 D["员工提问"] -->|"Town 智能体发起检索"| E["内部权限与隐私网关"] E -->|"智能体间安全对齐"| F["同事智能体人脉图谱"] F -->|"即时返回"| G["精准对接目标联系人"] end ``` - 传统协作模式的缺陷在于将人类当作“低速路由器”（Humans are slow routers），沟通成本极高且反馈周期长达数日。 - Town 采用 AI 介导（AI-mediated）的安全与隐私框架，使员工智能体能在受控权限内跨部门自动匹配商业信息。 - 创业公司的差异化核心在于保持数据中立性，避免像 Harvey 或 Rogo 那样直接面临大模型实验室（如 OpenAI）亲自下场做垂直场景的战略挤压。 --- ### `42:48` 实验室的垂直挤压与“适居带”博弈 Alex Rampell 与 Jean-Denis Greze 展开探讨了大模型实验室（AI Labs）与上层应用公司之间的生态博弈。Alex 提出了“适居带（Goldilocks Zone）”理论：如果一家公司的年营收达到数万亿美元，它会立刻成为 OpenAI 和 Anthropic 的头号围剿目标；如果营收为零且无人问津，实验室则完全不会在意。编程领域（Coding）目前正处于这个残酷适居带的靶心——Anthropic 依靠 Claude 在编程赛道全力以赴，每周斩获惊人增速，以至于各大模型厂商“不能不去争夺”。相比之下，法律、医疗等单点纵向领域虽然巨大且引来垂直布局，但针对海量中小微企业（SMB & Mid-market）日常工作流的横向产品，现阶段尚未被巨头全力封杀。 对于横向通用型产品是否能长期筑牢壁垒，Jean-Denis 提出了更加激进的市场预判。他坦言，虽然短期内实验室更聚焦于能迅速拉动顶线收入的编程等核心赛道，但帮助广大普通企业运转得更顺畅、让 AI 承担日常商业事务的市场体量，从长远来看远超纯编程市场。Jean-Denis Greze 坚决地表示：“对于中小企业和中端市场，我认为 AI 帮助企业做更多事情并取得成功的市场，绝对比编程市场更大。对我来说，这只是时间问题，而不是会不会发生的问题。”他预计在未来 12 个月内，这种横向智能体市场就会迅速展现出颠覆性的商业穿透力。 - 编程（Coding）赛道被推向风口浪尖，源于 Anthropic 孤注一掷押注代码能力并大获全胜，迫使 OpenAI 等巨头全面跟进。 - 实验室难以在现阶段全面吞并所有横向场景，给中立型 AI 应用公司留出了抢占企业入口的窗口期。 - 长期来看，为广泛的企业级长尾场景赋能的商业价值将显著高于代码生成本身。 --- ### `45:00` “修补者”匮乏瓶颈与海岛冲浪板的真实神话 讨论转入 AI 技术在现实世界的扩散鸿沟（Crossing the Chasm）。Alex 观察到一个极为割裂的现象：硅谷在狂热地宣传甚至被质疑过度炒作 AI，而世界上其他绝大部分地区却在严重“低估”AI。Town 作为一个基于产品驱动增长（PLG）的横向平台，并未刻意进行跨国市场推广，却自发在日本等海外地区走红。Jean-Denis 分享了一个令他极其意外的真实用户案例：一家位于海岛上的冲浪板与沙滩越野车租赁小公司，使用 Town 来对比游客租还前后的装备照片，由 AI 自动鉴别磨损或损坏程度，进而自主裁定是否退还押金。 这个完全不在团队“128×128 预测大转盘”上的案例，让 Jean-Denis 看清了 AI 扩散缓慢的真正病根所在：这项技术能否在企业中落地，完全取决于业务现场是否存在一个理解 AI 且愿意花时间鼓捣工具的“修补者（Tinkerer）”。现实中绝大多数企业根本没有这样的人才或闲暇，因此大多数人依然只是把 ChatGPT 或 Claude 当作更高级的“Google 搜索替代品”。Jean-Denis Greze 犀利地总结道：“你不需要更多智能了，智能早已摆在那里。真正的问题是如何将其高效且规模化地部署下去。” - 海岛租赁商借助视觉比对自动完成冲浪板损坏判定与押金结算，展现了长尾小微企业对 AI 自动化的原生需求。 - 硅谷内部因为工程师密集、全员皆是“修补者”，因此能迅速榨取 AI 的前沿价值，但外部普通商业世界极度缺乏这类技术探索者。 - 靠文本输入框（Text box）无法实现 AI 的大规模普及，用户需要的不是被教育如何编写 Prompt，而是开箱即用、直击痛点的自动化结果。 --- ### `47:35` 抹平集成鸿沟：从“提示词框”走向自适应交付 针对这一扩散困境，Jean-Denis 阐述了 Town 提升用户留存（Retention）的核心路径。数据显示，能真正长期留在平台上的核心用户通常满足一个条件：他们接入了自己的邮箱、日历以及至少一个关键的业务记录系统（System of Record），并且亲手搭建过一个开箱即用功能之外的“定制化集成（Custom integration）”。然而，要求普通用户自己去动手组装定制工作流是一种过高的门槛，Town 正在努力消除这个门槛——让软件在接入记录系统的那一刻起，就自动根据企业特征主动交付深度定制的价值，无需用户成为极客。 Alex 对此深表赞同，并将其定义为一场“集成与产品营销（Integration & Product Marketing）”维度的挑战。许多企业决策者对 AI 的认知停留在几年前的旧印象中。Alex 举了一个生动的例子：他曾拜访哈佛大学一位顶尖女学者，对方起初对 AI 嗤之以鼻，称自己几年前试过觉得不过如此。Alex 随即利用当前的最新模型，将她二十多年前撰写的博士论文结合当今最新的学术进展彻底重写了一遍，学者看后震惊不已，追问这是什么前沿新技术。Alex 借此指出，世界绝大部分人群是在用静态眼光审视当前的工具，而硅谷敏锐捕捉到的是其惊人的演进斜率；缩短这一认知鸿沟的关键，在于直接把成熟的技术产品化，交到非技术人员手中。 - 用户在 Town 上的关键留存闭环：连接邮箱 + 连接日历 + 连接核心记录系统 + 触发一项定制工作流。 - 智能体软件的演进目标：消灭手动搭建集成的环节，在数据接入瞬间自动呈现符合用户业务特质的自动化交付。 - 大众往往依据过时的体验给 AI 的能力下死刑，忽视了其技术迭代曲线的陡峭斜率，直观的降维打击演示是扭转认知的最快手段。 ### `50:00` 早期尝试者的认知固化与计算机控制能力（Computer Use）的爆发 Alex Rampell 与 Jean-Denis Greze 讨论了一个耐人寻味的现象：许多人之所以至今没有充分利用 AI，反倒是因为他们两年前就是“早期尝鲜者”。他们在两年前尝试过早期的模型，遭遇了幻觉和失败，得出了“这玩意儿不好用”的结论，从此将对 AI 的认知彻底“固化”在当时的水平。Alex 分享道，自己做过至少十次类似的实验，把最新模型的表现展示给这些人看时，他们往往惊呼：“你这是从哪弄来的 AI？我用的 AI 根本不是这样。”而事实是，他们用的就是同一个底座，只是从未再去尝试。 Jean-Denis Greze 指出，底座模型的演进速度实在太快，这种认知脱节会不断发生。关于未来的赢家，Jean-Denis 强调：“我认为获胜的平台将是那个能够不断让业务去适配最佳 AI 的平台（I think the winning platform is one that continues to adapt the business to the best AI）。”AI 能力会无休止地提升，人们对于“什么可行、什么不可行、什么简单、什么复杂”的直觉将被彻底颠倒。 以最近两三个月的“计算机使用与浏览器操作（Computer Use / Browser Use）”为例，AI 助手目前能做到的事情在半年前几乎是天方夜谭，任务成功率也从过去的 30% 飙升到了 90%。然而，大众对这种跃升毫无概念。当他们在 Town 中看到 AI 能帮自己跨系统去购物，或者直接操控已有软件完成原本以为必须走复杂 UI 界面的任务时，都会大为震撼。 向普通用户推广这类能力不能靠抽象的专业术语，比如宣传“Town 具备了 Computer Use 能力”是毫无意义的。真正有效的沟通必须落到极具象的业务场景中： - **场景切入**：“我能用你现有的 Canva 订阅，为你接下来的活动生成一组海报，风格和你平时做的一模一样，你想让我试试吗？” - **低成本试错**：耗费不到 25 美分和 5 分钟，AI 就能呈现极具价值的高水准产出。 - **主动触发自动化规则**：“以后只要在你的日程表里检测到同类活动，都需要我提前自动做一套海报吗？” - **自主感知品牌变化**：当中小企业日后更换了品牌设计，AI 下次生成传单时应当自动同步新视觉，无需用户再次提醒。 - **沿技术曲线扩展边界**：当底层视频模型成熟时，AI 应该主动提议：“需要我顺便把这些素材制作成宣传短视频发布到 Instagram 吗？” ### `52:41` 智能体流量颠覆互联网传统契约与中立性困境 Alex Rampell 指出，广义上的 Agent 流量如今已经大幅超越了人类自然流量，这在互联网行业引发了激烈震荡。当前主要分化为两大阵营：一种是对 Agent 严防死守的“对抗派（Adversarial）”，另一种则是拥抱 Agent 的“协同派（Synergistic）”。 典型的对抗派代表就是亚马逊（Amazon）。长久以来，互联网维系着一套默认的商业契约：用户去雅虎财经（Yahoo Finance）免费查看行情，雅虎花钱向纳斯达克购买数据流，但用户必须把注意力（眼球）留在网页上看 E-Trade 的广告；雅虎通过广告变现支撑免费服务。然而，当人类眼球被后端直接抓取数据的 Agent 取代时，广告曝光彻底失效，旧契约随之破裂。亚马逊本质上是一家广告巨头，封杀 Agent 抓取符合其短期商业利益。但在 Alex 看来，Agent 时代的到来就像当年的 SEO 浪潮，如果一家平台选择成为 Agent 流量的接收端而不是阻截端，将迎来巨大的增量市场，“亚马逊的损失很可能成为沃尔玛的收益”。 ```mermaid flowchart TD accTitle: Internet Value Covenant vs Agent Disruption accDescr: 互联网传统广告契约与智能体流量冲击对比流程图 subgraph Traditional ["传统互联网契约（以眼球变现）"] A["用户访问平台 (如 Yahoo / Amazon)"] --> B["获取免费内容或商品数据"] A --> C["人类视线接触展示广告"] C --> D["广告主付费支撑平台免费生态"] end subgraph Disrupted ["智能体时代的契约破裂"] E["自主 Agent 代理查询与交易"] --> F["直接提取商品/数据 API 结果"] E -.->|"无人类眼球接触"| G["广告展示与变现完全失效"] F --> H{"平台的战略分歧"} H -->|"对抗派 (如 Amazon)"| I["封锁 Agent 流量以保护广告盘子"] H -->|"协同派 (如 Walmart)"| J["开放对接 Agent 流量成为新型供货接口"] end ``` 这种分歧随之引出了 Agent 本身的中立性与信任危机。Jean-Denis 提出一个尖锐的问题：如果你使用的 Agent 本身背负着庞大的广告商业模型，分发机制会变成什么样？例如，使用 Meta 推出的 Agent 订机票，而美联航（United Airlines）恰好是 Meta 的头部广告主，用户还能相信 Agent 推荐的航班是客观最优解吗？ 关于个人 Agent 的归属与定位，Jean-Denis Greze 坚定地表示： > “我坚信你需要一个属于你自己的中立 Agent，这个 Agent 唯一的收入来源就是作为用户的你的钱包（I'm a firm believer in the like you want a neutral agent, that is yours, that you have no, like the only source of money for the agent is your wallet as the user）。” 如果创造 Agent 的机构其激励机制受到外部营销资本的扭曲，它提供的信息就无法保持客观纯粹。唯有完全由用户付费支撑的私有 Agent，才能无偏见地为人类的真实品味与需求服务。 ### `55:00` 享受的阻力 vs 厌恶的阻力：电商巨头的 Agent 反击战 Jean-Denis 进一步剖析了人性的非理性：构建 Agent 时人们往往执念于“消灭一切摩擦、让一切效率最大化”，但人类并非在所有事情上都追求极致效率。人类体验中存在两类截然不同的阻力： 1. **享受的摩擦（Friction you enjoy）**：比如在纽约曼哈顿的苏豪区（Soho）逛街，走在店铺之间、在人群中排队观察别人的穿搭与风格，这种阻力构成了逛街的乐趣本身。 2. **厌恶的摩擦（Friction you don't enjoy）**：比如家里卫生纸用完了要补货，用户只想以最快速度、最便宜的价格将经常买的那个牌子送货上门，过程越无感越好。 正因如此，商业世界依然有营销和展示的容身之地。当一位父亲想在亚马逊上给孩子买乐高积木时，他反而欢迎相关拼插玩具的广告推荐，因为他可能会偶然发现此前从未听过、但极具创意的优质小众品牌，并向 Agent 询问评测。但如果面对纯粹的标准消耗品，广告展示就会让人心生厌烦。 针对亚马逊每年高达 700 亿至 800 亿美元的广告业务，Jean-Denis 认为封锁 Agent 绝非长久之计： - **单纯的供给聚合者（Aggregator of Supply）地位正在失效**：在 Agent 横行的未来，仅靠陈列货架坐收广告费的护城河极易被摧毁。一个摆脱传统前端界面的“无头亚马逊（Headless Amazon）”体量完全有机会超越现在的亚马逊，但前提是他们敢于自我革命。 - **业务解耦与跨场景变现（以 Meta 与 Muse 为例）**：市场热议某些免费的个人 Agent（如 Muse）未来是否也不得不插入广告。但 Jean-Denis 分析，它们完全可以将 Agent 体验与变现解耦——Agent 保持纯净且免费，但它沉淀的个性化数据可用于在 Instagram 等娱乐沉浸场景中实现极高精度的广告推荐；当用户在 Instagram 想要消费时，直接调用 Agent 积累的支付钱包完成闭环。 - **亚马逊的最佳应对**：亚马逊不该对抗未来，其真正的反制手段不是封杀第三方的 Muse，而是为自己打造一个体验更好的原生 Muse，并将其整合进 Prime 会员权益中，通过专有购物闭环建立护城河。 在谈到电商价格监控与辅助决策的原型时，Alex Rampell 提到了运营长达 18 年的亚马逊历史价格追踪网站 CamelCamelCamel。用户只需输入商品（例如一箱 8 美元的胡椒博士汽水）并设置一个心理预期价位（如降到 6 美元时通知我），该网站就会在降价瞬间给用户发送邮件提醒。令人惊叹的是，这样一个仅靠联盟营销佣金维持、甚至无法直接替用户完成结算操作的纯功能性网站，全美居然有整整 2000 万活跃用户。Alex 借此说明，用户让机器代替自己盯梢消费的底层需求早已极其庞大，而真正的交易型 Agent 将彻底重构这一规模。 ### `01:00:00` 破裂的商业默契：AI Agent 如何颠覆“时间与金钱”的区隔 Alex Rampell 提到了他与 Jesse 共同撰写的一篇文章，核心观点是“预测未来最好的方式就是观察当下”，因为未来早已到来，只是分布不均。他从“时间多于金钱”与“金钱多于时间”这两类人群的划分切入，探讨了 AI Agent 将带来的商业模式冲击。过去，时间充裕但资金有限的人会去使用像 CamelCamelCamel（亚马逊价格追踪工具）这样的比价省钱软件；而资金充裕但时间宝贵的人则根本懒得花时间折腾。但如果现在让 AI Agent 代劳，将所有省钱、套利工具直接赋予那些“金钱多于时间”的高净值人群，他们必然会毫无犹豫地使用。“Like, of course they're going to use that. It's an IQ test.” Alex Rampell 如此断言。 这种变革正在打破传统商业中一条隐形的契约与区隔。过去商家发放优惠券（Coupons），富人不去搜集是因为机会成本太高，不值得花两个小时去翻找，因此优惠券成了一种精准筛选价格敏感型客户的自然区隔。同理，像运通（American Express）的高端会员卡或大通（Chase）信用卡，其商业逻辑很大程度上依赖“未兑现收益/磨损模型”（breakage model）——银行给出了大量高价值权益，但知道大部分高端用户根本没时间逐一去激活和兑现。 ```mermaid flowchart LR accTitle: AI Agents breaking the breakage model accDescr: 传统模式依赖人工精力限制保持权益沉淀，Agent化后导致100%权益兑现，迫使商业模式重构。 subgraph Traditional["传统商业区隔模式"] A["高净值用户: 时间少 / 资金多"] -->|"人工搜寻成本极高"| B["放弃优惠券与权益兑换"] B --> C["商家获利: 权益磨损沉淀 (Breakage)"] end subgraph Agentic["Agent 驱动的新常态"] D["AI Agent 自动化代劳"] -->|"零精力成本兑现"| E["100% 榨干 Amex / Chase 等权益"] E --> F["打破沉淀模型: 商家被迫改制 / 对抗加剧"] end ``` AI Agent 的普及将彻底宣告这种沉淀模型的终结。当每个人都能借助智能体，毫不费力地薅尽所有信用卡和平台的羊毛时，商家与用户之间的关系就会从原本互利协同走向直接对抗，迫使机构彻底修改规则。Alex Rampell 总结指出，贯穿这一切的核心主线在于：供给聚合（Supply Aggregation）的价值正在大幅缩水。过去供给聚合之所以极具价值，是因为人类的注意力是极其有限且稀缺的；而如今，人类的注意力可以被 AI 全面自动化代劳，过去构建在信息不对称和精力瓶颈之上的护城河正在瓦解。在节目的尾声，Alex 对 Jean-Denis 的到访表达了感谢并为 Town 送上祝福，Jean-Denis 也致谢回应，随后节目播出了常规的免责声明与播客订阅指引。 - Alex Rampell 指出：“So, Jesse and I wrote a post about this, where the best way to predict the future is to observe the present.” - 关于商业沉淀机制的失效与供给端价值的转变，Alex Rampell 直言：“The through line of all those is, I just think supply aggregation is a lot less valuable than it used to be very valuable because human attention was limited. And now supply aggregation is a lot less valuable because human attention can be automated away.” ## Notable quotes > `45:00` 你不需要更多智能了，智能早已摆在那里。真正的问题是如何将其高效且规模化地部署下去。 — Jean-Denis Greze > `26:40` 我认为软件的价值单元已经从 App 变成了计算机本身；但其真正含义是，它已经从一个 App 变成了人类所能做的一切事情。 — Jean-Denis Greze > `42:48` 对于中小企业和中端市场，我认为 AI 帮助企业做更多事情并取得成功的市场，绝对比编程市场更大。对我来说，这只是时间问题，而不是会不会发生的问题。 — Jean-Denis Greze > `40:00` 我认为未来的人在回路会比今天少得多。我们的职责就是让企业适应这种环境，因为由此带来的经济收益远远超过他们当前感受到的风险。 — Jean-Denis Greze ## Mentioned Jean-Denis Greze, Alex Rampell, Ben Horowitz, Town, Plaid, a16z, Opus 4.5, Dropbox, Steve Ballmer, iPhone, BlackBerry, Good Technology, James Watt, Geico, OpenAI, Anthropic, Claude, Google, Cisco, Akamai, Amazon, eBay, Yahoo, Slack, LinkedIn, Microsoft Teams, Harvey, Rogo, ChatGPT, Canva, Instagram, Yahoo Finance, E-Trade, Walmart, Meta, United Airlines, Muse, Prime, CamelCamelCamel, American Express, Chase ## Video description Why hasn't AI transformed everyday work as quickly as coding? Town CEO Jean-Denis Greze argues the bottleneck isn't intelligence, but making AI useful to people who don't have the time or technical skills to build their own automations. Jean-Denis joins a16z General Partner Alex Rampell to share why he pivoted Town from tax software to a personal AI assistant built around email, calendars, and the workflows that run businesses. They discuss why consumer-grade experiences may win in enterprise AI, how Town discovers and automates tasks users wouldn't think to delegate, and why model neutrality and data ownership matter as AI labs increasingly compete with application companies. They also explore what happens when agents begin acting on our behalf, from managing everyday work to shopping and commerce, and why the next generation of software may need to adapt to each user rather than the other way around. [Source](https://mgln.ai/e/1344/afp-848985-injected.calisto.simplecastaudio.com/3f86df7b-51c6-4101-88a2-550dba782de8/episodes/c3d1e5f9-5c39-4faf-8b09-0939ca33695d/audio/128/default.mp3?aid=rss_feed&awCollectionId=3f86df7b-51c6-4101-88a2-550dba782de8&awEpisodeId=c3d1e5f9-5c39-4faf-8b09-0939ca33695d&feed=JGE3yC0V)

[45:00] Jean-Denis Greze: 你不需要更多智能了，智能早已摆在那里。真正的问题是如何将其高效且规模化地部署下去。 [26:40] Jean-Denis Greze: 我认为软件的价值单元已经从 App 变成了计算机本身；但其真正含义是，它已经从一个 App 变成了人类所能做的一切事情。 [42:48] Jean-Denis Greze: 对于中小企业和中端市场，我认为 AI 帮助企业做更多事情并取得成功的市场，绝对比编程市场更大。对我来说，这只是时间问题，而不是会不会发生的问题。 [40:00] Jean-Denis Greze: 我认为未来的人在回路会比今天少得多。我们的职责就是让企业适应这种环境，因为由此带来的经济收益远远超过他们当前感受到的风险。

[26:44] Jean-Denis Greze: I think the unit of value in software has changed from the app to the computer. But what that really means, it's gone from an app to anything that a human can do. [50:00] Jean-Denis Greze: The winning platform is one that continues to adapt the business to the best AI. [13:16] Jean-Denis Greze: It is almost impossible to go from an experience that starts from the compliance IT enterprise requirements and then figures out how to make the product experience incredible if you take that as the baseline. [60:00] Alex Rampell: And now supply aggregation is a lot less valuable because human attention can be automated away.

[0:00] Impact in the world is really, really important. Really wanted to build something that mattered and helped people. I get pinged by Ben Horowitz. I found this amazing tool. It's called Town. It changed my life. It's so great. We're trying to build a product that works with people to find more and more opportunities for them to deploy AI to be better at their job. No one had built great AI around these very common knowledge worker workflows like meetings, meeting prep, getting people to be prepared for their day. better than OpenAI and Anthropik, it is risky always to be competing with your suppliers. You get to ask yourself, is that naturally the right state of the world? What if OpenAI is not the best model provider for the next three to five years? And so the first thing is we're just like, look, we're neutral. We will use whatever supplier we think is the best for you in the moment. So that's like one reason. I think the second reason, which is like very important is... AI is getting remarkably good at coding. So why are so many everyday business tests still done manually? In this episode, A16Z general partner Alex Rampell sits down with town founder and CEO Jean Denigrez, the former CTO of Plaid, to discuss what it takes to make AI genuinely useful at work. Jean Denis shares why he walked away from a growing tax software business to build Town, an AI assistant that works across email, calendars, and everyday business workflows. They discuss why AI adoption is still concentrated among people who know how to experiment with the technology, how agents could make useful automation successful to everyone, and why businesses may want AI assistants that aren't tied to a single model provider. Jean Denis believes the next big opportunity is helping people discover what AI can already do for them. Thanks for having me. So what I find cool is that we worked together when you were the CTO of Plaid, and I had led our investment in Plaid here. Yes. And I heard you were leaving and going to start a company. And then I remember you told me you're working on SMB tax tool. Yes, we had coffee. We had coffee here in Manlo or Palo Alto. And I got a little not happy about it. I was like, ooh, ooh. Like, I was so excited. And it's like SMB tax tool. That's rough. All cards on the table, right? You didn't tell me that at the moment. I know, I know, but I want to, you know, I have to maintain a very happy face at all times. You never know. And then this is what I find so funny about this is, you know, this whole story. I get pinged by Ben Horowitz about, like, I found this amazing tool. It's called Town. It changed my life. It's so great. And I'm like, that's good, right? It's like an app, so therefore it falls under my fund. And I was like, wait, that's Jean-Denis' company. But he does SMB tax. That doesn't make any sense. I'm so confused. And that's what I texted you. And then next thing, a week later, we became partners, which I'm extremely excited about. But maybe tell us what Town does. point. From the origination story, it's like you left Plaid, you started doing SMB tax. Kind of walk us through the founder's journey from there to here. I left Plaid. I took a few months off with family because I knew that if I was going to be a founder, I was going to have no more time off for a while. Yeah, and then started exploring in the very B2B way what would be interesting markets where AI could really change the product experience. And for a variety of reasons, tax made a lot of sense. in tax, and LLMs are very good at text. You deal with a lot of structured formats as inputs, like financials and lease documents. And it just felt like a medium where we could take this process of putting together a business tax return, which is pretty different from a personal tax return, and automate it. And I did what engineers do, which is we did technical due diligence. And to prove to ourselves that it could be done. And the feeling was like, yes, within the next two years, the process for a small to medium-sized business of going from your financials and a bunch of documents about your business and a finished tax return that could be automated with AI. And we spent a year building that. And I think… There's a knife day in it, but it worked from a product perspective, but not from a go-to-market perspective. Meaning, even though... I have no doubt that tax prep will be automated in, How do you reach millions of businesses efficiently and sell that product? That had a lot more question marks than we anticipated. And basically, after one year, we had what I would describe as a good business pre-AI. That is, if I showed you the revenue number and the revenue growth in our first year, like a little bit below a million, and if I showed you what I thought we could do in year two, we thought we could get to like two and a half to three million in revenue, I think pre-AI you would have been like, That makes a ton of sense. [5:00] Stay high. this was not going to be a venture-scale company. And also, I think even more importantly, we realized that the incumbents were going to be able to use the technology to automate tax prep for their existing customers faster than we could grab market share. And that's a key thing if you're a founder. You have to think about your advantage if you're a new entrant is you're faster moving. You can build a different product. You can use different go-to-market approaches, but there's a ticking clock on the wall, which is how quickly can you get real word-of-mouth penetration before the incumbents who have other advantages, they have distribution advantages, right, for example, before they can catch up and copy what you're doing effectively enough. And it felt for tax, like, we wouldn't be able to get to... hundreds of millions of revenue before our competitors, the incumbents could, because you can only sell tax once a year. Yeah. And that's like a, you know... They don't only sell it once a year, it's like the interaction. It's not like a... It just fundamentally cannot be a DAU tool. That's right. And it also doesn't have a ton of word of mouth because it's rare for people to brag about their tax prep situation. And when they do again, they can only do it once a year. After they've done the tax return, it's... When you brag about how good your tax prep experience is, usually it's also not the time when people are purchasing or looking for a new tax preparer account. And so they're going to forget in the next year. That's right. Those discontinuity make it really difficult to hyperscale. And so that's why it was not a great idea as a business. And it's funny, one of my best friends has built a successful ed tech company. And one of the things you always complain about with ed tech, it's the same thing. You can only sell in ed tech twice a year. window where you can sell at the end of the fall and winter season before the new year. Those are the only two times where you can really sell software to schools and to school districts. And he was like, that really limited my ability to hyperscale. And I chose to forget that when we were doing tax. Yeah, so we built tax. And after one year, after the first tax season, my co-founder and I, we went for a lot of walks with a lot of coffee for a week and really pondered the meaning of life and entrepreneurship and capitalism and decided we should do something a more fruitful market to play. If we want to impact millions of people, there are better ways to get at that goal. Which is a pretty brave thing to do. Most people probably wouldn't do that, especially because it wasn't not working. in the moment it felt like failure, like we were failing. But I think, what are your core motivations for building a business? As a founder, there are these moments, they happen more often than you wish they did, but you really ask yourself the question, why am I doing this? And for us, it was like, Impact in the world was really, really important. We really wanted to build something that mattered and help people. So we stopped believing that we could have as much impact in the world as we ought to have because we didn't feel like we could scale the business quickly enough. The second reason why I think you build a company for us was really enjoy working with our team. But everyone on our team has an opportunity cost They could be working at town or they could be working at a lot of other great companies in the valley. And so if you stop believing that you can have a great outcome, then it's really hard to come into the office every day and look at people in the eye. When they tell you, like, hey, I'm having a hard time with this, or when you're trying to convince a candidate to come join your company and they have an offer from someone else that's liquid, and you're like... "Oh, well, who am I as a human being?" - Yeah, I know, it's very hard. - I have to believe in it. - I know you're paying, I was there. - So there's a few things that we looked at, and we were like, "Well, we gotta make the decision." You know, the counter on the other side is, maybe we would've figured a way, and there's always this thing, you never know for sure, right? And, We're very resourceful humans, so you can always try something else. But it just felt we didn't have the idea in our brain to make it a really breakthrough business. And the hard part of it is admitting to other people that you failed. And to some of your employees, maybe like our tax team wasn't going to have a future at their company for when I'm going to do tax. And you face that and you're like, first of all, who cares what other people think of you? My friends will still be my friends regardless of whether I succeed or fail. We can treat them really well in terms of severance and other things and make it okay. And then you realize it's not as hard a decision as you might have emotionally reacted to initially. And it's the right decision. You start to get that conviction and then you just do it. One last little bit is because tax is annual, I think... Making the decision at that point in time, it was the only time we could make the decision. You can't pivot like April 16th after you've told a bunch of people that you're going to do their tax return. You can't pivot in January after you've signed half of your partners, right? We had a window between, let's say, right around October 15th of probably a month. where we could live in both universes, the universe where we kept doing tax and the universe where we couldn't. And after that, the door was closing. And so I think That time pressure helped us make the decision more efficiently than otherwise. I think if you're not in a cyclical business, I think you can always tell yourself, well, let me try the next idea for a month. Let me try the next idea for a month. [10:00] try the next thing for the next three weeks. Maybe that'll work. Maybe this will work. We had a short window. where we could change the direction potentially. And once the window was gone, it was like, hey, we have to decide right now. Otherwise, we're going to be here a year from now and a year from now. We won't have the capital to make the change. So from there to, obviously the name didn't change, it's still Town. It's a good domain. Town.com. It's a very good domain. But how did you get from there to where you are right now? We basically explored... about four ideas pretty seriously. And it's not worth going all four. What led us to what became Town was Um, this question, which is why is no one built... a great agent that works out of email and calendar. And we started asking that question, one, because when we're building the tax product, a lot of customers were telling us like, hey, it's amazing that you're automating tax. But I also spent a lot of time setting up. I set them up meetings with my customers. When I go into the meetings, I'm trying to, or my clients, when I go into the meetings, I forget what the meeting was about. Clients always ask me the same tax questions. Wouldn't it be awesome if you could pre-draft answers for me? to be prepared for their day. So we just started exploring that idea, and we did it at the time that Opus 4.5 was happening, and agents, at least for coding, were able to do more and more autonomous work. And I would say... That very simple insight plus the model suddenly being able to be more authentic meant that we got product market fit very quickly. It's a magical thing. once you have it, you kind of want to hang on for dear life and keep pulling at the string forever and ever. And so we didn't sit, we didn't actually sit back and ask ourselves like, how big could the tan be? Or like, will everyone want to do this? Or, you know, like, what does it mean to win? We were just like, hey, we're getting a ton of pull from users. Let's go after this for a few months really hard and learn about the market and see what the opportunity could look like. And, you know, as we did it, we started, you know, we start to realize, well, Most businesses... run out of email on their calendar, right? Especially any role at a business that's external, that's like sales or business development or recruiting. The way companies work together is through email, mostly. So we're like, "Oh, the TAM's got to be big." And it's not just in the US, it's like everywhere in the world. And then when our customers start telling us what they liked about the product, we realize, "Oh, there's a depth to what they're asking us to solve for them. This isn't as simple as just hooking up an LLM to your email. There's a lot more that goes into making this experience better and better. So we start feeling like on the product end, I'm not saying necessarily we had a line of sight to emote, but there was enough complexity that we thought, hey, we could really dig into this and build a great product experience. And basically, we did that from December of last year until April, and we just kept going every day. Like every day, the product grew by a percent or a percent and a half. And our customers kept asking us for more, and the roadmap was very clear. And we got pulled until the point where we're like, hey, enough with this closed beta. We're ready to launch the product out into the world. And one of the things that I thought was very bold about what you did as well, but you're also we talked about this a lot of, do you go PLG, right? Or do you go, like you have big companies that say, hey, give me your product, I want it. But you kind of realized that if the customer is not the user, the customer is the person that writes the check, the user is the person that has to deal with the product. If the customer is not the user, you might end up building a bad product because you're making the CISO happy. You're not actually making the user happy, you're making the customer happy. Yeah, my co-founder and I worked at Dropbox, and I think one of the big learnings for us from Dropbox is you – You can always take a world-class product experience and and figure out how it works for the enterprise. It's really easy to go in that direction. Let me rephrase that. It's tough to go in that direction, but it is possible. It is almost impossible to go from an experience that starts from the compliance IT enterprise requirements and then figures out how to make the product experience incredible if you take that as the baseline. we felt, and I think until now, Until maybe a couple months ago, we felt like there was so much alpha in just making the product better. That should be our focus. growth was just happening naturally. We're like, if we figure out the product experience, and if... Our foundation from an enterprise perspective, from a data security and privacy perspective is good. We can figure out later. [15:00] week. shape the admin side of the product to be perfect for an enterprise. I think that was the right call for us. you know, especially as consumers' expectations around what AI can do, Those are going to be consumer-grade experiences. Our thesis is that CIOs CEOs of companies are going to tell themselves, hey, this amazing personal assistant experience that I get on the personal side of the world. I want that for my company. And so whoever has... an enterprise experience that feels like a consumer experience, they're going to be set up to win. And we were like, "That's where we want to be as a business." And I think we've mostly actually succeeded in getting there, and now the question is, First, we have large companies that are using the product. So, it's like, how do we go from just one enterprise deal to being able to do dozens effectively? How do we take the data? privacy, security requirements of not just like a couple of customers but you know How do we satisfy those for every enterprise company? That feels like a much more knowable problem now and we can go after that with our heart out because the core product experience is good. Yeah, I agree. I know you and I have talked about this a bit where I remember having a BlackBerry. - Yes. - And call it 2007. The iPhone launches and then Steve Ballmer's like, nobody's gonna buy an $800 phone, this is gonna be a flop, it's such a joke, blah, blah, blah. Eventually Microsoft buys Nokia. But the BlackBerry died pretty quickly because my theory on this at least, is that every CEO got an iPhone for Christmas that year or something, was taking pictures of their kids. They're like, holy crap, this is really good. "This thing's not that good. "Hey, I'm the boss. "You go figure out how to make this thing work here." - Yep. - Right? And we're probably, I talk to a lot of enterprise leaders, and it's funny, the ones that have seen the future are like, I need this. There are some that are like, oh, yeah, like, AI doesn't work that well. I use Watson or something. Like, there are still plenty of those where there are skeptics. But, like, it feels like the iPhone has emerged in terms of, like, people have been exposed to the wonders of AI in a personalized setting, but not in a work setting. And it's kind of an open question as to why. And part of it is the permissioning part of it. But I'd love for your thoughts on kind of if you agree with that framing and kind of what the next few years of, - One, I agree with the framing. I think a little side note, there's this company, I believe it was called Good. - Yeah, Good Technologies. - Yeah, which was, it was an enterprise-ready email client for the iPhone. - Yeah, it was an app, it was like the ugliest thing in the history of the world. - That's right. - Yes, I used to use it. I think it's not just a BlackBerry story. I don't want to be good. Because good eventually is not a good business because people are like, yeah, sure, you're enterprise on the iPhone, but you're not at the level of quality of every other iPhone app. and the enterprise features that a business is going to care about. So Yeah, so... I often think about this... things that happen at the speed of a human, right? And First of all, Assistance as a category is used by a few million folks. worldwide today, it's still very early in the adoption curve. So I don't even think all of the CEOs... have gone there. And just like the iPhone, it will probably take 6, 12, 18 months for most people to experience this magic, even in their personal life. Once they do, then there's a question of how do you adopt that within the company? And I think it is more complicated than just the iPhone because – The way agents work, is they're more and more effective... if they have access to the data And the APIs... that run your business. If you have access to those things in a secure, privacy-preserving way, then they can do lots of things for you. But if they don't, they're only really as smart as the data they have the access to. So I think The way we think about the journey is, first, uh, be able to go to any company, and get a few people at that business simply by connecting their email and their calendar. So not the core systems of record yet, to feel enough magic that they're like, oh, this is going to bring a lot of value in my work. And as I said, for external facing roles, The calendar and email are real sources of insight about how you can automate part of someone's work. And so then you can show up with the assistant. You can be like, I can help prepare you for these meetings. I can tweak the decks for that meeting to use the design language. [20:00] and to be updated with the latest numbers that are in this PDF that was sent to you. And just from that, people will start to feel the magic. And then you ask yourself, well, Have I earned the right to ask you to connect your CRM? Have I earned the right for you to go to your IT team and ask all these other tools to be connected? And I think that journey will happen more at the speed of humans. And you have to have... you have to have the other side of the coin. You have to have... the audit logs, you have to have the posture that allows, you know, privilege information not to be shared with the agent. You need all these, you need the cost control for the CFO, so that they understand that if their whole team operates on town, they're understanding where money is being spent, and that tokens are being used efficiently, so that It's all that side of things that will take more human time to adapt because you're still selling into a Fortune 100 company. has to be run with a certain level of diligence. As you move down market, right, as you move towards SMBs and mid-market firms, I think you have fewer of those requirements, right, and the adoption will be much faster. And so our current bet right now is we are really focusing on from a product experience and self-serve growth the SMB market. the mid-market, and then from a what security compliance finance features do we need, focusing on the enterprise. And we think we can. knock out that whole part of the space. Yeah. I mean, I think there's also a Javon's paradox to this whole assistant space, if you think about it, right? So I have an executive assistant here, but still, I don't, I mean, I trust her. I like her. I don't want her reading all of my email. And actually, this just came up. I was in London. I thought I had a dinner to go to, I could have sworn I got an email about it. I texted Snap, my townie, and I was like, don't I have a dinner to go to? And he's like, yes, you do. You do. And it wasn't on your calendar? It was not on my calendar. It was my fault. Not my assistant's fault. It was my fault because I didn't respond to an email. They didn't CC her. That's not their fault. That's my fault. And it caught it. But now I have it automatically going through. It reads my email all the time. And that's something where, so I can, or my firm is nice enough to give me an executive assistant. There are plenty of people that don't have one. And it turns out, you might just say, well, the market for executive assistants is rather small because most people don't have them. Turns out if you could actually produce it via silicon, the market's quite large, and then you find things to do. Like, I'm going to have somebody read my email all the time, which I would never do, both from a privacy – it's actually – it's like a twofold – it's like a Javon's Paradox of Privacy that, obviously, like for those that don't know, Javon's Paradox is about how when the steam engine became more efficient – Like James Watt made a more efficient steam engine. And then the assumption is that the demand for coal would go down Because what used to take like a pound of coal now took like an ounce of coal, but actually it turned out that it went up like a thousandfold because now you had something that could do it very efficiently. So it feels like that probably exists for the assistant market. But I'm wondering, how do you define your market? Is it like assistant? Because that seems like it casts too narrow of a brush on it. I think of the market as... We're trying to build a product that works with people to find more and more opportunities for them to deploy AI to be better at their job. And I think in some case, so I think of our product as a token allocation, right? machine. We're trying to give options to the human about where they could apply tokens and then we choose the human can choose whether they want the outcome of that. And I think as the tokens get cheaper, they're are more and more places where you would apply tokens, where you would never have done it with today's expensive tokens, and you definitely would not have done it with humans before. Like, A very, very basic example is writing prep notes for every meeting that you walk into. Even if you have an EA, your EA is not spending... for every meeting, every 30 minutes that you go into writing perfect prep notes. It's just not a good use of time. But an LLM can do that for, I don't know, a penny, right? And so then suddenly for anybody feeling like you have – prep notes for every meeting that does some research on who else is in that meeting, have said anything in X in the last week that might be relevant for the meeting. All this long-tail research is now worth it because it's so much cheaper. So our job is to find things that look like that more and more and to tell you as the user. hey, do you want us to do that for you? So I give you another example that I'm recently pretty infatuated with. It's like if you are a small business owner and you run a restaurant, in America, your wholesalers that is the people that you buy the produce and other things that go [25:00] into making food at your restaurant. uh, They... Email you. probably once a month. maybe once a quarter, with an Excel spreadsheet or a PDF of the prices for all their inventory. So like thousands, if not 10,000 of items. And the prices are always changing. And if you're restaurant owner and you have a finance person part-time, what that person should do for you is once a quarter, Look at how the prices have changed to see if you could order some things from different wholesalers to see if you're maybe getting ripped off or paying too much because the price of something has gone up 20 or 30 percent. Mostly you don't do it. Because it's like a rare thing, and even though it could save you money, it's a pain in the ass, it's not part of your operations, you don't have a full-time person to do it. with the town. having a thing run once a month and do that analysis. Like literally you can prompt your way to having that exist for you in five minutes. And if it finds nothing, 11 out of 12 months, that's worth doing. It's going to cost you like a dollar a month to run that thing. But the interesting thing is the one month out of a year where it finds something, it's going to find like hundreds, if not thousands of dollars for you, right? And so now this thing that may or may not have been worth it, you should do it. at least once, you should do it whenever there's a change to the price. Maybe you should do it daily if you have a wholesaler that changes prices daily, it doesn't happen. Our job, though, is to show that opportunity to the restaurant owner because they won't It's not that they aren't very smart or don't understand their business very well. They do. They're not going to think that that's valuable. So what you kind of want to tell them like, hey, listen, Do you want me to always make sure that your wholesalers aren't changing prices in a way that's negative for your business? Sure, do that. You don't really care actually after how often it runs. What you care about is once in a while you get an email that tells you like, hey, you... you should probably move all these things to this other wholesaler. They have a discount. Or this one is really raising prices on these things, and it's getting a little too expensive. Maybe you should remove those menu items. It's not rocket science to do this, right? But you have to have a product that tells you that this thing is possible that makes it easy. Well, the interesting point around that example is, if you went to a restaurant owner and said, do you want an executive assistant? Right. No, they would have no idea. Whereas if you said, do you want to save, it's like Geico, right? It's like, do you want to save 15% on your car? Do you want to save money on iceberg lettuce and I can save you $500 a year and charge you $5 a year? The answer is yes. But how do you know, I mean, this is kind of the product marketing question. Because I remember the two things that we talked about when we first talked about the new direction, the future direction of the business post-tax were PLG versus sales-led growth. It's like if you want to build a product that is designed for the user, you have to build it for the user, not for the corporate overlords, all the lawyers at the company. And eventually you can work your way backwards into enterprise sales and sales-led growth. But the other thing is this kind of horizontal versus vertical. And this is a question that a lot of companies struggle with a lot. And this kind of actually will bring us to a different point on, like, what do you think about the big guys, right? Like, obviously, like, Anthropic, OpenAI, and everybody else who's kind of building things both for enterprise and for consumer. But how did you kind of think about horizontal versus vertical? Because what you said, hopefully I'm going to capture it correctly, was we're going to start off as horizontal as possible. Because we don't want to be like, you know, we don't want to be siloed into like something that might be, we don't want to be siloed and branded as like the save you money on iceberg. lettuce company. We want to do that. And we're going to learn. This is my sales pitch to your candidates, by the way, now. So hopefully I'm getting this right. We're going to learn from hopefully thousands, tens of thousands of businesses and people that use our product. And then eventually we can verticalize from there because we're going to let the market define that. And we're not going to presume that we know about the iceberg lettuce problem. We wouldn't know that until somebody actually downloads town, uses town, they learn from their townie. And then we can that can be a vertical, but I guess, how do you think about vertical versus horizontal? So the way I think about most AI products... today is that the capabilities will be very similar. meaning most harnesses in the work context will be able to do... the same things. And so the differentiator I think what has been the main differentiator in Silicon Valley for a long time, which is you have different product features that are ahead of market, it's going to be more and more difficult to sustain that kind of an advantage. And we'll put a pin in this one, but we can have an interesting conversation about, I think the unit of value in software has changed from the app to the computer. But what that really means, it's gone from an app to anything that a human can do. And a human can do almost anything with a computer. So we're abstracting away the app layer totally. And I think the value will drop in the app layer. So, So if the capability [30:00] aren't different than how do you build a successful company. And I think Our take on that is... We're betting that as AI gets better, It can better customize. the product for each user, like very specifically to their need. Meaning if you're an electrician that signs up for town and you're onboarding, the things we recommend to you that we can automate are very specific to electricians. And if you're a venture capitalist, they're very specific to a venture capitalist. And if you're an EA, it's specific to an EA in that industry and so on and so forth. And – So how do you do that? Well... The models aren't yet good enough where when you just ask the model, hey, recommend for Alex, who is a VC, things that he could automate with town. I think you get pretty good surface level suggestions, but not really deep ones. So what we do, Because we don't look at user data. Part of being enterprise-ready is you can't look at user data. So what we do is we look at aggregate metrics of... Which, who are our power users on the platform today? And then we go talk to them. And when we talk to them, they tell us the kinds of things they've automated. And so we might talk to 10 power users that run restaurants. And what we'll see when we talk to that, there's five things in common that they all do. And then we'll sit internally and we'll build prototype models automation, as we call routines, for restaurant owners that we think are really good, like the one about wholesalers. And we'll use that as the seed. And so when someone joins, we look at your email and inbox and we're like, you seem to be a restaurant owner. And then we have templates of what we think really good problems and potential solutions are for restaurant owners that we've kind of hand carved from those interviews. And we put those in front of you. Yeah. You sign up and you start using the product. And then AI is getting better. And we're talking to more restaurant owners. And now we don't have just five. We have like 15 things that we can do for restaurant owners. Still a little bit manual, but you can speed it up. You can schedule interviews automatically. You can ask the restaurant owners. You give them an Amazon email. You get a free Amazon card, and you're doing an AI recorded interview of them and what they do with your piece of software. So you're no longer operating at the speed of humans. You're operating at the speed of AI. And this library is getting better. And then the models get better. And so now the recommendations use the library less. And what we think is over time through this process. through this approach, we'll be able to offer 80-20, very customized software for customers for people and people aren't just one role in the future. Yes, you're a restaurant owner, But also you run finance for your restaurant too because you don't have a separate person doing that. And also you have a vacation rental that you rent. And so really you want your townie to be good at those three things. And also you have two kids at school and it can help you schedule their activities for the summer. So it's custom software all the way through. Our belief is that you can use AI to build that. I think the other strategy that you see in the market, which is I think what the labs do, is an FDE approach. And this is very different. FDE is like, I'll have a human go into your business, understand how your business runs, then they'll build custom software with AI that uses my lab's AI in order to sort of view better. I think the FDE approach… If you're an enterprise that's willing to spend $5 million a year paying OpenAI for FDs, it looks really good right now. But I think... Over time, as the models get better, the FDE approach actually gets much worse. And our approach, which is much more around automation and using AI to build the custom software, will scale much better. And so the bet we're making is we start with PLG, SMB, mid-market enterprises that are very tech forward because their approach scales really, really well. We'll let other people fight in the enterprise for this much more handcrafted software. Thank you. in the long term than these more custom approaches. But maybe also define why is Town better than OpenAI and Anthropic? They now also have connectors like you can hook up your Slack, you can hook up your LinkedIn. I mean, obviously, I believe I'm both an investor and a power user of Town. So I have my answer, but I'd love to have you expound on why use Town. And also just like business. I mean, so many people that are probably listening to this that might be entrepreneurs are also wondering like, uh-oh, are those guys going to eat my, not you, Town. layer and you have the application layer. That's always been the case. Like, you know, Cisco and Akamai were infrastructure companies in the 90s, and then Amazon and eBay were application companies. But never in a million years did Cisco say, you know what, we're going to build a bookseller. We're going to build an auction website. We're going to build a directory like Yahoo that competes with... No, they were just happy selling their switches and load balancers and everything. So I'd be curious, like, you know, maybe define how you think about town versus the people that you often buy the tokens from, and then... [35:00] How should others maybe think about this kind of weird time that we're living in where it's like infrastructure is becoming application? it is risky always to be competing with your suppliers, right? And we are competing in a way. Because they have cheaper cocks, right? That's right. We're competing as a supplier. So you've got to ask yourself, is that... naturally the right state of the world. So, The small answer, which is another good one, is... We believe very different things about the experience of the user initially around email and the calendar that, for whatever reason, none of the other players have really focused on. And there's a lot of complexity to email and calendar, which makes our product much, much better there. But that's a small answer because Obviously, someone at OpenAI, Adentropic, at Google, could wake up and decide to do that. So the better answer, actually, is to think about our customers and what they want. And... When we talk to a mid-market firm, so that's a firm with 500 to 1,000 employees about town, What we tell them is like, hey, listen, do you know which model provider is going to be the best model provider over the next three to five years? Like, what's your opinion on that? And they don't know. And they don't know. And you're like, well... You know, if you buy... the OpenAI product, well then... you'll always have open AI models in that product. And what if open AI is not the best model provider, doesn't have the best intelligence for you for the next three to five years? Then as AI does more and more for you, you may be at a competitive disadvantage. Or what if they hike the rent? Yeah. All of that. And so the first thing is we're just like, look, we're neutral app. We will use whatever supplier we think is the best for you in the moment. And that's like a big part of it. It's also not just general intelligence, like which supplier is better at voice, is better at video, is better at slides at that time. Like we feel like it is our job to determine from... all the AI capabilities, frontier model, open weight, what is the right thing for you as a business. And I think you can save money there, but also you can de-risk ever having made the wrong partner. And there are companies that... picked OpenAI like two years ago. And then Anthropic was better at coding for a 12-month period. And those companies had to use both because they realized to be competitive, they needed Claude, right? And they needed those models. And I don't think, I think if you're a smart business, you really think about the neutrality of who's providing intelligence for you and who's routing. So that's like one reason. I think the second reason, which is like very important, is our stance is that you own and truly own as a business... the data and memories and context that's being, um, that's being created on town. And so, you know, in the old world, pre, pre, computing, if you're a business, you had different assets. You had physical assets, and you had maybe human assets. And increasingly now, maybe people call it context, but there's all the things that AI knows about your business to help you run your business effectively. And I think you should own that entire bucket. And I think you should own it so much that if there was ever a better solution in town, switching to that other solution... would take, you know, a day. That's what I'm selling. I'm selling you want to own that asset. And I think if you look at anyone who's deploying FDEs, to build software for you, they actually, what they're trying to do is they're trying to make it more difficult for you to switch at any point to someone else. So I think that's the second. So first you want neutral intelligence, but I think you also want the app layer. on top of the intelligence, to be one that you feel comfortable, you could move to something else. And in the new world, that's really like, you know, it's owning this context layer and having someone who makes it available to you. Now, that means town, the risk that I'm taking is that the switching cost for my customers is easier, right? But that's my selling point. And look, if I can't have the best product in the market, then that's okay. Then that just means I'll lose. But that's like my competitive stance against the labs. me is just on training and I think you look if you're a large enough company that labs won't train on you but it's the way you operate your business, If that ever makes it into the weight of someone that's reselling to someone else, that basically you're selling what makes your business successful to someone else. And you really don't want to do that. So I think we are on the training layer. We are... We don't train, that's just not part of our philosophy as a business. And then the final one on competition is... We believe different things about the world. and about I think the relationship between a human and AI, where we are probably more forward-thinking than even though [40:00] So, you know, the team's product for town has a mode around basically building an internal CRM where we're trusting agents to decide what information at the company that is available to other employees at the company. Let me just give you an example. You're in town and you say, does anyone know someone in the finance team at this potential customer? And town's able to talk to the agents of other employees at your current company to find someone who's connected to that feed that you want. That's like a really powerful product, but from a privacy perspective, you really have to design it really, really well to work. And the way I think about that is we are really thinking about how do you get humans to operate at the speed of AI and not at the speed of humans. So the speed of humans version of that question that just asks is you would go in Slack or Microsoft Teams and be like, does anyone know someone at that team? And then everyone at your company would have to read that message and think about it and you get an answer like a daily. Yeah, exactly. Humans are slow routers. That right And for us you going to get that answer immediately because we decided that for that kind of a question we will trust AI to mediate the security and privacy of that request within a company And we believe things like that that can really speed up a company that I not seeing anyone else do And it's like when you're building a startup, you always have to think about what does the future look like? And I think the future looks like a lot less human in the loop than we see today. And I think our job is to get companies to be comfortable with that universe because the economic benefit for them far outstrips any risks that they would see in that experience today. So those are the differentiators. I think that data ownership and neutrality are so important. And look at what's happening to Rogo, what's happening to Harvey. If you look in OpenAI, they're building vertical companies that are competing with that. And from a competitive standpoint, you're like, oh man, those labs are really unstoppable. But if you're a neutral business choosing who to work with, you've got to think about, well, how far up the stack are they going to go? So I just think strategically it shouldn't make sense for most large companies to really be beholden to any one lab. It seems like it also like there the Goldilocks zone of relevance here which is if you had trillion in revenue I would obviously be very happy as a shareholder you be very happy as a CEO and co but you probably be the number one target for OpenAI and Anthropic if you have $0 in revenue and nobody uses your product, they just wouldn't care. But if you're, and this is kind of like coding fell in the wrong side of the Goldilocks zone, arguably, like the way that you, like if you were a coding agent right now, like what is, who's adding $10 billion a week or who knows what it is for, like Anthropik's adding so much, it's so relevant for them. They cannot, to use a double negative, they cannot not ignore it. And, or sorry, they cannot ignore it, just single negative. Whereas for what, you know, you're doing where, again, this sounds like an insult. It's not, right? It's just like, this is probably not the biggest thing. I think it actually is the biggest thing, longer term, but short term, it's not coding, right? It's not like, oh my God, I have to do this or else it's game over for me. Because like OpenAI, it seems like they learned that lesson from Anthropic where OpenAI had like 50 bets and then Anthropic went all in on code and that did really, really well. So, you know, there's also this element of just kind of like focus and relevance, which it seems like they not going to focus Like you know healthcare really big like law really big Coding really big These big bets they almost have to make whereas these kind of horizontal ones that are spread out it seems a little bit harder for them to commit to. I think for the SMB and MIG market I would say I would agree with you but the market is the market for AI doing more for a business and helping the business be more successful is bigger than the coding market. For sure. And so for me, it's a question of when, not if. I think in the current environment, it's not going to move the needle at those companies, but I think in 12 months it will. And so it's going to move very, very quickly. It's just right now you cannot not be. This is the one that I was looking for. You cannot not be in the coding market. Correct. Whereas this is something where it is like the iPhone, BlackBerry. There's a diffusion point here. And actually, this is another question I'd love to ask you. I have a theory, I'd love to see if you agree with it. Silicon Valley almost overhypes AI, if that's even possible. I don't think it is possible to overhype AI, because every week it's better. The rest of the world, I think, is almost underhyped AI. And I'm kind of curious, how do you think about, people talk about crossing the chasm, but this is not a great example, or it's not like this is a technology [45:00] that only exists here. It's nowhere else in the rest of the world. No, people are using it throughout the world. It's not evenly distributed. But I guess, how do you think about where are we kind of short AI or what pockets of the economy, either geographically or by profession? There's certain areas where—because this is the cool thing about having a very horizontal product and having a very horizontal PLG product where people are finding town from all over the world. Like that wasn't intentful. It wasn't like, let's go get big in Japan. It just became big in Japan. So you're finding this everywhere. Like, are there pockets that have surprised you? And there are pockets where you're like, wow, there's actually a lot more opportunity here than people have thought of. I'm thinking about individual anecdotes. I'm thinking about this like this business that rents like surfboards on an island to people and then uses town to inspect like surfboards and, and, little beach buggies, whether they've been damaged, to decide whether afterwards the person should get their deposit back. That's pretty cool. Yeah, but it's like... I'm like, that is not on my bingo card. It's not on my 128 by 128 board bingo card. It's just nowhere on there. Really what's happening is there is a tinker at that point. at that small business who there was something that was annoying to them. And then they figured out with town how they could automate comparing these pictures and deciding whether to release people's deposits. And And, What makes it work there? is the tinkerer And so my answer to you is actually we're seeing use in all the weird spots of the economy and the The unlock is, is there someone over there that gets AI enough to want to tinker a solution to their problem? The problem with that is that there aren't enough tinkers in the world or there aren't enough people that have the time to be tinkers to help their business. And I think that's why AI is not diffusing. The reason why it's not diffusing is because even though in Cloud or in ChatGPT, has existed for the last 12 months. features that a lot of small businesses you could use to save you so much more time. But instead what you're doing is you're using them as a replacement tool. as like a better Google search. And you're like, what's happening? And it's because the product experience can't just be a text box. And what we have to do is we have to, you can't teach someone. You have to just show the value prop to someone and do it for them. I think that's like the real unlock. And so whenever, you know, for our product, like the, What makes someone sick with town? It's this. They've connected. Their email, their calendar, and ideally one really important system of record for their business. and then they've built one custom integration. That means they've built one thing, That didn't come out of the box. That was magical. And I'm trying to just get rid of the custom integration. I'm going to get each person, which is as soon as you connect your system of record, can I deliver value to you in a way that feels very you, that really helps you? And that's the opportunity, and that's why you don't see it spread anywhere. evenly, it's because in Silicon Valley, We're all working at companies next to engineers. And so we're much closer to this innate problem solving and tinkerer. Like we have tinkers everywhere in our firms. So yeah, our firms are using AI effectively. And most of the world just doesn't have that. And so that's the real problem. And if you can fix that, if you can make a non-tinkerer into getting the same value as a tinkerer in the world of AI, like... you could be a ginormous business today because AI can solve so many business problems. You don't need more intelligence. The intelligence is there. It's just how do you deploy it effectively at scale? Yeah, it feels like it's an integration problem. It's a product marketing problem. Yes. Because... Everybody. And it's also, I think this is really interesting to me as well. I met a professor at Harvard. Very, very smart woman. She's like the number two person there. And I'm talking to her about AI. She's like, yeah, I tried this a few years ago. It was okay, not great. And I was like, no, no, it's really good now. So I had ChatGPT rewrite her dissertation. that she wrote 20 something years ago, but everything is now current. Take everything today, and she was like, this is amazing. What AI is this? And I'm like, it's chat. She's like, how is that? It's because this is moving so quickly. And a lot of people, they judge things on the present. I think that's the other thing is Silicon Valley tends to judge where it could be. And that's where sometimes you oversell reality and reality distortion field this, that, and the other. today, and it's like the slope of this stuff is also such where it's like if you follow this curve, I mean yes, where it is today, it feels like people aren't taking [50:00] - Taking advantage of it partially because they were almost, ironically, early adopters. - Yes. - They tried something two years ago. They were like, "Yeah, this hallucinated, this did that. It didn't really work that well." And then they've kind of ossified that version of the reality. And if they try it again now, it's just, I've done this experiment like 10 times. They're like, "Where did you get that AI?" That's not the AI that I have. It's like, "Yes, it is. You just have to try it again." - Yeah. Because the product space is moving so fast, this is going to keep happening, right? I think the winning platform is one that continues to adapt the business to the best AI. That's the problem because the AI will just keep improving and improving and improving and improving forever. And people's mental model of what's possible and not possible, what's difficult and not difficult will be upside down. I mean, just with computer use and browser use today, like in the last two and a half months, the assistants can do things that were near impossible six months ago. And they can do them 90% of the time, whereas before they were doing it 30% of the time. Most people, they've not experienced that. And that's why when you do a demo with town of it's shopping for something for you, or it's using your software. to do something that you're like, wait, don't you need a UI to do that? People are like, wow, I had no idea I could do this. But I don't think, I don't know how you market that to people, because I think if you told people like, oh, town now has computer use, It doesn't mean anything. It needs to be much more concrete, like, hey, like, I can make posters for this event that look just like the ones that you've been making in Canva. Hey, I can do it using your Canva subscription, right? Because you're already paying for it. And do you want me to just do that? And the person will be like... "Yeah, let's do a pass at that." And then five minutes later, when it was 25 cents, it's showing you something that's super valuable, then it needs to go like, "Hey, do you want me to do that "before every event like this that I notice in your calendar?" And then what happens is you separately change the design of your small business or your mid-market firm and then it should be like You shouldn't even have to ask, but the next time it makes... the flyers should look like the new design. Because it should have noticed that for you. When it can do video effectively, it should be like, hey, do you also want me to make video assets so you can advertise this on Instagram? It just needs to suggest this increasing... Like, things that feel increasingly impossible in today's technology, but really it's just following the curve of what the models will be able to do for you. Well, you've probably seen that agent traffic, and this is largely not assistant agent, but just kind of agent traffic writ large has vastly surpassed human traffic. So that's kind of point one. And then the other thing that's been in the news a lot is about this kind of adversarial versus call it maybe synergistic approach to agents. Yeah. kind of this new generation of agents, not agents as in like, you know, ChatGPT is, you know, just searching something on the internet, but actually doing proactively things on your behalf. So like, Amazon has obviously adopted a somewhat adversarial approach, which makes sense because they're an advertising company. And like, there is this covenant that existed on the internet for a long time. I go to Yahoo Finance to look up a stock quote. Yahoo pays money to go get a feed of stock quotes from the NASDAQ or something. But my eyeballs then hit an E-Trade ad. And the I, Yahoo, give you Alex free stock quote if you give me your eyeball. And that covenant has been broken. So people that have kind of hit that broken covenant are adversarial, but it feels like there's a big opportunity to be synergistic. It's like, wow, it's like the new SEO. If I can be on the receiving end of agent traffic and not block it, not figure out how do I stop this, but how do I welcome it, it seems like a massive opportunity on the other side there too. Yeah. I mean, you know, Amazon's loss should be Walmart's gain. Yeah. Right? I mean, Walmart... They don't have the advertising business. And so... their incentives are going to be different. And I think some... I mean, to your point, distribution has changed. I think the part that's interesting about the agent, by the way, is how does distribution change if you no longer trust your agent? themselves not being interested in things is like a little bit of an interesting question mark. So say you use the meta agent. And Meta has large advertisers. And then you ask it and say like United Airlines advertises a lot more on Meta than, you know. Nobody airlines. And then you ask it to do a flight. What happens if you feel like it's no longer giving you the neutral thing? Because this is the thing is there's always a... There's what is... There's still human incentives... to promote products and marketing. And so what does that look like in the world where you the human's eyeballs are no longer looking at things is a very open question. I'm a firm believer in the like you want a neutral agent. that is yours, that you have no, like the only source of money for the agent is your wallet as the user, and then who [55:00] Versus creating that agent, their incentive is to give you the best information for you as a human with no distortion in terms of the great forces of marketing. And then you as a human, you have taste and you express your taste as you make the choices that you do. And you can ask your agent like, hey, what are people talking about on Instagram or like what's being advertised on it? You can ask that question. Because there you want to be flavored by the world of advertising and marketing. But again, you want your agent to be neutral vis-a-vis that. You talk about adversarial versus not. I think the other thing that is interesting to me conceptually are experiences. like friction that you enjoy as a human, for that friction that you don't. So, me walking around Soho in New York City shopping, It's friction that I enjoy. I like that I have to walk between stores. I like that there are other humans waiting in line because I can copy their styles. I like the friction. It's a part of that shopping that is enjoyable. I like doing it. Then there's friction that I don't enjoy. I want to like order toilet paper at home. I don't You know what I mean? I just want the same toilet paper that I've been ordering for a while. Just get it in there as quickly as possible and as cheaply as possible. And so I think sometimes when we think about agents too much, we think too much about reducing friction and toil and making things as efficient as possible. But humans don't always optimize things. you know, for efficiency. And so, you know, the Amazon example for me on the ad is like, actually, I think there's times when I'm on Amazon, like when I'm buying Legos for my kids. Where I do want ads, of other construction like toys for my son. Because I may realize when it's not Lego branded, and I don't know about it, and it's interesting, Ask my agent for some reviews of it, and I'm like, that's interesting. Is it popular with the kids? And maybe I want to buy it. And in that experience, I think Amazon can make a ton of money, and there will still be value for places that are aggregates of supply with advertising and marketing. But I think there's a lot of times where you're optimizing for friction where you don't want that stuff. So if I were Amazon and I had the $80 billion or whatever, $70 billion a year advertising business, the question that I would ask is like, hey, I don't think you want to block agents. I I think they may be a bigger than Amazon company, like a headless Amazon could be a huge, huge business, right? But I think it puts more pressure on them. to create a separate experience where you do want the ads, and that's really good, and making that worthwhile to go on on its own. And I think I would still go to Amazon for those things. Right. At the end of the day. AI is moving so fast that it's going to change distribution and human habits at a rate where I don't think anyone can predict what the winning is. platforms will be in those spaces. And there's going to be huge losers, right? And Amazon might lose a huge chunk of their advertising business trying to fight the future, because they can't predict what it would look like without. The other one essentially is Muse for me. I don't know if you've seen the discussion around Muse. It's like free. Okay, cool. It's free personal agent. Cool. People are like, well, one day they're going to have to put ads in Muse. But actually, maybe they never have to put ads in Muse. Maybe Muse just gives them data that allows to target you better on Instagram. And when you go to Instagram, you're on Instagram because you don't want efficiency if you're on Instagram. You want to be entertained. And if they can target better on Instagram because of the information from Muse, And Muse is still free. And on Instagram, when you see the good that you like, they can also get you to pay for it. using a wallet that they've accumulated on you from use, Or Muse can look online for different variants of the thing. It's like a really nice... And what it means is that you've divorced the monetization from the agent experience for this personal use case. And that's where a property like Amazon is at a disadvantage, I think. It might be with agents. Having just been the aggregator of supply, just may not be a good business model in the world of agents. It just may no longer be. And so there's really not much they can do about it. They're going to lose. And probably the counter move for them, honestly, is they should have a personal agent. They should have a Muse. That's actually the counter move for them. It's not to block Muse. It's to have a Muse. Have a better Muse. A better Muse and have the exclusive on shopping for that and hope that's enough of an incentive and put it part of Prime. But I don't know if they're able to do that. Well, the fun thing is with Amazon, there's a website called Camel, Camel, Camel. It's a price engine for Amazon. So you basically say, "I want to know when Dr. Pepper is... right now it's $8. Tell me when it's $6." And then buy it, basically. The second it's $6, no, it doesn't do the buying. It just emails you and then you have to do the buying. 20 million people use this website. 20 million people? Yeah, 20 million people in America. And this is just a... And it's an affiliate website, and this has been around for 18 years. [60:00] So, Jesse and I wrote a post about this, where the best way to predict the future is to observe the present. Again, going back to like it's not evenly distributed. And like this thing is the future because like the other way of thinking about agents is there are people that have more time than money. Yeah. And there are people that have more money than time. Yep. And people that have more money than time, they're not going to use camel, camel, camel. People that have more time than money, They are. But wait a minute, what if you now can give everybody who has more money than time, Give them all of the tools and the agent does it. Like, of course they're going to use that. It's an IQ test. So these things are definitely going to be a lot more pervasive. And again, sometimes you don't even have to like picture what might happen. You just have to look at like, you know, coupons. Coupons, like you shouldn't use a coupon if you're wealthy because it takes too much time. You're going to spend like two hours like hunting for the right coupon. But guess what? And that's another covenant that's being broken because it was a very neat way of dividing the world. Yeah. I mean, MX premium membership, it's the same way. all of your, all of your, the breakage model is dead. Yeah. So we're all going to abuse all of our chase and MX benefits and then I'm going to have to change it. It's not cost effective because then it's going to be adversarial. This is kind of the adversarial versus synergistic thing. So by the way, the, Bye. The through line of all those is, I just think supply aggregation is a lot less valuable than it used to be very valuable because human attention was limited. And now supply aggregation is a lot less valuable because human attention can be automated away. This has been an awesome conversation. Thank you again for coming in today and go town. Yeah, thanks so much, Alex. It's always great to chat. Thank you. you Thanks for listening to this episode of the A16Z podcast. If you liked this episode, be sure to like, comment, subscribe, leave us a rating or review. and share it with your friends and family. For more episodes, go to YouTube, Apple Podcasts, and Spotify. Follow us on X at A16Z and subscribe to our Substack at a16z.substack.com. Thanks again for listening, and I'll see you in the next episode. As a reminder, the content here is for informational purposes only. It 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. Please note that A16Z and its affiliates may also maintain investments in the companies discussed in this podcast. For more details, including a link to our investments, please see a16z.com forward slash disclosures. you