在哈里·斯泰宾斯（Harry Stebbings）主持的20VC节目中，Town.com的联合创始人兼首席执行官、Plaid前首席技术官简-丹尼斯（JD）深入探讨了AI助手的演变格局。JD将Town.com定位为AI助手市场的前三位优先事项之一，称其作为一个AI助手存在于用户的电子邮件和日历中，推荐并自动化用户通常会自己完成的任务。该公司最近从AI报税业务转型，此前他们意识到在日常通信平台中运行的代理模型具有巨大潜力。 JD公开谈到了激烈的竞争，承认谷歌（Google）和苹果（Apple）等巨头也将这一领域视为重中之重。他认为“护城河”对于比Town.com发展更成熟的公司来说是一种奢侈品，强调当前迫切需要实现深度的产品与市场契合度，目标是拥有10万到100万付费用户。他相信，制胜产品将具备代理层面的网络效应，并重点介绍了Town的“代理到代理”功能，即助手之间相互通信以检索信息，他认为这是一个真正的差异化因素。JD概述了该领域护城河的几种理论：定制模型、个性化上下文数据和分销渠道，并指出Meta/WhatsApp因其现有的用户基础而构成重大威胁。 JD预测，人与代理的互动未来将涉及信任代理来决定与他人共享哪些数据，即使没有明确的规则。他认为，AI模型在尊重隐私边界方面将比人类更熟练，并举例说明了目前看似“疯狂”但将变得司空见惯的公司内部数据共享以实现业务成果的情况。他还提出，虽然人类监督对于设定代理的目标和预算仍然至关重要，但代理本身将处理大部分执行甚至监控工作，这可能比人类驱动的流程带来更少的错误。 在技术方面，Town.com采用模型路由，根据成本效益和任务需求利用不同的模型（例如Gemini、OpenAI、11 Labs）。JD指出，虽然用户体验（如语音个性）的一致性限制了直接用户互动的“任意路由”，但纯推理任务允许更大的灵活性。他坦言，由于定制工作流程的复杂性，Town.com严重依赖前沿模型，并将其视为增长的必要补贴，预计许多任务的成本会随着时间的推移而降低。他强调，他们的首要任务是最大限度地提高产品价值和增长，而不是立即进行成本优化。 JD分享了Town“惊艳时刻”（time to wow）的见解，将他们高达15%的支付转化率归因于用户通过连接电子邮件和日历所体验到的即时价值，这使得个性化自动化建议成为可能。他还透露，尽管其核心重点是企业，但他们意外地发现了与家庭，特别是管理大量学校和儿童相关通信的父母，具有产品与市场契合度。 谈及苹果的潜在进入，JD表达了怀疑，原因在于苹果缺乏云原生基因，以及他们对设备端处理和严格隐私的承诺，他认为这落后于AI能力“前沿”。他还谈到了AI驱动世界中数据泄露和网络威胁的更广泛担忧，将其比作化工行业早期学习管理风险的阶段。 在定义成功时，JD将用户持续的月度支付放在首位，表明正在提供足够的价值。他将企业（AI直接推动业务价值）的长期营收增长潜力与消费者使用场景的有限范围进行了对比，倾向于拥有大量支付较少用户的模式，而非少数支付较多用户的模式。他表达了对对话式AI进步的兴奋，以及AI在显著减少日常劳务、促进各行业创新，并最终导向一个更繁荣和充实世界的潜力。

In an insightful discussion on 20VC with Harry Stebbings, Jean-Denis (JD), co-founder and CEO of Town.com and former CTO at Plaid, delved into the evolving landscape of AI assistants. JD positioned Town.com as a top-three priority in the AI assistant market, stating it operates as an AI assistant living within users' email and calendar, recommending and automating tasks they would typically do themselves. The company recently pivoted from an AI tax preparation business after realizing the potential of agentic models operating within daily communication platforms. JD openly addressed the fierce competition, acknowledging that giants like Google and Apple view this space as a top priority. He views "moats" as a luxury for companies further along than Town.com, emphasizing the immediate need for deep product-market fit with a target of 100,000 to a million paying users. He believes the winning product will feature a network effect at the agent level, highlighting Town's "agent-to-agent" feature where assistants communicate to retrieve information, a concept he sees as a true differentiator. JD outlined several theories for moats in this space: custom models, personalized context data, and distribution channels, noting Meta/WhatsApp as a significant threat due to their existing user base. The future of human-agent interaction, JD predicts, will involve trusting agents to decide what data to share with others, even without explicit rules. He argues that AI models will become more adept at respecting privacy boundaries than humans, citing examples of intra-company data sharing for business outcomes that currently seem "insane" but will become commonplace. He also suggested that while human oversight will remain crucial for setting goals and budgets for agents, the agents themselves will handle the bulk of execution and even monitoring, potentially leading to fewer errors than human-driven processes. On the technical front, Town.com employs model routing, leveraging different models (e.g., Gemini, OpenAI, 11 Labs) based on cost-effectiveness and task requirements. JD noted that while consistency in user experience (like voice personality) limits "wild routing" for direct user interaction, pure reasoning tasks allow for more flexibility. He confessed that Town.com heavily relies on frontier models due to the complexity of custom workflows, accepting this as a necessary subsidy for growth, with the expectation that costs will decrease over time for many tasks. He stressed that their priority is maximizing product value and growth rather than immediate cost optimization. JD shared insights on Town's "time to wow," crediting their high 15% payment conversion rate to the immediate value users experience by connecting their email and calendar, which enables personalized automation suggestions. He also revealed an unexpected product-market fit with families, particularly parents managing numerous school and children-related communications, despite their core focus being enterprise. Addressing Apple's potential entry, JD expressed skepticism due to Apple's lack of cloud-native DNA and their commitment to on-device processing and strict privacy, which he believes lags behind the "frontier" of AI capabilities. He also touched on the broader concern of data leakage and cyber threats in an AI-driven world, likening it to the early days of chemical industries learning to manage risks. When defining success, JD prioritized consistent monthly payments from users, indicating that sufficient value is being delivered. He contrasted the potential for long-term revenue growth in enterprise (where AI directly drives business value) with the more limited scope of consumer use cases, favoring a large base of users paying less over fewer paying more. He expressed excitement for the advancements in conversational AI and the potential for AI to dramatically reduce daily toil, foster innovation in various sectors, and ultimately lead to a more prosperous and fulfilling world.