Fin Hits $100 Million ARR, Sells to Salesforce After AI Agent Bet

押注AI Agent，Fin实现1亿美元ARR并出售给Salesforce 随着生成式AI将软件执行成本降至零并抹平功能护城河，产品的防御壁垒仅留存于底层专有数据与高维战略品味之中。 Thom Rimmer详细阐述了Fin（前身为Intercom）如何成功摆脱传统按席位付费的SaaS模式，转型为Agentic AI领域的核心力量，不仅年经常性收入（ARR）达到1亿美元，还达成了被Salesforce收购的协议条款。他指出，自然语言界面正在消解传统软件类别，通过直接的对话式目标达成就此消除浪费的“工具时间”。随着产品设计从构思视觉界面转向编排自主智能体的思考，内部工作流也同步映射了这一演化——Fin超过90%的Pull Request均由AI编写。因此，功能护城河已荡然无存，真正的竞争壁垒仅存在于两端：底层专有数据与顶尖的战略设计品味。 Intercom在意识到AI将淘汰其按席位计费的商业模式后彻底转型，将Fin打造成AI Agent，最终达到1亿美元ARR并与Salesforce达成收购协议。 (0:04) 软件价值正从传统的UI层流失，并同时向下转移至底层原始模型能力、向上转移至交付与打包环节。 (0:04) AI作为一种聚合力量正在瓦解传统软件的品类边界，迫使产品横向扩张并吞并相邻领域。 (2:58) 对话式界面并非激进的科幻构想，而是在经历了数十年GUI的弯路后，重返人类以对话为驱动的自然交互模式。 (2:58) 用户通过索取最终结果来操控现代软件，而非手动点击菜单，这实质上消解了熟练使用工具的职业价值。 (5:23)

Thom Rimmer, head of product design at Fin (formerly Intercom), argues that generative AI is dissolving conventional graphical interfaces and feature moats, shifting software's economic value away from intermediate UI controls and toward underlying model capabilities, strategic product taste, and autonomous agents. Dissolving Interface Boundaries and Eliminating Tool Time (2:58) Rimmer contends that graphical user interfaces were a multi-decade historical detour, asserting that natural language marks a return to direct conversational interaction where users state desired outcomes rather than navigating mechanical controls. As conversational AI advances, legacy software divisions separating search, chat, products, and services are dissolving. Citing design pioneer Jared Spool, Rimmer distinguishes between "goal time" (achieving the core objective) and "tool time" (wasted overhead navigating menus and settings). While traditional design optimized tool time through better button placement, conversational systems eliminate it entirely—such as requesting highlighted spreadsheet numbers without manually configuring conditional formatting rules. Fin enacted this shift by disrupting Intercom's existing per-seat SaaS business—which served roughly 35,000 customers and handled 500 million monthly messages—to deploy an AI agent performing customer service directly. The company subsequently crossed $100 million in ARR and agreed to definitive acquisition terms with Salesforce. To streamline maintenance, Fin introduced Operator, an internal agent that configures, tests, and manages the customer-facing Fin agent through dialogue. However, Rimmer qualifies that graphical controls remain superior for certain tactile actions, such as photo exposure sliders, indicating that future interfaces will blend conversation with dynamic visual components. Autonomous Engineering Velocity and Compounding Error Risks (13:57) Internally, Fin restructured its development workflows around autonomous pipelines: approximately 94% of pull requests (PRs) are authored by Claude Code, and nearly 20% are initiated, written, tested, and deployed end-to-end without human intervention. This acceleration enabled Fin to triple internal productivity in nine months—surpassing its initial target of doubling output in twelve—and ship nine major product launches across a two-month span while compressing planning cycles from two-year roadmaps to three-to-four-month intervals. Despite these throughput gains, Rimmer warns of a critical reliability hazard in multi-step agentic execution. In sequential workflows, individual steps with 95% to 96% accuracy compound down to an end-to-end reliability of roughly 80% to 88% across a five-step chain, which remains unacceptable for production enterprise software. Consequently, software organizations must abandon assembly-line mindsets and operate as experimental testing labs, identifying where human-in-the-loop oversight is mandatory and validating model performance with rigorous evaluations. Vanishing Feature Moats and Defensibility at the Edges (15:28) Because foundation models push the marginal cost and time of software development toward zero, traditional feature-level moats have evaporated. Rimmer points out that a Fin engineer cloned a competitor's product using Claude in days, while competitors replicate Fin's capabilities at comparable speed. When execution constraints disappear, rapid building without clear guardrails merely amplifies organizational noise and accelerates strategic misalignment. As a result, defensibility has shifted entirely to the outer edges of the system: proprietary data and domain-specific models at the bottom, and high-level product taste and strategic direction at the top. To safeguard product quality against erosion from rapid autonomous deployment, Fin codified human design taste and qualitative judgment into automated evaluation "skills" embedded directly within CI/CD pipelines. Designers are thus repositioned away from crafting static screen mockups in Figma toward directing the system's strategic trajectory and enforcing standards. Rimmer observes that human psychology, organizational risk aversion, and cultural inertia—rather than access to AI technology—are the primary bottlenecks slowing adoption, while cautioning that engineering a flashy two-hour prototype into a dependable enterprise product can still require up to 24 months.

Fin（原Intercom）产品设计负责人Thom Rimmer指出，生成式AI正在瓦解以图形界面为核心的传统SaaS模式，软件交互正重返自然语言对话，中间层功能与控件的护城河趋于归零；软件的核心竞争力正向底层的专有数据与顶层的战略品味两端迁移，产研组织必须转型为容忍试错的实验型体系。 交互范式演进：界面消融与语言回归 (2:58) Rimmer回顾了计算历史的三大控制范式：门槛高但功能强大的命令行、降低门槛但施加人为限制的图形用户界面（GUI），以及直接表达意图的自然语言交互。他援引一块记录着劣质铜交易争端、距今四千年前的美索不达米亚泥板指出，商业与沟通自古依托于人际对话，过去数十年充满菜单与按键的GUI只是一段弯路，AI正在促使交互“重返故乡”。在这一趋势下，原本被清晰划界的应用（如搜索与对话、产品与服务）正在相互渗透，垂直工具的传统边界正被横向瓦解。 这一演进深刻改变了人机交互的时间构成。借用设计学者Jared Spool对“工具时间”（穿梭于菜单和界面的非必要开销）与“目标时间”（真正完成任务的时间）的划分，传统软件只能设法优化工具时间，而新一代系统（如Fin推出的自主内部智能体Operator）则旨在彻底删除工具时间。用户不再需要熟记电子表格中“条件格式”的操作路径，只需陈述使负数标红的结果；以往靠熟悉繁复菜单排布建立的工具专长正在贬值。不过Rimmer也强调，GUI不会彻底消亡，在诸如调整曝光滑块等具身感知任务中，视觉滑块依然优于多次文字描述，未来的核心形态是对话系统动态调配合适的交互组件。 功能护城河归零与价值向两极迁移 (0:04) 在转型前，Intercom是一家拥有约3.5万家客户、每月处理5亿条消息的SaaS先驱。面对ChatGPT的冲击，团队认定AI将全面接管人工客服，原有按坐席计费的商业模式终将被淘汰。公司随即孤注一掷推翻既有架构，全力打造直接承担客服工作的AI智能体Fin。经过约三年半至四年的转型，Fin的年度经常性收入（ARR）突破1亿美元，并与Salesforce达成了最终收购协议。Rimmer据此强调，在AI变革期进行仅供社交媒体宣传的表面化功能修补注定失败，必须敢于承担结构性重构的风险。 随着软件边际构建成本逼近于零，建立在界面或离散功能上的壁垒已经瓦解。Fin工程师曾利用Claude在数天内完全克隆了竞品的核心产品，而竞品同样能迅速复制Fin的功能更新。Rimmer指出，中间层的按钮、控件与UI布局正在迅速流失商业价值，行业价值正发生双向迁移：向下沉降至基础大模型的能力封装、外挂框架（harness）与专有数据资产；向上攀升至对用户实际业务结果的打包交付与战略品味。产品团队必须转向掌控端到端的客户任务（jobs-to-be-done），而非守着单点功能。 设计逻辑重构：从屏幕布局到思维工程 (11:45) 过去半个世纪中，软件设计长期围绕CRUD（增查改删）界面展开，默认必须由人类完成全部思考并逐步完成多步操作流程。设计师耗费精力搭建新用户引导、渐进式信息呈现、空白状态与工具提示，试图诱导用户学习软件路径。Rimmer提出，自主智能体消解了绝大部分导航界面的存在必要，产品设计的核心命题已经从“界面看起来如何、如何操作”彻底转变为“智能体是如何思考的”，关注重点转移到了底层模型表现、决策参数与可解释的逻辑链条。 当AI接管了代码生成与界面布局等执行环节，产研瓶颈从“如何构建”转移到“选择构建什么”。若缺乏清晰的战略方向，高速生成的大量方案只会放大组织混乱。为避免自主生成流程导致产品体验退化，Fin的设计团队不再局限于在Figma中绘制静态界面原型，而是将人类的设计品味、定性判断与决策逻辑提炼为自动化的评估“技能”，直接嵌入CI/CD部署流水线中，在系统上线前自动拦截质量缺陷，实现对产品品质的规模化把控。 实验型研发：智能体协同与复合误差 (13:57) 智能体技术深刻改造了Fin内部的软件研发流程。目前Fin约94%的代码合并请求（PR）由Claude Code编写，约20%的PR实现了端到端完全自主——从需求提出、代码编写、QA测试到上线部署均无需人类介入。这套机制使Fin在9个月内实现了生产力提升三倍，远超原定12个月翻倍的目标；规划周期也从两年缩短至三到四个月，从而在两个月内高频交付了9次重大产品发布。 然而，速度优势背后存在严峻的复合误差陷阱。Rimmer指出，若一个五步骤的智能体调用链条中每步准确率为96%，端到端综合可靠性将衰减至约81.5%，这一失败率在企业生产环境中是无法接受的。尽管两小时的Prompt工程可生成出色的原型Demo，但将其打造成工业级系统仍需长达24个月的严格工程化测试。软件团队必须从可预测的线性装配“工厂”思维，转向持续进行压力测试与行为评估的实验“实验室”，精准设定人类介入监督的节点。Rimmer认为，高迭代速度更有利于尽早打磨出高品质产品，当前阻碍组织拥抱AI的瓶颈不是技术获取门槛，而是由风险规避和沉没成本引发的心理惯性。 Rimmer总结指出，软件行业当前仍处于AI创新S曲线的底部拐点，过去几十年在图形界面上的深耕只是技术演进中的一段插曲。随着软件复制成本接近于零，传统的功能护城河已不复存在；但在智能体开发范式下，如何在多步链路的复合误差与海量自主代码生成中，兼顾极致的迭代速度与稳定的企业级可靠性，依然是整个行业需要持续探索的核心课题。

## Background 本内容由 YouTube 频道 Ai4 于 2026 年 8 月 16 日发布，主讲人为 Fin 产品设计副总裁 Thom Rimmer。演讲主要探讨了生成式 AI 如何将复杂的软件交互坍缩为一个简单的对话框，并详细阐述了 Fin 团队在这一浪潮下如何重塑产品研发流程与定义软件的未来形态。 ## TL;DR 面对生成式 AI 对传统客服软件的颠覆，Intercom 全面转型为 AI Agent 产品 Fin，不仅实现了超 1 亿美元 ARR 并达成被 Salesforce 收购的协议，还彻底颠覆了人机交互与研发范式。演讲指出，传统 GUI 正在消退，软件价值正在向底层的模型能力与顶层的战略品味两端位移，而中间层的功能与界面极易被快速复刻。在内部，团队借助 Claude Code 实现了 94% 的 PR 由 AI 编写，生产力大幅飙升 3 倍，使得产品团队的职责从绘制界面转变为把控战略航向与死守质量门槛。 ## Atmosphere & style 整场演讲充满前沿洞察与强烈的危机紧迫感，主讲人以清晰从容且极具信服力的语调，带领听众回顾了一场孤注一掷的公司生死突围。叙事风格兼具宏大的历史文明视角与扎实硬核的工程实战反思，既有直面传统界面工艺瓦解的冷峻客观，又洋溢着拥抱范式转移的极客兴奋感，令人深切感受到软件黄金时代再次开启的激荡与澎湃。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=JbF_SbKJSEc&t=4s) 面对 ChatGPT 带来的商业模式颠覆危机，Intercom 孤注一掷全面转向 AI Agent（Fin），不仅 ARR 突破 1 亿美元并达成被 Salesforce 收购的协议，也证明在 AI 时代必须颠覆性重构而非表面修补。 - [02:10](https://www.youtube.com/watch?v=JbF_SbKJSEc&t=130s) 软件价值正发生双向位移：传统表层 UI 控件的价值被迅速掏空，价值向下沉降至底层推理模型与系统 Harness，向上攀升至端到端业务交付与高阶解决方案。 - [04:09](https://www.youtube.com/watch?v=JbF_SbKJSEc&t=249s) 自然语言对话界面并非凭空捏造的未来主义，而是人类在经历了数十年以屏幕为介质的探索后，重新回归最原始、最自然的直接沟通本能，软件操控全面由“点击”转向“发问”。 - [07:00](https://www.youtube.com/watch?v=JbF_SbKJSEc&t=420s) 从初代的辅助聊天窗口演进到自主型 Agent（如 Operator），软件正在实现“由智能体去制造与维护其他智能体”，从而摆脱繁琐的手动配置流程。 - [09:18](https://www.youtube.com/watch?v=JbF_SbKJSEc&t=558s) 自主型 Agent 的核心价值在于彻底消除了用户摸索界面和层级菜单的“工具时间（Tool time）”，让精力 100% 聚焦于交付实际业务价值的“目标时间（Goal time）”。 - [10:57](https://www.youtube.com/watch?v=JbF_SbKJSEc&t=657s) 图形用户界面（GUI）正在走向消退，传统以 CRUD 和界面寻路为核心的设计范式被打破，软件设计的第一命题从“它看起来如何、怎么操作”转变为“它如何进行思考”。 - [14:04](https://www.youtube.com/watch?v=JbF_SbKJSEc&t=844s) 虽然 94% 的代码 PR 已由 Claude 编写且边际成本趋零，但串联链路中的“复合错误率”会导致系统整体可靠性急剧衰减，必须建立严苛的 Eval 评估体系与人工监督边界。 - [15:47](https://www.youtube.com/watch?v=JbF_SbKJSEc&t=947s) “没有方向的速度只是纯粹的噪音”，团队在 9 个月内将生产力提升 3 倍，并通过将设计品味和审查标准固化至自动化管线，实现了近 20% 代码 PR 的全自主闭环部署。 - [18:25](https://www.youtube.com/watch?v=JbF_SbKJSEc&t=1105s) 产品交付周期已由“月和季度”压缩至“天和周”，快慢团队之间的鸿沟日益扩大，而阻碍提速的真正瓶颈在于组织的心理惯性与路径依赖，而非技术本身。 - [19:29](https://www.youtube.com/watch?v=JbF_SbKJSEc&t=1169s) 当单一功能与代码能在数天内被 AI 迅速克隆时，传统功能护城河已不复存在；真正的壁垒只剩两端：底层的专有数据资产/模型，以及顶层的产品品味与战略方向判断。 - [20:22](https://www.youtube.com/watch?v=JbF_SbKJSEc&t=1222s) 行业目前仍处于 S 曲线拐点的最底部，过去几十年精心打磨的繁复屏幕控件本质上只是发展过程中的一段“弯路”，AI 正在让软件重新回归解决问题的本质。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=JbF_SbKJSEc&t=4s) 界面工艺的消逝与颠覆性的公司豪赌 在过去整整二十年里，产品设计师的工作核心始终聚焦于方寸屏幕之间——精雕细琢每一个按钮、推敲每一处排版布局，以及打磨绝大多数用户甚至难以察觉的微小细节。然而，这项支撑了无数设计师职业生涯的传统手艺，如今正处于迅速瓦解的边缘。在当今任何软件产品中，用户真正触摸、感知与交互的界面层，正在发生根本性的蜕变。这不仅是产品形态的重塑，甚至被视为过去两三十年间软件行业所迎来的最激动人心的变革。 这场行业巨变的前沿阵地正是由 Intercom 演变而来的 Fin。在过去十余年间，Intercom 是一家体量庞大且极为成功的 SaaS 巨头，坐拥约 35,000 家客户，每月处理高达 5 亿条消息，是连接在线企业与终端用户的行业先驱。然而，ChatGPT 的横空出世引发了一场无法回避的生死拷问：生成式 AI 是否会彻底颠覆并取代人工客户服务？如果答案是肯定的，那么原先按人工客服席位收费（Per-seat）的商业模式将瞬间归零。面对可能遭遇的覆灭危机，公司极早便做出了一个孤注一掷的决断：赌上整家公司的命运，全面转向构建能够完全替代传统人工客服工作的 AI Agent。 经过三年半到近四年的全力突围，这场豪赌换来了惊人的回报：Fin 的年度经常性收入（ARR）近期跨越了 1 亿美元大关，业务呈现惊人增长，并在近期宣布了将被 Salesforce 正式收购的最终协议。但成功的代价是打碎重来——从战略方向、产品架构、定价策略、研发流程，到整个团队的自我认知，乃至门牌上的公司名字，一切都被推倒重写。这种生死转型的沉痛教训在于：面对 AI 时代，要么全力挥棒（Swing really big），要么被淘汰出局；那些为了迎合 AI 趋势而在 LinkedIn 或博客上发布不痛不痒的微小改良、走安全舒适路线的产品与团队，注定无法在全新纪元中生存。 - **原业务规模**：拥有约 35,000 家企业客户，每月消息处理量达 5 亿条。 - **转型成果**：AI Agent 产品 Fin 突破 1 亿美元 ARR，并达成被 Salesforce 收购的确定性条款。 - **战略教训**：摒弃表面化的修补适配，面对技术范式转移必须进行颠覆性自我重塑。 --- ### [02:10](https://www.youtube.com/watch?v=JbF_SbKJSEc&t=130s) 价值的双向位移与软件边界的融化 AI 正在深刻改变软件的终极形态以及软件的构建方式。而在这两重变革的底层，发生着一场关于“软件价值究竟沉淀在何处”的巨大位移。价值正在以一种肉眼可见的速度撤离表面层——那些曾经耗费团队数年心血打磨的开关、按钮和交互界面，其核心价值正大幅缩水。一方面，价值向下沉降（Dropping down），流向底层大模型、系统 Harness 以及原生的底层推理能力；但与此形成悖论的是，价值同时也在向上攀升（Moving upwards），流向如何将这种强大能力进行产品化包装并交付给最终用户的解决方案层。 伴随着价值的位移，传统软件那道维持了数十年的清晰物理边界正在被彻底瓦解。过去，软件开发者可以画出一个边界清晰的“盒子”，界定好软件内部的功能范围与外部边界，每个用户也心知肚明这个工具是用来做什么的。然而，AI 作为一种强大的聚合力量，正将所有刻意划定的品类与护城河统统融化。搜索与聊天的界线、产品与服务的鸿沟正消失殆尽，各类工具都在沿着外延横向扩张并蚕食邻近领域（例如 Notion 的横向功能扩展，或是 Gong 与 Guru 之间相互渗透的品类争夺）。这给当下所有产品团队带来了一个极为棘手的命题：“你的产品现在究竟该在哪里停下脚步？” ```mermaid flowchart TD accTitle: 软件价值位移与传统交互层消解 subgraph ValueUp ["向上位移：交付与包装"] A["高阶价值交付"] --> B["端到端业务结果与方案打包"] end subgraph UI ["表面层价值流失"] C["传统 UI / 控件 / 按钮"] -.->|"价值迅速被掏空"| D["对话化与无形交互"] end subgraph ValueDown ["向下位移：底层能力"] E["推理模型与架构 Harness"] --> F["原生认知与执行算力"] end ``` --- ### [04:09](https://www.youtube.com/watch?v=JbF_SbKJSEc&t=249s) 4000年前的泥板：交互界面的本能回归 当下软件从布满丰富按钮、层级深厚且复杂的图形界面，快速坍缩为一个看似极简的对话框。表面上看，通过自然语言直接向系统发出指令是一种充满科幻感的未来交互，但其实质却是一场重返人类文明基石的“复古倒退”。在距今约 4,000 年前的美索不达米亚平原上，诞生了人类历史上已知最早的一块刻在黏土泥板上的客户投诉信——一位买家愤怒地向商人控诉其运送的铜矿石品质恶劣。这个经典的商业案例揭示了一个本质规律：商业交易自诞生之初，便与“对话沟通”紧紧绑定在一起。在人类历史的大部分时间里，商业行为的促成均是通过人与人之间的直接交谈实现的。因此，转向对话交互绝非凭空臆造的未来主义，而是人类在经历了几十年以屏幕为介质的“歧路”后，重新回到了最原始、最自然的交互本能。 回顾过去数十年的人机交互史，人类掌控计算机主要经历了两条路线：第一种是强大但学习门槛极高的命令行界面（CLI）；第二种则是易于上手但限制极多的图形用户界面（GUI）。在 GUI 时代，产品设计师如同“守门人”一般，通过绘制固定的按钮、滑块和控件，严格限制了用户在软件中能做什么、不能做什么。而今第三种交互方式正式降临：人们只需使用大白话直接表达诉求即可。 以电子表格操作为例：如果用户希望将一列数字中的所有负数标红，在传统 GUI 逻辑下，用户必须先知道软件中存在一个叫做“条件格式”（Conditional Formatting）的专门术语，接着在一级又一级的嵌套菜单中翻找并完成配置。后来出现的浏览器自动化 Agent 虽然能模拟人手点击，但本质仍是在走老路。而在全新的交互语境中，用户根本不需要关心功能叫什么名字，也不需要学习复杂的菜单层级，只要说一句“把负数变成红色”，期望的结果便神奇般地呈现在眼前。“功能本身从来不是目的，最终的结果才是目的。”人们操控软件的动作，已经由机械的“点击”全面转向自然的“发问”。 - **控制计算机的三种范式**： 1. 命令行（Command Line）：能力极为强大，但学习难度极高。 2. 图形界面（GUI）：上手容易，但交互受到设计者设定的控件极大限制。 3. 自然语言界面（Plain Language）：用日常语言直述意图，由系统直接达成结果。 - **技能维度的瓦解**：以往简历上引以为傲的“Figma 大师”或“Sketch 专家”等对特定工具操作极其娴熟的技能，在“发问即可获得结果”的时代正失去原本的职业壁垒。 --- ### [07:00](https://www.youtube.com/watch?v=JbF_SbKJSEc&t=420s) 从极简界面到能够自我演进的 Agent 这一轮人机交互变革如同当年的 iPod 替代 MP3 播放器一样，看似在一夜之间爆发，实则是“漫长酝酿后的瞬间倾泻”（Arrived gradually, then arrived all at once）。软件交互在短时间内经历了从极限界面（Maximal UI）向极简界面（Minimal UI）的剧烈收缩：最初，AI 只是作为一个外挂聊天窗口嵌入在应用边缘；随后，它变成了帮助用户配置软件的向导；紧接着，AI 开始全面吞噬应用本身，剥离绝大多数界面的存在必要；直到最后，思考与实际任务由 AI 全权接管，界面本身已名存实亡，甚至引发了“如果工作已经全自动完成，软件应用本身究竟身在何处”的哲学追问。 这种演进逻辑直接体现在 Fin 的产品迭代进程中。三年前，团队推出了突破性的 Fin，彻底打破了过去基于呆板脚本的客服机器人，带来了人类水平的生成式 Agent。但在当时，它依然被包裹在一个挂在屏幕右下角的传统聊天弹窗内。到了今年四月，团队推出专为销售场景打造的 Fin for sales，对体验进行了大刀阔斧的重塑：将沟通窗口移至屏幕正中底部，设计更具存在感、响应更流畅优雅，但归根结底，它仍是一个挂载在网站或产品上方的对话层。 真正的分水岭是自主型 Agent——Operator 的诞生。如果说 Fin 是服务于外部客户的 Agent，那么 Operator 就是赋能团队内部成员的 Agent：它是一个“能够制造 Agent 的 Agent”（The agent that builds other agents）。与以往需要人类一步一步指导流程的工具不同，Operator 具有高度的目标导向性：它深知团队的最终业务目标，能够自主判断下一步的最优解、进行自我测试，并仅在需要关键裁决时带着成果向人类寻求审批。在 Operator 出现之前，客户为了调优 Fin，通常需要手动维护大量电子表格和知识文档，来回反复核验；而如今，这种繁琐沉重的工具操作与配置逻辑，正在被自主执行的 AI 彻底替代。 - **三代产品演进路径**： - **Fin 初代**：引入生成式大模型取代死板脚本，实现人类级别的对话能力，但外壳仍为传统客服聊天气泡。 - **Fin for sales**：交互界面升级，居中置底且视觉更加精致灵动，仍属于嵌入式对话窗口。 - **Operator**：自主性质变，负责知识库管理与 Fin 的维护构建，具备自主推理与测试能力，彻底摆脱了人类手动按步骤操作软件的繁琐流程。 ### [09:18](https://www.youtube.com/watch?v=JbF_SbKJSEc&t=558s) 告别繁琐设置：Operator 的登场与消除“工具时间” 在过去，即使对于最终用户而言 Fin 的表现十分惊艳，但对于后台团队成员来说，维护文档、进行配置依然是一个相当繁重且困难的手动过程。想要达到高自动化率往往需要付出巨大心力。而 Operator 的出现彻底颠覆了这种工作流——所有的手动配置工作全部消失，取而代之的是纯粹的自然语言对话。用户只需与它交谈，Operator 就会自动完成所有任务。市场对 Operator 的反响极其热烈，客户需求堪称“狂热（wild）”，完全超出了以往的任何预期。本周 Operator 正式进入 GA（全面可用）阶段，团队在现场展位 531 提供了专家咨询，并在当天下午 4:30 安排了实况 Demo 演示。 ```mermaid flowchart LR accTitle: Jared Spool 软件时间分类模型 subgraph S["用户在软件中的总时长"] direction TB TT["工具时间 (Tool Time)<br/>• 寻找按钮与层级菜单<br/>• 学习操作逻辑与路径<br/>• 价值产出：0 (纯消耗)"] GT["目标时间 (Goal Time)<br/>• 达成核心工作目标<br/>• 交付实际业务价值<br/>• 真正的核心体验"] end TT -- "Operator 彻底消除" --> GT ``` Operator 能够带来质的飞跃，其核心在于它消除了设计界传奇人物 Jared Spool 所提出的“工具时间（Tool time）”，让用户完全聚焦于“目标时间（Goal time）”。Jared Spool 将人在软件中花费的时间划分为两类： - **工具时间（Tool time）**：指用户在操作界面本身上耗费的时间，例如寻找按钮、摸索菜单层级。这一过程不会产生任何实际价值，过去衡量一个软件或团队是否优秀，核心指标就是工具时间与目标时间的比例。 - **目标时间（Goal time）**：指用户真正用来完成其最初预定目标的有效时间。 优秀的软件以往只能尽可能缩短工具时间，而 Operator 则将其彻底抹去。用户不再需要去“操控（driving）”应用程序，而是通过对话交互，界面仅在真正相关时才会按需呈现。 --- ### [10:57](https://www.youtube.com/watch?v=JbF_SbKJSEc&t=657s) 图形界面的消退与“思考模式”的设计重构 作为设计负责人，Thom Rimmer 明确表示，图形用户界面（GUI）虽然还算不上“彻底死亡”，但它确实已经“一只脚踏进了坟墓”。之所以另一只脚依然充满活力，一方面是因为用户的认知习惯与行为模式需要漫长的时间来转变；另一方面，某些特定场景下传统界面确实体验更好。例如在 Photoshop 中修图时，用户绝对不希望连打 5 遍提示词“请把这里调暗一点”，直接拖拽滑块（slider）要高效直观得多。最强大的对话式界面，正是当它与必要的图表、可视化组件、表格和滑块等 UI 元素完美配合时诞生的。 回溯过去三四十年的软件开发史，整个行业都在围绕以 CRUD（增删改查）为主导的界面苦苦奋战。这些软件建立在一种固有假设之上：**必须由人类去思考、记住操作流程，最后在屏幕上获取确认反馈**。为了实现让用户“无需指引即可直观理解”、“顺利发现新功能”以及“无视那些没人会读的更新邮件就能学会使用”，整个设计行业构建了一整套庞大的学科体系，包括引导流程（onboarding flows）、空状态设计（empty states）和渐进式呈现（progressive disclosure）。然而在智能体时代，软件的核心问题已经从“它看起来如何、怎么操作”彻底转变为“**它如何进行思考（How does this think?）**”。 - **传统界面反思**：审视现有产品，大量 UI 仅仅是为了辅助用户在系统内“寻路（wayfinding）”——包括繁复的导航栏、仪表盘、报表和设置页面。 - **价值层下沉**：当 Agent 能够直接执行任务并在上下文中按需汇报时，导航和设置页面便失去了存在的意义，软件的表层被极限压缩，真正的核心价值全部沉降到底层的模型中。 - **产品团队的新课题**：团队不该再问“我如何把 AI 加入产品”，而应该深思“**当 AI 能直接完成我的产品最初被发明出来要解决的工作时，会发生什么？**” --- ### [14:04](https://www.youtube.com/watch?v=JbF_SbKJSEc&t=844s) 研发革命：94% 的 PR 由 Claude 编写与复合错误率挑战 在 Fin 内部，AI 不仅重塑了所售产品的形态，也彻底重构了内部的研发流程。目前公司内部约 **94% 的 PR（Pull Request）都是由 Claude Code 编写的**。智能体可以并行探索多种实现方案，在短短数小时或几天内就能构建出极为复杂的全栈功能，使得软件构建的边际成本逼近理论上的零。然而，研发逻辑也随之发生了重大转变：团队必须将任务拆解为精细步骤，清晰界定哪些可以自动化、哪些需要人工介入（Human-in-the-loop），并对“可靠性（reliability）”保持极度严苛的警惕。 ```mermaid flowchart LR accTitle: 复合错误率与可靠性衰减 subgraph Steps["串联执行链路 (5个步骤)"] direction LR S1["步骤 1<br/>95-96%"] --> S2["步骤 2<br/>95-96%"] S2 --> S3["步骤 3<br/>95-96%"] S3 --> S4["步骤 4<br/>95-96%"] S4 --> S5["步骤 5<br/>95-96%"] end Steps --> Result["最终整体可靠性<br/>衰减至 ~88%"] ``` 在多步骤串联任务中，模型极易陷入“复合错误陷阱”：如果把 5 个单步成功率为 95%~96% 的环节串在一起，最终的整体可靠性将暴跌至 88% 左右。一旦系统每 10 次调用就会失败 1 次，就必须保留关键的人工监督。因此，精准判断 AI 在何种场景下“足够好”，已经成为一项核心专业技能。团队也吸取教训“切勿迷信 Demo”——在 Claude 中花 2 小时跑通一个惊艳的原型并不难，但要将其打磨至能交付给客户的稳健生产级别，往往需要后续长期的持续工程化。团队的运作方式变得更像“实验室”而非“流水线”：通过持续的 Eval 评估体系验证模型输出，保留最优方案。 --- ### [15:47](https://www.youtube.com/watch?v=JbF_SbKJSEc&t=947s) 生产力暴增 3 倍：将设计品味铸入自动化部署流 当执行与编码不再是瓶颈，产品团队面临的最大制约因素转移到了“方向的选择”——即决定构建什么、为什么构建，并确保其质量卓越。AI 带来了极大的生产力过剩，如果缺乏战略方向，管理 AI 就像在指挥一千个互不沟通、同时独立开工的人，最终只会制造巨大的噪音。“**没有方向的速度只是纯粹的噪音（Speed without direction is just noise）**”。在超高移速下，一个错误的决策会导致团队以更快的速度偏离目标。 在这种新常态下，设计团队的职责不再是在 Figma 里绘制理想化的产品蓝图，而是设定产品的战略航向，并死守最高的质量标准门槛。Fin 在这套全新研发机制下取得了突破性成果： - **生产力跃升**：原本设定的目标是在 12 个月内将公司生产力翻倍，最终仅用了 **9 个月就实现了生产力提升至 3 倍（tripled）**。 - **全自主开发闭环**：近 **20%（五分之一）的 PR 实现了完全无人工参与**——由 Agent 自主发起、自主完整编写、自主完成 QA 测试，并由 Agent 最终部署上线。 - **品味基础设施化**：为了防止全自动流水线导致产品质量在无形中下滑，设计团队将自身的“审美与设计品味（taste）”深度固化到底层系统和自动化部署管线中。除了传统的组件设计系统，团队还将设计审查逻辑与思考技能转化为自动化检查节点，确保质量门槛伴随系统自身生效，而非依赖个别人员的肉眼审查。正如团队在速度与质量权衡中所领悟到的：“**走得快的团队，会在走得好的团队变快之前，先一步把质量做好（Fast gets good before good gets fast）。**” ### [18:25](https://www.youtube.com/watch?v=JbF_SbKJSEc&t=1105s) 研发速度的代际差距与组织心理的真正瓶颈 那些在推进中不断试错、边走边修补质量的团队，如今已经远远领先于同行。在过去的传统 SaaS 时代，发布九次重大产品更新可能需要耗费数年时间；而现在，团队在短短两个月内就完成了九次重大产品发布，交付周期从以往的“月、季度、年”直接压缩到了“天和周”。传统的两年期规划周期已被彻底淘汰，取而代之的是三到四个月的敏捷规划周期。快团队与慢团队之间的鸿沟每一天都在持续拉大，正如演讲者所强调的：“在当下，原地踏步就等于倒退。”（"Standing still is moving backwards today."） 这种差距的根源并不在于技术本身。从根本上说，每个人能接触到的底层技术几乎是一样的，技术迭代极快且具有普遍性。真正构成阻碍的是人的因素：心理惯性、旧习惯、风险厌恶以及沉没成本。组织文化以及团队对变革的响应和适应方式，才是团队唯一拥有完全主动权、能够切实产生改变的关键领域。 - **极速发布能力**：2 个月内完成 9 次重大产品发布，彻底颠覆了以往 SaaS 时代的研发节奏。 - **规划周期重构**：从旧时代的 2 年长期规划缩短至 3 到 4 个月的敏捷迭代周期。 - **核心制约因素**：技术对所有人均等开放，真正的制约来自团队的心理防线、对既有工作流的路径依赖与组织适应力。 --- ### [19:29](https://www.youtube.com/watch?v=JbF_SbKJSEc&t=1169s) 软件贬值时代的“护城河”重构：数据底层与产品品位 面对软件构建成本无限趋近于零的新现实，产品团队的工作范式必须全面重塑。团队应当去全盘负责一项“完整的工作任务”（Own an entire job），而不是沉湎于打磨单一功能。传统的界面设计正在被对话设计取代，设计师必须拿出过去雕琢屏幕的专注度去设计人机对话。产品团队的核心精力必须向价值的两端（The Edges）迁移——即在源头决定“该构建什么”，以及在终端判断“产出的质量是否达标”。 ```mermaid flowchart TD accTitle: AI时代软件防御性价值模型 subgraph DefensibleTop["顶层防线：人类核心能力"] Taste["品位、方向感与审美 (Taste & Direction)"] Judge["价值定义与严苛验收 (Edges: What to build & Quality judgment)"] end subgraph Middle["极度商品化/无壁垒区"] Code["传统 UI 界面、功能代码与交互控件<br/>（极易被 Claude/LLM 在数天内克隆）"] end subgraph DefensibleBottom["底层防线：专有壁垒"] Apex["专有模型/算法 (Apex)"] Data["不可复制的独家业务数据资产"] end DefensibleTop --> Middle Middle --> DefensibleBottom ``` 当编写软件变得近乎免费时，传统的“功能型产品护城河”已不复存在。团队工程师曾利用 Claude 在短短几天内完整克隆了竞争对手的整套产品；同样地，自身最好的功能也会迅速被对手抄袭。在这种环境下，真正具有防御性的壁垒只剩下两端：底层是凭借独家数据训练出来的专属模型（如 Apex），这是竞争对手无法在一个周末内复制的资产；顶层则是团队独特的产品品味（Taste）、方向把控与评判标准。 - **设计重心的转移**：从屏幕布局全面转向人机对话链路设计；产品工作退守至两端——“决定做什么”与“判断好不好”。 - **功能壁垒的消失**：借助 AI 代码工具，完整竞品功能可在数天内被复刻，代码与功能本身不再构成商业壁垒。 - **双重核心护城河**： 1. **底层（Underneath）**：建立在独家私有数据基础上的专属模型架构（如 Apex）； 2. **顶层（On Top）**：团队不可替代的审美品味、战略方向感与严苛的质检判断。 --- ### [20:22](https://www.youtube.com/watch?v=JbF_SbKJSEc&t=1222s) S 曲线起点：告别界面弯路，迈入产品黄金时代 从宏观的创新 S 曲线来看，整个行业目前依然处于曲棍球棒拐点（Hockey stick inflection）最底部的一个极小点上，变革才刚刚拉开序幕。行业在过去几十年里倾注心血打造的繁复屏幕界面和控制控件，本质上并不是最终产品，它们只是技术发展路径上的一段“弯路”（"They were just the detour along the way."）。 随着 AI 抹平了交互障碍，软件正在回归其解决问题的本质。这对于整个行业、优秀的产品与设计团队而言，是长久以来发生过的最好机遇，也是职业生涯中最具影响力和趣味性的阶段。演讲最后，发言人邀请现场观众前往 531 号展位体验全新的客户服务方案，并欢迎同行在会后进行深入交流。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=JbF_SbKJSEc&t=4s) Swing really big — Thom Rimmer > [02:10](https://www.youtube.com/watch?v=JbF_SbKJSEc&t=130s) 你的产品现在究竟该在哪里停下脚步？ — Thom Rimmer > [04:09](https://www.youtube.com/watch?v=JbF_SbKJSEc&t=249s) 功能本身从来不是目的，最终的结果才是目的。 — Thom Rimmer > [07:00](https://www.youtube.com/watch?v=JbF_SbKJSEc&t=420s) Arrived gradually, then arrived all at once — Thom Rimmer > [07:00](https://www.youtube.com/watch?v=JbF_SbKJSEc&t=420s) The agent that builds other agents — Thom Rimmer > [10:57](https://www.youtube.com/watch?v=JbF_SbKJSEc&t=657s) How does this think? — Thom Rimmer > [10:57](https://www.youtube.com/watch?v=JbF_SbKJSEc&t=657s) 当 AI 能直接完成我的产品最初被发明出来要解决的工作时，会发生什么？ — Thom Rimmer > [15:47](https://www.youtube.com/watch?v=JbF_SbKJSEc&t=947s) Speed without direction is just noise — Thom Rimmer > [15:47](https://www.youtube.com/watch?v=JbF_SbKJSEc&t=947s) Fast gets good before good gets fast — Thom Rimmer > [18:25](https://www.youtube.com/watch?v=JbF_SbKJSEc&t=1105s) Standing still is moving backwards today. — Thom Rimmer > [20:22](https://www.youtube.com/watch?v=JbF_SbKJSEc&t=1222s) They were just the detour along the way. — Thom Rimmer ## Mentioned Thom Rimmer, Fin, Intercom, Salesforce, ChatGPT, Notion, Gong, Guru, Figma, Sketch, iPod, Fin for sales, Operator, Jared Spool, Photoshop, Claude, Claude Code, Apex, Ai4 ## Video description Thom Rimmer - VP, Product Design, Fin AI will change almost everything we do in some way, and some things we do in every way. Product design is one of the latter. Product designers have spent entire careers perfecting the art of interfaces and controls, only to have AI collapse interaction to a simple chat box. Fin's VP of Product Design Thom Rimmer describes how the Fin team changed everything about how they build, and how they're thinking about the software future AI is pushing us all towards. [Source](https://www.youtube.com/watch?v=JbF_SbKJSEc)

[0:04] Thom Rimmer: Swing really big [2:10] Thom Rimmer: 你的产品现在究竟该在哪里停下脚步？ [4:09] Thom Rimmer: 功能本身从来不是目的，最终的结果才是目的。 [7:00] Thom Rimmer: Arrived gradually, then arrived all at once [7:00] Thom Rimmer: The agent that builds other agents [10:57] Thom Rimmer: How does this think? [10:57] Thom Rimmer: 当 AI 能直接完成我的产品最初被发明出来要解决的工作时，会发生什么？ [15:47] Thom Rimmer: Speed without direction is just noise [15:47] Thom Rimmer: Fast gets good before good gets fast [18:25] Thom Rimmer: Standing still is moving backwards today. [20:22] Thom Rimmer: They were just the detour along the way.

[0:04] Thom Rimmer: value is draining away from the surface layer—the toggles, controls, and UI—and dropping down into the raw capabilities of underlying models and harnesses, while simultaneously rising up into how those capabilities are packaged and delivered to the end user. [2:58] Thom Rimmer: software is simply 'going home' to conversational roots. [2:58] Thom Rimmer: where does a product actually end, and where do its boundaries lie? [5:23] Thom Rimmer: Users no longer drive software by clicking; they drive it by asking. [8:30] Thom Rimmer: the agent that builds the agent [8:30] Jared Spool: Tool Time: The wasted overhead spent navigating the interface, browsing menus, and clicking buttons, which provides zero inherent value. Goal Time: The meaningful time spent actually accomplishing the intended objective. [11:45] Thom Rimmer: How does this look and operate? to How does this think? [15:28] Thom Rimmer: leading a thousand people who all started independently at the same time without context of each other. [15:28] Thom Rimmer: painting pictures of the product that they wish they could build in Figma [18:16] Thom Rimmer: fast gets good before good gets fast.

[0:04] Hey, it's great to be here. For 20 years  my job as a designer has been the screen, [0:08] like the buttons, the layouts, the small moments  of polish that most people don't even notice. [0:13] The craft that I've built my entire career on, and  I think it's kind of disappearing at this point. [0:17] If you build products of any type today, the  bit that your customers actually touch and feel [0:22] and interact with, that's changing fundamentally.  And it might be the best thing that's happened to [0:27] my career and the best thing that's happened  to software in two to three decades. [0:31] Today I'm going to talk to you about why and what  it means for what you build next. I'm Thom, I lead [0:35] product design at a company called Finn. You  might know us better by our old name, Intercom. [0:40] I've had a front row seat, we've had a front  row seat to this change for the last 3 years [0:44] because we blew up the  entire company to build Finn. [0:47] For over a decade Intercom was a pretty big and  successful SaaS company, about 35,000 customers, [0:53] half a billion messages a month, one of the  pioneers in connecting online businesses with [0:57] their customers. And then ChatGPT landed. And  there was one question we just couldn't not ask [1:03] ourselves. Could something like this just replace  the human customer service industry altogether? [1:09] And if the answer was yes, our entire  business model was completely done. [1:13] So we made a call, made a call very early, and we  bet the entire company. And we built an AI agent [1:18] that does the job that customers pay us per seat  for their human agents to do customer service. [1:25] If you fast forward to about 3 and 1/2 years, come  up to 4 years now, the bet worked, like it paid [1:29] off. Like Finn recently crossed 100 million ARR,  it's growing at an absolutely astonishing rate, [1:35] and a month ago we announced definitive  terms to be acquired by Salesforce. [1:38] But it wasn't easy to get to where we  are today. We had to change literally [1:42] everything. Like the strategy, the product,  the pricing, how we build, who we are, [1:46] right down to the name on the door, we changed  everything. And what did we learn? Like we learned [1:51] that you have to swing really big, like really  big. And you have to ignore the people that tell [1:56] you you're crazy when you are swinging big. The  people who make really small, safe, comfortable [2:02] adaptations to accommodate AI, the the site  type of thing looks great on a LinkedIn [2:05] post or like a blog post. Those people,  they're not going to make it in this era. [2:10] Because AI changes almost everything we do in some  way and some things we do in absolutely every way. [2:16] Software and how we design it is one of the things  it changes utterly. Like AI is changing the shape [2:21] of software, but it's also changing how we  go about building that software. Underpinning [2:25] both though is a a monumental shift in where the  value lies in software. It's leaving the surface, [2:30] it's leaving the toggles and the buttons and  the UI that we built for years, and it's moving [2:34] downwards, it's dropping down. It's moving  towards the model and the harness underneath, [2:38] the raw capability. But kind of paradoxically,  it's also moving upwards at the same time. [2:42] It's moving upwards towards how we deliver that  capability and how we package it to customers. [2:48] Uh so, I want to talk about both of those shifts,  uh starting with how AI is shaping the future [2:53] of software, the software that teams like  mine design and build every single day. [2:58] So, for years, like decades even, the product  that we built was pretty stable and it had [3:01] pretty clear edges. You could actually draw like  a pretty neat box around what the software did, [3:06] and everyone knew what it was for and what was  inside it and what wasn't inside that box. And [3:10] AI dissolves that box, it's a convergent force.  Like uh all the walls and the categories that [3:15] we invented to categorize software, they're all  melting away at this point. Like the line between [3:20] search and chat, the line between product and  service, uh between uh like they're all melting, [3:25] they're all gone like at this point. Uh every tool  is growing outwards and swallowing the next ones [3:31] along the line. We're kind of seeing this  already, like Notion is expanding horizontally. [3:35] Like Gong and Guru are kind of eating at each  other from different directions. This is happening [3:39] in a lot of different categories. Uh which  leaves us with one genuinely hard question, like [3:44] where does your product actually stop now? Like  what's its focus and what are its boundaries? [3:49] So, all the things on the left look really  familiar. and quite wildly, they already [3:53] look like the thing on the right in most pieces of  software. We've gone from having like very brief [3:58] like deep, broad, crude interfaces with lots of  specific buttons and toggles and UI to having [4:04] a conversational interface where you just ask  the system to do the thing you want it to do. [4:09] Uh this may seem totally new, but it's actually  a recession back to I guess like the base of [4:14] everything. Uh this is a clay tablet from  Mesopotamia. It's about 4,000 years old, and [4:19] it's the world's oldest customer complaint, very  relevant for our category. Uh it's a man furious [4:24] with his merchant about the quality of his copper  that he's been delivered. Uh I love this example [4:29] because it tells you something really important,  like commerce and conversation have been bound [4:32] together since almost the beginning of time. Uh  for almost all of that time, business and action [4:38] were completely connected in the same way. They  were connected directly through people, through [4:42] conversation. So, when I tell you the interface  is uh turning back to conversation, it's not [4:46] science fiction, it's not the future. It's just  a recession to back to the oldest method we ever [4:51] had. Uh the screens were the detour along the way.  We're going home almost to like our human roots. [4:57] So, for the past few decades, we've had two main  ways to control computer. The command line, which [5:02] was really powerful, but pretty difficult to learn  for people, and then the graphic interface, which [5:06] was easier to learn, but fairly restrictive. And  the gatekeepers, people like me, who designed that [5:12] interface, like they designed the controls and  the buttons and the switches and limited what you [5:16] could do with that software. Now, there's a third  way. You just say what you want in plain language. [5:23] So, if you take this spreadsheet  example, um I've got a list of numbers, [5:27] and I want all the negative numbers to turn red.  So, I need to know that there's a feature called [5:32] conditional formatting, if anyone's a bit of an  Excel jockey, it's called conditional formatting. [5:36] Uh you have to find it in the right menu.  You have to click your way through that menu, [5:40] and I guess we got agents who control  browsers at one point. We got like [5:44] agentive browsers who could do that for  us, but really they were just our hands. [5:48] They were still doing the same job. You just  weren't individually directing it to do it. [5:53] But pretty quickly we get to a place where users  stop caring about how the app works all together, [5:58] and they only care about the end result at the end  of it. And I just want the app to do it for me, [6:03] and it happens and I have no idea what  the feature was called. I have no idea [6:06] what conditional formatting was. I just  wanted it to be red. Magically it was red. [6:11] And because the feature was never really  the point, the outcome was always the point. [6:14] Uh so in this world people drive software not  by clicking, they drive software by asking. [6:21] Not everyone feels the same about this though.  Um it's pretty big change like especially for [6:25] people like me, especially for product teams who  go and build that software day day and day out. [6:29] It changes who has to learn your product. And  it changes what it means to learn a product at [6:34] the same time. Like knowing where every button  was uh and that was a real skill. Like it um [6:40] But when anyone gets the result by asking, it's  not really the skill anymore. Like mastery of [6:43] the tool is not the job. I'm a designer. Most  people on their LinkedIn used to put like I'm [6:48] a Figma ninja, or like I'm a Sketch ninja. It's  kind of like silly when you look back at it, [6:52] but it was actually a job at that point.  You were a master of an individual tool, [6:55] and that applied in every single category.  Uh and all made sense at that point. [7:00] And this all feels sudden like most  shifts do. Um the iPod felt brand new, [7:05] but it came after the Discman, after the mini  disc player, after tons of MP3 players. This [7:10] one's been rolling towards us for many years. It's  arrived gradually, then it's arrived all at once. [7:16] Uh we've moved from maximal UI right  down to minimal UI UI very very quickly. [7:21] So like first AI chat arrived as like a layer  bolted on top of the app. Then it emerged as [7:25] a way to configure your app for you, then  it starts to absorb the app all together. [7:30] Then it removes the mode the need for most of  the interface at all. Then the AI starting to [7:35] do the thinking and the work and the interface is  barely there. And then finally, you start to pose [7:40] the question like where is the app? Like where  the work's actually happening anyway anymore? [7:46] Uh we've seen this play out in our own product.  Like we 3 years ago we launched Finn, and that [7:51] was a real leap. Like a generative agent instead  of a rigidly scripted chatbot. It took chat, which [7:55] was frankly pretty pretty shitty at the time,  and made it really good, like human level good. [8:01] Uh but the interface in which it showed up was  still the same one that we used when there was a [8:04] human agent at the other end of the conversation.  It was just packaged in the same way. [8:09] And then this April we launched Finn for sales,  and we gave that new end user experience like [8:13] a complete overhaul. The messenger now sits in  the middle at the bottom. It's far more present. [8:17] It's not hidden away in the bottom corner. It's  sleeker. It feels beautiful. It responds really [8:22] well. But it's still a chat window bolted on top  of the product or bolted on top of your website. [8:27] The main thing that's changed is you're talking to  an agent. You're not talking to a human. Nothing [8:30] else has really changed in that world. And then  we built Operator. And Operator is very different. [8:37] Uh Finn is an agent for your customers. Operator  is an agent for your teammates. Uh it's the agent [8:42] that builds the agent. It's building Finn. It  like manages your knowledge. It keeps the whole [8:47] system running. It's the agent that builds  other agents. Uh and here's the difference, [8:51] like you don't walk Operator through the steps.  It works out the next best move on its own, and [8:56] then it tests it, comes back to you, and it comes  back for approval of that work. It doesn't wait [9:00] to be told. It's goal-orientated. It knows what  the end goal is and what your team's end goal is. [9:04] And it figures that out, and it figures out what  needs to be done, and it does the work for you. [9:09] Before Operator, improving Finn meant some  of our customers had spreadsheets and docs, [9:14] and they would do lots of manual processes, lots  of checks back and forth. They were clicking into [9:18] those docs. They were going back and forth.  Like Finn was incredible for the end user. [9:21] For the teammate, it was still a pretty  manual process, like pretty difficult, hard. [9:26] Getting that thing set up, getting it to like  a high automation rate was pretty difficult. Um [9:31] Operator's changed that. Like completely changed  that. All that manual work's gone now. It's just a [9:35] conversation. You chat to it, and Operator does  the work and our customers love it. Like the [9:39] demand for operator has been wild. Like absolutely  wild. Like nothing we've ever seen before. [9:45] Operator's moving to GA this week. Uh if you  make sure you stop by booth 531 and meet some [9:51] of our Finn experts. There's also a few people in  the room will be hanging out afterwards and I'll [9:54] be hanging around outside. We'll also be doing  a live demo at 4:30. So come on for that demo [9:58] but you can also book an individual demo using  the QR code. I'm going to go past this. I'll [10:02] put it back up at the end again. So the reason  why operator is so good. Like operator is so [10:07] good because it knocks out something called tool  time and focuses on goal time. So there's a guy [10:11] called Jared Spool. He's like the OG of design.  Like everybody knows him in the design industry. [10:16] Um and he's split the time he's spending software  into two categories. Tool time and goal time. Tool [10:20] time is uh the time that you spend operating the  interface. Like finding the button, finding the [10:24] menus. And goal time is the time that you spend  actually doing the thing that you wanted it to do. [10:29] Uh tool time adds nothing. Like adds absolutely  no value at all. And the ratio between the two is [10:34] basically a measure of how good your software  was and how good your software team was. [10:38] The more intuitive a feature is like the less  time it took for you to learn it and to get you [10:42] through it and to understand how it works.  And operator doesn't just shrink tool time. [10:47] It completely deletes it. There is none of  that anymore. Uh you're not driving the app. [10:51] You're having a conversation and the interface  only shows up when it's genuinely relevant. [10:57] Now I'm not going to stand here as a designer  as a design leader and tell you that my [11:00] industry is dead. That like the graphic  user interface has one foot in the grave. [11:03] It certainly has one foot in it. It doesn't  have both foot in. It's not fully dead. [11:07] Um the other foot's very much alive and  kicking and will be for a long time to come. [11:11] Partly because products and habits take  lots of time to change. Like people [11:16] quite funnily take a lot of time to change. Uh  mostly because some things are just better within [11:21] interface. Like uh if you imagine editing a photo  in Photoshop, you don't want to write five times [11:26] make this a bit darker. That would just be a silly  non-efficient process. You just want a little [11:30] slider. So you can drag that slider up and down.  So, sometimes the interface is actually better. [11:35] Like, conversational interfaces are most powerful  when they're accompanied by necessary UI. Like, [11:39] they're accompanied by charts, visualizations,  tables, sliders, that type of thing. [11:45] But, going through this all makes you realize  something. When you start designing for the [11:48] conversation as the main event, you suddenly  start to appreciate how hard building good [11:52] software was in the first place. Like, we fought  that fight for 30, 40, 50 years since the software [11:58] industry started to begin. Like, most interfaces  were CRUD, create, read, update, delete. And [12:04] they were built on the assumption that the human  does the thinking, remembers the process, and then [12:09] needs a confirmation on the screen afterwards.  And the craft of it was brutal. Like, building [12:14] something a person just intuitively understands  is really hard. Getting them to discover that that [12:20] feature is even harder. And then getting them  to learn it with no walk-through, no tool tip, [12:25] no like change log email that they're inevitably  never going to open and ends up in the trash can. [12:29] Like, that's even harder. And then getting them  to do that while you're still shipping other [12:33] features simultaneously at the same time, so the  product surface is changing. Almost impossible. [12:39] Agents like Operator need none of that. They don't  need any of that. Like, they need good AI to power [12:43] them. They need clear structure from the user and  the decision trail that they're able to follow. [12:48] So, the question flips from how does this  look and operate to how does this think? [12:53] We built entire disciplines around the old problem  though. Like, onboarding flows, empty states, [12:57] progressive disclosure, like whole careers spent  answering the question of how do you get somebody [13:01] to do something without actually just telling them  how to do it. The craft, the thing that we spent, [13:05] I spent absolutely decades learning, it might just  be a relic at this point. It may no longer exist. [13:11] Now, take everything that I've shown you and  turn it on your own product. Like, how much [13:15] of your UI exists to navigate somebody around  something? Like, how many of your screens are [13:20] just wayfinding? They're just menus, they're just  like navigations, dashboards, settings, reports. [13:25] What happens when an agent can just do the thing?  Uh the navigation disappears, the setting screen [13:30] disappears, the dashboard that existed to surface  information that the agent can now just tell you [13:35] in context, they just disappear. So, the question  every product team should be asking themselves is [13:40] not how do I add AI to my product, it is what  happens when the AI handles the job my product was [13:45] invented to do in the first place. So, the shape  of software is changing fundamentally, surface [13:51] is thinning, the value is dropping down into the  model, and what you build as a team is changing [13:57] pretty dramatically. But, AI also changes the  nature of how you build software as a team, too. [14:04] So, at Finn, we didn't just become an AI  company in terms of the shape of the product [14:08] that we build and sell, we also changed how we  build that product completely. Right now, about [14:13] 94% of our PRs are written by Claude. They're  written by Claude code. Agents explore many [14:19] ideas simultaneously, they build really complex  full-stack features in like hours or days now. [14:25] And the cost of building software is hitting  to like a theoretical zero this point. Like, [14:29] and that changes the shape of the work itself.  You start with the problem, you break down the job [14:34] into steps, you decide what can be automated, what  can't be automated, you decide where a human has [14:38] to stay in the loop. Uh and you become obsessed  with reliability cuz herein lies a massive trap. [14:45] Like, if you chain five steps together  that all have like a 95-96% reliability, [14:49] you end up with something like 88% reliability.  Like, errors compound. And when something only [14:53] works nine times out of 10, a human still has to  be in the loop really critical. Uh so, a tool like [15:00] that means that it's a real skill. Like, judging  where AI is good enough is a real skill now. [15:05] Uh you also learn pretty fast not to trust the  demo. This is throwing a bit of shade on Apple, [15:09] but like 2 hours in Claude and you get something  that looks really fantastic, and then the next 24 [15:14] months happen afterwards where it takes you to get  to a point where it's like robust enough to put [15:18] into a customer's hands. So, you You like a lab,  not a factory. Like, you try something, you test [15:23] what the model actually produces, you run evals,  like you understand whether it's good or it's not, [15:28] and you keep what wins. And the quality is  decided below that surface. It's decided at [15:32] the model layer. Uh so, execution for product team  used to be a major constraint. Like getting the [15:37] thing actual built was a bottleneck. Uh but um  that's gone. Like I just told you that's gone, [15:42] but constraints don't disappear altogether. They  often just move to somewhere else. And this one [15:47] moves to direction. Like choosing what to build  and why it matters. And ensuring that that thing [15:51] is actually good as well. Because when anyone can  build anything, building stops being the thing [15:56] that sets you apart from another company, from  your competitors. Choosing what to build in the [15:59] first place is the thing that sets you apart.  So, AI already owns the middle of the process. [16:04] Like it owns generating interfaces, it owns  generating code, like variations, like deploying. [16:10] That part's real and it's scaled incredibly  rapidly over the last like 6 to 12-month period. [16:17] But like AI more AI doesn't give you more  clarity. It just gives you abundance. Like [16:23] um it's less like leading a team of engineers,  it's less like leading your product team anymore. [16:27] It's more like leading a thousand people who  all started independently at the same time [16:31] without context of each other. Uh it just becomes  noise. And speed without direction is just noise. [16:36] Uh the faster everything moves, like a bad  decision becomes worse now because you can [16:41] end up further off target. Uh and that's where  a design team earns their keep. That's where [16:45] my team earns their keep. Uh it's not in painting  pictures of the product that they wish they could [16:49] build in Figma anymore. It's in uh setting the  direction, the strategic direction for where the [16:53] product should go, and then holding the bar to  the absolute highest level. So, does any of this [16:59] actually work? Like um look at these two shifts  and what they produced for us at Thin. Like we set [17:04] out to double our productivity as a company inside  12 months, and we tripled it in 9 months. Uh [17:09] and the part that still surprises me, nearly one  in five of our PRs now have absolutely no human [17:14] involved at all. They are initiated by the agent,  they are completely built by the agent. They are [17:19] then QA'd by the agent and then deployed by the  agent. There's no human involved at all. So, it's [17:22] like a cyclical loop of product development that  happens based on the guardrails that we give it. [17:27] So, but there's a lot of danger in that. Like  danger of quality slipping. Danger of that [17:32] happening quietly while you don't notice. So, my  team, the design team, have invested a lot of time [17:37] and energy in baking what we call our like taste  into the foundations into of the system itself. [17:42] Like everyone's contributed to this, absolutely  everyone in the company. Uh we put a lot of time [17:46] and energy into like components, what would be  known as like traditionally like a design system, [17:50] but we also spend a lot of time in building  skills. Skills that replicate the thought [17:54] processes and skills that replicate like the taste  and the judgment of our design team, and we put [17:58] them in the deployment pipeline. So, there are  checks now against the software that's deployed. [18:02] So, the bar still holds and it travels with  the system. It doesn't travel with the people. [18:06] Uh the craft of design hasn't left us. Like we're  still incredibly like diligent about the craft of [18:11] design. It's just moved into the foundations  of the way that we go about building software. [18:16] Uh so, here's what we learned about  the speed and quality trade-off though. [18:19] Like fast gets good before um good gets  fast. Like the teams that move really quickly [18:25] and fix quality as they went along are the teams  that are miles ahead of the other teams today. [18:32] We shipped nine major product launches in 2  months. That's worlds away from our SaaS era, [18:37] like absolute worlds away. It would have  taken us years to do that in the previous era. [18:41] And we're doing that now in days and weeks, not  months, quarters, years. We no longer have 2-year [18:45] planning cycles. We have 3-to-4-month planning  cycles. The gap's widening. Like standing still [18:50] is moving backwards today. Like the distance  between teams that move fast and distance [18:55] between teams that move slowly is growing  every single day as we're sat here today. [18:59] And here's why that gap exists. It's  not necessarily the technology, like [19:03] fundamentally everybody has access to almost  the same technology. Um the technology's [19:07] changing fast and that's happening to you, but  people change incredibly slowly. And psychology, [19:13] habits, risk aversion, like sunk costs, the real  constraints, like um you've got control over that. [19:19] Like you have control over the people and the  have the way the culture and the way that you [19:23] respond and adapt to this. That's the one thing  that you have complete agency to affect change on. [19:29] So what should product teams do in this world?  They should own an entire job, not a single [19:32] feature. Um they should guard the model and  the data underneath it. That's the moat now. [19:36] Design the conversation uh with the same  care and attention that you would have [19:40] designed the screen with. Um to do that you  have to move your attention to the edges, [19:44] like deciding what to build and judging whether  it's good enough. That's the work that's left now. [19:48] So let me pull all back together. Like what  you build changing, how you build changing, [19:52] and then both uh are at the point at the same  really difficult truth for us all to read. [19:57] When building software is nearly free, there's  no product moats left anymore. Like one of our [20:00] engineers clawed like used claw to clone an  entire competitor's product in a few days. [20:05] And the same happens to us. Like the inverse  happens to us. Like our best features are also [20:08] copied. And it's happening and flattering at  the same time. So what's actually defensible [20:13] now? Like two things, the edges the values move  to. Underneath our model apex, which is trained [20:17] on data nobody else has. You can't copy that in  a weekend. And on top the taste and direction. [20:22] Um but still very early though. Like in the  S-curve of innovation we're a tiny dot at the [20:26] bottom of the hockey stick inflection.  We're still super early in all of this. [20:30] Uh the screens, the features that we all spent  decades crafting, they were never really the [20:33] product. They were just the detour along the way.  Um that's why I meant what I said at the start. [20:38] Like this is the best thing that's ever happened  to our industry in a very long time. It's the [20:41] most interesting version of the job I've ever had.  And for the right like team, right product team, [20:46] right design team, it's the most impactful  they could ever be as well. Uh so thanks so [20:50] much for listening to me. Uh come and find us at  booth 531. Come and get a demo of how uh perfect [20:56] customer experiences are now possible within.  And we'll also be hanging around afterwards.