Zapier PM Says AI Tool Adoption Fails Without Orchestration Frameworks

Zapier产品经理：缺乏编排框架会导致AI工具落地失败 企业在独立的AI工具上浪费了数百万美元，因为它们未能在明确的管理层责任划分下，将模型直接编排融入现有的运营工作流中。 Reid Robinson 分享了 Zapier 推动全企业 AI 转型与编排的内部四步指南。他以 Remote.com 通过自动化支持工作流程节省 100 万美元等实际案例，概述了建立紧迫感、任命责任负责人、通过黑客周扩大应用规模，以及将 AI 熟练度融入公司惯例的策略。他还强调了建立明确的数据治理、将失败视为学习指标，并促成管理层达成共识，优先鼓励实验探索而非一味追求有保证的初期成果。 单纯采购 ChatGPT 或 Claude 等独立 AI 工具难以实现实质回报，企业必须转向 AI 编排（AI Orchestration），将模型与现有业务系统和工作流打通以获取可复制的商业 ROI。 (1:45) Remote.com 通过 Zapier 在 Slack 中编排 OpenAI，仅凭 3 人的 IT 团队便支持了 700 名员工的日常需求，使工单解决效率提升 25%，并节省了约 100 万美元的成本。 (4:26) 推动 AI 转型需自上而下建立紧迫感，Zapier CEO 通过宣布“红色代码”（Code Red）打破组织惯性，赋予团队暂停常规交付进行 AI 探索的权限，并设立专职负责人推进全员赋能。 (6:34) 规模化落地需要将 AI 转化为常态化仪式，Zapier 联合法务与数据团队建立安全护栏、指导非技术团队构建 MVP，并将 AI 熟练度硬性纳入全员绩效评估与招聘考核中。 (14:07)

Reid Robinson, Lead AI Product Manager at Zapier, argues that corporate AI initiatives frequently stall because purchasing standalone tools does not produce systemic transformation. Achieving measurable return on investment requires AI orchestration—connecting models directly into existing operational workflows—supported by executive urgency, clear data governance, and embedded organizational habits. Shift from Standalone Tools to AI Orchestration (1:45) Citing S&P Global Market Intelligence research covering North American and European IT leaders, Robinson points out that most corporate AI projects are eventually abandoned due to scaling and ROI failures. Rolling out isolated solutions like ChatGPT Enterprise, Claude, or Cursor checks an adoption box, but isolated prompts without integrated operational context fail to produce business transformation. In contrast, AI orchestration links foundation models and agents to core operational software used by subject matter experts. For example, Remote.com scaled to 700 employees supported by an internal IT team of just three people. By orchestrating OpenAI and Zapier within Slack, the company achieved a 25% increase in ticket deflection and resolution speeds, yielding an estimated $1 million in operational cost savings that could be redirected toward strategic growth. Organizational Urgency, Executive Ownership, and Prioritization (6:34) Systemic adoption requires explicit executive permission to disrupt routine work. In 2023, Zapier CEO Wade Foster instituted a company-wide 'Code Red,' signaling to employees that standard deliverables could be paused to focus directly on identifying where AI could resolve operational bottlenecks. Robinson emphasizes that driving AI enablement requires an accountable leader empowered to dismantle internal red tape and push projects forward. Sustaining this momentum demands what Robinson characterizes as ruthless prioritization by executive leadership. If executives expect guaranteed positive business outcomes from every pilot, teams are incentivized to prolong weak experiments rather than abandon them. Leadership must consciously open space in team roadmaps, allowing personnel to explore AI opportunities while accepting the risk of failure. Governance, Safe Data Access, and MVP Scoping (9:01) Scaling internal adoption requires early, rigorous alignment with legal and data teams. To ensure data privacy and maintain customer trust, Zapier established clear rules regarding acceptable data usage and integrated platforms like Databricks so employees could securely query corporate data without compromising security. For early-stage testing, teams leverage synthetic datasets generated by large language models alongside bounded internal data baselines. Equally critical is coaching non-technical departments, such as legal, human resources, and finance, on software delivery fundamentals. Because these teams often attempt to design complex, fully finished systems immediately, enablement leaders must define Minimum Viable Product (MVP) frameworks. Establishing objective criteria for internal testing allows organizations to safely validate simple prototypes before graduating them into enterprise-wide production. Embedding AI Fluency into Continuous Company Rituals (11:16) To prevent AI exploration from fading as a temporary novelty, organizations must institute permanent knowledge-sharing structures. Zapier implemented optional, weekly Friday morning 'AI All-Hands' sessions featuring breakout rooms where remote employees troubleshoot technical barriers and share both successes and failures. They also hosted company-wide Hack Weeks, prompting non-technical staff to build working prototypes. Robinson observes that peer demonstrations—such as a communications team member building an automated PR draft generator—proved far more effective at inspiring staff than engineering-led showcases. Finally, Zapier codified AI as an institutional competency across the entire talent lifecycle. The company created dedicated support channels spanning casual interest to specialized tooling like Claude Code, integrated AI fluency assessments into regular impact reviews for existing staff, and added fluency evaluations to candidate interviews for all new roles. Robinson concludes that long-term AI transformation hinges on normalizing failure, urging leaders to train teams to cut unviable experiments early rather than forcing stalled projects into production. Without explicit permission to fail and trade off standard deliverables, corporate AI initiatives risk remaining superficial and unscaled.

Zapier人工智能主管产品经理Reid Robinson指出，单纯采购独立AI工具无法为企业带来实质投资回报，真正的AI转型必须依赖于将模型与既有业务流程深度集成的“AI编排”；这一变革需要高层明确打破惯性、确立合规治理边界、赋能非技术人员自建工具，并将试错与能力评估深度嵌入组织日常机制。 从孤立工具转向业务流程编排 (1:45) S&P Global Market Intelligence对欧美IT主管的调查数据显示，大多数企业AI项目最终因无法在生产环境中规模化应用而遭到弃置。Robinson强调，企业采购ChatGPT Enterprise、Claude或Cursor等独立工具仅属于初期铺设，若缺乏系统性重构，往往无法兑现商业投资回报。真正的转型依赖于“AI编排”（AI Orchestration），即将各类AI模型、工具与代理深度嵌入企业现有的业务系统，并由一线业务专家（SMEs）主导定位核心运营瓶颈，让数据与上下文顺畅流入既有工作流。 以Remote.com为例，该公司在团队快速扩张至700人时，内部IT支持团队始终仅有3人。他们借助Zapier与OpenAI在员工日常使用的Slack中搭建了原生的工单编排系统，不仅实现了内部支持工单拦截与解决效率提升25%，还预计节省了约100万美元的运营成本。这一案例表明，只有把模型连接到实际业务网络中，才能变分散实验为可重复的规模化效益。 高管指令驱动与安全治理底线 (6:34) 推动全员AI落地的首要阻碍是组织惯性。2023年，Zapier首席执行官Wade Foster发布了公司史上首次“红色代码”（Code Red）全员通告，明确赋予员工暂停常规交付任务的自主权，为AI实验腾出思考与操作空间。Robinson将AI推进者比喻为推石上山的“西西弗斯”，而高管的强力指令能暂时抚平阻力；企业必须任命具备跨部门协调权限的专职负责人，专门负责扫除流程阻碍与推进落地。 在确立紧迫感的同时，清晰的治理与数据安全机制是赋能的前提。Zapier在推进过程中由赋能团队与法务、数据团队深度协作，明确可接受的数据来源和模型安全规范，例如通过在Zapier内部整合Databricks，使员工能够在确保安全合规的前提下调用企业数据库。此外，在数据赋能方面，Zapier根据场景区分测试路径，既利用大语言模型生成合成数据用于初期安全实验，也建立受控机制接入内部数据集，确保创新不以牺牲客户信任和数据安全为代价。 普及全民开发与MVP思维 (9:01) 规模化普及AI能力需要打破技术壁垒，让员工从“我不是构建者”转变为“我能解决自己的问题”。Zapier在发布红色代码后迅速推行全员黑客周（Hack Weeks），要求法务、财务和公关等非技术团队全部参与构建实用工具。实践证明，身边同事的示范比技术人员的演示更具启发性；例如公关社交媒体负责人自行搭建的新闻初稿起草机器人，使业务人员可在咨询公关团队前自助生成稿件，显著降低了跨部门协作成本。 针对非技术团队普遍缺乏敏捷软件开发经验、容易在初期追求“一步到位”的倾向，赋能者必须传授最小可行产品（MVP）理念。团队需要学习如何构建最基础的原型，明确内部工具的测试标准，并建立逐步将验证成功的MVP推广至全员或对外发布的渐进机制，避免因首日追求完美而陷入进度停滞。 嵌入制度惯例与培育容错机制 (14:07) 为避免AI应用沦为短期跟风，企业必须将其固化为长期的组织制度与文化惯例。Zapier构建了从日常轻量交流的“趣味AI”到针对特定工具（如Claude Code）及通用AI构建者的常态化互助社群，并固定在每周五上午举行全员AI分享会，通过线上分组透明复盘成功与失败案例。更具约束力的是，Zapier将AI素养评估硬性纳入在职员工的绩效考评体系，并在所有新增岗位的招聘流程中统一实施考核。 在文化层面，转型要求组织学会包容失败并建立敏捷止损机制。Robinson强调，验证假设后发现不可行并果断终止，同样属于积极有价值的产出，管理层不应将停止无前景的项目视为挫败。高管必须贯彻“残酷的优先级排序”（ruthless prioritization），在业务规划中为探索留出试错余地；若领导层苛求每个试点项目都必须取得确切回报，往往会导致团队被迫延长不良实验或将劣质项目强推上线。 Robinson的论点主要建立在Zapier自身作为全员远程无代码工具厂商的敏捷实践之上，其依赖全员自建工具与全面考核AI素养的激进模式，在面临严格行业监管、层级分明或数字化基础薄弱的传统重资产企业中能否顺利复制仍存疑问；此外，当常规业务交付考核与实验探索发生直接冲突时，如何量化分配资源与权衡产出风险，仍需管理层在实践中持续摸索。

## Background 该内容由 Ai4 于 2025 年 8 月 29 日发布在 YouTube 平台，由 Zapier 的 AI 首席产品经理 Reid Robinson 担任主讲。视频围绕企业如何摆脱单纯采购 AI 工具的困局，通过“AI 编排”（AI Orchestration）与系统化落地框架，将技术投资真正转化为可量化的业务投资回报率（ROI）。 ## TL;DR 演讲指出单纯部署 ChatGPT、Claude 等单点工具并不等于完成 AI 转型，大量企业因缺乏落地路径导致项目烂尾。Reid 结合 Zapier 自身及标杆客户的实战经验，提出了以“AI 编排”和赋能非技术人员（平民开发者）为核心的四步演进框架：制造紧迫感、机制赋能、规模化推广与内化为组织惯性。通过 Remote.com 节省百万美元 IT 运营成本等实际案例，他强调必须由深谙业务的一线人员主导工具搭建，同时管理层需实施“无情优先级排序”并包容失败，才能真正打破组织惯性实现规模化提效。 ## Atmosphere & style 演讲氛围幽默且极具共鸣，主讲人以“被领导推下水游向对岸海岛”的西西弗斯式比喻幽默自嘲开场，瞬间打破了高大上的叙事套路，引发全场会心苦笑与强烈认同。整体叙事节奏明快干练、直击痛点，既有打破“买了工具就算转型”行业幻觉的清醒洞察，又有倾囊相授的务实落地方法论，传递出强烈的一线实战者与产品人特有的真诚与务实感。 ## Key takeaways - [01:45](https://www.youtube.com/watch?v=VttVM1t0MmU&t=105s) 单纯采购和分发 AI 单点工具（如 ChatGPT Enterprise、Claude、Cursor）并不等于实现 AI 转型，大多数企业 AI 项目由于无法规模化和算不清 ROI 而最终遭弃置。 - [02:40](https://www.youtube.com/watch?v=VttVM1t0MmU&t=160s) AI 编排（AI Orchestration）是将 AI 工具、Agent 与企业现有业务系统及数据有机串联的关键，其核心在于向模型精准提供业务上下文而非仅做 Prompt 调优。 - [02:40](https://www.youtube.com/watch?v=VttVM1t0MmU&t=160s) Remote.com 借助 Zapier 将 OpenAI 模型嵌入 Slack 支持体系，在 800 人规模仅配 3 名 IT 工程师的情况下自主化解了 25% 工单，节省了约 100 万美元运营成本。 - [05:12](https://www.youtube.com/watch?v=VttVM1t0MmU&t=312s) 赋能平民开发者（Citizen Developers）并在合理治理架构下让一线业务人员亲自构建解决方案，是企业实现生产力真正突破的核心驱动力。 - [06:54](https://www.youtube.com/watch?v=VttVM1t0MmU&t=414s) 企业 AI 转型的第一步是自上而下制造紧迫感（如拉响红色警报），为团队松绑手头常规工作，并设立明确专职负责人（推石头的“西西弗斯”）扫清组织内部阻力。 - [09:00](https://www.youtube.com/watch?v=VttVM1t0MmU&t=540s) 推进赋能必须紧密结合法务与数据治理，既要为员工提供合规调取企业数据的安全通道，也要教育非技术团队掌握最小可行产品（MVP）与敏捷迭代思维。 - [11:15](https://www.youtube.com/watch?v=VttVM1t0MmU&t=675s) 规模化推广依赖固定的跨团队交流仪式（如每周 AI All-Hands）和全员黑客周（Hack Weeks），通过非技术岗位的同侪示范打破部门孤岛、激发全员落地信心。 - [14:07](https://www.youtube.com/watch?v=VttVM1t0MmU&t=847s) 转型的终局是将能力沉淀为制度惯性，Zapier 已将“AI 熟练度评估”全面纳入员工日常绩效考评和新员工招聘面试标准。 - [14:07](https://www.youtube.com/watch?v=VttVM1t0MmU&t=847s) 企业推进 AI 时应避免两大核心陷阱：缺乏专职权责归属（Lack of Ownership），以及执念于第一天就构建完美大系统而导致项目难产。 - [16:00](https://www.youtube.com/watch?v=VttVM1t0MmU&t=960s) 组织必须建立快速止损与包容失败的文化，验证并证伪一个业务假设本身就是有价值的产出。 - [16:33](https://www.youtube.com/watch?v=VttVM1t0MmU&t=993s) 在测试数据支持上，可通过 LLM 生成合成伪数据或接入合规内部业务基准数据，并建立清晰的数据分级使用指南保障安全探索。 - [17:20](https://www.youtube.com/watch?v=VttVM1t0MmU&t=1040s) 管理层必须践行“无情优先级排序”以腾出探索空间，若考核机制强行要求实验必须成功，会导致团队不敢止损、无限期拖延失败实验。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=VttVM1t0MmU&t=4s) 现场共鸣与“被推下水”的 AI 转型现状 Zapier 的 AI 首席产品经理（Lead AI Product Manager）Reid Robinson 首先介绍了自己的双重角色：他不仅负责 OpenAI、Anthropic、Microsoft 等大厂的战略合作伙伴关系，同时在过去两年半里，一直负责主导 Zapier 内部的 AI 赋能与工具落地。 为了切入主题，Reid 向台下抛出了两个极其真实的问题。首先，他询问全场：“你们当中有多少人，公司的 CEO 或高管曾对你们说：‘必须全面搞 AI，这是每个人的职责，这对业务至关重要’？”现场约有三分之二的人举起了手。紧接着，他又问：“又有多少人原本只是在一个具体的 AI 项目里，结果莫名其妙就变成了全公司 AI 赋能（AI Enablement）的负责人？”台下随即引发了一片苦笑与认同的骚动。Reid 感同身受地回应道：“我看到你们了，我完全懂这种感受（I see you. I feel you）。” Reid 用一个生动幽默的比喻描绘了当代企业 AI 转型的真实处境：这就好比你和全公司的人正困在一个贫瘠的海滩上挨饿，突然发现海对岸有一座物产丰饶的美丽岛屿，上面长满了苹果，跑着肥美多汁的野猪。全公司的人都在惊呼：“天呐，那太棒了，我们怎么过去？”此时，你只是好奇地走上前，把脚趾伸进水里试了试水温。你的领导转头看见了，立马拍着你的肩膀说：“看起来你挺会游泳的。去吧，游过去！顺便教会我们所有人怎么游！”这就是当下绝大多数企业推动 AI 变革时最真实的写照。 --- ### [01:45](https://www.youtube.com/watch?v=VttVM1t0MmU&t=105s) 破除迷思：AI 工具不等于 AI 转型 现实往往比美好的愿景更加骨感。根据标普全球市场财智（S&P Global Market Intelligence）对北美和欧洲 IT 领袖的调研数据显示，绝大多数企业发起的 AI 项目最终都被中途废弃（abandoned）。Reid 指出，虽然作为产品经理他非常崇尚快速试错与敏捷迭代，但对于传统企业而言，这种大面积的烂尾极具打击性和威吓感。许多客户向 Zapier 倾诉：他们满怀热情立项，但模型一上生产环境就崩溃，无法规模化落地，更算不清真实的投资回报率（ROI），最后只能选择将项目关停淘汰。 Reid 强调，当前行业最大的认知误区就是将“采购 AI 工具”等同于“完成了 AI 转型”： > “很多公司跑来跟我们说：‘Reid，我们买了 ChatGPT Enterprise，内部用了 Claude，工程师也在用 Cursor，我们已经把 AI 该做的事情都做完了啊！为什么看不到成效？接下来到底该干嘛？’” 单纯购买或分发工具并不能让业务发生质变，关键在于如何让技术真正穿透并贯穿到整家企业的实际日常运作中。 --- ### [02:40](https://www.youtube.com/watch?v=VttVM1t0MmU&t=160s) 从单点工具迈向“AI 编排”（AI Orchestration）与 Remote.com 案例 针对工具无法落地的痛点，Reid 提出了“AI 编排”（AI Orchestration）的核心概念。所谓编排，本质上不是去教员工怎么给 ChatGPT 写 Prompt，而是深入业务一线，识别具体团队所面临的最棘手痛点，并将各类 AI 工具、AI Agent 与企业现有的业务系统与数据资产有机串联。 在构建编排系统时，最大的技术挑战往往不在于模型本身，而在于如何精准获取并喂给模型正确的上下文信息。Reid 以全球知名远程雇佣平台 Remote.com 的实战为例：该企业在员工规模快速暴增至 700 到 800 人的阶段，其内部 IT 支持团队居然仅有 3 名工程师。面对雪片般飞来的内部工单，他们没有选择盲目招人，而是借助 Zapier 的编排能力搭建了一套自动化支持体系——将 OpenAI 的模型能力无缝嵌入到全体员工日常沟通的 Slack 平台中。 ```mermaid flowchart LR accTitle: Remote.com 的 AI 编排支持流程 subgraph 交互端 A["员工在 Slack 提问"] end subgraph Zapier 编排层 B["Zapier 工作流"] C["知识库与上下文检索"] D["OpenAI 模型推理"] end subgraph 交付与成效 E["Slack 实时自动解答"] F["复杂问题流转至 3 人 IT 团队"] end A --> B B --> C C --> D D --> E D -.未解决.-> F ``` - **工单分流与响应提速**：实现了 25% 的工单自主化解（Ticket Deflection），同时员工问题响应速度大幅提升。 - **直接经济效益**：通过避免团队规模被动扩张，据内部估算直接节省了约 100 万美元的运营成本，这笔资金被直接投入到更有价值的业务拓展中。 - **业务专家主导**：虽然架构看起来简洁，但背后依靠的是深谙 IT 服务细节的一线人员来梳理数据源与请求类型，亲自参与搭建贴合自身需要的工具。 --- ### [05:12](https://www.youtube.com/watch?v=VttVM1t0MmU&t=312s) Zapier 的核心理念：赋能平民开发者（Citizen Developers） 作为一家拥有 14 年历史的无代码（No-code）自动化平台，Zapier 从一开始的使命就是让没有技术背景的业务人员也能自行搭建所需系统。大约三年前，Zapier 率先生态集成了 OpenAI 等大模型。 Reid 指出，无论是 AI 时代之前还是当下，实现真正生产力突破的核心逻辑始终未变：当一家公司的非技术人员都能在合理的治理架构和安全防护下构建自己的解决方案时，“魔法”就会自然发生，企业才能真正渡过大海摘得果实。 --- ### [06:54](https://www.youtube.com/watch?v=VttVM1t0MmU&t=414s) 变革第一步：制造紧迫感（Create Urgency）与任命专职“西西弗斯” 进入具体的实战指南，Reid 分享了 Zapier 自身及标杆客户总结出的四步演进框架。 第一步是**制造紧迫感（Create Urgency）**。2023 年，Zapier 的 CEO Wade Foster 在全员群发布了一条重磅公告，召开了公司历史上首次“全员红色警报”（Code Red）动员会。当时绝大部分员工甚至不明白“红色警报”代表什么，CEO 本人也非常坦诚地承认团队还在探索之中。但这次警报在全公司起到了至关重要的两个作用：首先它敲响了警钟，让所有人意识到变革迫在眉睫；更重要的是，管理层通过行政手段“强行给团队松绑”。 > “我常把自己在 Zapier 的工作戏称为‘AI 领域的西西弗斯’——每天都在推着一块名叫 AI 的巨石吃力地爬坡。阻力就是整个组织的固有惯性。但当最高层宣布‘红色警报’时，相当于高管亲自出面，把眼前的斜坡暂时给削平了。他们给业务团队腾出了喘息的空间，允许大家暂时搁置手头的部分常规工作，停下来认真思考 AI 到底能帮自己解决哪些硬核难题。” 与制造紧迫感并行的关键动作，是明确委派负责人。在红色警报前，Reid 只是一个小规模 AI 研发团队的产品经理（他的开发搭档甚至是公司的两位联合创始人）。警报拉响后，CTO 立即找他谈话，要求他专职负责为全公司扫清障碍、推进 AI 赋能。企业必须设立一个明确的转型责任人，专门负责切除繁文缛节（Cut red tape），充当那个推石头的角色。 --- ### [09:00](https://www.youtube.com/watch?v=VttVM1t0MmU&t=540s) 变革第二步：机制赋能（Enablement）与优先级框架 推进赋能的关键并非盲目发放工具账号，而是建立清晰的使用边界与优先级标准： 1. **紧密绑定法务与数据治理**：Reid 在该阶段花费了大量时间同法务团队合作，明确界定哪些数据能用、哪些数据绝对不能碰；同时与数据团队协作搭建合规可用的底层库。例如，Zapier 在内部通过无代码方式打通了 Databricks，让普通员工在合规安全的前提下能够自由调用企业级数据。这不仅是内部安全红线，更是面向外部客户的必要交代——客户在做供应商评估时，会极其严苛地审视你的 AI 工具如何使用他们的数据。 2. **重塑优先级认知与落地 MVP 概念**：当 HR、公关或行政团队开始接触 AI 时，往往会陷入“一上来就想做一个尽善尽美的大系统”的误区。非技术背景人员通常缺乏敏捷开发的思维习惯。因此赋能的核心在于教会他们什么是“最小可行产品”（MVP），帮助他们理解从 0 到 0.1 再到 1 的演进节奏，并明确界定内部测试系统如何才能毕业升级为全员通用（GA）或对外开放的产品。 --- ### [11:15](https://www.youtube.com/watch?v=VttVM1t0MmU&t=675s) 变革第三步：规模化推广（Scale）——打破孤岛与全员黑客周 在谈及如何全面铺开时，Reid 幽默地贴出了一张经典的网络热梗图——《如何画一只猫头鹰》：“第一步，画几个圆圈；第二步，画出这只猫头鹰剩下的所有细节。”他坦言，对很多团队来说，所谓的规模化落地往往就像这幅画一样摸不着头脑，实际情况千差万别。 ```mermaid flowchart TD accTitle: Zapier 规模化推广飞轮 subgraph 组织仪式 A["每周五全员 AI 大会<br/>（AI All-Hands）"] --> B["分流讨论室聚焦攻坚"] end subgraph 实战驱动 C["全员黑客周<br/>（Hack Week）"] --> D["非技术岗亲自构建工具"] end subgraph 同侪效应 D --> E["业务人员自主宣讲成果"] E --> F["激发跨部门‘我也能做’的信心"] end B --> C F --> A ``` Zapier 摸索出的两大法宝包括： - **固定的跨团队交流仪式（AI All-Hands）**：尽管 Zapier 是一家高度推崇异步办公和文字文档的公司，但依然在每周五早晨设立了全员可选的实时“AI 大会”。参会人员可以自由分享踩坑教训、成功案例，并根据感兴趣的技术工具迅速切入专门的线上分流讨论室（Breakout Rooms）。 - **全员黑客周（Hack Weeks）**：在 CEO 的红色警报后，Zapier 要求全员上阵，从法务、财务到品牌公关团队，每个人都必须针对自己日常工作中的真实痛点亲手做出一款工具（哪怕只是在 ChatGPT 里配置一个专属的 Custom GPT）。 Reid 强调了同侪示范的威力：如果一个高级算法工程师做出了一款自动化 Demo，公关人员只会觉得“那是因为你懂写代码”；但当负责社交媒体的运营经理亲自上台，展示自己做的一款能根据产品需求自动生成新闻稿草稿的聊天机器人，并省去了反复向 PR 团队提需求的流程时，台下的其他文职同事就会瞬间被点燃：“天呐，和我背景一样的人都能搞定，这事我也能做！” --- ### [14:07](https://www.youtube.com/watch?v=VttVM1t0MmU&t=847s) 变革第四步：内化为组织惯性（Embedding）与常见避坑指南 转型的终局是将能力沉淀为制度和习惯。所谓转型（Transformation），就意味着“不可逆转，无法走回头路”。 - **搭建永久性的学习交流载体**：Zapier 在内部长期维护了专门的兴趣组织与答疑频道，包括面向工程团队的“Fun AI”与“AI Builders”研讨会、针对新兴前沿工具（如 Claude Code）的实战专区，以及方便任何不懂技术的员工随时“举手求助”的通用支持渠道。 - **将 AI 素养列入考核与招聘标准**：Zapier 已经在全公司范围内推行“AI 熟练度评估”（AI Fluency Assessments）。这一指标不仅被严格纳入现有员工的绩效考评（Impact Reviews），更成为针对所有新招募岗位（不论职能）的必考面试项。 在演讲最后，Reid 总结了企业最容易踩中的两大核心陷阱： 1. **权责归属缺失（Lack of Ownership）**：没有选派坚定的负责人去主导推进并切断组织内部阻力。 2. **执念于一步到位（Building the Perfect Version）**：试图在第一天就憋出一个完美无瑕的系统，最终导致项目难产放弃。 他建议企业不要孤立地采购单点产品，而是必须紧紧围绕具体业务问题与产出目标，让业务专家冲锋在前，组织体系全力赋能，在全公司培养起快速小步试错、迅速规模化复制成功经验的敏捷文化。 ### [16:00](https://www.youtube.com/watch?v=VttVM1t0MmU&t=960s) 快速止损与拥抱失败：Zapier 落地指南与资源分享 在演讲的收尾阶段，Reid Robinson 强调了组织在推进 AI 实践时必须建立的关键文化——学会快速止损，不要在无法规模化的项目上死磕。他指出，验证假设即使最终证明不可行（失败），本身依然是一个极具价值的产出，这种对失败的包容度必须作为一种核心能力在整个组织中得到赋能与鼓励。随后，他展示了现场幻灯片上的二维码，供观众获取包含演讲 Slide、Zapier 完整的落地执行指南（Playbook）以及更多延伸资源在内的整套材料，并随即进入最后 3 分钟的现场快速问答（Rapid Fire Q&A）环节。 - **假设验证与失败价值**：即使某个实验最终无法推进上线，证伪一个假设也是成功的业务探索，组织必须赋予员工勇于尝试和及时终止实验的权力。 - **资源与落地 Playbook**：通过现场二维码向参会者开放了 Zapier 内部实施 AI 编排的实操指南与演讲演示文稿。 --- ### [16:33](https://www.youtube.com/watch?v=VttVM1t0MmU&t=993s) 问答环节：如何为员工探索提供测试数据支持 现场观众提出的第一个问题聚焦于数据赋能：当要求员工自主发掘业务痛点并尝试用 AI 构建解决方案时，公司究竟该如何为他们提供数据支持？是直接开放 API 端点与生产环境中的实时真实数据，还是使用模拟数据？ ```mermaid flowchart TD accTitle: AI 测试数据选用决策流 subgraph S1["用例需求评估"] A["评估测试用例与风险级别"] end subgraph S2["数据准备模式"] B["利用大模型生成合成/模拟数据"] C["接入合规内部业务基准数据"] end subgraph S3["组织治理规范"] D["明确各系统与场景的数据使用安全指南"] end A --> B A --> C B --> D C --> D ``` Reid Robinson 回应道，这完全取决于具体的业务用例。在实践中，Zapier 会利用大语言模型（LLM）来为特定场景自动生成测试用的合成伪数据（Synthetic / Fake Data）；而在另一些场景下，则会引入部分内部真实业务数据作为初始测试基准。他强调，组织在推进此类探索时，最关键的是在内部建立清晰的数据使用指南（Guidelines），明确界定哪类系统、哪类测试场景可以使用何种级别的数据，让员工在安全合规的框架内进行探索。 - **合成数据生成**：利用 AI 模型自身的能力自动生成模拟测试数据集，加速概念验证。 - **数据治理与使用指南**：组织必须明确规则，指导团队在不同系统与测试阶段中合规使用内部数据或脱敏数据。 --- ### [17:20](https://www.youtube.com/watch?v=VttVM1t0MmU&t=1040s) 问答环节：争取管理层支持与“无情优先级排序” 现场观众提出了一个普遍面临的执行痛点：一线团队日常工作量本已完全饱和，要求他们额外腾出精力去思考和落地 AI 往往阻力重重，该如何说服管理层为团队减负、真正留出探索空间？ Reid Robinson 分享了一句多年来一直铭记在心的话：“如果你无法说‘不’，那么你的‘是’就毫无意义（If you can't say no, then your yes means nothing）。”他指出，管理层当前的核心职责在于进行“无情的优先级排序（Ruthless Prioritization）”。管理层必须认识到，团队的探索性实验并不保证一定能产生直接的商业回报；如果强制要求每个实验都必须产出既定结果，后果就是团队会把大量本该终止的失败实验无限期拖延下去，导致项目迟迟无法退出实验阶段。 ```mermaid flowchart LR accTitle: 管理层赋能与 AI 探索机制 subgraph M["管理层支持 Leadership"] A["无情优先级排序<br/>Ruthless Prioritization"] --> B["预留学习与探索空间"] end subgraph E["团队执行 Team Execution"] B --> C["明确实验假设与探索范畴"] C --> D{"允许实验失败<br/>不强制产出结果"} end subgraph R["机制落地 Roadmap"] D --> E1["产品路线图设立 AI 探索指标"] D --> E2["明确不启动某项目的具体原因"] end ``` 他进一步呼吁，管理层必须走出舒适区并正视现实：如果真心希望组织规模化采用 AI，就必须为团队预留学习与试错的空间。在实际管理机制上，必须在产品路线图（Roadmap）中明确各团队对 AI 探索的承接责任，或者要求团队在决定暂不推进某项 AI 探索时给出清晰充分的理由。 - **无情优先级排序**：管理层必须学会砍掉低价值任务，为 AI 创新腾出真实的排期空间。 - **解绑“实验必须成功”的枷锁**：若考核机制强行要求实验必须带来业务成果，会导致团队不敢止损、长期滞留在实验期。 - **将探索纳入路线图考核**：将 AI 探索明确写入团队 Roadmap，将其视为团队必须拥有或明确论证为何暂不执行的战略任务。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=VttVM1t0MmU&t=4s) 必须全面搞 AI，这是每个人的职责，这对业务至关重要 — Reid Robinson > [00:04](https://www.youtube.com/watch?v=VttVM1t0MmU&t=4s) 看起来你挺会游泳的。去吧，游过去！顺便教会我们所有人怎么游！ — Reid Robinson > [01:45](https://www.youtube.com/watch?v=VttVM1t0MmU&t=105s) 很多公司跑来跟我们说：‘Reid，我们买了 ChatGPT Enterprise，内部用了 Claude，工程师也在用 Cursor，我们已经把 AI 该做的事情都做完了啊！为什么看不到成效？接下来到底该干嘛？’ — Reid Robinson > [06:54](https://www.youtube.com/watch?v=VttVM1t0MmU&t=414s) 我常把自己在 Zapier 的工作戏称为‘AI 领域的西西弗斯’——每天都在推着一块名叫 AI 的巨石吃力地爬坡。阻力就是整个组织的固有惯性。但当最高层宣布‘红色警报’时，相当于高管亲自出面，把眼前的斜坡暂时给削平了。 — Reid Robinson > [11:15](https://www.youtube.com/watch?v=VttVM1t0MmU&t=675s) 第一步，画几个圆圈；第二步，画出这只猫头鹰剩下的所有细节。 — Reid Robinson > [11:15](https://www.youtube.com/watch?v=VttVM1t0MmU&t=675s) 天呐，和我背景一样的人都能搞定，这事我也能做！ — Reid Robinson > [17:20](https://www.youtube.com/watch?v=VttVM1t0MmU&t=1040s) 如果你无法说‘不’，那么你的‘是’就毫无意义（If you can't say no, then your yes means nothing） — Reid Robinson ## Mentioned Reid Robinson, Zapier, Wade Foster, OpenAI, Anthropic, Microsoft, S&P Global Market Intelligence, ChatGPT Enterprise, Claude, Cursor, Remote.com, Slack, Databricks, Custom GPT, Claude Code ## Video description Reid Robinson, Lead AI Product Manager, Zapier Abstract Companies are spending millions on AI tools but struggling to show business impact. Investing in AI tools and infrastructure is just the starting point, that still leaves you with tools you've rolled out, not tangible results. Orchestration is what takes you from strategy to real, repeatable ROI. This session reveals how real companies turned AI investments into operational success using orchestration. You'll see the exact framework that took them from "we have AI tools" to "we reclaimed $1M in efficiency gains." In this session you'll learn: - The 4 components that separate systematic AI adoption from scattered experiments that never scale - Real workflows that automated 3-day approval processes into 3-minute automations and how to identify similar opportunities in your organization - A proven playbook that moves companies from AI pilots to enterprise-wide implementation [Source](https://www.youtube.com/watch?v=VttVM1t0MmU)

[0:04] Reid Robinson: 必须全面搞 AI，这是每个人的职责，这对业务至关重要 [0:04] Reid Robinson: 看起来你挺会游泳的。去吧，游过去！顺便教会我们所有人怎么游！ [1:45] Reid Robinson: 很多公司跑来跟我们说：‘Reid，我们买了 ChatGPT Enterprise，内部用了 Claude，工程师也在用 Cursor，我们已经把 AI 该做的事情都做完了啊！为什么看不到成效？接下来到底该干嘛？’ [6:54] Reid Robinson: 我常把自己在 Zapier 的工作戏称为‘AI 领域的西西弗斯’——每天都在推着一块名叫 AI 的巨石吃力地爬坡。阻力就是整个组织的固有惯性。但当最高层宣布‘红色警报’时，相当于高管亲自出面，把眼前的斜坡暂时给削平了。 [11:15] Reid Robinson: 第一步，画几个圆圈；第二步，画出这只猫头鹰剩下的所有细节。 [11:15] Reid Robinson: 天呐，和我背景一样的人都能搞定，这事我也能做！ [17:20] Reid Robinson: 如果你无法说‘不’，那么你的‘是’就毫无意义（If you can't say no, then your yes means nothing）

[0:04] Reid Robinson: You look like you can swim! Go swim—and by the way, teach all of us how to swim! [6:34] Reid Robinson: the Sisyphus of AI [11:16] Reid Robinson: Step 1: Draw some circles. Step 2: Draw the rest of the owl [11:16] Reid Robinson: I'm not a builder [11:16] Reid Robinson: I can build solutions to my own problems [14:07] Reid Robinson: Fun AI [14:07] Reid Robinson: AI Builders [16:00] Reid Robinson: try to cut your losses soon [16:00] Reid Robinson: rapid fire questions [17:19] Reid Robinson: If you can't say no, then your yes means nothing.

[0:04] So yeah, you've got Reed here. I am the [0:06] lead product manager for AI at Zapier. I [0:08] am also fortunate that I get to do a lot [0:09] of our strategic partnerships. I work [0:11] with folks like OpenAI, Anthropic, [0:14] Microsoft, and as my second time at [0:17] Zapier, I also started with a lot of [0:19] what we do in AI enablement and tooling [0:22] internally. So I've been driving how [0:24] Zapier adopts AI for about two and a [0:26] half years now as well. And excited to [0:28] share some things Let's start. Number [0:31] one, this probably looks familiar to [0:33] some of you, but how many of you [0:34] internally have actually had your CEO or [0:36] a leader in your business be like, "You [0:38] need to do AI. This is like everyone's [0:40] responsibility. This is critical for the [0:42] business." Let's see. Quick hands. [0:45] Oh, actually, all right. About [0:47] twothirds. Follow-up question. How many [0:50] of you were on an AI project and are now [0:53] found yourself running AI enablement at [0:56] your company? There we go. I see you. I [0:59] feel you. And for a number of you, I [1:02] just want to recognize that is really [1:03] the state of AI transformation in [1:05] businesses today. I talked to a lot of [1:07] our customers and it is extremely [1:09] shared. A lot of what we're seeing today [1:11] is that businesses are trying to figure [1:13] this out at the same time that we're all [1:15] trying to figure this out as an [1:17] industry, as companies. Uh, a metaphor I [1:20] like to think about sometimes is you're [1:21] on a beach with your whole company and [1:24] you've been starving and all of a sudden [1:26] you see a beautiful island across the [1:28] water. It's got apples. It's got juicy [1:30] pigs. And your company's like, "Yo, this [1:31] looks great. How are we going to get [1:33] there?" And one of you decided to dip [1:36] your toes in the water. And your leader [1:37] looked at you and said, "You look like [1:39] you can swim. Go swim. And by the way, [1:41] teach us all how to swim." And again, [1:43] that's where we're at today. [1:45] So we see that though this is from you [1:48] don't trust us S&P global intelligence [1:50] brings us data from IT leaders across [1:52] North America and Europe that the major [1:54] a lot of AI projects are being abandoned [1:58] and sometimes this is great I'm a [2:00] products person I love rapid [2:02] experimentation and iteration we'll talk [2:04] about that but this can also be very [2:07] intimidating for companies we hear from [2:08] a lot from our customers that they [2:10] wanted to do things with AI they had [2:12] these projects but they didn't scale [2:14] fail. They didn't work in production. [2:16] They couldn't actually figure out ways [2:18] to see ROI from this. And they [2:21] ultimately either sunset it or they're [2:23] trying to figure out like is there more [2:25] to this AI thing? Uh and h what is [2:28] there? And one of the biggest things [2:31] that we see a lot of like mis mis I [2:33] don't know miscommunication on is just [2:36] AI tools not equaling AI transformation. [2:40] A lot of companies will tell us like we [2:42] bought chat GPT enterprise we used [2:44] claude internally we use cursor like we [2:46] did the AI thing Reed uh we're done like [2:49] why like what are we doing and a lot of [2:52] that will then come down to okay like [2:53] what is next like what should we be [2:56] doing more are companies doing more what [2:57] are they doing and critically how are [3:00] they getting that across the entire [3:02] company uh that's a really big thing for [3:05] a company like Zapier is thinking about [3:07] really that spreading about the entire [3:10] company and enablement. And some of the [3:13] ways that we start to see this really [3:15] work is when we move over to concepts [3:17] like AI orchestration. And for us, this [3:20] is that concept of really taking your AI [3:23] tools, agents, and finding the right [3:26] problems and how they can work together [3:28] with your existing business systems. So [3:31] some of the ways to think about this, [3:32] this is less about, hey, I need to [3:35] enable the company on how to use chatgpt [3:38] or claude and more about what are the [3:41] problems that individual teams in our [3:43] business face and let's understand them. [3:46] Let's have them think as the subject [3:48] matter experts. If an AI enabled system [3:51] could solve your toughest problems, what [3:54] would that be? What would that problem? [3:57] What is that problem you have? what is [3:59] your idea for a solution? And really [4:01] thinking about coming them with the [4:03] systems that can do because those [4:05] systems are going to have multiple [4:07] different types of AI. They're going to [4:09] have multiple different components. The [4:12] information is going to be absolutely [4:14] critical. We see a lot of businesses one [4:16] of the biggest struggles is actually [4:18] finding the right information to give to [4:20] the models so that they actually can [4:22] execute and perform for the problem. And [4:26] as we move along, like one of the [4:28] examples I love sharing is company [4:30] remote.com. Their struggle was they were [4:32] rapidly scaling to about a 700700 person [4:36] company. And at the time, their IT team [4:38] is still three people handling tickets [4:40] internally. And instead of scaling the [4:43] team, they wanted to challenge this and [4:45] try like could we come up with a better [4:47] solution for our teams internally. And [4:49] on Zapier, they built a system that [4:51] incorporated OpenAI. They put it into [4:53] Slack where their employees live and [4:56] they saw a 25% faster resolute or ticket [4:59] deflection and faster resolution for [5:02] their teams and they were able to save [5:04] in their estimates about a million [5:06] dollars in cost savings that they would [5:07] have instead of scaling that and put [5:09] that into other areas in the business. [5:12] And the cool thing here for me is this [5:14] is like this looks simple on a slide [5:16] like I made it look simple on a slide [5:17] here but this is not simple, right? [5:19] These are like core problems that are do [5:21] take a lot of knowhow from the folks [5:24] from it who are doing those jobs on okay [5:27] where do you get that information to [5:29] handle these requests what are the types [5:30] of requests you get and enabling them to [5:33] actually build their own tools to work [5:36] with if you have teams building the [5:37] tools then working together through that [5:40] and for us we see a lot of that [5:42] collaboration and a lot of that [5:43] enablement being really critical and [5:47] that's really why you know statement [5:48] here on empowering ing the people Zapier [5:50] for a little background for anybody that [5:52] doesn't know we've been a noode [5:53] platform. We're not a new company here. [5:55] We've been a noode platform for 14 years [5:58] and our whole thing has really been [5:59] about teams who don't have technical [6:02] engineers to build things. how can they [6:04] build the systems that they need? So [6:06] with AI, you know, we launched our [6:08] integrations about three years ago with [6:10] OpenAI as an example and we've been [6:12] rapidly seeing how companies adopt this. [6:14] And one of the critical ones is really, [6:16] and this is true prior to AI, it's [6:18] really when your whole company can [6:21] actually build their solutions and has [6:23] the governance, the structure, [6:25] everything they need to build [6:26] efficiently, uh, cool magical things [6:29] happen and you get some fruit and apples [6:32] and pigs. [6:34] All right, let's talk a little bit about [6:36] the journey. Uh I think a lot of you [6:38] here seem to be the people that are [6:39] tasked with doing that in your [6:40] companies. So I'll share from our [6:42] playbook on, you know, this is from our [6:44] own internal journey through this and [6:46] working with a lot of customers on what [6:47] we've seen has worked in how companies [6:50] kind of like get towards a state where [6:52] there are more citizen developers in [6:53] that company. Uh for that step one, [6:56] create urgency. This is probably an [6:58] obvious one is that first question I [6:59] asked saw a lot of hands. Um, surprised [7:02] actually there weren't all hands. Maybe [7:04] some shy people here, but that kind of [7:06] looks like this for us, right? In 2023, [7:10] our CEO Wade put out a post to the whole [7:12] company. We had a follow-up meeting for [7:13] this. It was a whole big thing. It was [7:15] our first ever code red. Most people [7:17] were very confused on what a code red [7:19] meant exactly. Um, which we didn't know. [7:21] He was very honest that we're figuring [7:22] this out. But the critical thing that [7:25] this led was number one, it act it stood [7:28] out, right? This was like an alarm bell [7:29] type of Slack emoji for those in Slack, [7:31] right? Moment for the business. And it [7:34] was also one where critically we allowed [7:37] teams to take a little bit of a [7:39] breather. I often sometimes describe my [7:42] role at Zapier as being the cisphace of [7:44] AI, just constantly pushing an AI rock [7:46] up a hill. And I think these moments in [7:48] companies are ones where your [7:50] organization momentum is that hill that [7:52] you're constantly pushing up. But if an [7:56] a leader can really come in and says [7:57] like, "We're going to flatten that hill [7:58] for a bit. We're going to give you [8:00] breathing room to pause the work that [8:01] you're doing and actually think about [8:04] the problems and where AI can help and [8:06] think about what are those solutions and [8:08] start blending those people together." [8:10] Uh it was a really critical moment. [8:12] We've seen some great results coming out [8:14] of this and I took that personally. Uh [8:18] my job prior to that code red was a PM [8:21] on a small AI R&D team that had two of [8:24] the three co-founders as my engineers [8:26] and you know at that moment our CTO came [8:28] to me and said Reed you know we're going [8:30] to do this holding for the company can [8:31] you keep unblocking and enabling things [8:34] for everybody across the business now [8:35] that's what we want you to do and we [8:38] also see that as being a pretty critical [8:40] thing of like nominate someone I think [8:42] these days we see these like nominations [8:44] for AI transformation leader who is that [8:47] accountable able individual that is [8:49] going to be, for lack of a better word, [8:51] a lot of it is like cutting red tape, [8:53] pushing that rock, right? Who's going to [8:55] be your internal cisphace uh for this? [8:59] And then let's start to do some [9:01] enablement. For, you know, a lot of [9:03] companies, the enablement is going to be [9:05] things like the guidelines of what you [9:07] can do, the tools that you can use, uh, [9:09] a lot of the different things that you [9:11] want to get into. For us, guidelines was [9:13] one of the most important things. And [9:16] oddly enough, it's what I spent most of [9:18] my time doing during this stage. Uh I [9:20] partnered really closely with legal. I [9:22] encourage a lot of our companies like [9:23] work with your legal team. Be very clear [9:26] on what you can use, what data you can [9:28] use, where does that data come from. Uh [9:31] work with your data team to make sure [9:33] that you have your databases set up in a [9:35] way that are usable and safe for the [9:37] individuals building. Um an example of [9:41] that for us like we built a we use data [9:43] bricks. We built a data bricks [9:44] integration into Zapier in a way that [9:46] every uh employee could actually build [9:48] that into their systems. Um but this was [9:52] also for our build teams. Our [9:53] engineering teams were doing a lot of [9:54] work and we needed to be very clear. But [9:56] it's also critical for your customers. I [9:58] mean almost all of your customers if [10:00] they're like ours, they want to know how [10:02] are you using AI. It's important in [10:04] vendor selection to be clear on how [10:07] those tools are using your data and how [10:09] they are creating safe systems with AI. [10:13] The other thing that we really needed to [10:14] enable teams on aside from tools and [10:16] everything was prioritization and [10:19] helping them figure out where to start. [10:20] Uh I'm talking here specifically on you [10:22] know when you start to enable your [10:24] communications, your HR, your all the [10:26] teams across your business, um they [10:29] often one of the struggles we had is [10:31] they start with like these big lofty [10:33] pictures of like well I need the system [10:35] to feel perfect. It's got to feel great. [10:37] These are not typically product builders [10:40] who might be more comfortable with like [10:41] the zero to 0.1 to 0 to one stage and so [10:44] forth. Um so giving them clear [10:46] guidelines on how are you going to [10:48] prioritize what what is an acceptable uh [10:52] like what does an MVP even mean for [10:54] anybody not like minimum viable product. [10:56] Um helping them be comfortable with [10:58] those concepts and giving them a [11:00] framework of also how do you graduate [11:02] from that? What does it mean to like [11:04] what is good in an MVP state for an [11:06] internal tool or system and how does [11:08] that graduate and what does it mean to [11:10] consider it generally available or [11:12] externally available and being very [11:14] clear on that. Um then scale this is a [11:17] fun part right we all want to scale [11:19] these things we want to make it across [11:20] the whole business. Um this is how you [11:23] scale you know if anybody doesn't know [11:25] this it's how to draw an owl. Step one [11:28] draw some circles. Step two, draw the [11:31] rest of the owl. Uh, that's the PG [11:33] version for those familiar. And [11:36] honestly, this is what scale I I chose [11:38] this slide because scale for us and [11:40] working with our customers, it varies [11:42] greatly uh across your business. I think [11:44] some of the critical things when it [11:46] comes to scale is the shared practices, [11:49] the share the just the sharing of [11:51] information across teams is probably one [11:53] of the biggest things. We established [11:55] early on an AI all hands is what we call [11:58] it. It's an optional meeting every [12:00] Friday morning. Uh even though we're we [12:02] operate very asynchronously as a company [12:04] of Zapier with a lot of documentation. [12:06] Uh we intentionally chose to do a live [12:08] meeting with people that they can come [12:10] in, share their challenges, share what's [12:12] working, share what's not working, uh [12:14] learn from each other and we do breakout [12:16] rooms. We divide into breakout rooms [12:18] depending on what different individuals [12:20] want to talk about during that meeting [12:21] uh to go through. Uh but it's also about [12:24] for us you know for scale became our [12:26] hack weeks right after that code red [12:28] from Wade one of the biggest things we [12:30] did was we told everybody literally [12:32] everyone in the company we need you to [12:35] think about your problems and try to [12:38] build something that addresses them and [12:40] I literally mean like our legal team was [12:42] building stuff our finance team was [12:43] building stuff our communications team [12:45] was building and these could be things [12:47] like they could just be GPTs in chat GPT [12:49] these could be like actual systems but [12:51] the critical thing was really just [12:53] getting teams who are on the ground [12:55] thinking about the challenges that they [12:57] have and thinking about themselves as [12:59] builders. Uh like when you with a lot of [13:02] things that growing I'm sure like vibe [13:04] coding and whatnot but if you start to [13:06] shift people's mindset from I can't [13:09] build like I'm not a builder to like oh [13:11] I actually can build these things now I [13:13] can build the solutions to my problems [13:16] across the business. Um, really cool [13:18] things happen, but critically sharing [13:20] because if an engineer shares, hey, I [13:23] heard you had this problem. I built this [13:24] demo for you and you, it's for a [13:26] communications person, they're not going [13:28] to think, I can build that, too. But [13:30] when someone like when the social media [13:32] manager says, I built this tool that's [13:35] going to help us do like from us of [13:37] comms did a a chatbot for draft press [13:39] releases. So, individuals across the [13:42] business didn't need to go to PR right [13:44] away with their request. they could [13:45] actually get a good draft, workshop it [13:48] on the product owner, and then work with [13:50] them with a much more final product. And [13:53] others within that team were like, "Oh, [13:55] that's how you do that. I can see [13:57] someone like me built something that is [13:59] compelling for a problem I'm familiar [14:01] with." And it's that just like making it [14:03] personal uh across the organization. [14:07] Um then embedding, embedding for us is [14:10] almost like establishing it as rituals. [14:12] Uh and this is like you know we talked [14:14] AI transformation hot topic but [14:17] transformation means like you don't [14:18] typically go back you don't untransform [14:20] I don't even know what the opposite [14:21] would be uh for that and there's a [14:25] couple things here number one for us [14:27] it's making spaces like making dedicated [14:29] spaces for people uh we have a number of [14:32] ones for like we have sessions like fun [14:35] AI and AI builders for engineers we have [14:38] ones specific for tools like clawed code [14:40] right now uh but we also have [14:42] generalized places that make it really [14:44] easy for anyone in the business to raise [14:46] their hand and say, "I need help with [14:48] this. Can you someone help me?" Um, [14:51] because it's it can be very confusing [14:52] figuring out what you're even asking for [14:54] help for if you're not familiar with how [14:56] to build these things. Uh, and as well, [14:59] Zapier, we started to do uh AI fluency [15:02] assessments for current employees during [15:05] impact reviews as well as for all new [15:07] hires for all roles. Uh so we have a [15:10] component and we have individualized [15:12] ones for roles on how we're evaluating [15:14] AI fluency for that uh because we [15:16] consider it a critical skill to be [15:18] working at Zapier. [15:21] And as we wrap up here, some of the [15:23] common pitfalls I mentioned lack of [15:25] ownership uh comes up quite a bit um and [15:29] building the perfect version. Those are [15:30] like probably the top two uh for that we [15:33] come into seeing. And I'll leave you off [15:36] with the before we switch to some [15:38] questions here. uh some summaries like [15:40] don't just buy a single solution. Try to [15:42] think about the problems and what are [15:43] those individual outcomes that you're [15:45] trying to get to. Work with your teams [15:47] who are the subject matter experts. Work [15:49] on enabling them. Work on letting them [15:52] come to you with problems and solutions [15:54] and help the whole organization learn to [15:56] iterate fast and scale what works. Uh [15:59] because there's going to be things that [16:00] don't scale and like try to cut your [16:02] losses soon. teach that failing with or [16:05] like testing a hypothesis is still a [16:07] good outcome even if it didn't move [16:09] forward and that really needs to be a [16:11] skill that is empowered in the [16:12] organization. All right, I leave here [16:16] this QR code leads to a website with [16:17] some resources including the slides. So [16:19] for anybody taking pictures, get the QR [16:21] code, you'll see all the slides. Uh [16:22] there's also our whole playbook on how [16:24] we implemented this and a lot more [16:27] information. Um I think we've got three [16:30] minutes rapid fire questions here. Let's [16:32] do it. [16:33] So you said you give uh employees have [16:36] to figure out problems and try to come [16:38] up with solutions when they're doing [16:39] that. Uh do you give them endpoints or [16:42] access to live data that is real to them [16:45] or how do you enable them on the data? [16:47] >> Yeah, good question on how do you enable [16:48] them with data and is it real or or fake [16:50] data does depend on the the use case [16:52] really. Uh one of the things that we do [16:54] use some of the models for is generating [16:55] fake data for use cases. So if we're [16:57] testing something we do generate with [16:59] some fake data set. Uh we have other use [17:01] cases where we leverage some like [17:02] internal data as like our starting data [17:05] on what to test with. Um but that does [17:07] like really depend that's one of those [17:08] things where honestly you probably need [17:10] to get internally familiar with like [17:11] what are the guidelines on what type of [17:13] data is good for what type of testing [17:15] with which systems uh for that. [17:19] >> Uh so obviously it's a big ask when [17:21] you're going to a team who already has a [17:23] full workload to say start thinking and [17:25] working on how to get AI to implement [17:27] this. What was your experience getting [17:28] leadership to buy in and be like these [17:30] teams need space we need to be sending [17:32] them less work to do because I went to [17:34] you know senior leadership like we need [17:37] them to do less right like [17:38] >> yeah this question on yeah getting [17:41] basically creating space and how do you [17:42] get leadership to create space for teams [17:44] to actually do this work uh many years [17:46] ago someone told me the expression [17:47] that's always stuck with me like if you [17:49] if you can't say no then your yes means [17:51] nothing and I think for leadership the [17:54] the task right now is like [17:55] prioritization and like ruth less [17:57] prioritization. I think teams need to be [17:58] clear on we want to experiment with this [18:01] and they also need to be comfortable [18:02] with the fact that experimentation for [18:04] teams might not lead to outcomes. Um, [18:07] and if you require them to lead to [18:09] outcomes, that's where you're going to [18:10] end up with a lot of experiments that [18:12] probably should not have left the [18:13] experimentation stage and you're going [18:15] to be experimenting for too long. So I [18:16] think it's one where leadership does [18:18] need to be you know challenged and [18:21] comfortable with the fact that if they [18:23] want their teams to truly be adopting [18:25] this then their teams need space to [18:26] learn they need to account for that um [18:29] and they need to include like in your [18:30] road maps if it's a product team like [18:32] some element that each team should be [18:34] owning some aspect of exploration with [18:37] AI or that they should be at least [18:39] prioritizing and coming up with reasons [18:41] of why they're not doing that project at [18:43] the moment.