返回列表
🧠 阿头学 · 💬 讨论题

Ramp 把全员推上 AI 的方法论

这篇文章最有价值的判断是：企业 AI 落地的决定性因素不是先写战略，而是用强预期、低摩擦工具和公开激励把 AI 变成组织默认操作系统，但它也明显把高使用率包装成高生产力，带有强 PR 色彩。
打开原文 ↗

2026-04-10 原文链接 ↗
阅读简报
双语对照
完整翻译
原文
讨论归档

核心观点

  • 先上车比先规划更重要 作者判断多数公司在“AI 战略”上想太多，这个判断有现实基础，因为 adoption 的真正瓶颈通常不是 PPT，而是员工能不能立刻做出东西；但他把“少规划”说得过满，放在高合规行业就未必成立。
  • Aha 时刻比培训有效得多 文中最站得住的部分是：培训不是核心，低摩擦接入、即时出结果、可分享 workflow 才能真正改变行为；这比一轮轮宣讲更符合组织真实学习方式。
  • 中心平台 + 边缘创新是可迁移框架 Ramp 先踩过“全中心化”和“全去中心化”的坑，最后落到“中心团队做平台、业务团队贴痛点构建”，这个组织设计是全文最稳的一块，不靠 Ramp 神话也能成立。
  • 作者实际上推的是制度化施压，不只是文化 文章嘴上说“文化”，但真正起作用的手段是强制使用、全员追踪、排行榜、绩效挂钩、招聘门槛和公开比较；这不是自然生长的文化，而是明确的管理机制。
  • 高使用率不等于高业务价值 6,300% 增长、99.5% 使用率、1,500+ 应用、12% 非工程师 PR 都很抓眼球，但这些主要证明“很活跃”，还不能证明“更赚钱、更少错、更高客户价值”；作者在这里有明显偷换。

跟我们的关联

  • 对 ATou 意味着什么、下一步怎么用 ATou 如果在做组织设计或产品落地，不该先追求“最完整 AI 战略”，而该先压低到首个有效结果的摩擦；下一步可以先做一个统一入口，把 2-3 个高频工作流打通，验证谁能最快拿到 Aha 时刻。
  • 对 Neta 意味着什么、下一步怎么用 Neta 如果关注 agent 或内部工具，文章说明护城河不在模型本身，而在连接器、权限、技能市场和反馈闭环；下一步应优先梳理“哪些上下文接入最值钱”，而不是继续卷通用聊天壳子。
  • 对 Uota 意味着什么、下一步怎么用 Uota 如果在看团队扩张或增长，这套方法本质是内部 PLG：先做可见成功案例，再靠同伴模仿扩散；下一步可以设计公开展示、模板复用和案例传播，而不是只发制度通知。
  • 对投资判断意味着什么、下一步怎么用 这篇文章强化了一个判断：企业 AI 的价值更可能落在“工作流接入层”和“组织 adoption 层”，不只在底模层；下一步看项目时，应追问它是否真正嵌入系统、能否复用 skill、是否降低非技术用户上手门槛。

讨论引子

1. 强制使用 AI、排行榜和绩效挂钩，究竟是在推动升级，还是在制造新的形式主义？ 2. 如果没有 Ramp 这种高速度文化和工程基础，普通公司最小可行版应该从哪里开始？ 3. “非工程师写生产代码”到底是效率革命，还是把技术债和治理风险前置了？

大多数公司还在争论自己的 AI 战略。他们想太多了。下面是我们用过的一套打法，让公司里的每个人都开始用 AI 做东西。

Ramp 的 AI 使用量比去年增长了 6,300%。团队里 99.5% 的人都在用 AI 工具。84% 每周使用编码代理。六周内，我们在内部平台上交付了 1,500+ 个应用，来自 800+ 位不同的构建者。非工程师现在占生产代码库里所有人工发起 PR 的 12% 他们每个月提交成千上万次 通过 Ramp Inspect 这套我们自研的编码代理。

做到这些，是因为我们对每个员工都拥抱这项新技术这件事有一种执念，像当年电脑进入职场一样。我们自建了一个 Claude Cowork 的版本，叫 Glass，给公司里任何人一个高度配置的 AI 代理，它和 Ramp 的系统全打通，也知道我们怎么做开发。我们办了史上规模最大的 AI 黑客松，700 名参与者来自销售、客服体验、法务、市场、财务，由 100 位最强的工程和产品同事做教练。他们一周交付的东西，比我们以前一年能做的都多。

我们改了招聘流程和人才管理流程。我们给每个人无限预算去做、去学、去探索。我们做了排行榜来激励使用。我们围绕那些看见未来的人重组团队。我们在全员会上庆祝胜利。我们不厌其烦地推动每一个人和每一个负责人去做东西。结果远超我的想象。

下面是怎么走到这一步的。有意思的部分不是数字，也不是工具。关键在于，我们一开始没有计划。只有文化和人才，然后不断加码那些在我们眼前、由这些人和这项技术带来的有效做法，看着它复利式增长。

1. 开始的第二好时机是今天。

在 Ramp，我们的文化是速度。它塑造了每一个流程和团队仪式。这种文化后来证明，是 AI 采用的单一最大加速器。

在 2025 年 1 月的公司 kickoff 上，我们告诉全公司，我们要成为世界上最具生产力的公司。我们相信凭借 Ramp 的文化能做到，但完全不知道怎么做。

我们从最显而易见的事情开始：

  1. 领导层明确表态，使用 AI 是一种预期

  2. 设立专门的 AI guild，能快速响应任何问题

  3. 建立 Slack 频道，让团队分享自己做了什么

  4. 在全员会上专门留时间庆祝构建者

  5. 对所有人强制使用并追踪 AI 使用情况

我们没有正式的变更管理项目，也没有强制培训课程。相反，我们搭建了基础设施，让大家能自学，也能互相教。现实是，只要给团队一次机会就行。每个人都想做东西。有了 AI，任何人都能。

2. 把 AI 熟练度当成学习曲线，而不是开关。

一年前，我们大多数人用 AI 的方式和所有人一样。浏览器里开个 ChatGPT 标签页。在 Notion 里做 AI 搜索。不错，但谈不上改变。

我们观察到，当人跨过某些舒适度阈值时，产出会出现跃迁。2025 年之前，除了一些特别优秀的工程师，几乎没人处在高阶水平。但在 2025 年末和今年，我们的加速非常明显，因为去年我们先花时间打了很扎实的地基。

我们把 AI 熟练度分成四个层级：

  • L0: 偶尔用 ChatGPT。工作流完全没变。如果还停在这里，也不主动起步，大概率不会继续留在公司。

  • L1: 做过自定义 GPT，用过 Notion agents，浅尝过 Claude Code。开始看到可能性，但还没形成复利。

  • L2: 做了一个能自动化自己部分工作的应用。提交过代码，或给别人的工作提过反馈。从这里开始才算进入状态。

  • L3: 系统构建者。他们不只是用 AI，还在搭基础设施，让所有人一起升级。这类人是倍增器。

我们的工作，是把所有人都往上推。三件事让这成为可能：

  1. 做工具时从人的现状出发。 我们先把全公司迁到 Claude 和 Notion AI，并把它们连接到所有办公工具，技术门槛很低，每个人都能参与，也能拿到实实在在的收益。这让大家从 L0 到 L1。

  2. 工具成熟后提高预期。 AI 熟练度进入了招聘筛选、入职培训和绩效对话。不是把它当作目的本身，而是作为明确的预期：想在 Ramp 把任何工作做好，熟练使用这些工具是必要条件。这会把 L1 推到 L2。

  3. 让强制要求匹配工具能力。 如果工具还做不到，就先提高预期，会烧掉信任，人们就不再听了。

3. 拥抱创造性破坏。

这部分让 Ramp 既令人兴奋，也同样让人不适。

我们在 2026 年 1 月交付的很多工具，已经过时了，被更好的版本替代，往往还是同一批构建者做出来的。我们已经习惯了以周为单位的生命周期，而不是以月。每一次 LLM 更新，每一次 Claude Code 或 Codex harnesses 的改进，每一批我们发布的新技能，都会重塑可能性。如果你三个月前的内部工具还像最先进的东西，那说明你走得还不够大胆。

我们的数据民主化之旅很好地说明了这一点：

  • Phase 1: Notion AI 是最佳选择，所以我们把重要数据导入 Notion 数据库，让我们的 agent 在上面跑。

  • Phase 2: 我们上线了 Ramp Research，一个基于 Slack 的 Snowflake 研究工具。

  • Phase 3: 随着编码代理成熟，我们把 Snowflake 研究编码成这些代理能直接使用的 skills。

  • Phase 4: 现在我们在把数据研究做成可交互、且能自我改进的东西。

每一代都打开了上一代打不开的门。每一代旧的都会被悄悄下线。我们现在在跑的工具？我们真心希望它们到 6 月就已经过时。

从外面看，这像一团乱。从里面看，恰恰相反。大家不依附工具，大家依附问题。 一旦出现更好的解法，就会立刻去用。

4. 从中心搭平台，用边缘驱动前进。

我们先把组织设计做错了，后来才做对。

最初的直觉是集中化，一个小团队给全公司做工具。需求很快就远超产能。然后我们又摆到去中心化，每个团队各做各的，结果是大量重复学习。

正确答案是两者都要：

  1. 一个小型中心团队 负责搭平台、连接器和底层管道，打通 LLM、数据、知识和工作流。他们也负责培训、赋能和变更管理。

  2. 职能团队 在这些平台上做自己的东西，并提供反馈，推动中心团队的路线图。

结果不言自明：

  • 一位风控分析师把每月 16 小时的手工财务建模自动化了。

  • 一位销售运营负责人在 48 小时内，把三个组织里基于表格的提成模型替换掉了。

  • 一位学习与发展负责人 15 分钟做出了一个培训模拟器。

  • 财务团队有人做了一个合同审阅器，每份合同能省 45 分钟 而 Ramp 的合同很多。

他们都不是工程师。

这些人没有提工单。他们自己发现痛点，做出原型，到需要上生产时才把工程拉进来 甚至有时候根本不需要。边缘推动中心的力度，和中心推动边缘一样大。

5. 给人舞台，而不只是命令。

命令会衰减。留下来的是文化。

所谓策略 如果真有的话 就是尽可能点燃很多小火苗，看看哪些能长成火：

  1. 一个 Slack 频道 (#ramp-uses-ai) 只是想看看会发生什么。现在成员超过 1,000。又衍生出 40+ 个团队专属频道，合计每月产生 20,000 条消息。

  2. 每周五的 AI office hours 经常有 40 到 50+ 人带着问题来。

  3. 给新人的 AI onboarding 过去一年随着野心增长，已经重做了四次。

  4. 专门的全员时间 从 CEO 到一线操作员都会演示他们用 AI 做了什么

最早那批转化者比什么都重要。每个团队里总会有一个人，野心勃勃的销售运营负责人，被流程折磨的产品运营，充满热情的数据科学家。他们先好奇，然后被吸进去，再变成团队里的传染源。我们让他们被看见，公司 All Hands 的 spotlight、做团队级工具的资源支持、在需要协作时把他们配对起来。

这种公开构建会形成一种所有人都能感到的竞争氛围。没人想成为那个什么都不做的团队。当一个 CSM 看到风控分析师交付了能省每月 16 小时的东西时，脑子里不是好厉害 风控做得真棒，而是我能做什么。

这个循环 做 分享 启发 再做 比任何命令或备忘录都更有效。**最意外的不是谁做得最多，而是有多少人一直在等一个允许他们开始做的许可 **

6. 尽快把人带到 Aha 时刻。

培训没用。Office hours 和 workshop 有帮助。但世界上最好的老师其实就摆在眼前，就是 AI。You can only lead a horse to water. 最大的解锁，是让一个人在第一天就体验到真实结果。

这点我们是吃过亏的。尽管公司里 AI 工具的采用率已经超过 90%，大多数人还是卡在基础聊天界面上。模型已经够好了。不够好的是 harness。终端窗口、npm 安装、MCP 配置，这些对大多数人来说都太难理解了。而那些真的硬扛过去的人，配置千差万别，学习沉在各自的孤岛里，没有形成复利。

所以我们做了 Glass，我们自己的 Claude Code Cowork 版本，基于 Anthropic 的 Claude Agent SDK 搭的。

Glass 安装时自动配置。通过 Okta SSO 认证一次，30+ 个工具就都亮起来 Salesforce、Snowflake、Gong、Slack、Notion、Google Workspace、Figma。不需要配置指南。不需要给 IT 提工单。如果用户需要自己调试，那我们已经输了。

四个人的团队，用不到三个月就做出来了。上线一个月内，日活 700。获得最大价值的人不是去参加培训的人，而是第一天装了一个 skill，马上就拿到结果的人。产品教会他们的速度，比我们快得多。

工具自己在手上，就能看到人到底卡在哪里，然后当天就把修复发出去。每一次会话都会产生信号，告诉我们非工程师到底怎么学会用 AI，哪些技能会被采用，人们在哪个点突破，什么区别了每周用一次的人和每天都用的人。

我们还做了一个叫 Dojo 的技能市场，任何人都能把一个工作流打包分享。全公司已经分享了 350+ 个 skills。一位销售找到最好的方式去分析 Gong 通话并写 battlecards，把它打包成一个 skill，现在每个销售都有了这项超能力。每分享一个 skill，都在抬高所有人的下限。

**结果就是，任何人在 5 分钟内都能做出任何东西。 **

7. 把它做成一场比赛。

人是好胜的。至少在 Ramp 是这样。

我们做了一个内部排行榜，追踪 Ramp 每个团队和每个人的 AI 使用情况，跑了多少 sessions，用了哪些 skills，交付了多少应用，连了哪些工具。所有人都能看到。怀疑者会说这是虚荣指标，追踪使用会激励忙碌而不是生产力。我们的结论相反。

Ramp 里最顶尖的 AI 使用者，往往也是最高绩效的人。 AI 熟练度和其他技能一样，练得越多就越强。重度用户在形成肌肉记忆，什么时候该用 AI，怎么写提示词更有效，该组合哪些技能，什么时候要覆写。他们在为自己的杠杆做复利。

排行榜带来了三个我们没完全预料到的效应：

  1. 健康的同侪压力。 没人想垫底。当你看到别的团队同级同事跑了 3 倍的 sessions，还交付了能给团队省下好几个小时的工具，你不需要命令才开始做。你只需要自己的胜负欲。

  2. 经理的问责。 团队级排名让经理没法忽视 AI 采用。如果你的团队在后四分位，那你一定会被拉出来聊。这让 AI 从可有可无变成评估团队是否在释放潜力的一部分。

  3. 通过模仿进行发现。 排行榜不只是计分板，它还是地图。当你看到某个人在顶端，你会想知道他在做什么。你会去看他的 skills、工作流和应用。

不衡量就无法管理。 不把它做成可见的，你就在放弃最强的采用杠杆。

这也延伸到了招聘和绩效管理。现在，加入 Ramp 的任何人都必须熟练使用 AI 工具，毫无例外。对 PM 候选人，我们有专门的一轮面试，给我做个产品，展示你怎么做的，带我走一遍它怎么工作。必须是完整原型，不是幻灯片。你如果不能证明你把这些工具内化了，就过不了门槛。

8. 清除人和 AI 之间的所有约束。

公司扼杀 AI 采用的第一方式，就是把它当成采购决策。预算审批。IT 评审。Token 限额。连接器请求排队好几周。这些都是把人和他们的 Aha 时刻隔开的墙。

我们做了相反的事。早期有三件事，重要性几乎压过一切：

  1. 把 AI 使用当成无限学习预算。 如果在大家还没学会用工具之前，就要求每个 token 都要有 ROI，你永远拿不到采用。我们给了大家探索空间，并明确预期，回报来自复利，而不是第一天。

  2. 干掉 token 限额和访问限制。 不设使用上限。不按角色分层访问。不搞你不是工程师 你不需要。每个人拿到同样的工具、同样的模型、同样的权限。最让我们惊讶的人，往往是传统审批流程下我们绝不会给权限的人。

  3. 移除连接器上的每一个 IT 瓶颈。 AI 代理有多有用，取决于它能访问什么。如果你的员工必须提工单，再等两周让 IT 批准 Salesforce 连接或 Snowflake 集成，他们会失去势头，再也不会回来。我们预连接了 30+ 个工具 Salesforce、Snowflake、Gong、Slack、Notion、Google Workspace、Figma，这样当人打开 Glass，所有东西都已经可用。一次 SSO 认证就开工。

下面这段成本算账，足以让任何 CFO 重新看待这件事，我们给员工付了很多钱。现在每个员工的 token 消耗，离他们工资的两位数百分比还差很远。但如果有人用 AI 生产力提升了 2 倍，你就该愿意再用 tokens 支付他一整份工资。如果你有能比一个人多做 10 倍工作的 agents，你为什么不愿意给它付出那个人两倍的成本？

看它复利式增长。

我们并不是一开始就比大多数公司有更好的策略。我们可能只是起跑条件更好，一种奖励速度和主动性的文化，一群人不等许可就去尝试的同事，一个支持大胆下注的领导团队，因为我们知道这对客户是好事。

没有总规划，我们就先开始。持续做工具，持续抬高标准，持续投资数据与 AI 基础设施，持续创造让人展示的场域。每条轨道各自复利增长。当它们彼此增强时，曲线就直冲天际。

AI 还处在非常早期的阶段。作为领导者，工作是给团队超能力，让他们相信自己。其他一切都会随之发生。

最重要的教训最简单，just get started.

We're hiring builders across all functions. If you want to work somewhere that non-engineers ship production code, internal tools have a shelf life of weeks, and nobody thinks that's weird -- come build with us at Ramp.

*Shoutout to @bleviathan for the collaboration on this piece. *

Most companies are still debating their AI strategy. They are overthinking it. Here is the playbook we followed to get everyone at the company building using AI.

Ramp's AI usage is up 6,300% from last year. 99.5% of the team active on AI tools. 84% using coding agents weekly. 1,500+ apps shipped on our internal platform in six weeks, from 800+ different builders. Non-engineers now account for 12% of all human-initiated PRs on the production codebase - thousands per month - using Ramp Inspect, our home-built coding agent.

We did this by being obsessed over every employee embracing this new technology - akin to how computers entered the workforce. We built our own Claude Cowork called Glass that gives anyone at the company a highly configured AI agent that is fully connected with Ramp's systems and aware of how we build. We hosted the largest AI hackathon ever - 700 participants across sellers, CX, legal, marketing, finance, coached by 100 of our most capable engineering and product teammates. They shipped more in a week than we previously could in a year.

We modified our hiring process and talent management process. We gave everyone unlimited budget to build, learn, and explore. We created leaderboards to incentivize usage. We reorged teams around those who see the future. We celebrated wins at all hands. We relentlessly pushed every single person and leader to build. And the result is far beyond what I could ever imagine.

Here's how we got here. The interesting part isn't the numbers or the tools. It's that we didn't have a plan. All we had was a culture and talent, and we kept doubling down on the things that were working with the people and technology in front of us. And watched it compound.

大多数公司还在争论自己的 AI 战略。他们想太多了。下面是我们用过的一套打法，让公司里的每个人都开始用 AI 做东西。

Ramp 的 AI 使用量比去年增长了 6,300%。团队里 99.5% 的人都在用 AI 工具。84% 每周使用编码代理。六周内，我们在内部平台上交付了 1,500+ 个应用，来自 800+ 位不同的构建者。非工程师现在占生产代码库里所有人工发起 PR 的 12% 他们每个月提交成千上万次 通过 Ramp Inspect 这套我们自研的编码代理。

做到这些，是因为我们对每个员工都拥抱这项新技术这件事有一种执念，像当年电脑进入职场一样。我们自建了一个 Claude Cowork 的版本，叫 Glass，给公司里任何人一个高度配置的 AI 代理，它和 Ramp 的系统全打通，也知道我们怎么做开发。我们办了史上规模最大的 AI 黑客松，700 名参与者来自销售、客服体验、法务、市场、财务，由 100 位最强的工程和产品同事做教练。他们一周交付的东西，比我们以前一年能做的都多。

我们改了招聘流程和人才管理流程。我们给每个人无限预算去做、去学、去探索。我们做了排行榜来激励使用。我们围绕那些看见未来的人重组团队。我们在全员会上庆祝胜利。我们不厌其烦地推动每一个人和每一个负责人去做东西。结果远超我的想象。

下面是怎么走到这一步的。有意思的部分不是数字，也不是工具。关键在于，我们一开始没有计划。只有文化和人才，然后不断加码那些在我们眼前、由这些人和这项技术带来的有效做法，看着它复利式增长。

1. The second best time to start is today.

At Ramp, our culture is velocity. It shapes every process and team ritual. That culture turned out to be the single biggest accelerant for AI adoption.

At our January 2025 company kickoff, we told the whole company we would become the most productive company in the world. We believed we could do it given Ramp's culture. We had no idea how.

We started with the obvious:

  1. Leadership clarity that AI usage is an expectation

  2. Dedicated AI "guild" that were responsive to any questions

  3. Slack channels where teams can share what they built

  4. Dedicated all hands time to celebrate builders

  5. Mandated AI usage and tracking for everyone

There was no formal change management program. No mandatory training curriculum. Instead, we built the infrastructure for people to teach themselves and each other. The reality is that teams just need to be given a chance. Everyone wants to build. And with AI, anyone can.

1. 开始的第二好时机是今天。

在 Ramp，我们的文化是速度。它塑造了每一个流程和团队仪式。这种文化后来证明，是 AI 采用的单一最大加速器。

在 2025 年 1 月的公司 kickoff 上，我们告诉全公司，我们要成为世界上最具生产力的公司。我们相信凭借 Ramp 的文化能做到，但完全不知道怎么做。

我们从最显而易见的事情开始：

  1. 领导层明确表态，使用 AI 是一种预期

  2. 设立专门的 AI guild，能快速响应任何问题

  3. 建立 Slack 频道，让团队分享自己做了什么

  4. 在全员会上专门留时间庆祝构建者

  5. 对所有人强制使用并追踪 AI 使用情况

我们没有正式的变更管理项目，也没有强制培训课程。相反，我们搭建了基础设施，让大家能自学，也能互相教。现实是，只要给团队一次机会就行。每个人都想做东西。有了 AI，任何人都能。

2. Treat AI proficiency as a learning curve, not a light switch.

A year ago, most of us used AI the way everyone did. ChatGPT in a tab. AI search in Notion. Fine, but not transformational.

What we observed was that productive output leaps when people clear certain thresholds of comfort. Almost nobody outside of some exceptional engineers was operating at the upper levels before 2025. But in late 2025 and this year, we accelerated massively -- because we spent last year building a strong foundation first.

We think about AI proficiency in four levels:

  • L0: Sometimes uses ChatGPT. Has not changed any workflows. If you're still here and not self-starting, you will most likely not be at the company.

  • L1: Built custom GPTs, used Notion agents, dabbled in Claude Code. Starting to see what's possible but hasn't compounded it yet.

  • L2: Built an app that automates part of their job. Committed code or contributed feedback to others' work. This is where things get real.

  • L3: Systems builders. They don't just use AI - they build the infrastructure that levels up everyone else. These people are force multipliers.

Our job is to get everyone up the ladder. Three things make that possible:

  1. Build tools that meet people where they are. We started by shifting the whole company to Claude and Notion AI connected to all of our workplace tools - a low technical bar where everyone could participate and get meaningful benefit. That got people from L0 to L1.

  2. Raise expectations as tools mature. AI proficiency moved into hiring screens, onboarding, and how we talk about performance. Not as an end in itself, but as a stated expectation: getting good at these tools is essential to doing any job at Ramp well. That pushes L1s to L2.

  3. Match the mandate to the tooling. If you raise expectations before the tools can deliver, you burn credibility and people stop listening.

2. 把 AI 熟练度当成学习曲线，而不是开关。

一年前，我们大多数人用 AI 的方式和所有人一样。浏览器里开个 ChatGPT 标签页。在 Notion 里做 AI 搜索。不错，但谈不上改变。

我们观察到，当人跨过某些舒适度阈值时，产出会出现跃迁。2025 年之前，除了一些特别优秀的工程师，几乎没人处在高阶水平。但在 2025 年末和今年，我们的加速非常明显，因为去年我们先花时间打了很扎实的地基。

我们把 AI 熟练度分成四个层级：

  • L0: 偶尔用 ChatGPT。工作流完全没变。如果还停在这里，也不主动起步，大概率不会继续留在公司。

  • L1: 做过自定义 GPT，用过 Notion agents，浅尝过 Claude Code。开始看到可能性，但还没形成复利。

  • L2: 做了一个能自动化自己部分工作的应用。提交过代码，或给别人的工作提过反馈。从这里开始才算进入状态。

  • L3: 系统构建者。他们不只是用 AI，还在搭基础设施，让所有人一起升级。这类人是倍增器。

我们的工作，是把所有人都往上推。三件事让这成为可能：

  1. 做工具时从人的现状出发。 我们先把全公司迁到 Claude 和 Notion AI，并把它们连接到所有办公工具，技术门槛很低，每个人都能参与，也能拿到实实在在的收益。这让大家从 L0 到 L1。

  2. 工具成熟后提高预期。 AI 熟练度进入了招聘筛选、入职培训和绩效对话。不是把它当作目的本身，而是作为明确的预期：想在 Ramp 把任何工作做好，熟练使用这些工具是必要条件。这会把 L1 推到 L2。

  3. 让强制要求匹配工具能力。 如果工具还做不到，就先提高预期，会烧掉信任，人们就不再听了。

3. Embrace creative destruction.

This is the part that makes Ramp exhilarating and uncomfortable in equal measure.

Many of the tools we shipped in January 2026 are already obsolete - replaced by better versions, often from the same builders. We've gotten comfortable with a shelf life of weeks, not months. Every LLM update, every improvement to Claude Code or Codex harnesses, every new batch of skills we release reshapes what's possible. If your internal tools from three months ago still feel state-of-the-art, you're not moving boldly enough.

Our data democratization journey tells the story well:

  • Phase 1: Notion AI was the best option, so we piped important data into Notion databases to run our agent over it.

  • Phase 2: We launched Ramp Research, a Slack-based Snowflake research tool.

  • Phase 3: As coding agents matured, we encoded Snowflake research into skills those agents could use directly.

  • Phase 4: Now we're making data research interactive and self-improving.

Each generation opened doors the previous one couldn't. Each former generation was quietly sunset. The tools we're running right now? We genuinely hope they're obsolete by June.

From the outside, this looks chaotic. From the inside, it's the opposite. People aren't attached to their tools. They're attached to their problems. When a better way to solve the problem shows up, they grab it.

3. 拥抱创造性破坏。

这部分让 Ramp 既令人兴奋，也同样让人不适。

我们在 2026 年 1 月交付的很多工具，已经过时了，被更好的版本替代，往往还是同一批构建者做出来的。我们已经习惯了以周为单位的生命周期，而不是以月。每一次 LLM 更新，每一次 Claude Code 或 Codex harnesses 的改进，每一批我们发布的新技能，都会重塑可能性。如果你三个月前的内部工具还像最先进的东西，那说明你走得还不够大胆。

我们的数据民主化之旅很好地说明了这一点：

  • Phase 1: Notion AI 是最佳选择，所以我们把重要数据导入 Notion 数据库，让我们的 agent 在上面跑。

  • Phase 2: 我们上线了 Ramp Research，一个基于 Slack 的 Snowflake 研究工具。

  • Phase 3: 随着编码代理成熟，我们把 Snowflake 研究编码成这些代理能直接使用的 skills。

  • Phase 4: 现在我们在把数据研究做成可交互、且能自我改进的东西。

每一代都打开了上一代打不开的门。每一代旧的都会被悄悄下线。我们现在在跑的工具？我们真心希望它们到 6 月就已经过时。

从外面看，这像一团乱。从里面看，恰恰相反。大家不依附工具，大家依附问题。 一旦出现更好的解法，就会立刻去用。

4. Build from the center, drive from the spokes.

We got the org design wrong before we got it right.

The initial instinct was to centralize: one small team builds tools for the whole company. Demand outstripped capacity almost immediately. Then we swung decentralized -- every team builds their own things. Tons of redundant re-learning.

The answer was to do both:

  1. A small central team builds the platforms, connectors, and plumbing across LLMs, data, knowledge, and workflows. They also manage training, enablement, and change management.

  2. Functional teams build on top of those platforms and give feedback that drives the central team's roadmap.

The results speak for themselves:

  • A risk analyst automated 16 hours per month of manual financial modeling.

  • A sales ops lead replaced a spreadsheet-based comp model across three orgs in 48 hours.

  • An L&D lead built a training simulator in 15 minutes.

  • Someone in finance built a contract reviewer that saves 45 minutes per contract -- and Ramp has a lot of contracts.

None of them are engineers.

These people didn't file a ticket. They found their own pain, prototyped a fix, and pulled engineering in when it was time to go to production -- when that was even necessary. The spokes drove the center as much as the center drove the spokes.

4. 从中心搭平台，用边缘驱动前进。

我们先把组织设计做错了，后来才做对。

最初的直觉是集中化，一个小团队给全公司做工具。需求很快就远超产能。然后我们又摆到去中心化，每个团队各做各的，结果是大量重复学习。

正确答案是两者都要：

  1. 一个小型中心团队 负责搭平台、连接器和底层管道，打通 LLM、数据、知识和工作流。他们也负责培训、赋能和变更管理。

  2. 职能团队 在这些平台上做自己的东西，并提供反馈，推动中心团队的路线图。

结果不言自明：

  • 一位风控分析师把每月 16 小时的手工财务建模自动化了。

  • 一位销售运营负责人在 48 小时内，把三个组织里基于表格的提成模型替换掉了。

  • 一位学习与发展负责人 15 分钟做出了一个培训模拟器。

  • 财务团队有人做了一个合同审阅器，每份合同能省 45 分钟 而 Ramp 的合同很多。

他们都不是工程师。

这些人没有提工单。他们自己发现痛点，做出原型，到需要上生产时才把工程拉进来 甚至有时候根本不需要。边缘推动中心的力度，和中心推动边缘一样大。

5. Give people a stage, not just a mandate.

Mandates decay. Culture is what remains.

The strategy, to the extent there was one, was to light as many small fires as possible and see which ones grew:

  1. A Slack channel (#ramp-uses-ai) - just to see what happened. Now over 1,000 members. Spun off 40+ team-specific channels that collectively generate 20,000 messages per month.

  2. AI office hours every Friday - regularly 40-50+ people show up with questions.

  3. AI onboarding for new hires - rebuilt four times in the last year as ambition grew.

  4. Dedicated all hands - we have everyone from our CEO to a first line operator demo what they have built with AI

The early converts mattered more than anything. On every team, there was one person - the ambitious sales ops lead, the frustrated product operator, the eager data scientist. They got curious, got sucked in, and became contagion for their teams. We made them visible: company All Hands spotlights, resources to build team-level tools, pairing them up when collaboration was warranted.

All this public building creates a competitive dynamic that everyone feels. Nobody wants to be the team that isn't building anything. When a CSM sees a risk analyst ship something that saves 16 hours a month, they don't think "good for Risk." They think "what can I build?"

That loop - build, share, inspire, build more - does more than any mandate or memo. **The biggest surprise wasn't who built the most. It was how many people had been waiting for permission to build at all **

5. 给人舞台，而不只是命令。

命令会衰减。留下来的是文化。

所谓策略 如果真有的话 就是尽可能点燃很多小火苗，看看哪些能长成火：

  1. 一个 Slack 频道 (#ramp-uses-ai) 只是想看看会发生什么。现在成员超过 1,000。又衍生出 40+ 个团队专属频道，合计每月产生 20,000 条消息。

  2. 每周五的 AI office hours 经常有 40 到 50+ 人带着问题来。

  3. 给新人的 AI onboarding 过去一年随着野心增长，已经重做了四次。

  4. 专门的全员时间 从 CEO 到一线操作员都会演示他们用 AI 做了什么

最早那批转化者比什么都重要。每个团队里总会有一个人，野心勃勃的销售运营负责人，被流程折磨的产品运营，充满热情的数据科学家。他们先好奇，然后被吸进去，再变成团队里的传染源。我们让他们被看见，公司 All Hands 的 spotlight、做团队级工具的资源支持、在需要协作时把他们配对起来。

这种公开构建会形成一种所有人都能感到的竞争氛围。没人想成为那个什么都不做的团队。当一个 CSM 看到风控分析师交付了能省每月 16 小时的东西时，脑子里不是好厉害 风控做得真棒，而是我能做什么。

这个循环 做 分享 启发 再做 比任何命令或备忘录都更有效。**最意外的不是谁做得最多，而是有多少人一直在等一个允许他们开始做的许可 **

6. Get people to the "Aha" moment as fast as possible.

Training doesn't work. Office hours and workshops help. But the world's best teacher is staring right in front of you: it's AI. You can only lead a horse to water. The single biggest unlock is getting someone to experience a real result on day one.

We learned this the hard way. Despite hitting 90%+ adoption of AI tools across the company, most people were stuck on a basic chat interface. The models were good enough. The harness wasn't. Terminal windows, npm installs, MCP configurations - these were simply too hard for the majority of people to grok. And the people who did push through had wildly different setups with siloed learnings that weren't compounding.

So we built Glass - our own version of Claude Code's Cowork, built on Anthropic's Claude Agent SDK.

Glass auto-configures on install. You authenticate once via Okta SSO and 30+ tools light up - Salesforce, Snowflake, Gong, Slack, Notion, Google Workspace, Figma. No setup guide. No ticket to IT. If the user has to debug, we've already lost.

A team of four built it in under three months. 700 daily active users within a month of launch. The people who got the most value weren't the ones who attended training sessions. They were the ones who installed a skill on day one and immediately got a result. The product taught them faster than we ever could.

When you own the tool, you see exactly where people get stuck and ship a fix the same day. Every session generates signal about how non-engineers actually learn to use AI - which skills get adopted, where people break through, what separates someone who uses it once a week from someone who uses it every day.

We also built a skills marketplace called Dojo where anyone can package a workflow and share it. Over 350 skills shared company-wide. A sales rep figures out the best way to analyze Gong calls and draft battlecards - packages it as a skill, and now every rep has that superpower. Every skill shared raises the floor for everyone.

**The result is that anyone in 5 minutes can create anything. **

6. 尽快把人带到 Aha 时刻。

培训没用。Office hours 和 workshop 有帮助。但世界上最好的老师其实就摆在眼前，就是 AI。You can only lead a horse to water. 最大的解锁，是让一个人在第一天就体验到真实结果。

这点我们是吃过亏的。尽管公司里 AI 工具的采用率已经超过 90%，大多数人还是卡在基础聊天界面上。模型已经够好了。不够好的是 harness。终端窗口、npm 安装、MCP 配置，这些对大多数人来说都太难理解了。而那些真的硬扛过去的人，配置千差万别，学习沉在各自的孤岛里，没有形成复利。

所以我们做了 Glass，我们自己的 Claude Code Cowork 版本，基于 Anthropic 的 Claude Agent SDK 搭的。

Glass 安装时自动配置。通过 Okta SSO 认证一次，30+ 个工具就都亮起来 Salesforce、Snowflake、Gong、Slack、Notion、Google Workspace、Figma。不需要配置指南。不需要给 IT 提工单。如果用户需要自己调试，那我们已经输了。

四个人的团队，用不到三个月就做出来了。上线一个月内，日活 700。获得最大价值的人不是去参加培训的人，而是第一天装了一个 skill，马上就拿到结果的人。产品教会他们的速度，比我们快得多。

工具自己在手上，就能看到人到底卡在哪里，然后当天就把修复发出去。每一次会话都会产生信号，告诉我们非工程师到底怎么学会用 AI，哪些技能会被采用，人们在哪个点突破，什么区别了每周用一次的人和每天都用的人。

我们还做了一个叫 Dojo 的技能市场，任何人都能把一个工作流打包分享。全公司已经分享了 350+ 个 skills。一位销售找到最好的方式去分析 Gong 通话并写 battlecards，把它打包成一个 skill，现在每个销售都有了这项超能力。每分享一个 skill，都在抬高所有人的下限。

**结果就是，任何人在 5 分钟内都能做出任何东西。 **

7. Make it a competition.

People are competitive. At least this is true at Ramp.

We built an internal leaderboard that tracks AI usage across every team and individual at Ramp. Sessions run, skills used, apps shipped, tools connected. It's visible to everyone. The skeptics will tell you this is a vanity metric - that tracking usage incentivizes busywork, not productivity. We found the opposite.

The top AI users at Ramp are often the highest performers. AI proficiency is a skill like any other - the more reps you get, the better you become. The power users are developing muscle memory for when to reach for AI, how to prompt effectively, which skills to combine, and when to override. They're compounding their own leverage.

The leaderboard created three dynamics we didn't fully anticipate:

  1. Healthy peer pressure. Nobody wants to be at the bottom. When you can see that your peer on another team is running 3x more sessions and shipping tools that save their team hours, you don't need a mandate to start building. You need your competitive instincts.

  2. Manager accountability. Team-level rankings made it impossible for managers to ignore AI adoption. If your team is in the bottom quartile, that's a conversation you're going to have. It shifted AI from "nice to have" to "part of how we evaluate whether teams are operating at their potential."

  3. Discovery through emulation. The leaderboard isn't just a scoreboard - it's a map. When you see someone at the top, you want to know what they're doing. You look at their skills, their workflows, their apps.

If you're not measuring it, you're not managing it. And if you're not making it visible, you're leaving the most powerful adoption lever on the table.

This extends to hiring and performance management. We now have an absolute requirement for anyone joining Ramp to be proficient with AI tools. No exceptions. For PM candidates, there's a dedicated interview session: build me a product, show me how you built it, walk me through how it works. It's a full-blown prototype, not a slide deck. If you can't demonstrate that you've internalized these tools, you don't clear the bar.

7. 把它做成一场比赛。

人是好胜的。至少在 Ramp 是这样。

我们做了一个内部排行榜，追踪 Ramp 每个团队和每个人的 AI 使用情况，跑了多少 sessions，用了哪些 skills，交付了多少应用，连了哪些工具。所有人都能看到。怀疑者会说这是虚荣指标，追踪使用会激励忙碌而不是生产力。我们的结论相反。

Ramp 里最顶尖的 AI 使用者，往往也是最高绩效的人。 AI 熟练度和其他技能一样，练得越多就越强。重度用户在形成肌肉记忆，什么时候该用 AI，怎么写提示词更有效，该组合哪些技能，什么时候要覆写。他们在为自己的杠杆做复利。

排行榜带来了三个我们没完全预料到的效应：

  1. 健康的同侪压力。 没人想垫底。当你看到别的团队同级同事跑了 3 倍的 sessions，还交付了能给团队省下好几个小时的工具，你不需要命令才开始做。你只需要自己的胜负欲。

  2. 经理的问责。 团队级排名让经理没法忽视 AI 采用。如果你的团队在后四分位，那你一定会被拉出来聊。这让 AI 从可有可无变成评估团队是否在释放潜力的一部分。

  3. 通过模仿进行发现。 排行榜不只是计分板，它还是地图。当你看到某个人在顶端，你会想知道他在做什么。你会去看他的 skills、工作流和应用。

不衡量就无法管理。 不把它做成可见的，你就在放弃最强的采用杠杆。

这也延伸到了招聘和绩效管理。现在，加入 Ramp 的任何人都必须熟练使用 AI 工具，毫无例外。对 PM 候选人，我们有专门的一轮面试，给我做个产品，展示你怎么做的，带我走一遍它怎么工作。必须是完整原型，不是幻灯片。你如果不能证明你把这些工具内化了，就过不了门槛。

8. Remove every constraint between your people and AI.

The number one way companies kill AI adoption is by treating it like a procurement decision. Budget approvals. IT reviews. Token limits. Connector requests that sit in a queue for weeks. Every one of these is a wall between your people and their "aha" moment.

We took the opposite approach. Three things we did early that mattered more than almost anything else:

  1. Treat AI usage as an infinite learning budget. If you demand ROI on every token before people have even learned to use the tools, you'll never get adoption. We gave people room to explore with the explicit expectation that the payoff comes from the compounding, not from day one.

  2. Kill token limits and access restrictions. No caps on usage. No tiered access based on role. No "you're not an engineer, you don't need this." Everyone gets the same tools, the same models, the same access. The people who surprised us most were the ones we would have never given access to under a traditional approval process.

  3. Remove every IT bottleneck on connectors. An AI agent is only as useful as what it can access. If your people have to file a ticket and wait two weeks for IT to approve a Salesforce connection or a Snowflake integration, they'll lose momentum and never come back. We pre-connected 30+ tools - Salesforce, Snowflake, Gong, Slack, Notion, Google Workspace, Figma - so that when someone opens Glass, everything is already live. One SSO authentication and they're working.

Here's the cost math that should reframe the conversation for any CFO: we pay our employees a lot of money. Token consumption per employee today isn't even close to double-digit percentages of their salary. But if someone is 2x more productive with AI, you should be willing to spend their entire salary again in tokens. If you have agents that can do 10x more work than a person, why would you not pay them twice as much as that person?

8. 清除人和 AI 之间的所有约束。

公司扼杀 AI 采用的第一方式，就是把它当成采购决策。预算审批。IT 评审。Token 限额。连接器请求排队好几周。这些都是把人和他们的 Aha 时刻隔开的墙。

我们做了相反的事。早期有三件事，重要性几乎压过一切：

  1. 把 AI 使用当成无限学习预算。 如果在大家还没学会用工具之前，就要求每个 token 都要有 ROI，你永远拿不到采用。我们给了大家探索空间，并明确预期，回报来自复利，而不是第一天。

  2. 干掉 token 限额和访问限制。 不设使用上限。不按角色分层访问。不搞你不是工程师 你不需要。每个人拿到同样的工具、同样的模型、同样的权限。最让我们惊讶的人，往往是传统审批流程下我们绝不会给权限的人。

  3. 移除连接器上的每一个 IT 瓶颈。 AI 代理有多有用，取决于它能访问什么。如果你的员工必须提工单，再等两周让 IT 批准 Salesforce 连接或 Snowflake 集成，他们会失去势头，再也不会回来。我们预连接了 30+ 个工具 Salesforce、Snowflake、Gong、Slack、Notion、Google Workspace、Figma，这样当人打开 Glass，所有东西都已经可用。一次 SSO 认证就开工。

下面这段成本算账，足以让任何 CFO 重新看待这件事，我们给员工付了很多钱。现在每个员工的 token 消耗，离他们工资的两位数百分比还差很远。但如果有人用 AI 生产力提升了 2 倍，你就该愿意再用 tokens 支付他一整份工资。如果你有能比一个人多做 10 倍工作的 agents，你为什么不愿意给它付出那个人两倍的成本？

Watch it compound.

We didn't start with a better strategy than most companies. We might have had better starting conditions: a culture that rewards speed and initiative, people who try things without waiting for permission, a leadership team that backs bold bets because we know it's good for our customers.

In lieu of a master plan, we just started. We kept building tools, kept raising the bar, kept investing in data and AI infrastructure, kept creating venues for people to show off. Each track compounded separately. As they reinforced each other, the curve went vertical.

We are in the very early innings of AI. Your job as a leader is to give your teams superpowers and make them believe in themselves. Everything else follows.

The most important lesson is the simplest one: just get started.

We're hiring builders across all functions. If you want to work somewhere that non-engineers ship production code, internal tools have a shelf life of weeks, and nobody thinks that's weird -- come build with us at Ramp.

*Shoutout to @bleviathan for the collaboration on this piece. *

看它复利式增长。

我们并不是一开始就比大多数公司有更好的策略。我们可能只是起跑条件更好，一种奖励速度和主动性的文化，一群人不等许可就去尝试的同事，一个支持大胆下注的领导团队，因为我们知道这对客户是好事。

没有总规划，我们就先开始。持续做工具，持续抬高标准，持续投资数据与 AI 基础设施，持续创造让人展示的场域。每条轨道各自复利增长。当它们彼此增强时，曲线就直冲天际。

AI 还处在非常早期的阶段。作为领导者，工作是给团队超能力，让他们相信自己。其他一切都会随之发生。

最重要的教训最简单，just get started.

We're hiring builders across all functions. If you want to work somewhere that non-engineers ship production code, internal tools have a shelf life of weeks, and nobody thinks that's weird -- come build with us at Ramp.

*Shoutout to @bleviathan for the collaboration on this piece. *

Most companies are still debating their AI strategy. They are overthinking it. Here is the playbook we followed to get everyone at the company building using AI.

Ramp's AI usage is up 6,300% from last year. 99.5% of the team active on AI tools. 84% using coding agents weekly. 1,500+ apps shipped on our internal platform in six weeks, from 800+ different builders. Non-engineers now account for 12% of all human-initiated PRs on the production codebase - thousands per month - using Ramp Inspect, our home-built coding agent.

We did this by being obsessed over every employee embracing this new technology - akin to how computers entered the workforce. We built our own Claude Cowork called Glass that gives anyone at the company a highly configured AI agent that is fully connected with Ramp's systems and aware of how we build. We hosted the largest AI hackathon ever - 700 participants across sellers, CX, legal, marketing, finance, coached by 100 of our most capable engineering and product teammates. They shipped more in a week than we previously could in a year.

We modified our hiring process and talent management process. We gave everyone unlimited budget to build, learn, and explore. We created leaderboards to incentivize usage. We reorged teams around those who see the future. We celebrated wins at all hands. We relentlessly pushed every single person and leader to build. And the result is far beyond what I could ever imagine.

Here's how we got here. The interesting part isn't the numbers or the tools. It's that we didn't have a plan. All we had was a culture and talent, and we kept doubling down on the things that were working with the people and technology in front of us. And watched it compound.

1. The second best time to start is today.

At Ramp, our culture is velocity. It shapes every process and team ritual. That culture turned out to be the single biggest accelerant for AI adoption.

At our January 2025 company kickoff, we told the whole company we would become the most productive company in the world. We believed we could do it given Ramp's culture. We had no idea how.

We started with the obvious:

  1. Leadership clarity that AI usage is an expectation

  2. Dedicated AI "guild" that were responsive to any questions

  3. Slack channels where teams can share what they built

  4. Dedicated all hands time to celebrate builders

  5. Mandated AI usage and tracking for everyone

There was no formal change management program. No mandatory training curriculum. Instead, we built the infrastructure for people to teach themselves and each other. The reality is that teams just need to be given a chance. Everyone wants to build. And with AI, anyone can.

2. Treat AI proficiency as a learning curve, not a light switch.

A year ago, most of us used AI the way everyone did. ChatGPT in a tab. AI search in Notion. Fine, but not transformational.

What we observed was that productive output leaps when people clear certain thresholds of comfort. Almost nobody outside of some exceptional engineers was operating at the upper levels before 2025. But in late 2025 and this year, we accelerated massively -- because we spent last year building a strong foundation first.

We think about AI proficiency in four levels:

  • L0: Sometimes uses ChatGPT. Has not changed any workflows. If you're still here and not self-starting, you will most likely not be at the company.

  • L1: Built custom GPTs, used Notion agents, dabbled in Claude Code. Starting to see what's possible but hasn't compounded it yet.

  • L2: Built an app that automates part of their job. Committed code or contributed feedback to others' work. This is where things get real.

  • L3: Systems builders. They don't just use AI - they build the infrastructure that levels up everyone else. These people are force multipliers.

Our job is to get everyone up the ladder. Three things make that possible:

  1. Build tools that meet people where they are. We started by shifting the whole company to Claude and Notion AI connected to all of our workplace tools - a low technical bar where everyone could participate and get meaningful benefit. That got people from L0 to L1.

  2. Raise expectations as tools mature. AI proficiency moved into hiring screens, onboarding, and how we talk about performance. Not as an end in itself, but as a stated expectation: getting good at these tools is essential to doing any job at Ramp well. That pushes L1s to L2.

  3. Match the mandate to the tooling. If you raise expectations before the tools can deliver, you burn credibility and people stop listening.

3. Embrace creative destruction.

This is the part that makes Ramp exhilarating and uncomfortable in equal measure.

Many of the tools we shipped in January 2026 are already obsolete - replaced by better versions, often from the same builders. We've gotten comfortable with a shelf life of weeks, not months. Every LLM update, every improvement to Claude Code or Codex harnesses, every new batch of skills we release reshapes what's possible. If your internal tools from three months ago still feel state-of-the-art, you're not moving boldly enough.

Our data democratization journey tells the story well:

  • Phase 1: Notion AI was the best option, so we piped important data into Notion databases to run our agent over it.

  • Phase 2: We launched Ramp Research, a Slack-based Snowflake research tool.

  • Phase 3: As coding agents matured, we encoded Snowflake research into skills those agents could use directly.

  • Phase 4: Now we're making data research interactive and self-improving.

Each generation opened doors the previous one couldn't. Each former generation was quietly sunset. The tools we're running right now? We genuinely hope they're obsolete by June.

From the outside, this looks chaotic. From the inside, it's the opposite. People aren't attached to their tools. They're attached to their problems. When a better way to solve the problem shows up, they grab it.

4. Build from the center, drive from the spokes.

We got the org design wrong before we got it right.

The initial instinct was to centralize: one small team builds tools for the whole company. Demand outstripped capacity almost immediately. Then we swung decentralized -- every team builds their own things. Tons of redundant re-learning.

The answer was to do both:

  1. A small central team builds the platforms, connectors, and plumbing across LLMs, data, knowledge, and workflows. They also manage training, enablement, and change management.

  2. Functional teams build on top of those platforms and give feedback that drives the central team's roadmap.

The results speak for themselves:

  • A risk analyst automated 16 hours per month of manual financial modeling.

  • A sales ops lead replaced a spreadsheet-based comp model across three orgs in 48 hours.

  • An L&D lead built a training simulator in 15 minutes.

  • Someone in finance built a contract reviewer that saves 45 minutes per contract -- and Ramp has a lot of contracts.

None of them are engineers.

These people didn't file a ticket. They found their own pain, prototyped a fix, and pulled engineering in when it was time to go to production -- when that was even necessary. The spokes drove the center as much as the center drove the spokes.

5. Give people a stage, not just a mandate.

Mandates decay. Culture is what remains.

The strategy, to the extent there was one, was to light as many small fires as possible and see which ones grew:

  1. A Slack channel (#ramp-uses-ai) - just to see what happened. Now over 1,000 members. Spun off 40+ team-specific channels that collectively generate 20,000 messages per month.

  2. AI office hours every Friday - regularly 40-50+ people show up with questions.

  3. AI onboarding for new hires - rebuilt four times in the last year as ambition grew.

  4. Dedicated all hands - we have everyone from our CEO to a first line operator demo what they have built with AI

The early converts mattered more than anything. On every team, there was one person - the ambitious sales ops lead, the frustrated product operator, the eager data scientist. They got curious, got sucked in, and became contagion for their teams. We made them visible: company All Hands spotlights, resources to build team-level tools, pairing them up when collaboration was warranted.

All this public building creates a competitive dynamic that everyone feels. Nobody wants to be the team that isn't building anything. When a CSM sees a risk analyst ship something that saves 16 hours a month, they don't think "good for Risk." They think "what can I build?"

That loop - build, share, inspire, build more - does more than any mandate or memo. **The biggest surprise wasn't who built the most. It was how many people had been waiting for permission to build at all **

6. Get people to the "Aha" moment as fast as possible.

Training doesn't work. Office hours and workshops help. But the world's best teacher is staring right in front of you: it's AI. You can only lead a horse to water. The single biggest unlock is getting someone to experience a real result on day one.

We learned this the hard way. Despite hitting 90%+ adoption of AI tools across the company, most people were stuck on a basic chat interface. The models were good enough. The harness wasn't. Terminal windows, npm installs, MCP configurations - these were simply too hard for the majority of people to grok. And the people who did push through had wildly different setups with siloed learnings that weren't compounding.

So we built Glass - our own version of Claude Code's Cowork, built on Anthropic's Claude Agent SDK.

Glass auto-configures on install. You authenticate once via Okta SSO and 30+ tools light up - Salesforce, Snowflake, Gong, Slack, Notion, Google Workspace, Figma. No setup guide. No ticket to IT. If the user has to debug, we've already lost.

A team of four built it in under three months. 700 daily active users within a month of launch. The people who got the most value weren't the ones who attended training sessions. They were the ones who installed a skill on day one and immediately got a result. The product taught them faster than we ever could.

When you own the tool, you see exactly where people get stuck and ship a fix the same day. Every session generates signal about how non-engineers actually learn to use AI - which skills get adopted, where people break through, what separates someone who uses it once a week from someone who uses it every day.

We also built a skills marketplace called Dojo where anyone can package a workflow and share it. Over 350 skills shared company-wide. A sales rep figures out the best way to analyze Gong calls and draft battlecards - packages it as a skill, and now every rep has that superpower. Every skill shared raises the floor for everyone.

**The result is that anyone in 5 minutes can create anything. **

7. Make it a competition.

People are competitive. At least this is true at Ramp.

We built an internal leaderboard that tracks AI usage across every team and individual at Ramp. Sessions run, skills used, apps shipped, tools connected. It's visible to everyone. The skeptics will tell you this is a vanity metric - that tracking usage incentivizes busywork, not productivity. We found the opposite.

The top AI users at Ramp are often the highest performers. AI proficiency is a skill like any other - the more reps you get, the better you become. The power users are developing muscle memory for when to reach for AI, how to prompt effectively, which skills to combine, and when to override. They're compounding their own leverage.

The leaderboard created three dynamics we didn't fully anticipate:

  1. Healthy peer pressure. Nobody wants to be at the bottom. When you can see that your peer on another team is running 3x more sessions and shipping tools that save their team hours, you don't need a mandate to start building. You need your competitive instincts.

  2. Manager accountability. Team-level rankings made it impossible for managers to ignore AI adoption. If your team is in the bottom quartile, that's a conversation you're going to have. It shifted AI from "nice to have" to "part of how we evaluate whether teams are operating at their potential."

  3. Discovery through emulation. The leaderboard isn't just a scoreboard - it's a map. When you see someone at the top, you want to know what they're doing. You look at their skills, their workflows, their apps.

If you're not measuring it, you're not managing it. And if you're not making it visible, you're leaving the most powerful adoption lever on the table.

This extends to hiring and performance management. We now have an absolute requirement for anyone joining Ramp to be proficient with AI tools. No exceptions. For PM candidates, there's a dedicated interview session: build me a product, show me how you built it, walk me through how it works. It's a full-blown prototype, not a slide deck. If you can't demonstrate that you've internalized these tools, you don't clear the bar.

8. Remove every constraint between your people and AI.

The number one way companies kill AI adoption is by treating it like a procurement decision. Budget approvals. IT reviews. Token limits. Connector requests that sit in a queue for weeks. Every one of these is a wall between your people and their "aha" moment.

We took the opposite approach. Three things we did early that mattered more than almost anything else:

  1. Treat AI usage as an infinite learning budget. If you demand ROI on every token before people have even learned to use the tools, you'll never get adoption. We gave people room to explore with the explicit expectation that the payoff comes from the compounding, not from day one.

  2. Kill token limits and access restrictions. No caps on usage. No tiered access based on role. No "you're not an engineer, you don't need this." Everyone gets the same tools, the same models, the same access. The people who surprised us most were the ones we would have never given access to under a traditional approval process.

  3. Remove every IT bottleneck on connectors. An AI agent is only as useful as what it can access. If your people have to file a ticket and wait two weeks for IT to approve a Salesforce connection or a Snowflake integration, they'll lose momentum and never come back. We pre-connected 30+ tools - Salesforce, Snowflake, Gong, Slack, Notion, Google Workspace, Figma - so that when someone opens Glass, everything is already live. One SSO authentication and they're working.

Here's the cost math that should reframe the conversation for any CFO: we pay our employees a lot of money. Token consumption per employee today isn't even close to double-digit percentages of their salary. But if someone is 2x more productive with AI, you should be willing to spend their entire salary again in tokens. If you have agents that can do 10x more work than a person, why would you not pay them twice as much as that person?

Watch it compound.

We didn't start with a better strategy than most companies. We might have had better starting conditions: a culture that rewards speed and initiative, people who try things without waiting for permission, a leadership team that backs bold bets because we know it's good for our customers.

In lieu of a master plan, we just started. We kept building tools, kept raising the bar, kept investing in data and AI infrastructure, kept creating venues for people to show off. Each track compounded separately. As they reinforced each other, the curve went vertical.

We are in the very early innings of AI. Your job as a leader is to give your teams superpowers and make them believe in themselves. Everything else follows.

The most important lesson is the simplest one: just get started.

We're hiring builders across all functions. If you want to work somewhere that non-engineers ship production code, internal tools have a shelf life of weeks, and nobody thinks that's weird -- come build with us at Ramp.

*Shoutout to @bleviathan for the collaboration on this piece. *

📋 讨论归档

讨论进行中…