SaaStr 运行 Salesforce Headless 6 个月:AI
We’ve Been Running Salesforce Headless for 6 Months on Our Own “Claudeforce.” We’re Never Going Back
提供了一手企业级 CRM 从“人操作”转向“Agent 驱动”的真实落地路径,包含具体的架构思路、跨系统集成方法及对 SaaS 计费模式的深刻洞察,可直接用于评估自家产品的 API 开放度与 AI 集成策略。
For the past six-plus months, almost no one at SaaStr has logged into Salesforce.
过去六个多月里,SaaStr 几乎没有人登录过 Salesforce。
The data is all still there. We still pay for it. We still depend on it. We just stopped using the UI as the primary way to get at any of it.
数据依然都在。我们仍在为此付费。我们依然依赖它。我们只是不再将 UI 作为访问这些数据的主要方式。
Some context on why we did this. SaaStr runs today on a very small number of humans and 20+ production AI agents, with revenue up 47% year over year after being down 19%. At that ratio, the constraint stops being “is the CRM interface good.” Nobody has time to click through anyone’s interface. The constraint becomes whether your agents can reach the data.
以下是我们这样做的背景。SaaStr 目前由极少数人类和 20 多个生产环境 AI 代理运行,在收入下降 19% 之后,同比增长了 47%。按照这种比例,瓶颈不再是“CRM 界面好不好用”。没人有时间去点击浏览任何人的界面。瓶颈变成了你的代理能否获取数据。
What we actually did
我们实际做了什么
There’s no clever hack here. We used the Salesforce API to go headless, and we built an agent on top of that API using Claude. Internally we call it Claudeforce, and the agent itself we call 10K.
这里没有什么巧妙的黑客技巧。我们使用 Salesforce API 实现了无头化(headless),并基于该 API 使用 Claude 构建了一个代理。内部我们称之为 Claudeforce,而代理本身我们称之为 10K。
That’s it. The CRM became a system of record with an API in front of it, and the interface became whatever surface each person happens to prefer.
就是这样。CRM 变成了一个带有前置 API 的记录系统,而界面则变成了每个人各自偏好的任何表面形式。
Six months in, here’s what we’ve learned.
六个月过去了,这是我们学到的东西。
#1. We would never go back. No chance.
#1. 我们绝不会回头。没门。
It’s infinitely better, Saleforce headless run by an AI agent running on Claude (10K) that can work with 20+ other APIs in real time. Not “it’s been mostly good.” We would not return to a world where the way you interact with your CRM is by clicking through someone else’s opinion about how records should be laid out.
这要好得多,由运行在 Claude 上的 AI 代理(10K)驱动的无头 Salesforce,能够实时与 20 多个其他 API 协作。不是“大部分时候还不错”。我们不会回到那种通过点击他人对记录布局的意见来与 CRM 交互的世界。
Amelia and I now log into the classic Salesforce interface maybe twice a month, and usually only to check something obscure or trigger a marketing workflow. Our sales lead still uses classic Salesforce, but more and more our AI Agents take over that work from him. Both of those things are fine, which turns out to be the whole point (more on that below).
Amelia 和我现在可能每月只登录两次经典的 Salesforce 界面,通常只是为了检查一些 obscure 的东西或触发一个营销工作流。我们的销售主管仍然使用经典 Salesforce,但越来越多的工作被我们的 AI 代理接管给他做。这两件事都没问题,而这恰恰就是整个要点(详见下文)。
#2. We’re running 10+ of our 20+ agents directly on top of Headless Salesforce
#2. 我们在无头 Salesforce 之上直接运行了 20 多个代理中的 10 多个
10+ of our 20+ production agents now touch the CRM layer. Some we built. Several are third-party agents that plug into the same API. Once the CRM is headless, adding an agent stops being a project and starts being a config change. You’re not negotiating for UI real estate or waiting for a vendor to ship an integration. You point the thing at the API and it works.
我们 20 多个生产环境代理中的 10 多个现在都接触到了 CRM 层。有些是我们自己构建的。有几个是第三方代理,它们接入相同的 API。一旦 CRM 实现无头化,添加代理就不再是一个项目,而是一个配置更改。你不需要为 UI 空间进行谈判,也不需要等待供应商推出集成。你把东西指向 API,它就工作了。
The number keeps going up because the marginal cost of the eleventh agent is close to zero.
数量还在不断增加,因为第十一个代理的边际成本接近于零。
#3. It let us build a “meta CRM” that actually knows everything
#3. 它让我们构建了一个真正知晓一切的“元 CRM”
This is the part I underestimated going in.
这是我一开始低估的部分。
Our CRM data was never the problem. The problem was that our CRM data lived in one place, our marketing data in another, and our financial reality in Bill, Brex, and QuickBooks — three systems that had no idea the other two existed and certainly no idea who the customer was.
我们的 CRM 数据从来不是问题。问题在于,我们的 CRM 数据存储在一个地方,营销数据在另一个地方,而财务现实则在 Bill、Brex 和 QuickBooks 这三个系统中——这三个系统彼此互不相通,当然也不知道客户是谁。
Going headless let us stitch all of it together into one layer that sees everything. Not a dashboard that pulls from four sources on a refresh cycle. A layer that can answer a question that spans the CRM, the marketing stack, and the finance stack in a single pass, because it has API access to all of them at once.
采用无头(headless)架构让我们能够将所有数据整合到一个统一的层中,从而实现对全局的洞察。这不是一个在刷新周期内从四个来源拉取数据的仪表盘,而是一个能够一次性回答跨越 CRM、营销栈和财务栈的问题的层,因为它能同时通过 API 访问所有这些系统。
Ask it whether a given account has actually paid us, what they’ve opened in the last month, and what stage the renewal is in, and it just answers. That question used to require three people and a spreadsheet.
询问某个特定账户是否真的向我们付款、过去一个月打开了什么邮件、续约处于哪个阶段,它都能直接给出答案。以前提出这个问题需要三个人和一张电子表格。
#4. We built an agentic quote-to-cash agent
#4. 我们构建了一个智能报价到收款(quote-to-cash)代理
Contract to signature to collection, running on top of Salesforce.
从合同到签名再到收款,运行在 Salesforce 之上。
Quote-to-cash is the classic example of a workflow that was theoretically automated for a decade and practically still involved a human copying numbers between systems. When the CRM is headless and the finance stack is on the same layer, the handoffs stop being handoffs. The agent generates the contract, chases the signature, and follows the collection.
报价到收款是工作流的经典案例:理论上已自动化十年,但实际上仍涉及人工在不同系统间复制数字。当 CRM 是无头的,且财务栈位于同一层时,交接环节就不再是交接。该代理生成合同、追踪签名并跟进收款。
#5. Our AI VP of Revenue pushes everything into Slack, daily and in real time
#5. 我们的 AI 收入副总裁每天实时将所有内容推送到 Slack
I don’t ask for the update. It shows up.
我不再请求更新。它会主动出现。
Pipeline changes, closed deals, things that moved, things that didn’t. Real time, in the channel where I already am. The daily rhythm of a revenue org used to be a person compiling numbers and a leader reading them a day late. Now it’s a push.
管道变化、已关闭的交易、发生变动的情况以及未变动的情况。实时推送到我所在的频道。收入团队的日常节奏曾经是有人汇总数据,领导人在一天后才查看。现在则是即时推送。
#6. Data usage is up 10x. The bill is up 40%.
#6. 数据使用量增加了 10 倍。账单增加了 40%。
Our data usage against Salesforce has gone up roughly 10x. We expect it to hit 100x. Our bill already went up about 40%. That’s OK, but it’s approaching our limit. If it goes up much more, we will stop storing as much data in Salesforce. Our AI Agents just consume and create so much more data than humans.
我们与 Salesforce 之间的数据使用量大约增加了 10 倍。我们预计将达到 100 倍。我们的账单已经增加了约 40%。这没问题,但正在接近我们的上限。如果增加得更多,我们将停止在 Salesforce 中存储大量数据。我们的 AI 代理比人类消耗和创建的数据要多得多。
Read that again if you’re a vendor. When we made the platform genuinely open and API-first, we consumed dramatically more of it and we paid meaningfully more for it. The seat count didn’t drive that. Consumption did. Agents don’t check the CRM twice a day the way a rep does — they hit it constantly, because there’s no cost to them in doing so and enormous value in being current.
如果你是供应商,请再读一遍。当我们使平台真正开放并以 API 优先时,我们对它的消耗大幅增加了,为此支付的金额也显著上升。驱动这一点的不是席位数量,而是实际消耗量。代理不像销售代表那样每天只检查两次 CRM——它们会持续不断地访问,因为这样做对它们没有成本,而且保持数据最新具有巨大的价值。
This is the same argument the Okta Enterprise AI Index makes from the outside. Account growth is a bad proxy for what’s actually happening with AI in the enterprise, because accounts get provisioned and then sit there. Usage is the number that moves. Our Salesforce line is that finding from the inside: our account count didn’t change at all, and our consumption went up 10x.
Okta《企业AI指数》报告正是从外部视角提出了同样的论点。账户增长是衡量企业AI实际发展情况的糟糕代理指标,因为账户被配置后往往就闲置在那里。使用量才是真正变化的数字。我们Salesforce的立场是从内部得出的这一发现:我们的账户数量完全没有变化,而我们的消耗量增长了10倍。
The vendors terrified that agents will collapse seat revenue are looking at the wrong line. The risk isn’t that agents reduce what you consume. It’s that a closed platform makes agents impossible, and someone else’s open platform becomes the substrate everything gets built on.
那些因智能体(agents)将导致席位收入崩溃而感到恐慌的供应商看错了方向。风险不在于智能体会减少你的消耗量,而在于封闭平台会让智能体无法实现,从而让别人的开放平台成为构建一切的基础设施。
The real lesson: everyone wants a different surface
真正的教训是:每个人都想要不同的交互界面
This is what six months taught me that I didn’t expect.
这是我过去六个月学到的、未曾预料到的经验。
Our sales lead uses classic Salesforce, because when you’re working deals all day the record view is genuinely the right tool. I check Slack, because I want the state of the business to find me. Amelia talks to 10K for an hour every morning, working through questions conversationally in a way no dashboard has ever supported. I ask it things when I happen to have a question, which is a completely different usage pattern from hers.
我们的销售负责人使用经典的Salesforce界面,因为当你整天处理交易时,记录视图确实是正确的工具。我查看Slack,因为我希望业务状态能主动找到我。Amelia每天早上花一小时与10K对话,以仪表板从未支持过的对话方式逐一解答问题。当我碰巧有疑问时,我会向它提问,这与她的使用模式完全不同。
Four people. Four surfaces. Same underlying data. Nobody is wrong.
四个人。四种界面。相同的底层数据。没有人是错的。
The old model assumed one canonical interface that everyone learned and adapted to, and the vendor’s job was to make that interface as good as possible. That assumption is finished. There is no single surface that wins, because the right surface depends on the job, the person, and the hour of the day.
旧模型假设存在一个所有人都学习并适应的标准界面,而供应商的职责是让该界面尽可能完善。这种假设已经终结。没有哪一种单一界面能通吃,因为合适的界面取决于任务、人员以及一天中的具体时段。
Support every surface. Throttle it, overcharge for it, and ultimately you lose. Even if it helps make the quarter
支持所有界面。对其进行限制、对其收取过高费用,最终你会失去客户。即使这有助于完成季度目标
Our #1 learning: support every surface, and mean it. If Salesforce hadn’t, we’d be gone.
我们最重要的经验教训是:支持所有界面,并且要真心实意。如果Salesforce当初没有这样做,我们就已经离开了。
Your UI is one surface among many now, and for a growing share of your users it won’t be the primary one. That’s not a threat to your product. Your product is the data model, the workflow logic, and the system of record. The UI was always just one way to reach it.
你的用户界面现在只是众多界面中的一种,对于越来越多的用户来说,它将不再是主要界面。这对你的产品并非威胁。你的产品是数据模型、工作流逻辑和记录系统。用户界面始终只是访问这些内容的一种方式。
The platforms that win the next decade will be the ones where a customer can go headless without asking permission, build whatever agent they want on top, and consume 10x more as a result. The ones that fight it will find their customers building the meta layer somewhere else — and then discovering that the CRM underneath is replaceable.
赢得下一个十年的平台,将是那些允许客户在不请求许可的情况下实现无头化(headless),在其之上构建任何他们想要的智能体,并因此将消耗量提升10倍的平台。而那些试图阻挠的平台会发现,它们的客户会在其他地方构建元层(meta layer)——然后发现底层的CRM是可以被替换的。
We went headless six months ago. Our usage is up 10x, our spend is up 40%, and we’re not going back.
我们在六个月前实现了无头化。我们的使用量增长了10倍,支出增加了40%,而且我们不会回头。
更进一步:量化金融体系
看懂新闻只是起点——沿量化金融路径,把它变成能交付的工程能力