SaaS向Agent收费的陷阱：定价过高导致客户自建数据副本与流失
Almost Every Pre-AI Vendor We Use Is Raising Prices for Agent Access. They May Be Building an Agentic Death Spiral
为SaaS创始人提供了极具价值的定价避坑指南：揭示了按次计费若脱离成本基础将直接导致客户数据外流和护城河失效的机制，并给出了可执行的合理定价框架。
Almost every pre-AI B2B product we use has now told us they’re raising prices for agent access.
几乎我们使用的每个前AI时代的B2B产品现在都告诉我们，他们将为代理访问提高价格。
Salesforce and HubSpot are the two latest, and they’re doing it differently. HubSpot’s increases are mainly for their own agents: Breeze credits, per-resolution pricing, custom agents metering since July. Their MCP server is still free for the agents you bring. Salesforce is going the other way and metering third-party agents. Every successful call an agent makes through MCP or the API becomes a Flex Credit charge, agents have to be registered, and existing customers migrate to the new billing at renewal.
Salesforce和HubSpot是最新的两家，而且它们的做法各不相同。HubSpot的涨价主要针对其自有代理：Breeze积分、按解决次数定价、自7月起对自定义代理进行计量。其MCP服务器对于你引入的代理仍然免费。Salesforce则走向另一个方向，对第三方代理进行计量。代理通过MCP或API发起的每次成功调用都会产生Flex Credit费用，代理必须注册，现有客户将在续订时迁移到新的计费方式。
It isn’t just those two.
不仅仅是这两家。
A niche CRM we’ve used for five-plus years just told us we have to pay extra for agent use of their API. Another product we use deprecated the API 10K, our AI VP of Marketing, was running on. No replacement, no notice worth the name.
我们使用超过五年的一个细分CRM刚刚通知我们，我们必须为代理使用其API支付额外费用。另一个我们使用的产品废弃了API 10K——这是我们AI营销副总裁正在运行的基础。没有替代品，也没有值得称道的通知。
The pattern is consistent enough to say out loud: the apps that aren’t doing this are the ones that started agentic. If your pricing was built for agents, you don’t need to bolt a meter onto a seat model. If your pricing was built for humans clicking around in a UI, and the humans are logging in less, the meter has to move somewhere. It’s moving to the API call.
这一模式足够一致，可以直言不讳地说：那些尚未这样做的应用，是那些从一开始就面向代理构建的应用。如果你的定价模型是为代理设计的，就不需要在席位模型上附加计量功能。如果你的定价模型是为在UI中点击操作的人类设计的，而人类登录的频率降低了，那么计量就必须转移到别处。它正转向API调用。
{“model_id”: “unified-v1/prod/20260818-030536”}
{"model_id": "unified-v1/prod/20260818-030536"}
What actually happens when we get the email
当我们收到邮件时实际会发生什么
Here’s the part I don’t think the vendors have thought through.
这里是我认为供应商尚未深入思考的部分。
When a vendor tells us they’re charging for agent API access, the immediate reaction of our agents is: how can we work around this?
当供应商告诉我们他们将收取代理API访问费用时，我们的代理的第一反应是：我们如何绕过这一点？
Not cheat. Nothing shady. Literally move the data off the platform so there’s no API tax for the agents to pay. Sync the records to our own database, read from there, write back to the vendor only when something actually changes. Agents read far more than they write. Most of the calls a vendor would meter are lookups against data we already have a copy of.
不是作弊。没有任何不正当行为。字面意思上将数据移出平台，这样代理就不必支付API税费。将记录同步到我们自己的数据库，从中读取，仅在真正发生变化时才写回给供应商。代理的读取操作远多于写入操作。供应商会计量的大多数调用都是针对我们已经拥有副本的数据进行的查询。
That’s a weekend project now. Moving a database used to be very hard. Now it’s just annoying.
这现在成了一个周末项目。以前迁移数据库非常困难。现在只是令人烦恼。
And the humans react the same way, one step later. Nobody wants their agents running on apps that charge steeply for the privilege. When we pick a new tool, “how does this price agent access” is now a real question in the evaluation, and a bad answer is disqualifying.
而人类的反应也是如此，只是晚一步。没有人希望他们的代理运行在对这种特权收费高昂的应用上。当我们选择新工具时，“这个工具如何定价代理访问”现在已成为评估中的一个实际问题，而糟糕的答案会导致资格取消。
The death spiral
死亡螺旋
So here’s what I worry these vendors are inadvertently building.
所以，我担心这些供应商无意中正在构建的是以下情况。
They meter agent access. Usage drops, because we route around the meter. New agents get built against other systems from day one, or against our own copy of the data. Less data flows through the platform. Less of our work happens there.
它们对代理访问进行计量。使用量下降，因为我们绕过了计量机制。新代理从第一天起就针对其他系统构建，或针对我们自己的数据副本构建。流经平台的数据更少。在我们平台上完成的工作也更少。
And then the thing that made them sticky goes away. The moat for a system of record was never the software. It was that everything lived there and everything touched it. Price the touching, and less of it touches.
然后，使它们具有粘性的东西消失了。记录系统的护城河从来不是软件本身。而是因为所有事物都存在于那里并都与它交互。如果为这种交互定价，那么与它交互的事物就会减少。
Less usage. Less stickiness. Less of a moat. Then a renewal conversation where the vendor has a smaller footprint and we have more options than we did last year.
使用量减少。粘性降低。护城河变浅。然后是续约谈判，此时供应商的足迹更小，而我们拥有的选择比上一年更多。
What would actually work
真正可行的方案
The meter isn’t the problem. Agent traffic is real load and somebody pays for it. I don’t expect any of this for free, and agent identity, with scoped credentials per agent, is something I want regardless.
问题不在于计量。代理流量是真实的负载，并且有人为此付费。我不指望这一切都是免费的，而且我希望拥有按每个代理限定范围的凭据来标识代理身份。
The problem is the additive version: keep the seat, keep the storage premium, keep the API tiers, and add a per-call meter on top. Three bills for one piece of work.
问题在于叠加版本：保留席位、保留存储溢价、保留 API 层级，并在顶部添加按次计费的计量。同一项工作却收到三张账单。
The vendors who get this right will do one of three things. Publish the rate and cap it, so we can budget. Let a registered agent replace a seat instead of stacking on top of one. Or price the outcome rather than the call, the way some of these same vendors already price their own agents.
正确理解这一点的供应商将采取以下三种措施之一：公布费率并设定上限，以便我们进行预算；允许注册的代理替代一个席位，而不是在席位之上叠加；或者按结果而非按调用次数定价，就像这些供应商中的一些已经为其自身代理定价的方式一样。
The ones who leave it vague and uncapped will get exactly what they’re asking for: our agents reading from a copy, and writing back as little as possible.
那些保持模糊且无上限的供应商将获得他们应得的回报：我们的代理从副本中读取数据，并尽可能少地写回数据。
The honest truth: these API calls cost almost nothing
诚实的说法：这些 API 调用的成本几乎可以忽略不计
A record read from a database that’s already running is close to free. The rest of the infrastructure world prices it that way:
从正在运行的数据库中读取记录的成本接近于零。基础设施世界的其余部分也以此方式定价：
Look at the two Salesforce rows together. Salesforce already sells extra API capacity today, for integration traffic, at about $83 per million calls. The proposed agent meter lands somewhere between $5,000 and $100,000 per million. Same endpoint, same record, same write, 60x to 1,200x, decided by whether a human’s integration or an agent made the call.
将 Salesforce 的两行放在一起看。Salesforce 目前为集成流量出售额外的 API 容量，价格约为每百万次调用 83 美元。拟议的代理计量价格介于每百万次调用 5,000 美元到 100,000 美元之间。相同的端点、相同的记录、相同的写入操作，差距为 60 倍至 1,200 倍，这取决于调用是由人类集成还是由代理发起的。
Charging per call isn’t the problem. Firebase has charged $0.06 per 100,000 document reads for years and nobody calls it a tax. The question is the multiple, and what the same vendor charges for the identical call when a human’s integration makes it.
按调用次数收费并不是问题。Firebase 多年来一直按每 10 万次文档读取收取 0.06 美元的费用，但没有人称之为税。关键在于倍数，以及当人类集成发起相同调用时，同一供应商收取的费用是多少。
Atlassian is doing a version of this too, and arguably doing it better. Rovo credit usage explicitly includes calls made through the Teamwork Graph CLI and the Atlassian Rovo MCP server, and overage billing starts December 3, 2026, at $0.01 per credit. A basic action is 10 credits, so $0.10, the same rate as an Agentforce action. Two things make it land differently: every paid plan includes an allowance (25 credits per user per month on Standard, 70 on Premium, 150 on Enterprise, pooled org-wide), and reads don’t draw credits today. The allowance is thin, and extra usage is on by default unless an admin caps it. But Atlassian published the rate, published the date, and gave admins a switch. That’s the difference between a meter you can plan around and one you can’t.
Atlassian 也在做类似的事情，而且可以说做得更好。Rovo 的信用额度使用明确包括通过 Teamwork Graph CLI 和 Atlassian Rovo MCP 服务器发出的调用，超额计费从 2026 年 12 月 3 日开始，每信用点 0.01 美元。一个基本操作消耗 10 个信用点，即 0.10 美元，与 Agentforce 操作的费率相同。有两点使其体验不同：每个付费计划都包含一定的额度（标准版每位用户每月 25 个信用点，高级版 70 个，企业版 150 个，组织范围内共享），并且目前读取操作不消耗信用点。额度很有限，除非管理员进行限制，否则额外使用默认开启。但 Atlassian 公布了费率和日期，并给管理员提供了开关。这就是你可以据此规划的计量方式与你无法规划的方式之间的区别。
Storage tells the same story. Salesforce extra data storage runs about $125/month per 500MB, roughly $3,000 per GB per year. Neon charges about $0.35 per GB-month, roughly $4.20 per GB-year. Several hundred times, for the same bytes on the same kind of hardware.
存储情况也说明了同样的问题。Salesforce 的额外数据存储费用约为每月 500MB 125 美元，即每年每 GB 约 3,000 美元。Neon 的费用约为每 GB-月 0.35 美元，即每年每 GB 约 4.20 美元。对于相同字节、相同类型的硬件，价格相差数百倍。
None of this means don’t charge. Vendors can price on value, and an API call that returns a record worth something to us is worth more than a raw S3 read. Salesforce isn’t selling disk, it’s selling the data model, the permissions, and the ten years of history around that record.
这并不意味着不应该收费。供应商可以基于价值定价，对我们而言返回有价值记录的 API 调用，其价值高于原始的 S3 读取。Salesforce 出售的不是磁盘，而是数据模型、权限以及围绕该记录积累的十年历史。
What it does mean is that the gap between cost and price is so wide that our agents will see it immediately and work around it. When one lookup costs thousands of times more through the vendor than through a copy of the same data, the agent isn’t making a hard call. It’s making an obvious one.
真正意味着的是，成本与价格之间的差距如此之大，以至于我们的代理会立即察觉并绕过它。当一次查询通过供应商的成本是通过同一份数据副本成本的数千倍时，代理做出的不是艰难的决定，而是显而易见的决定。
We’ll Keep (Most of) Our Current Stack. But No New Apps That Aren’t Agent-Friendly
我们将保留（大部分）现有技术栈。但不再选择对代理不友好的新应用
We’ll stick with our current stack. Ten years of context lives in it, and that context is why our agents are any good.
我们将继续使用现有的技术栈。十年的上下文信息存在于其中，而这些上下文正是我们的代理表现出色的原因。
But we won’t pick another vendor that charges a significant tax on our agents. 10K alone does 30,000+ API calls a day. He won’t put up with it.
但我们不会选择另一个对我们的代理征收高额税费的供应商。仅 10K 每天就要执行 30,000+ 次 API 调用。他无法忍受这一点。
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