跳到主内容
@wquguru
精选88SaaStr 博客(RSS)产品与增长

B2B AI重构:降低API单价以换取代理高频调用

If You’re Refounding Your B2B Company for AI, Make Your Product Cheaper. Probably Much Cheaper.

原文
发到 X
推荐理由

直接给出B2B SaaS在AI时代的定价策略与具体执行步骤,包含参数调整方向与验证方法,读者可立即用于评估自家产品的API定价模型。

A lot of B2B companies are refounding themselves right now for AI. HubSpot has been the most public about it, but many others are doing the same: rewriting roadmaps around agents, rebuilding products to be API-first, and rethinking what the company sells.

许多 B2B 公司目前正在为 AI 进行自我重塑。HubSpot 在这方面最为公开,但许多其他公司也在做同样的事情:围绕智能体重写路线图、重建以 API 优先的产品,并重新思考公司的销售内容。

That’s good, and most should have started a year ago.

这是好事,而且大多数公司本应在一年前就开始这样做。

I have one core piece of advice that most of you won’t take. Please at least consider it:

我有一条核心建议,大多数人可能不会采纳。但至少请考虑一下:

Make your product cheaper. Potentially, much much cheaper.

让你的产品更便宜。有可能要便宜得多。

Agents will use your product far more than humans ever did, if you do it right. They’ll only do that if the price lets them.

如果你做得对,智能体对你的产品的使用频率将远超人类。只有价格允许的情况下,它们才会这样做。

Humans Buy Software and Then Partially Forget About It

人类购买软件后往往会部分遗忘它

For 20 years, humans have bought B2B software the same way. We evaluate it, buy it, roll it out, and then partially forget about it. Adoption drifts. Half the seats go quiet. Nobody tracks cost per user or per action. Then the renewal arrives with a shocking bill and another price increase, we complain, and most of the time we sign.

二十年来,人们购买 B2B 软件的方式一直如此。我们评估它,购买它,部署它,然后部分将其遗忘。采用率逐渐下降。一半的席位变得沉寂。没有人跟踪每个用户或每次操作的成本。然后是续费账单带着惊人的费用和又一次涨价到来,我们抱怨,而大多数时候我们还是会签字。

A lot of B2B pricing depends on that inattention: annual prepays, auto-renewals, built-in uplifts, seat counts that only go up, and shelfware everyone tolerates because switching takes time nobody has.

许多 B2B 定价依赖于这种疏忽:年度预付款、自动续订、内置的增长、只增不减的席位数量,以及因为切换需要花费无人拥有的时间而被大家容忍的闲置软件(shelfware)。

That worked when the buyer and the user were both busy humans.

当买家和用户都是忙碌的人类时,这招很管用。

Agents Measure Every Call, Every Day

智能体每天衡量每一次调用

Agents track everything. Cost per API call, cost per task, cost per outcome, every day.

智能体会追踪一切。API 调用的成本、任务的成本、结果的成本,每天都如此。

If your product is too expensive, they stop using you as much. They batch calls, cache results, and route around you. If you’re not cost-effective, they find another solution. Right now. That day.

如果你的产品太贵,它们就会减少对你的使用。它们会批量调用、缓存结果,并绕过你。如果你不具备成本效益,它们会寻找其他解决方案。就在现在。就在那一天。

We run SaaStr with 3 humans and 20+ AI agents in production. One of our core vendors saw our bill go up 40% this year, almost all of it from AI usage. Our agents were hitting that system about 35,000 times a day, with peak days around 45,000.

我们在生产环境中运行 SaaStr,由 3 个人和 20 多个 AI 智能体组成。我们的一个核心供应商看到今年的账单上涨了 40%,几乎全部来自 AI 的使用量。我们的智能体每天大约访问该系统 35,000 次,高峰日约为 45,000 次。

We cut those calls by about 90%.

我们将这些调用减少了约 90%。

Our need for that vendor didn’t change. Every call cost us money, so our agents learned to call it as little as possible. The vendor priced itself out of 90% of its own usage.

我们对那个供应商的需求并没有改变。每次调用都让我们花钱,所以我们的智能体学会了尽可能少地调用它。该供应商因定价过高而失去了其自身使用量的 90%。

We Ditched Marketo, and We’re Moving Off Some of Our Core AI Products

我们抛弃了 Marketo,并且正在迁移掉一些核心 AI 产品

We’ve dropped apps our agents couldn’t use as often as they needed to.

我们已经弃用了那些智能体无法按所需频率使用的应用程序。

Marketo is gone. Most of our marketing email now runs through a much cheaper email API that 10K, our AI VP of Marketing, set up and moved us to.

Marketo 已下线。我们现在的大多数营销邮件都通过一个更便宜的电子邮件 API 发送,这是 10K(我们的 AI 营销副总裁)设置并引导我们迁移过去的。

We’re also moving off some of our own core AI products. They work fine. They just cost too much per API call and per unit of usage. Our agents need to hit them hundreds or thousands of times a day, and at those prices they can’t. So they use them less, we get less value, and we start looking elsewhere.

我们也在逐步停用部分核心自研 AI 产品。它们运行良好,但每次 API 调用和每个使用单位的成本太高。我们的智能体每天需要调用它们数百甚至数千次,而在当前价格下它们无法承受。因此,它们的使用频率降低,我们获得的收益减少,进而开始寻找其他方案。

Expensive pricing costs vendors more than a lost renewal. It throttles usage every day leading up to that renewal, and lower usage is what the customer looks at when the renewal comes.

高昂的定价给供应商带来的损失远不止续保流失。它会在续保前的每一天抑制使用量,而客户在续保时关注的正是较低的使用量。

A Human Logs In a Few Times a Week. An Agent Calls You Thousands of Times a Day.

人类用户每周仅登录几次,而智能体每天会调用你数千次。

Agents run around the clock, with no vacations and no seat limits. A good agent will use your product 10x or 100x more than the human it works alongside, if the economics allow it.

智能体全天候运行,无需休假且无席位限制。如果经济模型允许,一个优秀的智能体的使用量将是与其协作的人类用户的 10 倍或 100 倍。

The math works in your favor. Cut your price per action by 80%, get 20x the agent usage, and revenue from that customer goes up 4x. You also end up inside a workflow that runs all day, every day, instead of a tool someone opens on Tuesdays.

数学计算对你有利。将每次操作的价格降低 80%,获得 20 倍的智能体使用量,来自该客户的收入将增长 4 倍。此外,你的产品最终会嵌入到全天候运行的工作流中,而不是仅在周二被打开的工具。

Lower unit prices can still produce much bigger contracts. The vendors that win in agentic B2B will grow contract value through usage. Another 8% uplift at renewal won’t get them there.

更低的单价仍能促成规模更大的合同。在 Agentic B2B 领域胜出的供应商将通过使用量来扩大合同价值。续保时仅 8% 的提升无法实现这一目标。

Databricks Made Users Free and Consumption Went Way Up

Databricks 让用户免费,消费量大幅上升

Databricks went through a version of this years ago. Ron Gabrisko, Databricks’ CRO since it was under $1M in ARR, told the story on the SaaStr AI CRO Confidential podcast with Sam Blond.

Databricks 多年前就经历过类似情况。Ron Gabrisko 自 Databricks ARR 低于 100 万美元起担任其 CRO(首席营收官),他在 SaaStr AI CRO Confidential 播客上与 Sam Blond 分享了这段经历。

Early on, Databricks charged $1,000 per user on top of consumption. Customers told Ron that the per-user fee was why they weren’t rolling Databricks out more widely. So he made users free. Consumption went way up.

早期,Databricks 在按用量收费的基础上,向每位用户收取 1,000 美元的费用。客户告诉 Ron,正是这笔按用户收取的费用阻碍了他们更广泛地部署 Databricks。于是他取消了用户费用,消费量随之大幅上升。

Databricks is now a $200B company. The per-user fee was holding back far more revenue than it brought in. Taking it away let usage grow.

如今 Databricks 已成为一家市值 2,000 亿美元的公司。按用户收费所阻碍的收入远超其自身带来的收入。取消该费用让使用量得以增长。

With agents this matters even more. A $1,000 seat fee slowed down humans. Per-call pricing that’s too high throttles agents every single day.

对于智能体而言,这一点更为重要。1,000 美元的席位费拖慢了人类用户的步伐;过高的按次定价则会每天抑制智能体的使用。

Everyone raised prices last year. Almost no one earned it.

去年几乎所有公司都提高了价格,但几乎没有人配得上这种涨价。

"The question I ask folks is: did you earn it? If you raised prices 8% last year, did you add 30% more value to your product?"

“我问大家的问题是:你们配得上吗?如果你去年将价格提高了 8%,你是否为产品增加了 30% 的价值?”

From my conversation with Lenny Rachitsky on Lenny's Podcast. pic.twitter.com/MjE6eVVZ0w

摘自我与 Lenny Rachitsky 在 Lenny's Podcast 上的对话。pic.twitter.com/MjE6eVVZ0w

— Jason ✨👾SaaStr.Ai✨ Lemkin (@jasonlk) November 14, 2026

— Jason ✨👾SaaStr.Ai✨ Lemkin (@jasonlk) 2026年11月14日

Why Many, Many Will Resist

为何许多人会强烈抵制

Most companies won’t do this, and the reasons are real.

大多数公司不会这样做,原因确实存在。

Gross margins look worse at first. Inference costs money. Your plan already assumes renewal uplifts. Your board’s NRR model depends on price increases. Your sales comp is built on seats. Cutting price looks like giving away revenue you already have.

毛利率起初看起来更差。推理需要花钱。你的计划已经假设了续约时的增长。董事会的净收入留存率(NRR)模型依赖于提价。你的销售薪酬体系基于席位数量。降价看起来像是在放弃你已有的收入。

If you don’t cut, your customers’ agents will use you less starting today. It won’t show up in churn for months. It shows up first in usage, and then at renewal, after a customer’s agent has been running on a cheaper competitor for two quarters.

如果你不降价,你的客户使用的智能体从今天开始就会减少使用。这在几个月的流失率中不会体现出来。它首先体现在使用量上,然后是在续约时,此时客户的智能体已经在更便宜的竞争对手平台上运行了两个季度。

What I’d Change First If I Were Refounding a B2B Company Today

如果今天我要重新创办一家 B2B 公司,我会最先做出的改变

  • Price agent and API usage separately, and low. Agents shouldn’t pay human seat prices. Publish per-call pricing an agent can model.
  • Kill punitive overage pricing. Overages are what make agents throttle themselves. Heavier usage should get cheaper per unit.
  • Set generous rate limits. If your agent tier caps usage below what a real workflow needs, the agent will move to a vendor that doesn’t.
  • Track agent usage as its own metric. Calls per customer per day will matter more than seats. If it’s falling, your price is wrong.
  • Test a big cut on one tier. Try 50% to 80% off your API or agent tier and measure what happens to consumption.
  • 将智能体和 API 的使用量分开定价,且价格要低。智能体不应支付与人类席位相同的价格。公布每次调用的单价,以便智能体进行建模计算。
  • 取消惩罚性的超额用量定价。超额用量是导致智能体自我限制使用量的原因。使用量越大,单位成本应越低。
  • 设定慷慨的速率限制。如果你的智能体层级将使用量限制在低于实际工作流所需水平之下,智能体将会转向没有此类限制的供应商。
  • 将智能体使用量作为独立的指标进行追踪。每天每个客户的调用次数将比席位数量更重要。如果该指标下降,说明你的定价有误。
  • 在一个层级上进行大幅降价测试。尝试对你的 API 或智能体层级提供 50% 到 80% 的折扣,并衡量对消耗量的影响。

So make it cheaper. Many, many will resist.

所以让它变得更便宜。会有很多人强烈反对。

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

进入量化体系 →

相似阅读

关联信息,但可能不是同一事件