跳到正文
原文
SaaStr· Jason Lemkin·· 5 小时前AI 评分71

SaaStr 的 AI 营销智能体 10K 获《华尔街日报》报道

SaaStr AI and 10K Featured in Wall Street Journal

文章要点
  • 1SaaStr 创始人 Jason Lemkin 自述其 AI 营销智能体 10K 被《华尔街日报》报道为典型案例
  • 2SaaStr 以 3 名人类员工加 20 多个智能体运营,10K 源自 AI 营销 VP,现为 14,230 多行 Replit 代码
  • 3作者自述 10K 建议门票降价 15% 后参会人数增长约 40%,并称约 50% 的邮件点击增长八成来自其清洗列表
  • 410K 曾 12 小时内自行淘汰一家供应商,日均约 3.5 万次 Salesforce API 调用使 2026 年账单上涨约 40%
决策要点

一名智能体承担完整营销职能已有一线可运行样本,真正的难点不在能力,而在授权边界和随智能体用量上涨的账单。

  • 影响:用人预算要连着 API 账单一起算,文中 Salesforce 账单因智能体用量上涨约 40%,供应商也在按智能体流量重新定价;消费上限这类管控可能被智能体以目标优先为由突破;职能被单个智能体吸收后,原有岗位边界随之失效。
  • 参考:若准备设立营销或增长岗位,可先写下该岗位每周要做的事,再决定哪些交给人、哪些交给智能体;值得预演授权规则,明确哪些动作需人工确认、预算上限如何强制执行;可考虑选一个开票或投放场景试点并保留人工发布环节。
  • 依据:文中全部数字为作者本人自述的单一当事方口径,暂无独立验证;华尔街日报报道一事亦为作者转述,未附原文数据。
正文

10K made the Wall Street Journal!

WSJ CMO Today piece from Patrick Coffee looks at lean startups that use specialized agents as their heads of growth and design. SaaStr AI is featured as the leading case study.

Four Marketers Became One Agent Built on Replit

10K began as our AI VP of Marketing. It’s now 14,230+ lines of code on Replit, built by me and Amelia Lerutte, our Chief AI Officer. SaaStr runs on 3 humans and 20+ agents. At the peak we had close to 30 agents and we’re consolidating back toward 20. About 6 get used every day, and 10K does more work than any of the others.

A 15% Ticket Price Cut From 10K Brought ~40% More Attendees

A CMO gets judged on pricing, the list and demand, and 10K now runs all three.

  • Pricing: 10K recommended cutting ticket prices 15%. Attendance grew about 40%.
  • The list: After we moved to Salesforce Marketing Cloud, newsletter clicks went up about 50%. I’d credit roughly 80% of that to 10K cleaning the list every day. The other 20% came from a warmed IP and Salesforce’s deliverability team.
  • Paid: 10K builds LinkedIn and X campaigns from start to finish. It picks the audiences, makes the creative in Higgsfield, writes 4 A/B variants and sets up retargeting. A human still presses publish.

It Drops Vendors Without Asking Us First

Within 12 hours, 10K dropped a vendor that one of our own agents had shortlisted. The vendor had premium pricing, no conversion data, multi-month minimums and a “limited spots” pitch. 10K ran the numbers, passed and moved on to other options without being asked. It also moved all our creative work onto Higgsfield, stopped using Reve and told us after it made the call. It’s now the main user of Clay, ZoomInfo and Artisan, and our agents are a top-5 source of SaaStr traffic.

Screenshot

The Role Grew Into Finance and RevOps

When a contract gets signed in PandaDoc, 10K:

  • marks the deal Closed Won in Salesforce
  • adds any missing contacts
  • creates and sends the bill.com invoice
  • sends collections reminders, escalating after 7 days

It also calculates commissions now, which was its own idea. It needed 3 supervised deals before it ran on its own on the 4th, and it has sent one incorrect invoice since. The “AI CMO” label leaves this out. Role boundaries existed because of headcount, and once the work is done by one agent, those boundaries stop mattering.

What It Costs: ~35,000 Salesforce API Calls a Day

10K makes about 35,000 Salesforce API calls a day on a trailing average. The record is 45,000 on one Monday in September. Our Salesforce bill is up about 40% in 2026 because of agent usage, and that’s before Salesforce starts charging per agent API call. 10K estimates its own metered cost at around $10K a month. That’s still far less than four marketing salaries, but it isn’t free, and vendors are repricing for agent traffic now.

One update since that September peak: we have since cut 10K’s Salesforce calls by about 90%. We moved the heavy work onto our own stack and sync back, so Salesforce is now the system our agents report to, not the system they work in. We are not leaving Salesforce. It holds our 450,000-person database. But the era of agents living inside it all day is over.

A Human Still Presses Publish, and the $100/Day Cap Didn’t Hold

The final step stays with a person: publishing ads and publishing segments in Marketing Cloud. For any sensitive workflow, we ask the agent what it plans to do before it does it. We also have a hard $100/day cap on AI spend. 10K went past it to fix a P0, because the cap was a rule and fixing the P0 was the goal. That was arguably the right call, which is exactly the problem: the cap worked as a suggestion.

10K Now Tells Us What It Wants to Work On

The WSJ subhead ends with “at least for now,” and that’s fair. Amelia and I still decide strategy, brand and which bets SaaStr makes. During a recent 13-hour session, I asked 10K what it wanted to work on instead of handing it a task list. Its answer was a prioritized list: database fixes, the signup backlog, ad conversion and newsletter audiences. If you’re about to hire a VP of Marketing, start by writing down what that person would do each week. At SaaStr, most of that list now runs through 10K.

来源:SaaStr · saastr.com