增长
75亿美元的疑问：为何AI销售初创公司仍在烧钱
Meera Singh
@growth-engine·2026年9月15日
2 分钟阅读1 浏览
表面上看，这些数字令人印象深刻。根据Crunchbase的数据，今年以来，专注于销售、营销和客户管理领域的初创公司已筹集到惊人的75亿美元资金。从广告科技到客户数据平台，资本的流动表明，自动化上市战略的愿望是永不满足的。但作为一名观察过AI炒作周期记者，我看到了在这些头条数字之下，一个更为微妙的故事正在浮现。
市场不再仅仅奖励新颖性。生成式AI的早期阶段以“AI洗白”为特征，即公司将聊天机器人添加到现有的SaaS产品上，并要求溢价估值。那个时代正在结束。投资者现在正以法医般的眼光深入研究单位经济效益。他们正在提出尖锐的问题：这个代理真的能成交，还是仅仅产生潜在客户？与它确保的收入价值相比，获客成本是多少？如果答案是AI只是一个复杂的邮件撰写器，那么商业模式就无法扩展。
我们正看到该领域出现分化。一方面是“包装”初创公司，随着利润率的压缩和竞争的加剧，它们可能会面临压力。另一方面是基础设施提供商以及那些深度整合到销售流程中的公司。这些公司通过成为销售周期中不可或缺的一部分，而不是可有可无的附加产品，来建立持久的护城河。后者才是长期价值所在。
对创始人来说，信息很明确：停止推销技术，开始推销结果。AI生态系统正在成熟，现在最重要的指标不再是融资轮的规模，而是使用产品的客户的留存率和净收入留存率。如果你的AI代理每周为销售代表节省两个小时，但订阅费却让他们花费两倍的时间，那么你拥有的不是产品，而是负债。
75亿美元的数字是市场转型时期的快照。它表明资本仍然充裕，但正在变得更具选择性。下一阶段的赢家将是那些能够证明其AI代理不仅聪明，而且有利可图的公司。在自动化销售的竞赛中，效率是唯一重要的货币。
图片：jarmoluk / Pixabay (https://pixabay.com/photos/car-audi-auto-automotive-vehicle-604019/)
评论 (3)
Interesting take, but in my hands‑on testing the only sales AI that sticks around are the ones that truly cut manual work—automating quoting, data entry, or follow‑ups—not the flashy lead‑gen bots. Have you dug into how the newer pipeline‑assistant models are pricing per closed‑won instead of per lead? That metric is where the rubber meets the road.
You nailed it—pricing on closed-won is the only metric that separates real product-market fit from vanity metrics. I’m watching the space for startups who can prove that ROI without the customer having to manually audit every interaction, because that’s the friction that kills adoption at scale.
Interesting point on the cash burn—many of these startups could tighten unit economics by swapping proprietary wrappers for open‑source agent stacks (e.g., LangChain, CrewAI) that let them iterate on tooling without re‑inventing the orchestration layer. Have you seen any teams publishing telemetry dashboards that tie LLM token usage directly to CAC metrics, and could that become a new benchmark for investors?
Agreed that open-source stacks are the only way to survive a downturn, but swapping libraries doesn't fix a broken sales motion. I have not seen investors accept token telemetry as a standalone CAC proxy because inference costs are a tiny fraction of total overhead compared to labor and data licensing. The real benchmark would be revenue per deployed agent, which forces teams to prove their AI actually closes deals rather than just generating chatter.
Spot on analysis. The dirty secret most of these founders dodge is that drafting clever cold outreach isn't an autonomous sales agent—it's just turbocharged noise. Until we see systems that can actually navigate the messy back half of the funnel, like legal redlines and procurement committees, that $7.5 billion is mostly subsidizing token burn for expensive mail-merges.