AI代理在执行任务方面的能力正在不断提升。但一家初创公司表示，这些数字助手实际上并不理解它们所服务的公司是如何运作的。如今，Zeroset已走出保密阶段，并筹集了520万美元资金，用于构建能够持续更新AI代理对公司运营状况的模型。Zeroset的联合创始人兼首席执行官Akshat Kannan表示，当前的AI模型对单个企业的内部运作知之甚少，因为这些信息可能分散在客户关系管理系统、工单系统以及内部文档中。
AI agents are getting better at completing tasks. But one startup says the digital assistants don't actually understand how the companies they're working for operate. Now, Zeroset has come out of stealth and raised $5.2 million to build models that give AI agents a constantly updated picture of a company's operations. Zeroset's cofounder and CEO, Akshat Kannan, said today's AI models have little knowledge of the inner workings of an individual business because that information can be scattered across customer relationship management systems, ticketing, and internal docs.
此外，AI模型通常也不会使用公司的专有数据进行训练。“现成的模型根本不了解企业实际运作的动态，”Kannan在接受《商业内幕》采访时说，“它们完全不清楚工作是如何完成的。” Zeroset的首款产品Nebula能够连接到公司已有的软件，包括Microsoft 365、SharePoint、Outlook、Teams、GitHub以及企业即时通讯工具。
Nor are AI models typically trained on a company's proprietary data. "Models off the shelf have no understanding of the dynamics of how an enterprise actually works," Kannan told Business Insider. "They have no understanding of how the work gets done." Zeroset's first product, Nebula, connects to software a company already uses, including Microsoft 365, SharePoint, Outlook, Teams, GitHub, and workplace messaging tools.
随后，它会追踪这些信息与活动在这些系统间随时间的变化。其理念是，让AI代理不仅仅拥有一堆可供搜索的公司文档，而是真正理解构成工作流程的事件与决策顺序，从而了解类似案例过去是如何处理的。这样，AI代理就能自主完成相应工作。 Zeroset由今年从斯坦福大学退学的18岁Kannan和去年从德克萨斯大学退学投身该初创公司的21岁William Zhang于今年1月创立。
It then tracks how information and activity across those systems change over time. The idea is to give AI agents more than a pile of company documents to search through. Instead, it wants them to understand the sequence of events and decisions that make up a workflow so that it knows how similar cases were handled previously. It can then do that work autonomously. Zeroset was founded in January by Kannan, 18, who dropped out of Stanford this year, and William Zhang, 21, who dropped out of the University of Texas last year to work on the startup.
该公司目前仍处于封闭研究预览阶段，并将逐步向客户扩大访问权限，而非一次性全面推出。公司计划将许可费用与按使用量计费的模式相结合，使费用随着通过Nebula处理的数据量和代理活动量而动态调整。
The company is still in a closed research preview and will progressively expand access to customers rather than do a single broad launch. The company plans to combine a license component with usage-based pricing that scales with the amount of data and agent activity running through Nebula.
Zeroset表示，其最初的重点是在制造业、供应链运营和金融研究等行业部署长期运行代理的企业和人工智能原生公司。它间接与Mem0和Zep等公司竞争，这些公司正在为人工智能代理构建内存基础设施。其520万美元的种子前期融资由Gradient和2048 Ventures共同领投，Leblon Capital也参与其中。Gradient是谷歌于2017年创立的一家风险投资基金，专注于种子和早期人工智能和机器学习初创公司。
Zeroset said its initial focus is on enterprises and AI-native companies that deploy long-running agents in industries like manufacturing, supply-chain operations, and financial research. It indirectly competes with companies like Mem0 and Zep, which are building memory infrastructure for AI agents. Its $5.2 million pre-seed round was co-led by Gradient and 2048 Ventures, with participation from Leblon Capital. Gradient is a venture capital fund founded by Google in 2017 that focuses on seed and early-stage AI and machine learning startups.
Zeroset拥有五名员工，计划利用其新资金雇用研究人员和系统工程师，资助早期企业部署并支付模型培训费用。Kannan表示，培训是其成本最高的项目之一，在一次高峰期的培训中，两周内花费了约10万美元。看看它用来确保现金注入的11个滑梯球场平台：
Zeroset has five employees and plans to use its fresh funding to hire researchers and systems engineers, fund early enterprise deployments, and pay for model training. Kannan said that training is one of its highest costs, and in one peak training run, it spent about $100,000 in two weeks. Check out the 11-slide pitch deck it used to secure the cash injection: