AI is starting to "feed on" the chat records of bankrupt companies.

What is the most valuable asset after a startup collapses? In the past, the answer might be code, patents, customer lists, or the founding team. Today, it could be the chat history on Slack, an office collaboration software similar to DingTalk, Feishu, and WeCom.

Mercor, the AI "data factory" behind OpenAI and Anthropic, has recently begun seeking out defunct or acquired startups to purchase their internal datasets. These data include not only code repositories, but also Slack chat records, Google Drive cloud files, emails, and Zoom online meeting recordings.

At the end of June, AI agent startup Warmly announced that it would be acquired by HubSpot. Only 8 days later, Max Greenwald, founder and CEO of Warmly, received an email from Mercor in his mailbox, in which the company hoped to purchase or license Warmly's internal data including Slack chats, GitHub records, content on Asana (a project and work management software), Google Drive files, and employee meeting transcripts, with an offer of up to 300,000 US dollars.

In the following days, Warmly received multiple similar contacts from AI data companies including Mercor one after another, but Greenwald rejected all of them in the end.

It is not uncommon to see news that AI data companies purchase books and startup codes. However, as AI giants accelerate the development of agents that can perform tasks independently, the "appetite" of AI is constantly growing. Public content on the Internet, books and code are no longer enough to "feed" it.

Bobby Samuels, CEO of data company Protege, said that AI labs have basically "scraped the entire Internet", and what they need now is real human-to-human interactions.

Protege mainly helps enterprises judge whether their internal data meet the conditions for transaction and sale. Since the beginning of this year, the total value of data transactions processed by the company has reached at least 100 million US dollars, compared with about 30 million US dollars in the same period last year.

Why have the chat records of defunct companies suddenly become so valuable?

When AI begins to evolve from a "chatbot" to a real "digital employee", it not only needs to know "what the answer is", but also needs to know "how humans actually complete the work".

For example, how human employees will respond when a customer raises a question; how engineers will troubleshoot step by step when the IT system fails; how financial staff will handle problems with invoices; how sales staff update customer information in the CRM system. Even, where an employee will click and what to enter when the software suddenly reports an error.

These real work experiences are difficult to find on the public Internet.

For example, an IT work order may record that "an employee cannot log in to the website", then the discussion of engineers on the cause of the fault appears in the Slack chat records, the final solution is left in the email, and GitHub records a related code modification.

For humans, this is just an ordinary fault handling process.

But for AI agents, this may be a complete set of highly valuable "work tutorials", which is exactly what AI agent training needs most.

Of course, this business is far less simple than imagined. Relevant companies must properly handle issues such as privacy and intellectual property rights. For example, Slack may contain customer information, emails may have employee names, trade secrets may appear in meetings, and data from medical enterprises even involve patient privacy. All these data need to be anonymized before they can be sold or licensed to AI companies.

This article is from the WeChat official account "Cailian Press", author: Zhu Ling, published with authorization from 36Kr.