AI LABS MINING DEFUNCT STARTUPS FOR TRAINING DATA

Defunct startups are being liquidated for their operational data—a practice that transforms years of internal communications into what AI labs call "reinforcement learning gyms." The data includes Slack message histories, project management tickets from Jira, email threads, and other records of workplace activity. AI researchers use this material to train agents capable of performing business tasks autonomously, from project coordination to customer support. ■ Why This Matters Traditional AI training relies on public datasets or synthetically generated data. Real workplace archives offer something different: authentic patterns of human decision-making, communication style, and problem-solving within organizational contexts. A decade of a startup's Slack history provides millions of data points on how teams actually collaborate. The approach addresses a key challenge in AI development. Creating realistic simulations where agents can practice and improve requires massive amounts of contextual, structured data. Startup archives provide exactly that—complete operational records that show cause-and-effect relationships between actions and outcomes. ■ The Supply Chain When startups fail, their assets typically go to liquidators or investors. Previously, communication archives had minimal resale value. Now, AI labs are specifically acquiring these records, sometimes as part of broader asset purchases. The practice sits in a legal gray area. Data ownership varies by jurisdiction and company policy. Some startups may have kept backups that employees couldn't access; others explicitly retained communication data. Acquisition terms between liquidators and AI labs remain largely opaque. ■ What's Next As AI agents move from research projects toward commercial deployment, demand for high-quality training data will likely increase. This could create new market dynamics around startup liquidation, potentially changing what assets acquire value during business failures. The trend also raises questions about data privacy and consent. Workers who created these archives—many now at other companies—may not realize their professional communications are training machines to automate their former roles.