Here's something you don't see every day: bankruptcy proceedings becoming vendor sales pipelines. Defunct startups are now being liquidated not for their code, their patents, or their user bases—but for their Slack messages, Jira tickets, and email threads. AI labs are paying real money for this operational exhaust, treating the accumulated workflow debris of failed companies as premium training data.
The logic, once you think about it, is sound. These aren't just random text dumps—they're structured records of how people actually worked together. Decision threads, conflict resolutions, feature requests that never shipped, sprint retrospectives that led nowhere. It's the fossil record of organizational decision-making, and it's proving invaluable for training AI agents that need to operate in realistic workplace environments.
"We're building reinforcement learning gyms out of these archives. The patterns of how teams collaborate, how decisions get made, how priorities shift—it's all there."
Why Now?
Two forces are colliding here. First, the open internet is getting picked clean as a training source—there's only so much public data to go around, and quality matters more than quantity when you're trying to build reasoning systems. Second, AI labs are getting serious about agentic workflows, which means they need to understand how humans actually execute multi-step tasks in realistic settings.
The acquisition approach solves both problems. You're getting real-world organizational data (not synthetic or scraped) that's already structured in ways that map to the tasks you want your agents to learn. Someone's Jira tickets for a feature that never shipped? That's a perfect example of a partially-executed project with documented rationale, tradeoffs, and context—all useful signals for an AI trying to reason about software development workflows.
What This Means
There's an uncomfortable irony here: the companies that failed are now providing the training data that might help the next generation of AI companies succeed. But it's also a reminder that the digital exhaust we produce at work—the messages we send, the decisions we document, the back-and-forth that makes up most of our jobs—has value we never imagined.
The implications stretch further than you'd think. If AI labs are training on this data, they're implicitly learning the patterns of startup culture, of tech industry workflows, of how software teams actually operate. That's going to produce agents that are deeply shaped by the assumptions and practices of a very specific slice of organizational life. Whether that's a feature or a bug depends on where you sit.
The data graveyard is being mined, and the ore is surprisingly rich.