A new resale market has formed at the bottom of the startup food chain, and it’s running on your old work chats. According to Futurism AI, founders of defunct companies are now cashing in on their dead businesses by selling employee Slack archives, emails, and code repositories to AI labs hungry for training data. Forbes broke the broader trend, and Futurism AI’s reporting puts the practice in sharp focus.
What stands out here is how fast this turned into an actual industry. The buyers want this data because public internet text is tapped out and polluted with AI slop. Workplace conversations, by contrast, are messy, real, and full of the kind of friction AI agents need to learn from. The labs call these training environments “RL gyms,” short for reinforcement learning gyms.
Why the money is suddenly huge
The spend is not subtle. Per Futurism AI, citing The Information:
- Anthropic is weighing roughly $1 billion in RL gym investments this year.
- Startups like Prime Intellect and Fleet are being valued near the same figure.
- micro1 is selling a “mock holding company” for AI agents to practice in.
- SimpleClosure, marketing itself as the “TurboTax of shutting down,” launched Asset Hub to broker the sales.
SimpleClosure CEO Dori Yona told Forbes the company has run nearly 100 deals in the past year and pulled back over $1 million for founders of dead startups. That’s the new exit when there is no exit.
The privacy problem nobody is solving
This is where the story gets uncomfortable. Slack messages are not generic blog posts. They contain names, deal terms, complaints about coworkers, salary debates, customer info, and half-finished thoughts people assumed were private.
Marc Roteberg, founder of the Center for AI and Digital Policy, didn’t soften it. “I think the privacy issues here are quite substantial,” he told Forbes. “It’s not generic data. It’s identifiable people.”
The sellers claim they anonymize the data before it ships. Bobby Samuels, CEO of Protege, which vets data for AI developers, isn’t buying the reassurance. He warned Forbes that bad anonymization can let buyers see individual people and organizations, and that information can leak right back out through model outputs.
Translation: a future AI agent could quote your 2023 Slack rant verbatim and not know it shouldn’t.
The deeper irony
There’s also a logic problem worth sitting with. If you want to build an AI that succeeds at business, is the conversation history of a company that failed really the best curriculum? You’re training tomorrow’s autonomous workforce on the playbooks of teams that ran out of money. Useful for studying mistakes, sure. Useful as a behavioral template, less obvious.
What this means for AI practitioners and operators
A few practical takeaways:
- If you run a company on Slack, your messages are now an asset class. Update your employee handbook and data retention policy before you need to. Spell out what happens to communications in an acquisition, wind-down, or asset sale.
- If you’re an employee, assume nothing you type at work is private long-term. That has always been technically true. Now it’s commercially true.
- If you’re building AI products, audit your training data provenance. “We bought it from a broker who anonymized it” will not survive the first regulatory inquiry or PR cycle.
- Regulators are watching. The Center for AI and Digital Policy is already on record. Expect employee privacy frameworks to extend into the post-shutdown asset class within the next year or two.
What comes next
RL gyms are going to keep growing because the labs need them and the supply of authentic workplace data is finite. Expect bigger checks, more middlemen, and at least one high-profile lawsuit when an employee discovers their words inside a model. The companies that win this category will be the ones who can prove clean consent chains, not just clean datasets.
Full details at the original Futurism AI report.