AI labs are buying Slack, Jira, and email archives from defunct startups to build “reinforcement learning gyms” and train AI agents in simulated workplaces
Defunct startups are being liquidated for their Slack archives, Jira tickets, and email threads—operational exhaust that AI labs now treat as premium training data.
Context & Ripple Effects
Related coverage traces AI labs’ expanding search for task-specific training inputs: from contractors teaching software-engineering work to marketplaces seeking professionals’ past materials and apps paying users for usable data.
The archive purchases extend that search from isolated examples to connected records of how work is assigned, discussed, revised, and closed. They arrive alongside a push to win B2B usage and deploy engineers to help customers operationalize models.
First-order effects
- AI labs obtain linked Slack, Jira, and email histories that can be turned into simulated workplace tasks and feedback loops for training agents.
- The liquidated startups’ operational records become an asset class for buyers, while labs gain training environments closer to multi-step business workflows than standalone prompts.
Second-order effects
- Labs competing for enterprise adoption will have more reason to differentiate on agents that can navigate tickets, communications, and handoffs rather than only generate text or code.
- The value of distressed-company data may rise, while buyers and sellers face sharper questions over what rights attach to employee communications, customer information, and work product in archived systems.
Third-order effects
- If this sourcing model persists, agent development may increasingly depend on privileged workflow datasets and evaluation environments, concentrating an advantage with labs able to acquire or license them.
- Enterprise AI adoption could make data provenance and controls over historical workplace records a more central governance issue, particularly where archives contain mixed ownership or sensitive operational context.
The trend: This is part of the shift from training general-purpose models on broad corpora toward building embedded workplace agents on proprietary, task-connected operational data.