Mercor chases shuttered startups' Slack and Jira archives

Eight days after AI agent startup Warmly agreed to be acquired by HubSpot in late June, CEO Maximus Greenwald opened his inbox to find a note from Mercor offering to buy or license Warmly's code base and the records of its tasks, according to The Information. Mercor, a San Francisco startup that pays contractors to train AI models for OpenAI, Anthropic and Google, is treating the operational exhaust of dying startups as a discrete raw material, not just a curiosity.

The category the buyers are building is what people in the field are calling reinforcement learning gyms: simulated workplaces stocked with a real company's Slack threads, Jira tickets and internal emails so an agent can practice being an employee. The Information reports that leaders at Anthropic have discussed spending more than $1 billion on RL environments over the next year, and Mercor is not the only bidder. Forbes reported in April that Shanna Johnson, CEO of shuttered transcription and captioning company Cielo24, sold 13 years of Slack, Jira and email archives for 'hundreds of thousands of dollars', and that SimpleClosure, a startup that helps founders wind down, has processed roughly 100 such data deals in the past year at payouts of $10,000 to $100,000.

For a founder staring down a soft landing or a full shutdown, that changes the math of the exit. The customer list and the IP were the obvious things you sold; now the wiki, the ticket queue and the arguments in #eng-help are line items too. For an acquirer like HubSpot, it changes the diligence checklist: if a target's messy internal history has a separate street price, silence on data-disposition in the deal papers is a gap, not a formality.

None of the retrieved reporting says whether Warmly accepted Mercor's overture, whether HubSpot's deal terms would have let it, or how these gyms scrub employee names and customer conversations before a model sees them. Fast Company's write-up and the Forbes piece both lean heavily on the sellers' side of these transactions; the labs doing the training are quieter about what they filter. Mercor also confirmed a security breach earlier this year in which stolen samples reportedly included Slack and internal ticketing data, a useful reminder that this new inventory sits in a supply chain that has already leaked once.

The forward interest is who ends up owning the standard. If Mercor and micro1 lock in exclusive pipelines to shutdown platforms like SimpleClosure, the agent story we keep tracking turns from a modelling problem into a sourcing one, and the moat sits with whoever holds the richest catalogue of real work.