Dead Startups' Slack and Email Archives Are Being Sold as AI Training Data

That "what's for lunch?" message you left on Slack at your old job? It's being sold as AI training data. For up to $100,000 a pop. forbes

What's Going On?

AI companies have struck a new vein of training gold: Slack archives, email threads, Jira tickets, and internal documents from companies that no longer exist — what the industry calls "operational exhaust." forbes

Why dead companies specifically? Simple. Early LLMs trained on public internet data — Wikipedia, Reddit, digitized books. But according to former OpenAI chief scientist Ilya Sutskever, all of that was exhausted by late 2024. What the industry is building now is "agentic AI" — models that actually do work. Training those models requires data showing how real people actually work: making decisions, collaborating, solving problems. forbes +1

Enter "reinforcement learning gyms" (RL gyms) — simulated office environments built from real company data where AI agents practice doing work. Anthropic has reportedly discussed spending up to $1 billion on RL gyms this year, and there are already 50+ startups in the space. gizmodo +1

AfterQuery, for example, sells off-the-shelf "worlds" like "Big Tech World" and "Finance World" to AI labs. A sample task: plan a coworker's birthday party — except another coworker is secretly planning one too, and the AI agent has forgotten when the birthday actually is. It has to message colleagues, investigate, and decide whether to collaborate or bail. forbes

Why Does This Matter?

The market is being driven by startup shutdown specialists. SimpleClosure launched Asset Hub this week, and competitor Sunset offers similar services. forbes

	Public Web Data	Defunct Startup Internal Data
Data type	Wikipedia, Reddit, news articles	Slack chats, emails, Jira tickets, code commits
Availability	Exhausted by late 2024forbes	Steady supply from ongoing startup failures
Work context	Fragmented	High-context: decisions, workflows, collaboration linked together
Agentic AI suitability	Low	High — reflects real work processes
Price per deal	Nearly free (crawling)	$10K–$100K per companyforbes
Privacy risk	Public data	Identifiable employees, no consent obtainedforbes

According to Sunset CEO Brendan Mahony, pricing depends on company size, age, and "data richness" — how well internal data links together. A Jira ticket tied to a specific code commit is worth more than a standalone document. Healthcare and finance data commands a premium. forbes

Companies promise anonymization, but a 2020 OpenAI/Google study showed that LLMs can memorize training data verbatim and regurgitate it with the right prompts. Scrubbing PII from a career's worth of work data "isn't an on-off switch," industry experts warn. forbes

What You Should Do About It

1. Check your data rights when leaving a job
   Re-read your employment agreement and NDA. IP assignment and "selling internal comms to third parties" are separate issues. If there's no explicit consent clause, you may have grounds to push back.
2. Stop putting sensitive info in Slack and email
   Review your habits around sharing personal data — SSNs, health info, salary details — in work messengers. Digital footprints outlive companies.
3. Companies: Establish data disposal policies
   Plan how internal data gets handled during shutdown. GDPR, CCPA, and emerging regulations should inform your deletion and sales policies.
4. Founders: Know what your data is worth
   If you're winding down, explore platforms like SimpleClosure's Asset Hub or Sunset. But verify their anonymization is truly "rock solid" before signing off. forbes
5. Watch the RL gym market
   This is core infrastructure for the agentic AI era. Key players: Prime Intellect (valued at $1B+), Fleet (in talks at $750M valuation). forbes

Go Deeper