AI labs want records that show how real people do real work: the email thread, the ticket, the document revision and the code change that together explain a decision. Public web text has been largely used up for training, and it rarely shows that chain. Company records do. This guide explains what buyers ask for, why, and what they turn down. To check how your own records would be valued, get a confidential estimate.
Why do labs want company records at all?
To train AI agents to do multi-step work. Buyers use company records to build reinforcement-learning environments. In plain terms, these are practice settings where an AI system tries a task, such as resolving a support ticket or closing the books, and is scored on the result.
The buyers say so directly. micro1 says it uses de-identified company data to build reinforcement-learning environments that mirror real business operations (micro1, Oct 9, 2026). MIT Technology Review reported that Surge AI buys company archives to build reinforcement-learning “gyms” for agents (Oct 13, 2026). Collinear licenses code and business app history to build training and evaluation environments (Collinear). aitrainer.work describes the “data wall”: public internet text has been exhausted for training (Jun 29, 2026).
Turing puts it in one line on its Project Lazarus page: frontier AI labs “train on how real companies operate, not the polished product” (Turing, observed Oct 16, 2026).
Which records do buyers ask for?
Buyers list the same families of records, by tool. All observed on Oct 16, 2026.
- Email and chat. Slack, Teams, Gmail and Outlook. Mercor names these first among 100+ connectors (Mercor). Nyne AI lists chat, email and calendar (Nyne).
- Tickets and project history. Jira, Zendesk, ServiceNow and Linear. Turing lists tickets, wikis and runbooks.
- Documents and SOPs. micro1 asks for SOPs, knowledge bases, project histories, internal processes and decision-making patterns (micro1).
- CRM and finance. Salesforce, HubSpot, QuickBooks and NetSuite. Mode buys CRM and billing data with email, documents, chat and calls (Mode).
- Code with its history. Collinear wants pull requests with the tickets behind them, CI logs and coding-agent sessions.
- Industry records. The General Data Company buys completed projects, orders, maintenance records, drawings and CAD models from industrial firms (Who Sells Data, Oct 14, 2026). Pewter lists rate quotes, shipment documents, CAD drawings, workpapers and reconciliations (Pewter).
- Hiring and process logic. Handshake AI names business process flows, hiring logic and document-level communications (Who Sells Data).
Which industries are buyers asking for?
Most programs take any sector, but several name the ones they want. Mercor lists legal, healthcare, finance and insurance, professional services, manufacturing, retail and e-commerce, real estate, and logistics and supply chain (Mercor). Turing’s Project Lazarus lists software and engineering, finance and accounting, customer support, sales and CRM, legal and contracts, and operations and logistics (Turing). Centific names finance and accounting, healthcare operations, sales and deal workflows, people operations and IT service management (Who Sells Data, Oct 14, 2026). aitrainer.work lists software engineering, finance and accounting, revenue and CRM, and operations and logistics as the high-value verticals (Jun 29, 2026).
Lock also listed what labs were asking him for in October 2026: circuit board CAD files, architecture firms, heavy construction and 3D video game assets (NGMI podcast, unverified).
What makes one archive worth more than another?
Records that connect
Ooak Data prices on how many tools connect “and how deeply” (Ooak). A support ticket that links to the email, the code fix and the refund shows a whole decision. Ryan Lock, a data broker, made the same point on the NGMI podcast in October 2026: a customer email that moves into Slack and is resolved in a ticket is a complete story, and a company that switched chat tools and lost that chain is worth a fraction. He named three quick measures buyers use: how many software tools a company uses, how many years of connected records it has, and its headcount. His views are his own and unverified.
The reasoning, not only the outcome
Turing asks for “the reasoning behind decisions.” Review comments, rejected drafts and escalation threads show reasoning.
Years and people
Collinear says prices rise with 15+ people at peak and several years of history. Most buyers set floors of 20 to 40 staff and three or more years.
Work done by humans
Collinear does not want AI-generated code. It lists human-written code among the things that raise its price.
A field few others can supply
micro1’s top tier, $1M+, is for “highly unique, proprietary operational data.”
What do buyers turn down?
- Data you cannot prove you own. Collinear lists it as not needed. Every buyer asks.
- Personal information they cannot strip. Buyers de-identify before use, and some records are too personal to clean. See how to de-identify support tickets and email.
- HR, payroll and health records. SimpleClosure does not sell sensitive employee and health records (SimpleClosure, Sep 3, 2026).
- Loose documents without context. Collinear does not want them.
- Customer or supplier records under confidentiality terms. The General Data Company takes them only where no confidentiality agreement stands in the way.
Is this the same as selling our customers’ data?
No. The buyers in this guide want records of how your team works. Consumer personal data is a separate market, and registers such as Who Sells Data leave it out. Most buyers remove personal information before data is used, and you should ask how. Our guide on whether it is legal covers the employee and customer questions.
What does this mean for your company?
Three questions decide most of it. Does your work run through several tools that reference each other? Do you have three or more years of it? Do you own it outright? If yes to all three, you likely hold what these buyers describe. Our guide on what company data is worth sets out the published price bands, and can you sell your company’s data to AI companies? covers size floors, legal checks and the steps. Marin Data takes it from there at no cost to you.
Sources
- NGMI podcast, interview with data broker Ryan Lock, Oct 16, 2026 (his figures are his own and unverified)
- micro1, $1B enterprise data commitment, Oct 9, 2026; micro1, data partnerships, observed Oct 16, 2026
- MIT Technology Review, AI labs are buying data from dead startups, Oct 13, 2026
- aitrainer.work, AI labs paying for enterprise workflow data, Jun 29, 2026
- Turing Project Lazarus, Collinear, Mercor, Nyne AI, Mode, Ooak Data, Pewter, observed Oct 16, 2026
- Who Sells Data, buying programs, observed Oct 14, 2026
- SimpleClosure release via Webull, Sep 3, 2026