What data do AI companies buy, and why now?

Epoch AI estimated in June 2024 that the effective stock of quality public human text is about 300 trillion tokens, and that at current trends language models will fully use it sometime between 2026 and 2032. MIT Technology Review reported in October 2026 that demand for private company data jumped as labs ran short of public and synthetic training data.

The buyers are foundation AI labs, big tech firms training their own models and companies building simulated workspaces, called RL gyms, where AI agents practice real tasks. Surge AI, for example, buys data to build reinforcement-learning gyms for agents. They all want what the public internet does not show: how work actually gets done inside a company. These are the seven categories SourceX looks for when it licenses data from companies.

1. Customer support conversations

Support is one of the most obvious jobs for AI agents, and models need thousands of real cases to learn it. Each conversation works as a training example and an evaluation test in one.

The value is in the full arc: a customer problem, the agent's response and a clear resolution. Tickets that end in "solved" or a satisfaction score beat chats that trail off. This data typically lives in Zendesk, Salesforce, HubSpot or shared inboxes in Microsoft 365 and Google Workspace.

Watch out: customers type everything into support chats, from home addresses to account numbers. De-identification has to happen before anyone outside your company sees a record.

2. Workflow and ticket histories with outcomes

AI agents are being trained to do multi-step work, and the hardest thing to find is a clean record of tasks, decisions and outcomes in sequence. RL gyms are built from exactly this kind of material.

Outcomes are what make it valuable. A ticket that shows who picked it up, what they decided and whether it worked teaches a model to connect actions to consequences, something plain text rarely does. Look in Jira, NetSuite, Procore and Salesforce.

Watch out: workflow systems often hold vendor and partner details. Check whether your vendor agreements restrict how shared material can be used.

3. Sales and CRM data

Sales runs on judgment calls: which lead to chase, how to answer an objection, when to discount. Agents built to help sales teams need real deal histories to learn those patterns, and evaluators need them to test whether a model's advice holds up.

The richest records connect years of notes, emails, stage changes and won or lost results to the same account. Most of it sits in Salesforce or HubSpot.

Watch out: a CRM is mostly personal data about people at other companies. Names, titles, phone numbers and email addresses must be removed or transformed, and some customer contracts limit what you can do with their information.

4. Call recordings and transcripts

Voice AI needs real conversations with real interruptions, accents and pauses. Transcripts paired with outcomes, such as whether the issue was solved or the deal moved forward, are useful both for training and for testing how a model handles a live exchange.

Labels multiply the value. Thousands of hours tagged with call reasons and dispositions beat a folder of unlabeled audio. Recordings often live in call platforms, Microsoft Teams and CRM attachments in Salesforce or HubSpot.

Watch out: a voice can identify a person on its own, and call recording consent rules vary by state and country. Confirm what callers were told and whether recordings can be used beyond their original purpose.

5. Software engineering code and tickets

Coding agents are one of the busiest areas in AI. Real repositories linked to the tickets, reviews and bug reports behind them show how engineers reason through a problem, and they make honest evaluation tests. MIT Technology Review reported that one founder's shut-down crypto startup got about $25,000 for its code base, with roughly $15,000 more expected for its workspace data.

History is the value: commits, pull request discussions and the ticket that started it all. Look in GitHub and Jira.

Watch out: open-source code inside your repos carries its own licenses, and contractor-written code may have separate ownership terms. Separate what you own from what you only use.

6. Internal documents and knowledge bases

Agents need to look things up the way employees do. Wikis, policies, playbooks and how-to docs teach a model to find and apply company knowledge, and they make strong evaluation sets for retrieval.

The best documents are the ones people actually used, linked to the work they describe, like a runbook cited in the tickets that followed it. Expect to find them in SharePoint, Google Workspace, Microsoft 365 and Slack.

Watch out: knowledge bases quietly collect confidential material such as customer case studies, pricing, HR policies and board decks. Decide which spaces are in and which are out before preparation starts.

7. Legal contracts

Legal AI tools need real agreements to learn how clauses are drafted, negotiated and amended. A clean set of contracts with redlines and final versions is excellent material for both training and testing.

Variety and versions matter most: NDAs, master service agreements, vendor contracts and amendments, ideally with the negotiation history. They usually sit in SharePoint, Google Workspace or Salesforce.

Watch out: contracts almost always include confidentiality clauses that protect the other party, and some may bar sharing even in de-identified form. Your counsel should review them before they go anywhere.

Which Types Do You Hold?

☐ Support tickets with resolutions

☐ Jira or workflow histories

☐ CRM data and call recordings

☐ Code, docs and contracts

How to turn this list into a licensing decision

Pricing is opaque. MIT Technology Review reported that buyers quote wildly different numbers, and that broker Sunset values data by headcount, years in operation, data volume and how a team used its tools. That is a fair summary of what matters across all seven types.

SourceX licenses these data types directly from companies, usually for a one-time fee with no royalties, and the company keeps ownership of its original data. The best fit is 50 or more full-time employees at peak, with several years of operating history, whether the company is still operating, was acquired or has wound down. Nothing leaves your company until you sign.

Frequently asked questions

What kind of data do AI companies pay for?

AI companies pay for private data that shows real work: support conversations, workflow histories with outcomes, CRM records, call recordings, code, internal documents and contracts. MIT Technology Review reported that buyers include foundation labs, big tech firms and companies building RL gyms for AI agents.

Why is outcome data so valuable for AI agents?

Because it shows whether an action worked. A ticket that records a problem, a decision and the result teaches an agent to connect steps to consequences, and it gives evaluators a known right answer to test against. Public web text rarely includes outcomes.

Do AI companies buy data from smaller businesses?

Sometimes, but scale matters. Buyers weigh headcount, years in operation and data volume, according to MIT Technology Review. The SourceX best fit is 50 or more full-time employees at peak, not counting contractors, with several years of operating history.

Does business data need to be anonymized before AI companies use it?

It needs to be de-identified before it is used or shared. De-identification removes or transforms names and other identifiers so records cannot reasonably be linked to a person. No method removes all re-identification risk, so a good agreement acknowledges that risk plainly.

Sources

• Will we run out of data? Limits of LLM scaling based on human-generated data (Epoch AI, June 2024)
• AI labs are buying data from dead startups to train their models (MIT Technology Review, Oct 13, 2026)
• NIST SP 800-188, De-Identifying Government Datasets
• ArticlesBase contact page (editorial enquiries)

About the author

See which of these seven data types you hold with the five-question fit check at sourcex.si. No uploads, no commitment, and you can walk away at any point before signing. https://sourcex.si/

Noah Loul is the founder and CEO of SourceX, which licenses business data directly from companies for AI training and evaluation. A four-time founder, he previously built B2B Rocket, which was acquired by Blackpearl Group.