Who sells AI training data? The five kinds of supplier explained

The short answer

Licensed AI training data comes from five kinds of supplier. Publishers license archives, online platforms grant access to user-generated content, data vendors assemble and resell datasets, human-data firms produce labels, rankings and expert examples to order, and operating companies license records of their own work.

The first four run sales teams and standard contracts. The fifth, established companies whose records show how real work gets done, mostly has neither, even though those records are among the hardest for AI developers to find anywhere else. That gap is where a referral partner comes in.

AI developers also collect data themselves, by crawling the web and generating synthetic examples. Those are sources rather than suppliers, and they are covered in where AI training data comes from.

Who are the five kinds of supplier?

Two public examples show the first two categories at work. In July 2023 the AP agreed to license part of its text archive, going back to 1985, to OpenAI, with financial terms undisclosed, according to the AP report carried by PBS NewsHour. In May 2024, OpenAI announced that it would access Reddit's Data API for real-time, structured content, again with no financial terms disclosed. Both are public market events, and none of these companies is a SourceX buyer or partner. Disclosed deal values are compared in enterprise AI data licensing deals: beyond the media headlines.

How do the suppliers differ in what they offer buyers?

Each category answers a different need. Publishers and platforms supply breadth and freshness. Vendors supply convenience. Human-data firms supply exactly the examples a developer specifies, written for the purpose. Operating companies supply something none of the others can: an authentic record of work as it happened, with the tools used, the decisions taken and the result.

That last property is why developers building agents care. A model learning to work a support queue, run a sales cycle or manage a project needs to see those jobs carried through to an outcome, and that evidence sits inside private systems.

Can you sell AI-generated data for training?

Not through SourceX. Records generated with AI tools in order to sell them are on SourceX's red-flag list, and a company offering them will not qualify.

The logic is simple. A developer can generate synthetic text itself, in any volume. What it cannot generate is evidence of what really happened inside a business: the ticket that escalated, the quote that lost, the approval refused and why. A buyer licensing company records is paying for that provenance, and fabricated records undermine it.

The red flag is fabrication for sale, not ordinary use of AI tools at work. If staff drafted some documents with an assistant, mention it during qualification so it can be assessed with everything else.

Why are operating companies the least organized suppliers?

Publishers have rights departments, platforms have API teams, and vendors and labeling firms exist to sell. A 150-person freight broker or managed service provider has none of that. Its records sit across many systems (most strong companies run 10-15 or more), nobody owns the question of licensing, and nobody has checked which records the company has the right to license.

The pool is large. The Census Bureau counted 5.58 million US employer firms with fewer than 500 employees in 2023. Only part of that pool meets SourceX's baseline, which looks for US companies with 50+ full-time employees at peak (contractors excluded), several years of documented operations, the right to license their records and an authorized sponsor. The rights questions are set out in the legal framework for licensing business data.

What does this mean for a referral partner?

Partners matter most in the fifth category, because no sales team will reach these companies unprompted. Someone who already advises, invests in or works alongside an owner can raise the idea in an ordinary conversation.

A four-point screen for spotting one in your network:

• A US business that reached 50+ full-time employees at its peak
• Years of history in ticketing, CRM, project and finance tools
• Records of its own work rather than files it holds for clients
• An owner or executive you can speak to directly

If a company passes, the company fit checker gives a preliminary, non-binding read. From there the introduction runs like this:

1. You register and send the owner your referral link, or submit the company through the referral form.
2. SourceX checks headcount, history, breadth of systems and rights with the sponsor.
3. The company inventories its systems. You never see, export or describe its records.
4. The owner settles price and terms with SourceX before any buyer sees the opportunity, then decides whether to sign once a buyer is interested.
5. The company is paid once after delivery, and your reward follows once SourceX has received its fee.

Partners earn 25% of the eligible platform fees SourceX actually collects from the referred company's licensing deals, capped at $100,000 per referred company. Rewards become payable only after the buyer pays and SourceX receives its fee, and no reward is guaranteed. How it works has the full sequence.

Who is not a supplier?

• Sellers of lead lists or scraped contact data.
• Businesses built mainly on consumer personal data or medical records and claims, with no basis to license them.
• Firms whose records belong to their clients, without those clients' consent.
• Anyone offering records made with AI to look like real business activity.
• A company that has already licensed the same records for AI training.

For definitions of the data involved, see what is AI training data.

Next step

If a company you know passes the four-point screen, register as a partner and make the introduction, or point the owner to sourcex.si/apply with your referral link attached.