Sourcing guide: how AI labs buy company data
AI labs buy company data by licensing it. A specification is matched against a metadata-only manifest, a redacted sample is reviewed under NDA, terms are signed directly with the owning company, and de-identified records are delivered. Published timelines are six to twelve weeks and published payouts run $20,000 to over $3 million per company (checked November 2026).
By BestDataForAI.com · Content reviewed 10 November 2026
Step 1 — Write the specification before looking at supply
The specifications that close name the task the model must learn, the industry, the record types (tickets, case files, code history, chat threads), the years of coverage and whether exclusivity is required. “More business data” is not a specification; supply matched to it disappoints in both directions.
Step 2 — Review a manifest, not the data
A manifest is metadata: systems in use, years of history, approximate record counts, languages, and a rights summary. It lets a buyer judge fit without any raw record changing hands — the standard first step at micro1, Mercor and every broker in this market, and the step that protects the owner if no deal follows.
Step 3 — Sample under NDA and check the reasoning trail
Buyers pay for records that show how work was done: the ticket thread and its resolution, the drawing and its revisions, the deal file and its outcome. A warehouse of final PDFs with no trail is worth a fraction of the same volume with the trail attached.
Step 4 — Licence terms that decide the price
| Term | What it controls | Price effect |
|---|---|---|
| Scope of use | Training, evaluation, internal only vs product use | Wider use, higher price |
| Term | Fixed years vs perpetual | Perpetual costs more |
| Exclusivity | Whether the owner can license the same records again | Roughly 2.5–5× in published analysis |
| De-identification | Who removes personal data, to what standard, who warrants it | Buyer-side de-id lowers the owner's cost, not the standard |
Step 5 — Routes compared
| Route | Suits | Watch-out |
|---|---|---|
| Direct to a lab | Very large / unusual datasets | One buyer, one price — no comparison |
| Training-data supplier | Defined programmes (micro1, Mercor, Scale, Handshake) | Published floors exclude smaller companies |
| Broker / introducer | Comparing offers, or sourcing several owners to one spec | Fee must be in writing before submission |
| Marketplace | Sellers who can package and describe a dataset themselves | Commission (Datarade publishes 30%, falling with paid plans) |
Frequently asked questions
How do AI labs buy training data from companies?
Through a licence. The lab or its supplier scopes a metadata-only manifest, reviews a redacted sample under NDA, agrees term, scope and exclusivity in a written licence with the company that owns the records, and receives de-identified data after signature. Most published timelines run six to twelve weeks from first contact to payment.
What makes company data good enough to buy?
Four things: it shows decisions and their outcomes (not just final documents), it covers several years so patterns repeat, the seller holds the rights to license it, and personal data can be removed without destroying its value. Data that fails the rights test is worth nothing to a careful buyer, however large it is.
Should a lab buy direct, use a supplier, or use a broker?
Direct suits very large or unusual datasets and teams with in-house licensing counsel. Suppliers (micro1, Mercor, Scale AI, Handshake AI) suit defined programmes with published floors. A broker suits a buyer who wants several qualified owners compared on one specification — or an owner who wants buyers compared instead of accepting a single offer.
What does exclusivity do to price?
It raises it materially and removes the owner's ability to license the same records again. Published pricing analysis (ddiy.co, 2026) puts exclusive deals at roughly 2.5 to 5 times a non-exclusive licence for comparable data. Buyers should price exclusivity as the separate asset it is.
Have a specification?
Send it on the request-supply form. You will get an honest read on whether qualified, rights-cleared supply exists for it.
Sources — checked November 2026
- micro1, Mercor, Scale AI and Handshake AI published programme pages, checked November 2026.
- ddiy.co, “How to Sell Company Data to AI Companies (2026 Guide)” — routes, timelines and exclusivity analysis, checked November 2026.
- Datarade published commission terms, checked November 2026.
Cite this entry: Best Data for AI. “Sourcing Guide: How AI Labs Buy Company Data — Best Data for AI.” Content reviewed 10 November 2026. Figures on this page are published buyer ranges with sources and check dates shown — they are not offers.