How do AI labs buy data?
AI developers buy data much as large companies buy any critical input: a team defines the need, finds suppliers, checks them, tests a sample, negotiates a contract and takes delivery. The details differ by buyer, and most contract terms are private, so this page describes the process in general terms only.
Public job postings at AI developers often use titles such as data acquisition, data partnerships and data operations, which suggests a dedicated function. The steps below are a general model of what such roles imply, not any named buyer's internal playbook.
What are the stages of a data purchase?
| Stage | What the buyer does | What the supplier needs ready |
|---|---|---|
| 1. Need definition | Decides which skills or domains the model lacks | Nothing yet; the buyer is scoping |
| 2. Sourcing | Looks for suppliers who hold matching records | A clear description of systems, years and volume |
| 3. Screening | Checks rights, privacy exposure and the supplier's authority to license | Evidence of ownership, notices and policies |
| 4. Sample review | Evaluates a sample for quality, structure and outcomes | A cleaned, authorized sample |
| 5. Negotiation | Agrees price, term, exclusivity and usage limits | An authorized signer and a view on exclusivity |
| 6. Delivery | Receives data in an agreed format | Exports, redaction done, a handover method |
| 7. Payment | Pays under the contract | An invoice and payment details |
For a company, most of the effort lands in stages 3, 4 and 6: proving rights, preparing a sample and producing clean exports.
What do buyers look for in a supplier?
Buyers reduce risk before they chase volume. The questions are consistent.
- Rights. Did the supplier create this material and may it license it? Data belonging to customers, clients or third parties without consent is a stop sign.
- Privacy exposure. How much personal information is in it, and what redaction has been agreed?
- Structure and outcomes. Do records show steps, decisions and results, not only finished documents?
- Provenance. Were the records produced by people doing real work, rather than generated to be sold?
- Delivery ability. Can someone actually export the data, in volume, in a usable form?
- Exclusivity. Will the owner consider an exclusive AI-training license for an agreed term?
These map to the screens a partner can run early; see the company fit checker.
Why is direct selling hard for a mid-sized company?
A company with 50+ full-time employees at peak (contractors excluded) is unlikely to have anyone whose job is negotiating data licenses. Selling directly means finding the right buyer team, preparing a legal review, building an inventory, producing a sample and answering diligence, all on top of a normal workload.
A company that goes direct also negotiates without a benchmark for what scope and terms are typical. And a supplier approaching buyers one at a time has to find the right team at each one, with no guarantee of a reply.
What does an intermediary change?
A managed intermediary packages the work. In SourceX's process, the company completes one data inventory, SourceX qualifies the company and handles rights review, price and terms are agreed with the company before buyers see anything, and buyers review the opportunity. Buyers set their own interest and timing, and no outcome is promised. Once a company is deal-ready, buyers typically respond within about two weeks. The company receives one all-in price, SourceX's fee included, with no separate charges.
The full sequence is in how it works. Two points are worth stating plainly: the company keeps ownership because data is licensed, not sold, and nothing is binding until the company agrees price and terms and signs. Data is delivered only after an executed agreement and the company's authorization. For the larger picture, read enterprise AI data licensing deals.
What does this mean for a referral partner?
You are not a procurement agent. Your role is the introduction and basic fit information. You do not negotiate, you do not touch records, and you do not describe them.
Useful ways to frame the buyer side for a company owner:
- Buyers care less about a sales pitch than about clean rights and a clear inventory.
- Preparation is the main workload, and the same preparation helps with internal projects; see licensing preparation as an AI readiness assessment.
- The owner can ask questions before signing anything, and there is no obligation to proceed.
- A plain-language handout is available: the client explainer for licensing operational data to AI labs.
Limits of what is public
Contract values, exclusivity periods and buyer preferences are generally confidential, and this page does not state any. Procurement practice is also changing as the market matures, so treat the stages above as a general model. If an owner asks "what will I get", the honest answer is that price depends on the data, the term and buyer demand, and that is agreed only after review.
How rewards work
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. The reward is paid only after the buyer pays and SourceX receives its fee, and no reward is guaranteed.
Next step
If you know a US company with years of records across many systems, register as a partner and make the introduction. Owners who prefer to start themselves can apply at sourcex.si/apply.