Guide / procurement
How AI labs procure enterprise data
AI labs procure enterprise data through licences: a supplier grants named uses of a dataset for a term, after the lab has set a spec, screened the supplier, reviewed rights and privacy, tested a sample and received documentation. The work that takes longest is diligence, not data processing. This page sets out the licence models, the steps and the paperwork.
Licence models
Enterprise data is licensed, not sold. The organisation that produced the records keeps them; the lab receives the right to use a processed copy for named purposes. Four variables define a licence.
- Exclusivity. Non-exclusive licences let the supplier license the same data again. An exclusive window keeps it from competitors for a fixed period and costs more.
- Scope of use. Pre-training, fine-tuning, evaluation, RL environments and internal research are separate rights. Name each one.
- Payment shape. A one-off fee, an access fee followed by recurring payments, or a price per unit of data such as records or tokens.
- Term and survival. How long the data may be held, and what happens to trained models when the term ends.
Three public reference points
What disclosed licences look like
$203.0 million
Aggregate contract value of data licensing arrangements Reddit entered into in January 2024, with terms of two to three years.
Large licences run for fixed multi-year terms, which is why renewal and end-of-term clauses matter.
Source: Reddit Form S-1, SEC, checked 19 October 2026
$10m+
Initial data access fee in Informa's agreement with Microsoft, followed by a recurring payment across 2025, 2026 and 2027, for non-exclusive access.
An access fee plus recurring payments is a common shape, and non-exclusive access is stated explicitly.
Source: Informa market update, RNS, checked 19 October 2026
$49.1 million
AI licence revenue reported by John Wiley & Sons for the year ended 30 April 2026.
Licensing to AI developers is now a line item rightsholders report, so expect them to negotiate it as a product.
Source: Wiley annual report, SEC, checked 19 October 2026
Ten steps
From spec to signed licence
- 1. Spec
- Data type, languages, period, functions, volume after filtering, intended use, exclusions. See the spec sheets.
- 2. Supplier screen
- Who the supplier is, who owns it, where it operates, and a sanctions check (EU sanctions map).
- 3. Rights review
- The chain of title from the organisation that produced the records to the party granting your licence.
- 4. Privacy review
- Method, evidence and residual risk, and whether your own assessment is needed. See the checklist.
- 5. Sample
- A representative slice loaded into your own pipeline and measured against the spec.
- 6. Security review
- How files are transferred and stored on both sides, who has access, and when raw copies are deleted.
- 7. Documentation
- A completed datasheet, including the AI Act inputs.
- 8. Contract
- Licence grant, warranties, re-identification ban, withdrawals, term, deletion and liability.
- 9. Acceptance
- Checksums verified, schema validated, counts compared with the datasheet, then sign-off.
- 10. After delivery
- Corrections, withdrawal requests, updates and deletion at the end of the term.
Where the time goes
Processing a dataset is the predictable part. The unpredictable part is getting answers: who owns code written by contractors, whether a mailbox export included private mail, which de-identification recall was measured and on what. Send the datasheet template with your spec so the supplier collects those answers while it works, and bring counsel and privacy in before the sample, not after.
Red flags
- No written basis for the right to license, or a basis that stops at "the customer agreed".
- Data obtained by scraping, or through third-party apps that pulled it from a platform's API.
- No sample before signing, or a sample hand-picked rather than drawn at random.
- De-identification described in adjectives instead of numbers.
- Reluctance to complete a datasheet or to warrant what it says.
Requesting a dataset here
If you would rather not run the sourcing yourself, send a dataset request. Datasets are assembled to your spec, de-identified, documented and delivered under a licence with a datasheet, following the steps above.
FAQ
Procurement questions
How do AI labs procure enterprise data?
Through a licence rather than a purchase. The lab writes a spec, screens the supplier, reviews rights and privacy, tests a sample, collects documentation, and signs a licence that names the permitted uses, term and end-of-term obligations.
What licence models are used for training data?
Non-exclusive licences, which let the supplier license the same data to others, and exclusive windows for a fixed period. Payment is usually a one-off fee, an access fee plus recurring payments, or a price per unit of data.
What slows down an enterprise data purchase?
Rights and privacy diligence more than data processing. Unclear ownership, missing evidence of de-identification and late involvement of counsel are the usual causes of delay.
Who should be involved on the lab side?
Data acquisition to own the deal, research to set the spec and judge the sample, legal for rights and the licence, privacy for personal data, security for transfer and storage, and compliance for AI Act documentation.
What happens to a trained model when the licence ends?
Only what the contract says. Settle it explicitly: whether models trained during the term may stay in use, and what must be deleted or returned.
Why are so few enterprise data deals public?
Because the licensing organisations rarely disclose them. Public figures mostly come from listed publishers and platforms that must report material contracts, which is why the reference points on this page come from filings.
What are red flags in a data supplier?
No written rights basis, data obtained by scraping or through apps that pulled it via an API, no sample before signing, de-identification claims without evidence, and reluctance to complete a datasheet.
Keep reading
Next step
Have a spec? Send it over
Name the data type, languages, period, volume and intended use. We answer by email with what can be assembled and how it would be documented.