A data licensing agreement is the contract that sets what an AI lab, or the company in between, may do with a copy of your records and what it pays for them. The clauses that decide most of the outcome are scope, permitted use, term and deletion, payment, warranties and indemnity. We found two such agreements published by programs in this market. This guide reads both clause by clause, alongside what buyers say on their own pages, all checked on Oct 20, 2026. If you want a range for your records before reading contracts, get a confidential estimate.
This is general information for US companies, not legal advice. Have your own counsel review any agreement before you sign it.
Is a data licensing agreement the same as selling your data?
Owners usually say “sell,” and most programs we read describe a license. A license gives the buyer defined rights over a copy and leaves ownership with you. Anthroly puts it as “A license is not a sale” and offers “defined material, defined term, defined use” (Anthroly, read Oct 20, 2026). Collinear says “Ownership stays with you” (Collinear). Some deals are closer to a purchase: Mode says it “buys the agreed, filtered copy directly from you” while your company keeps its original records (Mode). Either way, the agreement sets what the buyer may do with that copy.
Who is the agreement actually with?
Often it is not with the AI lab. Data partners and marketplaces sign with you and then license on to labs under their own contracts. Under the pre.dev Codebase Agreement, pre.dev signs each buyer license “as principal in Pre.dev’s own name,” sets the price and needs no further approval from the owner for non-exclusive deals. The buyer’s identity stays confidential, and the owner is “not entitled to an accounting” of what the buyer paid (sections 6.1 and 13.8, read Oct 20, 2026). Endless says its published terms “describe what we do, not promises about the labs” (Endless).
Ask early who signs with you, who ends up with the data, and which of your terms flow down to them.
What exactly is being licensed?
The scope clause defines the material. pre.dev’s definition of “Codebase” is wide: repositories and version history, documentation, test suites, CI/CD pipelines, “issue trackers, tickets, and engineering notes,” and the intellectual property in all of them (section 2). Ooak Data annexes your exclusions to the contract “before a single record moves” (Ooak). Labelbox Prism says participation can be scoped to “specific workflows, metadata layers, or historical datasets” (Prism).
Name the systems, date ranges and fields that are in. Annex what is out: HR, legal, health, and anything a client owns. Our guide on what AI labs want from company records covers which records usually belong in scope.
What can the buyer do with the data?
This is the permitted-use clause. The two published agreements, from pre.dev and the marketplace Defined.ai, sit at opposite ends.
| Term | pre.dev Codebase Agreement (owner signs) | Defined.ai Data License Agreement (buyer signs) |
|---|---|---|
| Grant | Irrevocable, non-exclusive, worldwide, transferable, sublicensable | Non-exclusive, non-sublicensable, non-transferable |
| Use | Buyers may use the code “for any lawful purpose,” including training and fine-tuning | Internal business use; buyers may commercially exploit models trained on the data “as long as the Data itself is not made available” |
| Ending it | Either side on 30 days’ notice; closed buyer licenses survive | Either side on 60 days’ notice |
| After it ends | Deals in progress have a 90-day tail; backups and licensed copies are kept | Buyer must delete the data and certify it in writing; trained models are not addressed |
| Law | Delaware | Washington |
Sources: pre.dev Codebase Agreement and Defined.ai Data License Agreement, effective Nov 30, 2022, both read Oct 20, 2026.
Ooak’s contract “licenses them for training only.” Training, evaluation and resale to other customers are separate uses, and each one should be named or excluded. If you still operate, ask for a bar on use in a product that competes with you. Frankfurt Kurnit Klein & Selz (FKKS) warns a buyer “could ultimately use it to train an AI that develops a competing product” (FKKS, Oct 1, 2026).
How long does it last, and what happens at the end?
A deletion clause can only reach copies of the data. FKKS puts the limit plainly: “Once data is used to train a model, it cannot practically be pulled back out.” Of the two agreements and five deletion terms we read, none reaches models already trained on the data.
What buyers published on deletion, read Oct 20, 2026:
- Ooak deletes your raw archive within 30 days, “a contract obligation, not a policy,” with written confirmation, and keeps only the de-identified dataset for the license term.
- micro1 retains data “only for agreed periods” and deletes it on request or at the end of an engagement (micro1).
- SuperAnnotate retains data “for an agreed period, then deleted” (SuperAnnotate).
- Endless lists “Deletion at the end of the agreed term.”
- pre.dev returns or destroys materials on request, except backups, copies needed for existing buyer licenses and copies it licensed for itself (section 10.5).
Ask for a deletion date for raw data, a written certificate, and a clear statement of what the buyer may keep. Our guide to exclusive and non-exclusive licenses covers how term and exclusivity interact.
How and when do you get paid?
The payment clause sets the amount, the trigger and the timing. Published structures, read Oct 20, 2026:
- A fixed sum agreed first. Endless agrees “a fixed sum with Endless before anything is shared.” Mode’s written agreement “sets the amount and payment schedule before export,” and Mode expects about three months from first conversation to payment.
- Paid after delivery and acceptance. Nyne AI pays “Once your dataset is delivered and accepted” (Nyne). Collinear says it will “confirm delivery is complete, then pay according to the contract.”
- Paid per license, after the buyer pays. pre.dev pays $100 per million source tokens and $40 per million git-history tokens per non-exclusive license, within 15 days of the buyer’s funds clearing. It notes buyers commonly pay on 60-day terms, it can claw back refunded amounts, and the owner “has no guaranteed minimum compensation” (sections 13.2, 13.5, 13.6 and 13.10).
- Upfront plus revenue share. Prism pays “an upfront licensing fee at signing plus a revenue share on downstream use.”
Write down what “acceptance” means and what happens if the buyer rejects the data. If any part of the payment is a revenue share, ask for per-sale statements and audit rights. Our market index and guide to what company data is worth set out the published price bands.
What will you be asked to promise?
Warranties are the promises you make about the data, and a buyer relies on them if a claim arrives. pre.dev’s owner warrants that it owns the code or holds sufficient rights; that every founder, employee and contractor assigned their IP to it; that no client holds rights in the code; that open-source components are disclosed; and that the code contains no personal data, credentials or malware “except as expressly disclosed” (section 5). Those promises are repeated each time a buyer license is granted (section 13.9) and may be passed through to buyers (section 7.2).
Compare the other direction: Defined.ai, as licensor to its buyers, provides the data “AS IS” and disclaims implied warranties, including non-infringement.
You can ask to limit warranties “to the knowledge” of named officers, to schedule known exceptions, and to stop repeating them after signing. Our guide on whether it is legal to sell company data to AI lists the ownership and consent checks to run before you promise anything.
Who pays if something goes wrong?
That is the indemnity and liability cap. Under pre.dev’s agreement the owner indemnifies pre.dev and its buyers for IP, privacy and ownership claims, and those obligations are “not subject to any cap or exclusion” under the agreement. pre.dev’s own liability is capped at what it paid the owner in the prior 12 months (sections 9.1 and 9.4). If code is found to infringe, the owner may have to refund payments for the affected licenses (section 9.3). In the other direction, Defined.ai caps its own liability to buyers at the total paid to it under the agreement.
Ask for a cap on your indemnity, usually tied to what you were paid, with any exceptions named.
What privacy terms should the agreement contain?
If the records include personal information of California residents, the CCPA’s test for “deidentified” data includes a contract term. The business must take reasonable measures, publicly commit not to re-identify, and it “Contractually obligates any recipients of the information to comply with all provisions of this subdivision” (Cal. Civ. Code §1798.140(m), as posted by the California Privacy Protection Agency, January 2025). FKKS advises requiring certification of de-identification “to the company’s satisfaction.”
In practice: name who de-identifies and where, set the standard (SuperAnnotate removes identifying detail “to the standard your team sets”), see a sample first, and add a ban on re-identification with a remedy.
What other clauses are easy to miss?
- Confidentiality and publicity. pre.dev’s confidentiality runs five years after disclosure and bars either side from announcing the deal without consent (sections 3.6 and 10.4).
- Non-circumvention. pre.dev’s owner may not knowingly close a deal directly with a buyer that pre.dev introduced during the term or in the 12 months after it ends (section 11).
- Exclusivity. Prism’s licenses are exclusive by default; pre.dev’s are non-exclusive unless both sides sign a separate confirmation (section 13.4). See exclusive vs non-exclusive licenses.
- Assignment. pre.dev may assign the agreement without the owner’s consent (section 12.2). If you might sell your company later, read how a data license affects M&A due diligence before you agree to assignment and exclusivity terms.
For software owners, our page for software companies covers code-specific checks.
Is there a sample data licensing agreement?
The two above are the clearest public examples we found, and both are written for the platform that publishes them. Use them to learn the vocabulary and spot one-sided terms. Our comparison of AI data partnership programs lists which programs publish terms at all.
Marin Data does not draft contracts or give legal advice. We read buyers’ terms against each other and point out where you need counsel. It costs your company nothing up front, and nothing at all if no deal closes; Marin Data is paid by the AI lab.
Sources
- pre.dev Codebase Agreement (PDF), read Oct 20, 2026
- Defined.ai, Data License Agreement, effective Nov 30, 2022, read Oct 20, 2026
- California Consumer Privacy Act, Civil Code §1798.140(m), California Privacy Protection Agency, posted January 2025
- Frankfurt Kurnit Klein & Selz, selling company data for AI training, Oct 1, 2026
- Ooak Data, micro1, SuperAnnotate, Labelbox Prism, Mode, Anthroly, Endless, Nyne AI, Collinear, all read Oct 20, 2026