The AI data license agreement, clause by clause: what each term means for the seller

A data license agreement is where an AI data deal is actually decided. The price gets the attention, but the clauses decide what the buyer can do with your records, what you keep, what happens when the term ends, and who pays if something goes wrong. This is a walk through each clause in the order it usually appears, with what it should say from the seller's side and the red flags worth catching early.

How the agreement is put together

Most data licenses are a main agreement plus a few exhibits. The main agreement holds the legal terms. The exhibits hold the specifics: a data description (what records, which years, what format, how many), the de-identification plan, the security requirements, and the fee schedule. If personal information is involved at any point, a separate data processing agreement usually sits alongside it.

The exhibits matter as much as the main text. A clause that says the data was "de-identified in accordance with Exhibit B" is only as good as Exhibit B.

The clauses, one by one

1. Definitions

The definitions section looks like boilerplate and is the part most worth reading slowly, because every later clause depends on it. Four terms carry most of the weight.

• Licensed Data: the records you deliver, described precisely enough that nobody can argue later about what was included.
• Model: the machine learning models the buyer trains or improves with the data.
• Derived Data: anything the buyer creates from your records that is not a model, such as embeddings, labels, summaries, extracted features or synthetic records generated to imitate yours.
• Outputs: what the buyer's models produce when used.

The red flag is a missing or narrow definition of Derived Data. If synthetic copies of your records are not covered, the restrictions and deletion duties later in the contract may not reach them.

2. The license grant

The grant is the permission itself. A seller-friendly grant is specific: a non-exclusive, non-transferable, non-sublicensable right for a named buyer (and only the affiliates named) to use the Licensed Data to train, fine-tune and evaluate Models, for the term, within a defined field of use.

Each of those words does work. Non-transferable means the right does not move if the buyer is acquired or reorganizes without your consent. Non-sublicensable means the buyer cannot pass the data to contractors or partners beyond those the agreement allows. A field of use (for example, models for customer service) keeps the data out of products you did not have in mind. The red flag is a grant to use the data "for any purpose" or "for the buyer's business," which turns a license into something close to a sale.

3. Exclusivity

If the license is non-exclusive, this clause is short or absent. If the buyer wants exclusivity, scope it narrowly: a defined field, a fixed period, a specific subset of records, and a price that reflects what you are giving up. We cover the trade-off in selling vs. licensing data. The red flag is exclusivity that is broad in scope and open-ended in time while the price is a non-exclusive price.

4. Restrictions

The restrictions list what the buyer may not do, and they are the seller's main protection. Expect, at a minimum:

• No resale, publication or display of the records, and no product that lets users search or retrieve them.
• No sublicensing or sharing beyond the parties the agreement names.
• No attempt to re-identify any individual, and no combining the data with other sources to defeat the de-identification.
• Reasonable measures to keep Models from reproducing records in their Outputs.

That last point is specific to AI. Language models can memorize and repeat fragments of their training text; researchers showed that verbatim sequences, including personal contact details, could be extracted from a trained model. A restriction against reproducing the Licensed Data in Outputs, backed by the buyer's testing, closes that gap. The red flag is a restrictions clause that covers the raw data but says nothing about Derived Data or Outputs.

5. Data description, delivery and acceptance

This clause, and the exhibit behind it, says what you will deliver, when and how. It should name the format, the delivery method (encrypted transfer, never email attachments), the schedule if delivery happens in batches, and an acceptance window: a fixed number of days for the buyer to check the delivery against the description.

Acceptance protects both sides. Without it, a buyer can raise quality complaints months later and hold back payment. With it, a delivery that is not rejected for a stated reason within the window is accepted. The red flag is acceptance based on whether the data is "satisfactory to the buyer," which is not a standard anyone can meet on purpose.

6. Fees and payment

The fee clause sets the structure (one-time, recurring, or per record or per hour delivered) and, as importantly, when money moves. Common triggers are signing, acceptance of each delivery, and each anniversary of the term. Payment terms, late charges and who bears taxes belong here too.

Pricing models and public examples are covered in what your data is worth. The red flag is payment tied to something you do not control, such as the buyer's model reaching a performance target or the buyer launching a product.

7. Term, termination and what survives

The term is how long the buyer's rights last; one to five years is common. Termination covers how either side can end it early, usually for an uncured material breach after written notice and a cure period.

The clause that matters most is what happens at the end. A model trained during the term cannot unlearn the data, so most licenses separate three things:

• Trained Models: commonly the buyer keeps the right to use Models trained during the term.
• Licensed Data: the buyer deletes every copy, including backups on a stated schedule, and certifies the deletion in writing. No new Models may be trained on it.
• Derived Data: deleted along with the Licensed Data, or treated as Licensed Data for every restriction that survives.

Restrictions on re-identification, confidentiality and reproduction in Outputs should survive termination indefinitely. The red flag is an end-of-term clause that mentions only "the data," leaving embeddings and synthetic copies in the buyer's hands with no restrictions attached.

8. Privacy, de-identification and security

Here the seller commits to having de-identified the data according to the documented plan, and the buyer commits to keeping it that way. The buyer's side usually includes access controls, encryption at rest and in transit, access logging, use only in the buyer's own controlled environment, and notice to you within a fixed number of hours or days if the data is exposed. Our security page sets out the baseline we require.

The red flag is a seller warranty that the data "contains no personal information." De-identification lowers risk; it does not make a promise of zero personal information safe to sign. Warrant that the plan in the exhibit was followed, and let the plan describe the methods.

9. Audit and certification

Audit rights let you, or an independent assessor, verify that the buyer is keeping its promises: on reasonable notice, no more than a set number of times a year, and more often after a breach. The certification duties sit alongside: a signed statement from an officer of the buyer at the end of the term confirming deletion. The red flag is an audit right so limited (for example, only the buyer's own self-assessment) that it cannot catch anything.

10. Confidentiality and publicity

Most data licenses keep the deal terms, the data description and the identity of the parties confidential, and neither side may name the other publicly without consent. That is how Knowledge Fuel works too: we never name a buyer, and buyers do not name the companies they license from unless the company agrees.

One wrinkle is regulatory disclosure. Under the EU AI Act, providers of general-purpose AI models must publish a summary of the content used for training, following a template the European Commission published in July 2025, which asks providers to indicate whether they used commercially licensed content. If your buyer is subject to that rule, agree in advance how your data may be described, at what level of generality, and that you will not be named. The Commission's explanatory notice and template shows what these summaries contain.

11. Warranties

Warranties are promises about facts. A seller should warrant only what it knows: that it holds the rights described, that it has the authority to sign, and that the data was prepared under the documented de-identification plan. A seller should not warrant that the data is accurate, complete, or fit for the buyer's purpose, or anything about how the buyer's Models will perform. Those disclaimers are standard and buyers expect them.

The buyer's warranties are about its conduct: that it will use the data only as licensed and in compliance with applicable law.

12. Indemnities and the limit of liability

An indemnity is a promise to cover the other side's costs if a specific kind of claim arises. The usual split is that the seller covers claims that it did not have the rights it warranted, and the buyer covers claims arising from its misuse of the data, its Models or its Outputs.

Buyers care about rights claims for a reason. The Federal Trade Commission has ordered companies to delete models built on data they were not entitled to use: in 2021 it required a photo app developer to delete facial recognition models and algorithms developed with its users' photos, and in 2022 an order required WW International and Kurbo to destroy algorithms derived from data collected from children. A buyer whose model depends on your rights being sound will ask for strong promises about them. That is exactly why the rights review comes before any buyer conversation: so the promises you make are ones the facts support.

The limit of liability caps what either side can owe, often by reference to the fees paid over a period, and lists exclusions that stay uncapped. Reasonable exclusions include breach of confidentiality, attempts to re-identify individuals, and use of the data outside the license. The red flag is a cap that applies to the seller's rights warranty but not to the buyer's misuse, or the reverse.

13. Governing law and disputes

The last clause names the law that governs the agreement and where disputes are heard, by a court or an arbitrator. It rarely changes a deal, but it changes the cost of enforcing one. A seller should know before signing where it would have to go to enforce the restrictions above.

The clauses at a glance

What to have ready before the first draft

The agreement goes faster, and ends up better, when four things exist before anyone drafts it: a rights review of the systems the data comes from, a written description of the dataset, a de-identification plan, and a lawyer who has read the plan. Our guide to licensing your data safely walks through each.

To see where the agreement fits in the process from first conversation to payment, read how it works. To find out what your records could be worth before any of this, start with a free estimate.