AI data license agreements: the terms that matter

An AI data license agreement decides what you’re actually giving up and getting. The terms that matter most are scope, permitted use, exclusivity, retention and deletion, rights in models and outputs, security, payment, warranties and indemnities, and what happens if either side is acquired or shuts down. Two offers with the same price can be worth very different amounts once these are read closely.

This guide explains each term from the seller’s side. It’s general information, not legal advice. Every agreement is different, so work with counsel.

The key terms at a glance

Term	What it controls	What sellers usually want
Scope	Which records are licensed	A precise list of systems, date ranges and exclusions
Permitted use	What the buyer may do with them	Named purposes only; everything else reserved
Exclusivity	Whether you can license elsewhere	Non-exclusive, or exclusive only for a limited time and a price
Term	How long the license runs	A defined end date
Retention and deletion	How long data is kept	Deletion of raw and prepared data, with certification
Models and outputs	Rights in what’s built	Limits on reproducing records and competitive use
Security	How data is protected	Specific controls, access limits and breach notice
Audit	How you verify compliance	Certifications, and audit rights for cause
Payment	Amount and timing	Clear amounts, dates and limited conditions
Warranties	Promises you make	Narrow, knowledge-qualified promises
Indemnities and caps	Who pays if something goes wrong	Mutual indemnities, with your liability capped
Assignment	What happens if the buyer changes hands	Your consent required; deletion on shutdown

Scope: exactly what’s licensed

Scope should be specific enough that a stranger could tell what’s in and what’s out. List the systems, record types and date ranges, and attach the exclusions: legal correspondence, HR and payroll matters, personal tax records, bank statements and anything you flagged. Say whether the license covers a one-time delivery or a recurring feed, and how future records are handled.

Permitted use: training, evaluation or both

This is the most important clause in the agreement. Name the purposes the buyer may use the records for, such as training AI models, evaluating them, or building simulated work environments, and reserve everything else. Then address the edges:

• May the buyer combine your records with other data?
• May it sublicense or resell them, or anything derived from them?
• May it use them to build products that compete with your business or your clients?
• Must it keep records out of public datasets and benchmarks?

Evaluation-only rights are narrower than training rights, since the buyer uses the records to test models rather than to build them. That difference belongs in the price.

Exclusivity and term

Exclusivity stops you from licensing the same records to anyone else. If a buyer wants it, limit it to specific records and a set period, and price it in. A defined term, with a clear end date, gives you a natural point to revisit the arrangement.

Be careful with no-shop clauses in letters of intent. They can stop you from comparing offers before the main terms are settled.

Retention, deletion and derived data

Spell out how long the buyer may keep raw records, prepared datasets and backups, and require deletion or return at the end, with written certification. Cover contractors and cloud copies too.

Be clear-eyed about one limit. As lawyers at Frankfurt Kurnit Klein & Selz note, once data is used to train a model it can’t practically be pulled back out. Deletion clauses protect the underlying records. Exclusions and use limits protect you from what a model might retain.

Rights in models and outputs

The buyer will usually own the models it trains. You can still negotiate protections:

• No outputs that reproduce your records or identify people in them.
• No use of your company name, trademarks or client names.
• A commitment not to attempt to re-identify anyone.
• Limits on using the records to train systems aimed at your market.

These matter because models can reproduce parts of their training data, a risk Forbes highlighted in its reporting on workplace data.

Confidentiality, security and transfer

Require specific security controls rather than a general promise. That means encryption in transit and at rest, access limited to named roles, prompt notice of any incident, and a secure transfer method. If the buyer will receive raw records to de-identify itself, the contract should name the protocol, who may see raw data and when it’s deleted. For what a solid protocol looks like, see how business records are de-identified.

Decide too whether the deal itself is confidential. California’s AB 2013 requires generative AI developers to publish high-level information about their training data, including whether it was purchased or licensed, the Frankfurt Kurnit post notes. Assume the existence of the license may become known even if its terms stay private.

Audit and verification

You need a way to confirm the buyer is keeping its promises. Common tools include annual written certifications, deletion certificates at the end of the term and audit rights triggered by a suspected breach.

Payment structure and timing

Common structures include a lump sum on signing or delivery, installments, payments tied to delivery milestones and recurring fees for ongoing feeds. Whatever the structure, check:

• Exact amounts and due dates.
• Conditions that could reduce or delay payment, such as open-ended quality reviews.
• Who pays for preparation and de-identification.
• Interest or remedies if payment is late.

To judge whether the price itself is fair, see what business data is worth.

Representations and warranties

Buyers will ask you to promise things about the data, most often that you have the right to license it. Keep these promises narrow:

• Limit them to the records actually delivered.
• Qualify them by your knowledge where possible.
• Exclude anything you disclosed during rights review.

The buyer should make promises too: about its security, its compliance with law and its use of the data only as permitted. Privacy and security representations deserve particular attention, the Frankfurt Kurnit lawyers note.

Indemnities and limits of liability

Indemnities decide who pays if a third party makes a claim. Aim for mutual indemnities, where the buyer covers claims arising from its use of the data and you cover a narrow set of claims arising from your breach. Cap your total liability, often by reference to the amount you were paid, and avoid open-ended exposure for consequential damages.

Termination and what survives

Set out when either side can end the agreement, such as for material breach or non-payment, and what happens next. Deletion obligations, confidentiality, use restrictions and indemnities should survive termination.

Assignment, change of control and shutdown

Data outlives companies. Require your consent before the buyer assigns the agreement, and decide what happens if the buyer is acquired. If the buyer shuts down or enters bankruptcy, the records should be deleted, not sold on. That risk is real: the most prominent data sale of 2026 came out of an airline’s bankruptcy, as Business Insider reported.

A pre-signing checklist

Before you sign, confirm that:

1. Scope and exclusions are listed precisely.
2. Permitted uses are named, and everything else is reserved.
3. Exclusivity, if any, is limited in time and paid for.
4. Retention, deletion and certification cover raw data, prepared data and backups.
5. Outputs and re-identification are restricted.
6. Security controls, incident notice and the transfer method are specific.
7. Payment amounts, dates and conditions are clear.
8. Your warranties are narrow and your liability is capped.
9. Assignment requires your consent, and shutdown means deletion.
10. Your counsel has reviewed the final draft.

If a buyer has already reached out, start with our list of questions to ask before you say yes. For the full process, read how to license your business data to AI labs.

Negotiating these terms is a large part of what Cascade does. We compare offers, negotiate price and terms on your behalf and bring you in for the decisions that matter. You approve the terms, and we’re paid a success fee only. Start a conversation to talk it through.

This guide is general information, not legal advice. Have your own counsel review any agreement before you sign.