AI data licensing agreements: a clause-by-clause guide for sellers

By Cody Houser, Founder · Last updated October 28, 2026

The short answer: an AI data licensing agreement gives a buyer defined rights to use your company’s records to train, fine-tune, or evaluate AI. As the seller, the clauses that matter most are the exact scope of data, each permitted use named separately, whether trained models survive termination, who de-identifies the data, deletion with certification, and how far your warranties and liability reach. Have counsel review the buyer’s draft.

What a data licensing agreement is, and why the wording matters now

In most deals the buyer sends its own draft. That draft decides what you are actually giving up, so it pays to read it as carefully as the price. Frankfurt Kurnit, which reports a wave of client questions from companies approached to sell internal data, says the agreement “should restrict how the buyer can use the data, allocate privacy risk, and address what happens if something goes wrong.”

Vague terms cause real disputes. Proskauer describes a 2025 federal lawsuit, Fastcase v. Alexi, over a 2021 data license limited to “internal research purposes.” Fastcase alleges the licensee used the data to train and run a commercial generative AI product; the licensee says that use was consistent with the agreement. The lesson Proskauer draws: name specific AI uses, such as retrieval-augmented generation (RAG), fine-tuning, training, pre-training, and inference, rather than relying on general terms.

Clause-by-clause: what to look for as the seller

Use this table with the buyer’s draft open. A red flag marks a term to raise with your lawyer.

Clause	What it does	What to look for as the seller	Red flag
Definitions and scope	Names the systems, record types, date ranges, and volume covered, and what is excluded.	A dataset schedule you wrote or approved, with exclusions listed by category.	“All company data” or scope the buyer can expand later by notice.
License grant (training, fine-tuning, evaluation, RAG)	Lists what the buyer may do with the data.	Each use named separately: pre-training, fine-tuning, evaluation, retrieval-augmented generation, display to end users.	“Any purpose,” or uses defined only by vague words like “internal research.”
Exclusivity	Says whether you may license the same or similar data to others.	Non-exclusive by default; any exclusivity limited by data set, field, and time, and priced.	Open-ended exclusivity over “similar” data, which can block future deals.
Term and termination	Sets how long the license runs and how either side can end it.	A fixed term, termination for breach, and clear rights on change of control.	Perpetual, irrevocable rights with no termination for misuse.
Trained-model survival	Says whether models already trained on your data can keep being used after the license ends.	An explicit answer, and any limits on new training after termination.	No clause at all, leaving the question to a later dispute.
Deletion and certification	Requires the buyer to delete or return the dataset and copies.	A deadline, coverage of backups and contractors, and a signed certificate.	Deletion “where commercially reasonable” with no proof.
De-identification standard and who performs it	Sets how personal and confidential information is removed, by whom, and how it is checked.	A named standard, the buyer’s team or its third party doing the work, written confirmation it was done, and a ban on re-identification.	No named standard, or no written confirmation that the work was done.
Security	Sets how the data is stored, accessed, and transferred, and breach notice.	Specific controls, a secure transfer method, and no buyer tools inside your systems.	Generic “industry standard” language with no notice deadline.
Sublicensing and affiliates	Controls who else may receive or use the data.	No sublicensing without consent; named affiliates and contractors bound by the same terms.	Free transfer to affiliates, partners, or “successors” without limits.
Representations and warranties	Your promises about your rights in the data and how it was collected.	Promises limited to your knowledge and to the dataset as delivered.	Absolute warranties that the data infringes nothing and contains no personal information.
Indemnities	Who pays if a third party sues.	Mutual indemnities: you for breaches of your warranties, the buyer for its models and outputs.	You indemnify for anything the buyer’s model does or outputs.
Limitation of liability	Caps what each side can owe.	A cap tied to the fees paid, with matching carve-outs on both sides.	A cap on the buyer only, or uncapped seller liability.
Payment and acceptance	Sets the fee, timing, and any acceptance test for the dataset.	Objective acceptance criteria, a short review window, and deemed acceptance after it.	Payment only after open-ended “satisfactory” review.
Audit and reporting	Lets you check how the data is used.	Usage reports, a right to audit or certify compliance, and notice of breaches of the use limits.	No way to verify anything after transfer.
Confidentiality and publicity	Covers the deal terms and whether the buyer may name you.	No use of your name without consent, consistent with the buyer’s legal disclosure duties.	A buyer right to publicize you as a data source.
Governing law and disputes	Picks the law and forum for disputes, and sometimes where data may be processed.	A forum you can realistically use and limits on where data is stored.	A distant forum plus mandatory arbitration with costs on you.

The clauses that carry the most risk

The license grant. Proskauer recommends separately identifying AI uses and, where needed, distinguishing generative uses that create content from machine-learning uses that only detect patterns. Watch for “technologies now known or later developed” language: Proskauer notes its placement can change which rights it extends.

Trained-model survival. According to Frankfurt Kurnit, once data is used to train a model, it cannot practically be pulled back out, and models can reproduce portions of their training data in outputs to other users. Deletion clauses cover the dataset, not what the model learned. Decide what you’re willing to have survive before you sign.

De-identification. Buyers handle this, either with their own team or, more often, through an independent third party they bring in. Sellers aren’t expected to redact records themselves. Frankfurt Kurnit advises sellers to have a say in the provider and protocol and to require certification; at minimum, get the standard and written confirmation into the agreement. It also warns that de-identification addresses whether a record links to a person, not whether its content is confidential: a pay complaint stays sensitive without a name on it.

Representations and warranties. Your promises about your rights in the data are where most seller liability sits. Email and chat contain other companies’ confidential information under NDAs, and vendor-generated data may come with competing ownership claims. Even extraction can be restricted: Slack’s API terms say an app provider may not “use API Data to train a large language model.” Know what you can truthfully promise before you promise it.

The legal context buyers and sellers work in

California privacy law. Under the CCPA, Cal. Civ. Code § 1798.140, a “sale” includes making a consumer’s personal information available to a third party “for monetary or other valuable consideration.” Data counts as “deidentified” only if the business takes reasonable measures against re-association, publicly commits not to re-identify it, and contractually binds recipients to the same. A good agreement carries those obligations through to the buyer. See can companies sell employee data.
Privacy promises. The FTC has warned that it may be unfair or deceptive for a company to start using data for AI training or sharing it with third parties and tell people only through a “surreptitious, retroactive amendment” to its terms or privacy policy.
California AB 2013. By January 1, 2026, and before each new release, developers of generative AI systems released on or after January 1, 2022 and made available to Californians must post documentation of their training data, including the sources or owners of datasets and “whether the datasets were purchased or licensed by the developer.”
EU AI Act. Article 53(1)(d) requires providers of general-purpose AI models to publish “a sufficiently detailed summary about the content used for training,” using a template from the AI Office. Recital 107 says that summary may list the main data collections, “such as large private or public databases or data archives.” These general-purpose model rules apply from 2 August 2025.

The practical effect: your data may be described publicly at a high level even if the contract is confidential. Negotiate how you are named, and plan your own communications with that in mind.

Selling vs. licensing business data

People often say they want to “sell their data” to an AI company. Most of these deals are actually licenses, and the difference decides how much control you keep.

Term	License	Outright sale
What the buyer gets	Specified rights for specified uses	Broader rights in the material
Conditions on use	Set out in the agreement: uses, sharing, security	Few or none after transfer
Duration	Usually a defined term	Usually permanent
When it ends	Deletion or return, subject to model survival terms	Nothing comes back
Licensing to others	Possible if non-exclusive	Generally not, for the same material

Either way, the label on the agreement matters less than its terms. A “license” that is perpetual, irrevocable, exclusive, and sublicensable for any purpose works much like a sale. Read every proposal for what it actually grants.

Licensing data is also separate from selling your company. You keep the business, and with a non-exclusive license you keep the records too. Companies that are winding down face a different set of choices, covered in licensing data when your business is closing.

Before the agreement: what to settle first

The agreement can only protect data you had the right to license. Frankfurt Kurnit’s practical steps start well before drafting: assemble legal, privacy, security, HR, and communications; inventory systems, record types, volume, and time period; confirm your rights under customer, partner, vendor, and platform terms; and exclude high-risk data entirely, such as privileged material, HR and personnel files, payroll and tax records, health and benefits information, and union-related communications. Don’t share sample data before that review; a sample itself can be a disclosure.

If you’re still at the start, read how to sell data to AI companies and run through the data value checklist first.

How to negotiate a data licensing deal

Write the dataset schedule yourself. Describe what you’re offering by system, date range, and exclusion, and make the agreement reference your schedule rather than the buyer’s request.
Trade scope for price, not the other way round. Wider uses, exclusivity, and model survival are worth more to a buyer. If you grant them, they should show up in the economics.
Ask what “training” means. Pre-training, fine-tuning, evaluation, and RAG create different exposure. RAG and display to end users can surface your actual text; Proskauer flags display to end users as a question of “publication.”
Get de-identification in writing. The buyer’s team or its third party does the work; the agreement should name the standard and confirm the work was done.
Match warranties to what you know. Knowledge qualifiers, a defined dataset, and a liability cap tied to fees are reasonable asks.
Get something you can verify. Deletion certificates, usage reports, and breach notice are the only way to know the terms are kept.
Document the decision. Frankfurt Kurnit notes that a prior data sale will come up in M&A diligence, and buyers will expect to see the record.

Price is only one variable. See how much company data is worth for what drives it, and marketplaces vs. brokers for how deals reach you. Hyperdrive Data is a broker: the buyer’s proposed agreement sets the transaction terms, and you review it and decide whether to proceed.

Questions people ask

What is an AI data licensing agreement?

It’s the contract under which a company gives an AI developer or data supplier defined rights to use its records, for example to train, fine-tune, or evaluate models. It sets which data is covered, what the buyer may do with it, how long, what happens when it ends, who carries which risks, and how and when the seller is paid.

Is there a data licensing agreement template I can use?

We don’t publish one, and you should be wary of any generic template. The terms that matter most, such as which uses are permitted, whether trained models survive, who de-identifies the data, and what you warrant, depend on your records, your contracts, the privacy notices you gave, and the laws that apply to you. In most deals the buyer sends its own draft. Have a lawyer who handles data and technology deals review or draft it.

Is it better to sell or license company data for AI training?

For most operating companies a license gives more control: the buyer gets specific rights, for specific uses, with conditions on handling and deletion. An outright sale transfers broader rights with fewer conditions. The label matters less than the terms, so read what each proposal actually grants.

Can I get my data back or have it removed from a model after the deal ends?

You can require the buyer to delete the dataset and certify that it did. Removing what a model has already learned is a different matter: Frankfurt Kurnit notes that once data is used to train a model, it cannot practically be pulled back out. That’s why the survival clause for trained models matters so much.

Will anyone know my company licensed its data?

Possibly. California’s AB 2013 requires generative AI developers to publish a high-level summary of their training datasets, including their sources or owners and whether they were purchased or licensed. The EU AI Act requires general-purpose model providers to publish a summary of training content. Negotiate how you are described, but don’t assume the deal stays secret.

Who should de-identify the data, me or the buyer?

Usually the buyer, through its own team or, most often, an independent third party it brings in. Buyers don’t want sellers redacting records themselves. Your part is to flag categories that can’t be shared at all, and to make sure the agreement names the standard, bans re-identification, and confirms in writing that the work was done.

Sources

This guide is general information, not legal, tax, or financial advice. Review any proposed agreement with your own advisers.

Related guides

Check whether your company is a fit. Start with a description of your records; no files needed.