AI data licensing agreement: a seller’s checklist

This is a plain-English checklist, not legal advice. Have a lawyer review any agreement before you sign.

Why the wording matters more than the price

A data license looks simple: you hand over a copy, they pay. The trouble is that AI changes what “use” means. Once data trains a model, that model keeps what it learned, even after the raw files are deleted. So the agreement has to settle, up front, every way the buyer is allowed to benefit from your data.

A live case shows what happens when it doesn’t. Fastcase licensed its legal database to Alexi in 2021 for “internal research purposes” and barred commercial or competitive use. In November 2025 it sued, alleging Alexi used the data to train and power a commercial AI legal research product (LawNext). A law firm’s analysis noted the outcome may turn on how broad words like training, internal research and distribution are read (Proskauer). Alexi disputes Fastcase’s reading of the license and has countersued (BetaKit). Whatever the outcome, both companies ended up in court over what a few words in a 2021 license meant. Precise terms are cheaper than a lawsuit.

The seller’s checklist at a glance

The sections below explain each term and what to watch for.

1. Scope and permitted use: training, RAG or evaluation

This is the core of the deal. Name each use you allow, and say that everything not named is barred. The common ones:

• Training and fine-tuning. Your data changes the model’s weights. Permanent, so it should carry the highest price.
• Retrieval-augmented generation (RAG). Your data sits in an index the AI searches when it answers. Your text can appear in its output, so limit how much it may quote and who sees it.
• Evaluation. Your data is used to test whether a model can do real work. Lower risk, often a narrower license.
• RL environments and tasks. Your records become practice exercises for AI agents. Say whether the tasks built from your data count as your data under the agreement.

Also define what you’re licensing: which systems, folders and date ranges. Carve out sensitive customers and projects by name.

2. Exclusivity

An exclusive license means only this buyer may use the data, often for years. That can be worth it, but only for a clear premium. Ask for non-exclusive by default. If you grant exclusivity, limit it to a short period, a field (for example, coding models only) or a type of buyer, and write down when it ends. Without that, one deal can lock up an asset you could have licensed several times.

3. Term, deletion and what survives

Set a fixed term. Then accept the hard truth: models trained during the term stay trained afterwards. Price that in, and put your effort into deletion of everything else:

• The raw copy, every working copy and backups, within a set number of days of the end of the term or of training.
• A written certificate of deletion signed by someone accountable.
• Clear treatment of derived files: indexes, embeddings, cleaned datasets and tasks built from your records.
• Survival: the deletion duty, confidentiality and unpaid fees continue after the agreement ends.

4. Sublicensing, resale and who may receive your data

Many buyers are vendors, not labs. They package data and sell it on. That’s fine if you know about it. Ask for no sublicensing or resale without your written approval of each recipient or category of recipient, and a regular report of who received it.

Then add a blocklist. Name competitors who can never receive the data, directly or through a reseller. Some sellers also exclude certain countries or labs. That concern is real: Forbes reported that several large U.S. data vendors sell training data to Chinese labs, which spend roughly $500 million a year with U.S. vendors (AI Weekly). An AI data consultant told Forbes:

You can’t control what you didn’t write down.

5. Payment structure

Most offers fall into one of three shapes, or a mix:

• Flat fee. One price for a defined scope. Simple and certain.
• Installments or milestones. Paid in parts, for example at signing, delivery and acceptance.
• Revenue share. A percentage of each resale. Labelbox, for example, says its Prism program pays an upfront fee at signing plus a share of downstream use (Labelbox).

To compare offers, turn each into four numbers: guaranteed cash, likely extra from revenue share, days until you’re paid, and what you give up in exclusivity or control. The biggest headline isn’t always the best deal.

Watch the payment trigger. “Paid on acceptance” is only safe if acceptance has written, testable criteria and a deadline, with the delivery accepted automatically if the buyer doesn’t object in time.

6. Warranties and indemnities about rights

Buyers care more than ever about where data came from. In the Anthropic book case, a court found training on lawfully bought books was fair use but keeping pirated copies was not, and Anthropic settled for $1.5 billion (NBC News). In the EU, makers of general-purpose AI models must now publish a summary of their training content (Mayer Brown). So buyers will ask you to promise you have the right to license what you deliver.

Keep those promises narrow and true:

• Warrant that you own the data or have the right to license it, and that you can sign. Don’t warrant that it’s complete, accurate or fit for any purpose.
• Check how the data leaves each tool. Slack’s API terms, for example, bar using API data to train a large language model (Hunton), so the export method matters.
• Cap your indemnity, ideally at the fees paid, and make it mutual. The buyer should cover claims from how it uses the data.

7. De-identification and a ban on re-identification

Agree the redaction standard in writing before any data moves, and test it on a sample. Removing names is a start, not the finish. A rare job title, a city and a date can still point to one person, so review whole records and leave out what can’t be cleaned.

Then make the buyer promise not to re-identify anyone. The U.S. Federal Trade Commission’s privacy framework treats data as de-identified only if a company takes reasonable steps to de-identify it, publicly commits not to re-identify it, and contractually bars anyone it shares the data with from trying (FTC). Ask for the same ban, and make it flow down to every sublicensee. Your lawyer can tell you which privacy laws apply to the people in your data and whether you also need a data processing agreement.

8. Audit rights and reporting

If you can’t check it, you can’t enforce it. Ask for regular statements showing how the data was used, who it was resold to and what revenue share is owed. Add the right to audit those statements, through an independent auditor if the buyer prefers, with the buyer paying if the audit finds a meaningful shortfall. Tie the deletion certificate to the same reporting cycle.

9. Security

Your data will live on the buyer’s systems, and maybe a reseller’s. Ask for encryption in transit and at rest, access limited to named teams, logs of who touched it, and a short deadline to tell you about any breach. If you’re giving the buyer access rather than an export, make it read-only, narrow in scope and time-limited. Prefer your own admin exports where you can.

10. Governing law and disputes

The agreement will name whose law applies and where disputes go. That choice affects how hard it is to enforce deletion or collect a late payment, especially with a buyer in another country. There’s no universal right answer, so this is one to settle with your lawyer.

Before you sign

1. Know what you’re licensing. Inventory the systems and years, and confirm you have the right to share each one. Our guide on how to sell company data to AI companies covers the rights to check.
2. Get offers in writing, with the price, payment trigger, exclusivity and permitted use spelled out.
3. Raise the big three on every deal: exclusivity by default, vague acceptance and unlimited resale.
4. Have a lawyer review it. Especially warranties, indemnities, privacy duties and governing law.

With Smelted, you approve the buyer and the terms, can rule out competitors, and the agreement sets when the buyer must delete the raw data. Nothing is upfront: we’re paid from the license. Start with our 60-second estimate to see what your data could be worth.

Questions founders ask