Questions to ask an AI data buyer before you license your data

A data license is easier to negotiate when you have asked the right questions before the first draft arrives. The answers tell you whether a buyer is serious, what it plans to build and which terms it will fight for.

These 25 questions are grouped in the order most deals settle them. Each one explains why it matters and what a good answer looks like. Use them alongside our AI data licensing agreement checklist, and ask the buyer to answer against the rows of our data license term sheet.

What answers should worry you most?

The answers that should worry you most are refusals to name the signing entity, open-ended rights to share your data and no fixed date for deletion or payment.

What should you ask about the buyer's identity and intended use?

Ask about the buyer's identity and intended use first, because every later term depends on who receives the data and why.

1. Which legal entity will sign, and who owns and controls it?

Ownership decides whether the deal is lawful. The Justice Department's Data Security Program bars data brokerage of bulk US sensitive personal data with China, Cuba, Iran, North Korea, Russia, Venezuela and covered persons linked to them. A good answer names the registered entity, its parent, its beneficial owners and the signatory.

2. What exactly will you use the data for?

Pre-training, fine-tuning, evaluation and retrieval at inference are different uses with different value. A good answer lists the permitted uses, so the license can prohibit everything else.

3. Which models or products will it go into, and will they be sold?

Commercial deployment is worth more than internal research, and you may want to exclude a product that competes with yours. A good answer names the models or model families and says whether outputs will be commercialized.

4. Are you buying for your own use or for someone else?

A reseller or marketplace can relicense your records to buyers you never approved. Under New Jersey's 2026 data broker law, selling your consumers' data to a data broker can also make you a registered data collector. A good answer confirms own use, or names the end buyer.

What should you ask about rights and exclusivity?

Ask about rights and exclusivity before price, because they set what you are giving up.

5. Do you need exclusivity, and how narrowly can it be defined?

Exclusivity stops you licensing the same records again, so it should be priced and limited. A good answer limits it by field of use, model type or named competitors, for a fixed period. Our guide to exclusive vs non-exclusive data licenses compares the trade-offs.

6. What term do you want, and why?

Buyers often ask for perpetual rights because a trained model cannot give data back. A good answer separates a fixed term for raw data from a perpetual right to keep models already trained.

7. Can I keep using my own data, including for my own AI tools?

An exclusive license drafted loosely can stop you using your own archive. A good answer is a written yes, stated as a reserved right in the license.

What should you ask about derived assets and model weights?

Ask about derived assets and model weights because they outlast the license and can carry your data forward.

8. Who owns cleaned, annotated or embedded copies of my data?

A cleaned and labeled copy can be worth more than the original, and it can be licensed on. A good answer lets the buyer use derived copies only within the permitted uses and never license them to others.

9. What happens to models trained on my data when the license ends?

Most buyers keep trained models. The FTC has ordered companies to delete models built on improperly used data, including Everalbum in 2021 and Rite Aid in 2023. A good answer states what survives termination and what happens if a regulator orders deletion.

10. Will you share weights or fine-tuned models with partners?

Releasing weights or handing them to a model partner spreads what was learned from your data beyond the licensee. A good answer names any partners or says the weights stay in-house.

11. How will you stop the model repeating my data word for word?

Verbatim output can leak confidential text or personal data. A good answer describes deduplication, output filtering and testing for memorized content.

What should you ask about security?

Ask about security in writing, because the buyer will hold your raw data for months.

12. Where will the data be stored, and who can access it?

A good answer names the cloud region or data center, limits access to named teams and logs that access. Copies on personal laptops should be ruled out.

13. What independent security assurance can you show?

A buyer's own claims are weaker than an independent report. A good answer offers a current SOC 2 report, which under AICPA criteria covers security, availability, processing integrity, confidentiality and privacy, or an equivalent certificate.

14. How fast will you tell me about a breach?

You may have your own duties to notify customers or regulators. A good answer commits to notice within a fixed number of days and to sharing what was exposed.

Ask about retention and deletion because raw copies are what can be leaked, resold or misused after the training run.

15. When will raw copies be deleted?

A good answer gives a fixed date or event, such as 90 days after the last training run, and covers backups and subprocessors.

16. How will you prove deletion?

A promise is hard to check. A good answer offers a signed certificate naming the systems cleared and the method, for example one following NIST SP 800-88 Rev. 2, the media sanitization guidelines NIST published in September 2025.

17. Will you try to re-identify anyone or combine my data with other sets?

Combining datasets is the usual route to re-identification. The CCPA's de-identification standard requires contract terms binding recipients not to re-identify. A good answer accepts a ban on re-identification and on linking without your consent.

What should you ask about payment and acceptance?

Ask about payment and acceptance together, because a buyer who controls acceptance controls when you are paid.

18. When is each payment due, and what triggers it?

A good answer ties each payment to a date or a delivery, such as 50% on signing and 50% within 30 days of delivery, in US dollars. Our guide to tax on AI data license payments covers withholding.

19. How does acceptance work, and how long do you have?

A good answer sets a review period, accepts delivery against the data manifest and treats silence at the end of the period as acceptance.

20. Can you show you are able to pay?

For a large deal, a good answer offers financial statements, a parent guarantee or payment into escrow before delivery.

What should you ask about audit and warranties?

Ask about audit and warranties to set how much risk you carry after signing.

21. What warranties do you want from me, and what is the cap?

A good answer limits your warranties to facts within your knowledge, such as ownership and de-identification method, with liability capped at the fees paid.

22. Will you indemnify me for your use of the data and your model's outputs?

Claims about model outputs arise from the buyer's choices, not yours. A good answer accepts an indemnity for use outside the license and for output claims.

23. Can I check that you are keeping to the license?

A good answer allows an annual written compliance statement and an independent audit on reasonable notice if a breach is suspected.

What should you ask about publicity?

Ask about publicity because AI developers increasingly have to describe their training data in public.

24. Will you name me or describe the data in public?

California's AB 2013, in force since January 1, 2026, requires generative AI developers to post whether training datasets were purchased or licensed, and Article 53(1)(d) of the EU AI Act requires a public summary of training content. A good answer explains how you will be described and gives you approval over any press release.

What should you ask about termination?

Ask about termination last, because it tests every earlier answer.

25. What happens if either side breaches, or you stop paying?

A good answer lets you end the license for non-payment after a short cure period, ends any exclusivity at once, requires deletion of raw data and lists the terms that survive, such as confidentiality and the re-identification ban.

We ask buyers the first set of these questions for you. Every buyer is verified before it sees a sample: company identity, intended use, security and retention commitments, onward transfer limits and, for larger deals, ability to pay. You approve every buyer and every term. Read how we vet, or request a free data valuation.