A data licensing agreement for AI is the contract that lets a buyer use a defined copy of your company's records to train and test AI models, for a fee, while you keep the originals. The terms that matter most are scope, exclusivity, length, de-identification, deletion, payment and the promises you make about your rights. Published examples run from two-year licenses to perpetual ones; this is general information, not legal advice, so have your own lawyer review any agreement.
On this page
- What are the key terms in an AI data licensing agreement?
- What does the license actually let the buyer do?
- Exclusive or non-exclusive: which is better?
- How long should the license last?
- Who removes personal information, and to what standard?
- What gets deleted, and when?
- When do I get paid, and what can delay it?
- What will I be asked to promise?
- How is liability limited?
- What should I check before signing?
- Questions
- Sources
We use real published terms as examples where we can, mainly micro1's public Data License Master Terms, the most complete buyer template we found published online (micro1, checked October 2026). Data Licensors, the company that publishes this page, also licenses company records and signs its own agreements with owners.
What are the key terms in an AI data licensing agreement?
| Term | What it decides | What to look for |
|---|---|---|
| Scope or data manifest | Which records are covered | A written list of systems, date ranges, teams and exclusions |
| License grant | What the buyer may do | Training and testing allowed, no resale of the raw records |
| Exclusivity | Whether you can license the same records elsewhere | A higher price if exclusive, and a narrow field if possible |
| Term | How long the rights last | A fixed term, or a clear reason for perpetual |
| De-identification | Who removes personal details and how | The buyer's duty, a named standard, and your sample review |
| Acceptance | When the data counts as delivered | A short review window and clear rejection reasons |
| Payment | How much and when | A fixed amount, a due date, and what happens to partial rejections |
| Deletion | What is destroyed and when | Raw export deleted on a set date, with written confirmation |
| Warranties | What you promise | Promises limited to what you know and control |
| Liability and indemnity | Who pays if something goes wrong | A cap on your exposure and a matching promise from the buyer |
| Confidentiality | Whether your name is public | No naming you or your company without consent |
What does the license actually let the buyer do?
Read the grant sentence slowly. micro1's terms grant a "worldwide, royalty-free, fully paid-up, nonexclusive, irrevocable, perpetual license" with the right to sublicense through multiple tiers and to create synthetic or derivative versions. The same terms then set limits: the buyer may not use the data to build a product that directly competes with your business, may not knowingly try to re-identify any person or customer, may not reverse-engineer your pricing or customer relationships, and may not resell the data as a standalone product. It also may not name you as the source without your written consent (micro1, checked October 2026).
What you want is for the limits to be clear and to pass down to anyone who receives the data. micro1's terms require customers who receive the data to be bound by substantially equivalent obligations. Ours require every recipient to sign an agreement that limits use, forbids resale and forbids attempts to identify any person or company (Data Licensors, October 2026).
Exclusive or non-exclusive: which is better?
Neither is better in general. It is a trade:
- Exclusive means you cannot license the same records to anyone else. Labelbox Prism says its licenses are exclusive by default. Polyshares' published case studies are perpetual exclusive licenses, for example $450,000 for a roofing company's operating record (Polyshares, checked October 2026).
- Non-exclusive means you can license again. Endless says its license is non-exclusive, and micro1's template is non-exclusive. SuperAnnotate says it pays on delivery and again on each relicense.
If a buyer asks for exclusivity, ask what it would pay without it. Also check whether exclusivity covers only AI training or every use, and whether it applies to records you create after signing.
How long should the license last?
Terms vary from two years to forever. Grepped's founder said it paid around $500,000 to $1.2 million for two-year licenses from manufacturers with roughly 60 to 130 employees, a founder's claim we could not verify (X, October 2026). Polyshares' case studies and micro1's template are perpetual. Labelbox Prism lists "revocation with 90-day notice" among its terms (Labelbox, checked October 2026).
Perpetual model rights are hard to avoid, since a model cannot easily unlearn what it was trained on. What you can negotiate is how long the buyer keeps the raw records, and whether new uses beyond training need your approval.
Who removes personal information, and to what standard?
The agreement should say who does it, how, and how you check it. Spirit Airlines is the clearest public example: Google won the August 2026 bankruptcy auction for the airline's business records with a $10 million bid (Axios, August 2026), and the court hearing on the sale was moved to November 6, 2026, so the sale is not final (TravelPulse, October 2026). The sale agreement filed on August 14 lets the buyer choose and pay for the de-identification agent, requires the agent to certify its work against the California Consumer Privacy Act standard, and requires "referential integrity," meaning the same person keeps the same placeholder across email, tickets and code (The Next Web, August 2026). In its rival $12.5 million offer, micro1 proposed giving an independent reviewer a 1% sample of the de-identified data before any onward transfer (Fortune, September 2026).
Ask for: the categories that will be removed, the buyer's duty to remove them before any onward transfer, a sample you review before delivery, and a promise to tell you if something identifying is found later. Our guide on how de-identification works explains the process and its limits.
What gets deleted, and when?
There are three things to delete: your raw export, the de-identified copy, and anything built from them. micro1's terms require the buyer to delete the data at the earlier of 120 days after final delivery or termination, and to confirm in writing on request. They also say deletion does not require deleting or retraining models, weights or derivatives built from accepted data, and does not cover routine backups (micro1, checked October 2026). Endless says data is deleted at the end of the agreed term. We delete the original identifiable export within 30 days of delivering the de-identified version (Data Licensors, October 2026).
When do I get paid, and what can delay it?
Look for an acceptance clause. In micro1's terms, the buyer has 30 business days after delivery to reject data with written reasons; silence counts as acceptance, and you invoice only after acceptance. Rejected data is not paid for. A separate clause lets the buyer end the agreement with nothing owed, and get back money already paid, if the data manifest was materially inaccurate or misleading (micro1, checked October 2026).
So describe your records accurately, keep the manifest short and factual, and ask for payment terms in days, not "upon completion." Polyshares publishes net 30 to 60 after approval. We pay within 30 days of signing (Data Licensors, October 2026).
What will I be asked to promise?
Warranties are where owners take on the most risk. micro1's template asks the company to promise, among other things, that it has authority to grant the license, has not granted the same rights to someone else, has all consents or other legal bases needed for any personal information, and that the data contains no protected health information, payment card or bank account data, or children's data. It also asks for a promise, to the company's knowledge, that the data has no AI-generated content unless disclosed (micro1, checked October 2026).
Push for "to the company's knowledge" wording where you cannot check every message, and for the buyer's de-identification duty to sit next to your promises, so a slip in their process is not only your problem.
How is liability limited?
Check the cap and the carve-outs. micro1's terms cap each side's liability at three times the fees under the order form, except for fraud, gross negligence or willful misconduct. They exclude indirect damages, except for those same cases and for indemnification. The company licensing the data indemnifies the buyer for third-party claims about its promises, rights and privacy; the buyer indemnifies for use outside the license (micro1, checked October 2026). Ask your lawyer whether the indemnity falls inside the cap, and whether the buyer carries insurance.
What should I check before signing?
- The scope list matches what you meant to include, and every exclusion is written down.
- Training and testing are allowed; resale, re-identification and competing uses are not.
- You know whether it is exclusive and for how long, and the price reflects it.
- The de-identification duty, standard and sample review are in the contract.
- Raw data has a deletion date and a written confirmation.
- Payment has a fixed amount, a due date and clear rejection rules.
- Your promises are limited to what you know, and your liability has a cap.
- The buyer will not name you publicly without consent.
- An assignment clause covers what happens if either side is acquired. This matters if you may sell your business later; see company data and business valuation.
Before you sign, also confirm you have the right to license at all. Our guide to whether selling company data is legal covers ownership, employees, customers and privacy law. For who offers these agreements, see AI data partnership programs and the Grepped and micro1 profiles.
