AI data licensing agreements: what to include and what to watch for
Key terms in an AI training data licence (scope, exclusivity, price, warranties, re-identification bans, termination) and how to negotiate them.
By Databounties Editorial Team · Updated 6 October 2026 · 7 min read
Key points
- You license data for a defined use. You don't sell ownership.
- The terms that matter most are permitted use, exclusivity, term, price structure, warranties and liability caps.
- Always include bans on re-identification, onward sale and use outside the agreed purpose.
- Clarify what happens to trained models when the licence ends. Removing data from a trained model is usually impossible.
The licence agreement decides how much you're paid, what the buyer can do with your data, and how much risk you keep after the sale. Here are the terms that matter for a seller, in order of importance.
General information, not legal advice. Have a lawyer review any agreement before you sign.
1. Permitted use
Define exactly what the buyer can do: train models, fine-tune, evaluate, or all three. Say whether the data can be used to build products that compete with you, and whether outputs that reproduce the data are allowed.
2. Exclusivity
- Exclusive: only this buyer may use the data, usually for a set period. Commands the highest price.
- Non-exclusive: you can license the same data to others. Lower price per deal, more total upside.
- Field-limited exclusivity: exclusive for one use or sector only. A useful middle ground.
3. Term and what happens at the end
Set the licence length, renewal terms and deletion obligations. Most importantly, address trained models. Data generally can't be removed from a model once it's trained on, so contracts usually let the buyer keep using models trained during the term. Price the licence on that basis.
4. Price and payment structure
- A flat licence fee, often paid in stages (on signing and on delivery).
- Refresh fees for ongoing updates (monthly or quarterly data drops).
- Bonuses for additional volume or for meeting quality thresholds.
Benchmarks are in how much is company data worth?
5. Restrictions that protect you
- No re-identification, and no linking with other data to identify individuals or your clients.
- No onward sale, sublicensing or redistribution of the raw data.
- Security requirements, and notification if there's a breach.
- Confidentiality of the deal terms and of your identity, if you want it.
6. Warranties and liability
Buyers will ask you to warrant that you have the right to license the data and that it has been anonymised. Keep warranties specific and factual, tied to a described method, and cap your total liability. Avoid open-ended indemnities.
7. Delivery and acceptance
Agree the format, delivery method, data card and acceptance criteria. Acceptance should be tied to the agreed specification, so payment can't be held up by changing requirements.
Red flags
- "Any purpose, in perpetuity, worldwide" with no matching price premium.
- Exclusivity with no time limit.
- Uncapped liability or broad indemnities.
- Payment entirely on vague acceptance criteria.
Read the full process in how to sell company data to AI companies.
Sources
- 1.A Survey of Machine Unlearning
Nguyen et al., arXiv, 2022
- 2.Data sharing: a code of practice
Information Commissioner's Office (ICO)
- 3.Regulation (EU) 2024/1689 (Artificial Intelligence Act)
Official Journal of the European Union, 2024
Frequently asked questions
Do I lose ownership of my data when I license it?
No. A licence grants permission to use the data under specific terms. You keep ownership and can continue using the data in your business, and you can license it again if the licence is non-exclusive.
What happens to an AI model trained on my data when the licence ends?
Usually the buyer must delete the raw data, but models already trained on it can keep being used. Removing data from a trained model is generally impossible. Price the licence with that in mind, and say explicitly in the contract how trained models are treated.
Should I accept unlimited liability or broad warranties?
No. Limit warranties to things you can confirm (for example that you have the right to license the data and have anonymised it using a described method), and cap liability, typically at the fees paid.
