# AI data licensing agreement checklist

URL: https://sellmydatatoai.net/guides/ai-data-licensing-agreement-checklist
Reviewed: 2026-11-08

An AI data licensing agreement should define exactly which records are licensed, what the buyer may do with them and with anything trained on them, how personal data is de-identified and certified, when you are paid, when raw copies are deleted, and how far your liability goes. Use this checklist with your own lawyer before you sign any license.

## Key takeaways

- Attach the data manifest and the exclusions list to the contract as schedules, so the licensed data is defined precisely.
- Deal with derived assets separately: cleaned data, annotations, embeddings, evaluation sets, fine-tuned models, weights and outputs.
- Tie payment to written acceptance tests with a deadline.
- Require de-identification certification before each transfer and a contractual ban on re-identification.
- Limit your warranties to your knowledge and disclosed exclusions, and cap your liability.

Most AI data licenses are drafted by the buyer. That is normal, but it means the first draft protects the buyer. This checklist sets out each clause a seller should look for, what a fair position looks like and why it matters, drawing on Nixon Peabody's October 2026 diligence alert and guidance from Frankfurt Kurnit Klein and Selz.

> **Not legal advice** This checklist explains what to look for. It is not advice on your contract. Sell My Data To AI is a broker, not a law firm, and you should take independent legal advice before signing any license.

## What should an AI data licensing agreement include?

An AI data licensing agreement should include at least the 14 clauses below. If a draft is missing one, ask why before negotiating anything else.

| Clause | What to look for as the seller |
| --- | --- |
| Parties and authority | The right legal entities, and signatories who can bind them |
| Licensed data | Defined by an attached manifest, not by a general description |
| Exclusions | A schedule of what is not licensed, and what happens if it turns up |
| Permitted uses | Training, evaluation or fine-tuning, stated precisely |
| Prohibited uses | Re-identification, resale, competing products and any field you want to protect |
| Derived assets | Who owns or may use annotations, embeddings, evaluation sets, models and outputs |
| Exclusivity and term | Scope, length and what exclusivity actually covers |
| Price and payment | Amount, triggers, deadlines and late payment terms |
| Delivery and acceptance | Transfer method, acceptance tests and a deadline for acceptance |
| Personal data | De-identification standard, method, provider and certification |
| Security and deletion | Storage, access controls, retention period and deletion of raw copies |
| Onward transfer | Limits on affiliates, subprocessors and model partners |
| Warranties, indemnities and caps | Knowledge qualifiers, disclosed exceptions and a liability cap |
| Termination | What ends, what survives and what happens to trained models |

## How should the licensed data be defined?

The licensed data should be defined by a schedule that lists each dataset by system, record type, record count, date range and format. A vague definition such as all business records lets a buyer argue later that something was promised and not delivered.

- Attach your [data manifest](https://sellmydatatoai.net/glossary/data-manifest) as a schedule and make it the definition of the licensed data
- Attach a separate exclusions schedule, such as HR files, privileged material and named client matters
- Record file hashes or version numbers so both sides can prove what was delivered
- Agree in advance what happens if prohibited material is found after delivery, which Nixon Peabody recommends
- State whether future records are included, or only the archive at a fixed date

Our guide to [building a data manifest buyers will trust](https://sellmydatatoai.net/guides/how-to-build-a-data-manifest) shows the fields to include.

## What uses should the license permit and prohibit?

The license should permit only the uses you have priced and prohibit the uses you are worried about. Nixon Peabody suggests stating which models, systems and affiliates may use the data, and whether use covers training only or also testing, product development and commercial deployment.

- Permitted: name the purposes, such as pre-training, fine-tuning or evaluation
- Permitted: say whether commercial deployment of trained models is included
- Prohibited: any attempt to re-identify individuals
- Prohibited: resale or sublicensing of the raw data without your written consent
- Prohibited: use to build a product that competes with you, if that matters to your business
- Required: steps to stop model outputs reproducing confidential or personal information, which Nixon Peabody lists as a term to address

## Who owns what the buyer builds from your data?

Who owns what the buyer builds from your data depends entirely on the derived assets clause. Nixon Peabody lists the assets to address separately: cleaned datasets, annotations and embeddings, evaluation sets, fine-tuned models, model weights and outputs.

Most buyers will insist on owning trained models. That is usually reasonable. The items worth negotiating are cleaned and annotated copies of your data and evaluation sets built from it, because these can be resold or reused in ways that work like a second license of your records.

## How should exclusivity and term be written?

Exclusivity and term should be written narrowly enough that you know exactly what you are giving up. An exclusive license fetches about 2.5 to 5 times a non-exclusive one, according to AIData.Broker, so make sure the price reflects it.

- Define what exclusivity covers: all AI uses, a field of use, a type of model or a list of named competitors
- Set a term in years, or confirm in writing that the license is perpetual
- Check for first refusal or matching rights hidden elsewhere in the contract
- Agree whether you may still use the data internally, including for your own AI tools

Read [exclusive vs non-exclusive data licenses](https://sellmydatatoai.net/guides/exclusive-vs-non-exclusive-data-licenses) for the options in between.

## How should payment and acceptance work?

Payment and acceptance should work to a fixed timetable, so the buyer cannot hold payment open by delaying acceptance. Published terms show the range: micro1 takes up to 30 days to review and accept the data and then pays within 30 days of invoice, and Polyshares pays net 30 to 60 days after approval.

- Written acceptance tests based on the manifest, such as record counts, date ranges and file integrity
- A deadline for acceptance, after which the data is treated as accepted
- A deadline for payment after acceptance or invoice
- Milestones for staged deliveries, if the deal uses them
- Interest or a right to suspend delivery if payment is late
- For revenue shares, reporting dates and a right to audit the figures

## What should the license say about personal data?

The license should say exactly how personal data is removed, who removes it and how that is proved. Nixon Peabody recommends that transaction documents set the standard and the process, including an independent de-identification provider informed by NIST guidance, re-identification risk testing against external data and certification before each transfer.

- The de-identification standard, by reference to the law that applies, such as the CCPA's deidentified test, HIPAA's Safe Harbor or Expert Determination method, or UK GDPR for UK records
- Who chooses and pays the provider. Frankfurt Kurnit Klein and Selz recommends the seller selects or jointly selects it and approves the protocol
- Certification before each transfer
- A ban on re-identification, backed by controls, and the same ban passed to any recipient
- Record-level exclusions for privileged or highly sensitive material

Under the CCPA, data only counts as de-identified if the holder publicly commits not to re-identify it and binds recipients to the same, according to Nixon Peabody. See our guide to [de-identification for AI training data](https://sellmydatatoai.net/guides/de-identification-for-ai-training-data).

## What should the license say about security and deletion?

The license should say where your data will be stored, who can see it and when raw copies are deleted. Frankfurt Kurnit Klein and Selz advises against giving buyers direct access to your systems and recommends secure transfer methods.

- A fixed export delivered by an agreed secure method, never live system access
- Named storage locations and access limited to named teams
- A retention period for raw copies and written confirmation of deletion
- Incident notification within a set number of hours or days
- A statement of what deletion means for models already trained, since Nixon Peabody notes deletion may not remove the data's influence

## Can the buyer pass your data to anyone else?

The buyer can pass your data to someone else only if the license allows it, so the onward transfer clause needs to be specific. Nixon Peabody recommends stating whether data may be combined with other datasets or shared with subprocessors or research partners, and notes that restrictions must flow through the buyer's whole technical stack, including cloud and model vendors.

- A list of permitted affiliates, or a ban on sharing outside the licensee entity
- Subprocessors bound by terms at least as strict as the license
- No transfer to model partners or research partners without your written consent
- The buyer remains liable for anything its subprocessors do with the data

Training-data companies such as micro1, Mercor and Scale AI supply AI labs, so onward supply is their business model. With them, the question is not whether data passes on but in what form, with which protections and under whose name.

## How far should your warranties and liability go?

Your warranties and liability should go only as far as you can control. Buyers will ask for warranties on authority, privacy notices, contractual restrictions, IP rights, regulatory compliance, litigation holds, security and completeness of disclosed exclusions, according to Nixon Peabody.

| Buyer asks for | A fair seller position |
| --- | --- |
| Unqualified warranty that no personal data is included | Warranty that the agreed de-identification method was applied, with the provider's certificate |
| Warranty that you hold all rights in all records | Warranty given to your knowledge, subject to the exclusions schedule |
| Uncapped [indemnity](https://sellmydatatoai.net/glossary/indemnity) | Indemnity capped at the fee received, or a multiple agreed in writing |
| Large holdback for a long period | A smaller holdback released on a fixed date |
| Cost of retraining models if a problem appears | Removal of affected records from future use, with retraining costs excluded or capped |

Nixon Peabody notes that buyers may seek remedies covering the cost of removing data or retraining a model. That exposure can exceed the price of the license, so the cap matters more than almost any other term.

## What about confidentiality and publicity?

Confidentiality and publicity clauses should say whether the buyer may name you, and how. California's AB 2013 requires developers of public generative AI systems to post a summary of their training data, including whether datasets were purchased or licensed, according to Perkins Coie. Article 53 of the EU AI Act sets a similar duty for general-purpose AI models offered in the EU.

You cannot stop a buyer meeting a legal disclosure duty, but you can agree how you are described, ban marketing use of your name and get notice before any announcement. The [mutual NDA](https://sellmydatatoai.net/glossary/mutual-nda) you signed for the sample stage should be replaced or extended by these terms.

## What should happen on termination?

On termination, the buyer should stop using your raw data and certify deletion, and the contract should say clearly which rights survive. Trained models usually survive. Confidentiality, the re-identification ban, the security obligations and payment terms should survive too.

Before you sign, read [AI data licensing explained](https://sellmydatatoai.net/guides/ai-data-licensing-explained) for how each deal structure affects these clauses. If you want help comparing offers clause by clause, [book a free consultation](https://sellmydatatoai.net/book-a-consultation?side=seller).

## Frequently asked questions

### Who usually drafts an AI data license?

The buyer usually drafts an AI data license, because it signs many and has standard terms. That means the first draft favors the buyer. You can and should negotiate it. Ask your lawyer to focus first on scope, derived assets, warranties and the liability cap, since those carry the most money.

### Should the data manifest be part of the contract?

The data manifest should be attached to the contract as a schedule and used as the definition of the licensed data. That makes acceptance testable and protects you from claims that more was promised. Attach an exclusions schedule too, listing what is deliberately left out.

### Can I cap my liability under a data license?

You can usually negotiate a cap on liability, and you should. Nixon Peabody notes that buyers may seek remedies covering the cost of removing data or retraining a model, which can exceed the license fee. A common position is a cap at the fee received, though any cap is a matter for negotiation.

### What is a holdback in a data deal?

A holdback is part of the price that the buyer keeps for an agreed period, to cover claims if a problem with the data appears. Nixon Peabody lists holdbacks alongside indemnities, caps and insurance as ways of allocating risk. Keep any holdback small and release it on a fixed date.

### Do I need a data processing agreement as well?

You may need one if personal data is still included. Under the CCPA, a service provider processes personal information on your behalf under a contract that bars it from selling the data or using it for other purposes. A buyer that trains its own models uses data for its own purposes, so a license with de-identification terms is more typical. Your lawyer should decide which applies.

## Sources

- [Shopping for AI training data: five diligence questions before you buy](https://www.nixonpeabody.com/insights/alerts/2026/10/26/shopping-for-ai-training-data-five-diligence-questions-before-you-buy), Nixon Peabody, accessed November 2026
- [Selling company data for AI training: what to know before you say yes](https://technologylaw.fkks.com/post/102o455/selling-company-data-for-ai-training-what-to-know-before-you-say-yes), Frankfurt Kurnit Klein and Selz, accessed November 2026
- [Company data partnerships](https://www.micro1.ai/data-partnerships), micro1, accessed November 2026
- [How to sell your company's data to AI companies](https://ddiy.co/sell-company-data-to-ai/), Don't Do It Yourself, accessed November 2026
- [How much is company data worth to AI? (2026 prices and real deals)](https://ddiy.co/how-much-is-company-data-worth-to-ai/), Don't Do It Yourself, accessed November 2026
- [AB 2013: new California AI law mandates disclosure of GenAI training data](https://legacy.perkinscoie.com/insights/update/ab-2013-new-california-ai-law-mandates-disclosure-genai-training-data), Perkins Coie, accessed November 2026
- [California Civil Code section 1798.140 (definitions)](https://leginfo.legislature.ca.gov/faces/codes_displaySection.xhtml?lawCode=CIV&sectionNum=1798.140), California Legislative Information, accessed November 2026
- [Guidance regarding methods for de-identification of protected health information](https://www.hhs.gov/hipaa/for-professionals/special-topics/de-identification/index.html), US Department of Health and Human Services, accessed November 2026

---

Book a free consultation: https://sellmydatatoai.net/book-a-consultation