Data licensing agreements: what they should cover

A data licensing agreement is the contract that lets an AI company use a copy of your company's records on set terms: which data, for what uses, for how long, for what fee, and what happens to the data and trained models at the end. Below we go through each clause. This is general information, not legal advice.

  • Guide
  • By
    Keilian Knudsen
  • 9
    min read
  • Updated
    20 Oct 2026
Two framed white grids on black joined by one bright line of points

Key takeaways

  • A data licensing agreement decides what an AI company may do with a copy of your records, for how long, and what happens at the end. Vague wording causes disputes: a 2025 US lawsuit turns on whether a licence for "internal research purposes" covered training a commercial AI product.
  • The end of the licence deserves the closest reading. Once data has been trained into a model, its influence persists even if the files are deleted, so the agreement has to say whether trained models may be kept.
  • Expect to promise that you own the records and have the right to license them. Ask for a cap on your liability, and check which claims sit outside it.
  • We don't publish a template, because every deal is different. Use the clause checklist below, and have your own lawyer review the agreement before you sign.

What is a data licensing agreement?

A data licensing agreement, or data licence agreement, is the contract behind a data deal. You, the licensor, grant an AI company, the licensee, the right to use a defined copy of your records for stated purposes. You keep the original. Anything left out of the agreement is not part of the deal.

New to the subject? Our guide to AI data licensing explains how licences work, how they differ from a sale and how reported deals are paid. This page goes clause by clause through the document itself.

The same name covers other contracts. A market data licensing agreement, for example, licenses financial price feeds to traders. Here we mean licensing company records, such as Slack and email threads, CRM history, support tickets and documents, to AI companies for training and testing models.

Old contract forms rarely fit this use. Reviewing a 2025 US lawsuit over a data licence signed in 2021, the law firm Proskauer notes that older forms are likely silent on AI, and that both sides should now draft with AI uses in mind.

Which clauses does a data licensing agreement need?

People often search for a data licensing agreement template. We don't publish one. A template written for a market data feed or a software vendor will miss what matters when the buyer is an AI lab, such as what happens to trained models, and every deal sets its own terms. The checklist below is how we read a draft instead. Whatever the starting point, your own lawyer should review every agreement before you sign.

A 2017 practice note by lawyers at Fried Frank and Wilson Sonsini sets out the key points in a data licence: ownership and use, derived data, the scope of the licence and exclusivity, sublicensing, delivery, confidentiality and security, audits, warranties, indemnities, term and termination. AI training adds questions about models that older notes do not cover.

Which clauses does a data licensing agreement need?
ClauseWhat it decidesWhat to check
Data specificationWhich sources, date ranges and record types are licensedAn exact list, plus what stays out, such as data you hold for clients
Permitted useWhat the AI company may do with the copyEach use is named: training, fine-tuning, testing or retrieval. Anything not named is excluded
SublicensingWhether anyone else may use the dataOnly named parties, with the buyer responsible for them
ExclusivityWhether you can license the same records to othersWhich data, which uses and how long the restriction lasts
Term and terminationHow long the licence runs and how it can endYour right to end it for misuse, and what happens if the buyer is sold
End of licenceWhat happens to the copy and to trained modelsA deletion deadline, backups included, written confirmation, and a clear answer on models
Fees and paymentHow much, and when it falls dueWhat triggers payment, the schedule, and what is still owed if the deal ends early
WarrantiesWhat you promise about ownership and consentPromises limited to what you know and can show
Liability and indemnitiesWho pays if a claim arisesA cap, the claims outside it, and protection running both ways
Confidentiality and securityWho may know about the deal and how the data is heldA security standard, breach notice, and whether the buyer may name you

What can the AI company do with your data?

The permitted use clause is the heart of the agreement. It should name each use the buyer may make of your records and treat everything else as excluded. Broad words such as "research", "internal use" or "development" are where disputes start.

The 2025 lawsuit Proskauer describes shows why. A legal research company licensed its database in 2021 for "internal research purposes", with no commercial use, and now alleges the licensee used it to train and run a commercial AI product. The licensee says that use fits the agreement. Proskauer suggests naming specific AI uses, such as pre-training, fine-tuning, retrieval augmented generation and inference, and taking care with catch-all wording like "technologies whether now known or later developed".

Three further points belong in this clause:

  • Sublicensing: whether the buyer may pass its rights to affiliates, contractors or customers. The Fried Frank note says an agreement should either prohibit sublicences or set out exactly which are allowed.
  • Derived data: summaries, labels or datasets the buyer builds from your records. The same note calls their definition and ownership an extensively negotiated part of data licences.
  • Outputs: whether the buyer may show your records, or close copies of them, to its own users. A licence to train is not a licence to publish.

Should the licence be exclusive, and how long should it run?

Most data licences are non-exclusive: the buyer gets its rights, and you stay free to license the same records to others. An exclusive licence shuts out everyone else, sometimes including you. A sole licence sits between the two, keeping other buyers out while you can still use the data yourself. If a buyer asks for exclusivity, pin down what it covers: which records, which uses or fields, and for how long. Our guide to what company data is worth covers how terms like this affect value.

The term clause sets how long the buyer's rights last and what happens at renewal. A licence with no end date that cannot be withdrawn comes close to handing the data over for good.

Termination is where your protection sits if things go wrong. Licensors usually ask for the right to end the agreement if the buyer uses the data outside the permitted scope, breaks confidentiality or fails to keep the agreed security. Read the change of control wording too: if the AI company is bought, does the licence pass to the new owner? The Fried Frank note says the parties should consider addressing this expressly.

What happens to your data and the trained models when the licence ends?

This is the clause most likely to be thin in a first draft. For the copy of your data, the agreement should say whether it is returned or destroyed, by when, whether backups and contractors' copies are included, and whether the buyer must confirm deletion in writing. The UK Information Commissioner's Office (ICO) recommends that data sharing agreements set out how the arrangement ends, including deleting shared data or returning it to the organisation that supplied it.

Trained models are harder. The law firm Goodwin notes that once data is absorbed into a model's weights its influence persists, even if the contract requires the original files to be deleted. So the agreement has to deal with models directly. Options include letting the buyer keep models trained during the term, requiring them to be retrained or retired, or limiting their use after the licence ends.

Deleting a model can be required. In 2021 the US Federal Trade Commission ordered a photo app developer to delete not only users' photos and videos but also the models and algorithms it had built with them. That was enforcement, not a contract, but it shows a model built from data can be ordered destroyed.

Also check which clauses survive the end of the agreement. Confidentiality, payment and indemnity obligations normally should.

How do payment, warranties and liability work?

The payment clause should say how much is paid, what triggers each payment and when it falls due. In data licences, payment may be triggered by delivery, either once at the start or on a schedule. Tie payment to delivery you can prove, and settle what is still owed if the buyer ends the deal early. We don't quote prices here.

Warranties are the promises you make about the data. An AI company will usually ask you to confirm that you own or control the records, that you have the right to grant the licence, and that personal data inside them was handled lawfully. Promise only what you know and can show; qualifiers such as "to the best of our knowledge" help keep them fair. Whether you hold those rights at all is covered in our guide to whether selling company data is legal.

Buyers take this seriously. Goodwin lists training data warranties, and who is liable if they prove wrong, among the most consequential terms in AI model licensing agreements. Google told Reuters in 2024 it would "take immediate action, up to and including termination" of an agreement with a supplier if it found a violation.

Liability and indemnities decide who pays when a claim arises. Expect the buyer to want an indemnity against third-party claims that you had no right to license the data. In return, ask for a cap on your liability, for example linked to the fees paid, and an indemnity from the buyer if it uses the data outside the agreed scope. Check what sits outside the cap: breaches of confidentiality, privacy or security are common exceptions.

How is your data kept confidential and secure?

Your records describe how your business works, so the agreement should treat them as confidential information, with obligations that match the limits on use. The Fried Frank note adds that trade secrets should not lose protection when a time limit on confidentiality runs out.

Security terms should set a standard the buyer must meet, how the data is transferred, and a duty to tell you promptly about any breach. The ICO's guidance on data sharing agreements also covers common security arrangements and procedures for dealing with any breach of the agreement.

Two confidentiality questions are specific to AI deals:

  • Publicity: whether the buyer may name you as a data source, and whether you may name the buyer. Many deals stay private: Photobucket declined to name its prospective buyers to Reuters, citing commercial confidentiality.
  • Public summaries: under the EU AI Act, providers that place general-purpose AI models on the EU market must publish a summary of the content used to train them, on a template from the European Commission that covers private datasets as well as public ones. Ask how your data would be described.

What should you prepare and ask your lawyer before you sign?

A clear picture of your own data makes every clause easier to negotiate. Before a draft arrives, list the sources on offer, their date ranges and roughly what they contain: that becomes the data specification. Then list what stays out, such as data you hold for clients and records that need personal details removed. Our guide to the company data AI companies buy covers what usually has to come out.

Decide your red lines early: uses you will not allow, how long you would accept exclusivity, and whether trained models may be kept. Gather the customer, supplier and employment contracts that touch the records, and name one person who can sign.

This is also where we come in. Before any agreement, our free check tells you whether your records could qualify and what affects their value, with no commitment. When companies sell company data to AI companies with us, nothing is licensed until they've agreed the terms, and they decide which sources, date ranges and uses are included before they sign.

When a draft arrives, take it to your own lawyer with these questions:

  • Does the permitted use clause name each use and exclude everything else?
  • Can the buyer sublicense, and to whom?
  • If the licence is exclusive, which data, uses and period does it cover?
  • Does deletion cover backups and contractors' copies, and what happens to trained models?
  • Are our warranties limited to what we know, and do they match our records?
  • Is our liability capped, and which claims sit outside the cap?
  • What happens if the buyer is sold, and which law governs disputes?

Frequently asked questions

Generic data licence templates exist, but most are written for market data feeds or software services and say nothing about AI training or trained models. We don't publish a template, because every deal sets its own terms. Use the clause checklist on this page to read a draft, and have your own lawyer review every agreement before you sign.

Yes. A non-exclusive licence gives the buyer its rights while you stay free to license the same records to others, and lawyers note that data licences are often non-exclusive. If a buyer asks for exclusivity, the agreement should say exactly which data, which uses and how long the restriction lasts.

Whatever the agreement says. A good one requires the buyer to return or delete its copy by a set date, covers backups and copies held by contractors, and asks for written confirmation. Trained models need their own answer, because deleting the files does not remove what a model has already learned.

A data sharing agreement, in the UK regulator's sense, sets the rules when organisations share personal data: who the controllers are, the purpose, security, retention and deletion. A data licensing agreement is a commercial contract granting rights to use data for a fee. A licence that includes personal data may need terms of both kinds.

It is a contract to use financial market data, such as prices and quotes from an exchange or data vendor, usually on a subscription with limits on users and redistribution. It is a different contract from licensing company records to an AI company, though both need clear terms on scope and use.

Sources
7

Keep reading

Abstract white lattice of lines and points on black, with a bright square at the centre

What is AI data licensing?

Data licensing gives an AI company the right to use a copy of your records for a set purpose. See how it differs from selling and how deals are structured.

A bright white rectangular frame of points around a faint grid on black

Is it legal to sell company data to AI companies?

Ownership, customer and supplier contracts, and how personal data is handled decide it. Read the three questions to settle before selling any data.

Columns of white points rising from left to right on black

How much is company data worth to AI companies?

Depth, years of history and outcomes matter more than size. Learn what drives the value of company data for AI training, then check your own.