Guide
How to sell data to AI companies.
A practical guide for companies that hold valuable data: what AI companies buy, what has to be cleared before anything changes hands, and how a sale is structured, documented and protected.
WhatsApp and Telegram +44 7563 115980
[email protected]
A practical guide for companies that hold valuable data: what AI companies buy, what has to be cleared before anything changes hands, and how a sale is structured, documented and protected.
Since 2013 we have helped firms with due diligence and data services, in more than 60 countries.
The short answer
Companies sell data to AI companies by licensing it. A licence is a written agreement that lets the buyer use a defined dataset for named purposes, for a set period, in return for payment, while the company keeps ownership. AI companies pay most for data they cannot collect from the public web, that is cleared for use, free of personal information and well documented. Before a sale, the company confirms it has the right to license the data, removes personal and sensitive details, tests that no one can be identified, documents the dataset and then agrees the licence terms.
On this page
An AI model is software that has learned patterns from examples. Training is the process of showing it those examples. The buyers of training data fall into four groups.
The US Copyright Office reported in May 2025 that licensing agreements for AI training are fast emerging, and that direct licensing is most common with companies that hold catalogues of high-quality, easily identified content. In other words, buyers prefer to deal with an owner who controls a large, well-organised body of material and can grant the rights in one agreement.
Public web content is open to every model builder, so it gives none of them an edge. What buyers cannot collect for themselves is the material that only exists inside companies: the records people create while doing real work. Research published in Nature in 2024 found that training models on text generated by other models degrades them over time, and concluded that data about genuine human interactions will become increasingly valuable.
| Type of data | Why buyers value it | What must be removed first |
|---|---|---|
| Communications | Emails, chats, call transcripts and meeting notes show how people explain, negotiate and resolve problems | Names, contact details, signatures and anything that identifies a person in the text |
| Documents | Contracts, reports, policies and forms show expert writing, with versions that show how drafts improve | Client names, confidential terms and material owned by third parties |
| Workflow records | Tickets, cases and approvals follow a task from request to outcome, which teaches models to plan and act | Customer and staff identifiers, account numbers and free-text personal details |
| Structured data | Tables, logs and transaction records give models facts, numbers and cause and effect | Direct identifiers, rare combinations of values and precise locations |
| Labelled examples | Records with descriptions, categories or quality ratings attached save the buyer the cost of labelling | The same, plus any labels that reveal who a record is about |
Context raises value. Shutterstock reported that revenue in its data, distribution and services line grew to $203.3 million in 2025, driven mainly by the sale and delivery of metadata, meaning the descriptions and labels attached to each item. Buyers pay for data they can use without rebuilding it.
Yes, where four conditions are met. The company must own the data or hold the rights to license it. Personal information must be removed or handled under a lawful basis. Contracts with customers, suppliers and staff must not prohibit the use. And the licence must respect copyright and database right in the material.
Personal data. In the UK and the EU, anonymous information is outside data protection law, but the act of anonymising personal data is itself a use of that data. The Information Commissioner's Office (ICO, the UK privacy regulator) confirms that anonymisation counts as processing, so the company needs a lawful basis and must tell people in its privacy notice. Data where names are only swapped for codes is pseudonymised, and remains personal data. Our guide to anonymisation, pseudonymisation and de-identification sets out the difference.
United States. Under the California Consumer Privacy Act, deidentified information is not personal information, but only if the business takes reasonable measures to prevent linking it to a person, publicly commits not to re-identify it, and binds every recipient by contract to the same terms. Health information follows the separate rules covered in our guide to HIPAA de-identification.
Data broker laws. Selling personal information about people you have no direct relationship with can make a company a data broker under state law. New Jersey goes further: its 2026 law also requires data collectors, meaning businesses that sell or license personal data of their own customers to a data broker, to register. The first New Jersey registration period runs from 1 April to 30 June 2027. Our data broker registration page covers each state. Data that is properly de-identified before sale falls outside these regimes, which is why de-identification comes first.
We clear the rights and the privacy position, de-identify the data on your own systems and prepare the records a buyer will ask for.
Speak to an expertData de-identificationOur data de-identification service runs steps two to six inside the owner's environment, so raw records never leave the company.
Most deals are licences, not outright sales. The owner keeps the data and grants defined rights to use it, which means the same dataset can earn more than once if the licence is non-exclusive.
Our guides to AI data licensing deals and the AI data licensing agreement cover reported deal values and contract terms in detail.
Serious buyers run due diligence, meaning a structured check of the data and the rights behind it, before they sign. In the EU, providers of general-purpose AI models must publish a summary of their training content, and the European Commission's template asks whether licensing agreements were concluded with rights holders and which types of content they covered. Buyers therefore want evidence that every dataset is licensed and documented.
Sellers who prepare these records in advance close faster and negotiate from a stronger position. The buyer's side of this process is set out in our AI training data due diligence checklist.
Yes, where it owns the data or holds the rights to license it, personal information has been removed or is handled lawfully, and no contract prohibits the use. Most companies license data rather than sell it outright, so they keep ownership.
AI labs that build models, companies that fine-tune existing models for specific tasks, teams that test models before release, and data brokers and marketplaces that supply data on to model builders.
Private data that is not on the public web and shows real work: communications, documents, workflow records, structured records and labelled examples. Data with clear rights, no personal information and good documentation commands the highest prices.
There is no public price list. A few listed companies have disclosed figures, such as Reddit's data licensing arrangements with an aggregate contract value of $203.0 million over two to three years. Most agreements do not disclose financial terms.
Not if the data is truly anonymous, but anonymising personal data is itself processing and needs a lawful basis and notice. Customer contracts may also restrict use, so they must be checked before any sale.
It can if the data identifies people. Several US states require data brokers to register, and New Jersey also requires businesses that sell their own customers' personal data to a data broker to register. Properly de-identified data falls outside these laws.
License it. A licence lets you keep ownership, limit the uses, set a term and license the same data again if the first licence is non-exclusive. An outright sale gives all of that away.
The licence decides. It should require the buyer to delete or return copies and state whether models already trained on the data may continue to be used.
Each point of law and each reported figure on this page was checked against these sources on 27 October 2026. This page is general information, not legal advice.
Tell us about the data you hold and what you would like to do with it.