- AI companies license data they cannot get from the open web: private, well-structured records from real work, with clear rights.
- Most company data deals are licenses, not sales. You keep the data and grant a defined right to use a copy.
- Rights and privacy checks come before any sample is shared, because they decide what can be offered at all.
- Never give a buyer live access to your systems. Deliver a prepared, de-identified export after the license is signed.
What AI companies actually buy
AI teams have already collected most of the public web. What they lack is data that shows how specialist work is really done: support conversations and their resolutions, code with its history, recordings of people doing hands-on tasks, expert judgments, and documents from niche industries. Our data catalog describes each type, what buyers look for and how it is priced.
Signs your data may be valuable:
- It is not published anywhere online
- It covers a specialist domain, language or task
- It has structure and outcomes (who did what, and whether it worked)
- It spans years, not weeks
- You can show who created it and on what terms
Step 1: Take an inventory without moving anything
List each system that holds data (email, chat, CRM, help desk, code repositories, file shares, recordings), with record types, volumes, date ranges and who created the content. Do not export anything yet. The inventory is what a broker or buyer needs to judge interest. See how to build a data manifest.
Step 2: Check you have the right to license it
Ask four questions for each data set:
- Did we create it, or did customers, users or partners?
- What did our privacy notice and terms say when it was collected?
- Do contracts, NDAs or platform terms restrict it?
- Does a sector law apply (HIPAA for health, GLBA for financial institutions, FERPA for education records, BIPA for biometrics)?
Lawyers advising on these deals recommend involving legal, privacy, security, HR and communications early, and excluding high-risk material such as privileged communications, HR files, health data and third-party confidential data (Frankfurt Kurnit Klein & Selz). Read is selling company data legal?
Step 3: Decide what to exclude and how to de-identify
Mark each data set as included or excluded. For included data, agree a de-identification protocol: what gets removed or replaced, by whom, and how it will be checked. Removal lowers risk but does not by itself make data anonymous, so the license should also forbid re-identification. See de-identification for AI training data.
Step 4: Get a realistic valuation
Price depends on scarcity, quality, volume, exclusivity and rights clarity. Published reference points range from fractions of a cent per message in very large archives to dollars per image or minute of video. Read how to value a dataset and our AI data prices page.
Step 5: Find and vet buyers
Buyers include foundation model labs, specialist AI companies, data vendors that resell to labs, and marketplaces. Each has different terms. A data vendor may resell your data to several developers; a lab may want exclusivity. Ask who will hold copies, for what purpose, and whether resale is allowed. Read questions to ask an AI data buyer.
Step 6: Share a sample under NDA
A buyer evaluates a small, de-identified sample, typically a few hundred records, under a non-disclosure agreement. The sample should be representative of the full set, not the best-looking part.
Step 7: Negotiate the license
Key terms: named licensee, permitted use (training, fine-tuning, evaluation), exclusivity, term and territory, resale and sublicensing, retention and deletion, security, warranties, indemnities and payment. Read what goes in an AI data license agreement.
Step 8: Deliver securely and document everything
Deliver a prepared export by secure transfer. Keep records of what was included, excluded, how it was de-identified and who received it. That record protects you if questions come later.
Mistakes to avoid
- Giving buyers system access or API keys instead of an export
- Sharing samples before a rights review
- Quietly changing your terms to permit AI use after collection
- Agreeing to open-ended "affiliates" language that lets copies spread
- Assuming de-identified means risk-free
How we help
We review your inventory, tell you honestly whether there is buyer interest, help you prepare, introduce vetted buyers and help structure the terms. Sellers pay a success fee only when a deal completes, agreed in writing before any introduction. You sign the license directly with the buyer. List your data or book a consultation.
Questions
Do I sell my data or license it?
Almost always license. You keep ownership and the original data, and grant a buyer defined rights to use a copy for agreed purposes.
How long does a data deal take?
From first conversation to signed license, small deals can close in a few weeks. Larger deals with legal review and de-identification commonly take two to four months.
Is my company too small to sell data?
Not necessarily. Scarcity and domain matter more than size. A small specialist firm can hold more useful data than a large generalist one.
Will buyers want exclusive rights?
Some will, and pay more for it. Non-exclusive licenses let you license the same data to several buyers. See exclusive vs non-exclusive licenses.