How to sell your data to AI companies

Almost every organisation produces data as a by-product of its work: recorded calls, scanned documents, inspection videos, clinical notes, support tickets. For AI teams, some of that material is hard to find anywhere else. That makes it worth money.

This guide is for data owners: companies, hospitals, media groups, research institutes and agencies that are wondering whether their data has value for AI, and how to turn it into revenue without losing control of it. It is based on what we see across more than 3,000 commercial dataset listings on DataToLab.

The short version: you rarely sell data. You license it, for specific uses and for a set period, and you keep ownership. The work is in three places: proving you have the right to license it, making it easy to evaluate, and agreeing clear terms.

1. Is your data worth anything to AI teams?

AI teams pay for data they cannot easily collect, scrape or generate themselves. The value usually comes from one or more of these:

• Real-world behaviour. Genuine conversations, real documents, footage of real tasks. Synthetic data cannot fully replace it.
• Specialist domains. Medical, legal, financial, industrial or scientific material, ideally with expert annotation.
• Under-represented languages and accents. Most public data is English. The commercial catalogue we track covers 223 languages, but supply drops quickly outside the largest ones.
• Scale with consistency. Thousands of hours or documents collected the same way, with the same metadata.
• Clean rights. Data whose consent and ownership are documented is worth far more than a larger set with doubts attached.

Where is there already a lot of supply? In our catalogue, speech and audio is by far the largest category (about 1,700 listings), followed by text and language (about 780), finance and business (about 660) and images (about 590). Healthcare (about 225 listings) and source code (about 20) are much thinner, and so are many languages. A thin category with steady demand is where a new dataset stands out.

2. Check that you have the right to license it

This is the step that decides whether a deal can happen at all. Buyers increasingly ask for written answers before they even look at a sample. Work through these questions, and have a lawyer confirm the answers (this guide is not legal advice).

Ownership and contracts

• Did your organisation create the data, or did it come from customers, partners or suppliers? If it came from others, do your contracts allow you to license it to third parties?
• Does it contain third-party content (copyrighted text, images, music, software) that you have no right to pass on?
• Is any of it covered by confidentiality agreements or trade secrets, yours or your clients'?

Personal data

• Does the data identify, or could it identify, people: voices, faces, names, account numbers, health information?
• If so, under rules such as the EU GDPR you need a lawful basis that covers this new use, and sensitive categories such as health data face stricter conditions. Consent collected for "service improvement" often does not cover licensing to a third party for AI training.
• Properly anonymised data (where people can no longer be identified by any reasonable means) falls outside the GDPR. Pseudonymised data, where you only replaced names with codes, does not.

Documentation buyers will ask for

• Where and when the data was collected, and how.
• What participants agreed to, with the consent text if there is one.
• How personal data was removed or protected.
• Any restrictions (territory, field of use, competitors excluded).

AI developers are under growing pressure to document what their models were trained on; in the EU, providers of general-purpose AI models must publish a summary of their training content. Clear provenance is therefore a selling point, not just a legal box to tick.

3. Prepare the dataset so it can be evaluated

A buyer's first question is "what exactly is in it?". The easier you make the answer, the faster the deal.

• Write a one-page datasheet: what the data is, volume (hours, files, documents, rows), languages, time period, collection method, file formats, annotation (what was labelled, by whom, with what quality checks), and known gaps or biases.
• Prepare a representative sample: a few hundred items or a few hours that reflect the whole set, cleaned and de-identified, shared under a short evaluation agreement.
• Fix the basics: consistent file naming, documented formats (for audio: sample rate, channels; for images: resolution), a metadata file that describes every item.
• Measure quality where you can: transcription accuracy, annotation agreement between labellers, share of unusable items.

You do not need a perfect dataset to start a conversation. You need an honest description and a sample that matches it.

4. Choose a licensing model

Model	What the buyer gets	When it fits
Non-exclusive licence	The right to use a copy, while you can license it to others too	Most off-the-shelf datasets; lets you sell the same data several times
Exclusive licence	Sole right to use it, for a period, a field or a territory	Rare, strategic data; priced at a large premium
Custom collection	Data collected or annotated to the buyer's specification	You have the access (people, sites, equipment) but not the dataset yet
Subscription or refresh	Regular updates of a living dataset	Data that keeps growing: calls, transactions, documents

Whatever the model, the licence should spell out:

• Permitted uses: training, fine-tuning, evaluation, and whether commercial products may be built on the resulting models.
• Redistribution: usually forbidden; check whether the buyer may share data with contractors.
• Duration and what happens at the end: may models already trained keep being used? Must copies be deleted?
• Territory and field of use, if you want to keep some markets for yourself or other buyers.
• Warranties and liability: what you guarantee about rights and consent, and the cap on your liability.
• Attribution and confidentiality: whether the buyer may say they used your data, and whether you may say you supplied it.

5. Price it

There is no public price list for data, but listings give a sense of the range. Among the commercial datasets we track:

• About four in five are priced on request (2,403 of 3,044 published listings). Prices are negotiated case by case.
• 288 listings show a public US-dollar price, from 34 different providers. Entry prices range from about 40 to 150,000 dollars; the most common list price is 20,000 dollars, often a "from" price for a standard licence.
• For speech data where both a price and a number of hours are stated (89 listings), the median works out at about 33 dollars per hour, with half of them between roughly 17 and 75 dollars per hour.

Treat these as reference points, not a valuation: a few providers publish many prices, and list prices are starting points. What moves the price up:

• scarcity of the language, domain or recording conditions;
• verified consent and clean rights;
• expert or multi-layer annotation;
• exclusivity, or a limited number of licensees;
• commercial-use rights and long licence durations;
• volume delivered on a schedule, with updates.

6. Find buyers

There are four main routes. Many data owners combine them.

1. Direct deals with AI developers. The highest prices, but you need to find the right person, handle evaluation and negotiate contracts yourself.
2. Data vendors that resell or package data. They do the selling and often the preparation, in exchange for a large share of the revenue or a one-off purchase price.
3. Marketplaces and listing platforms. Your dataset becomes visible to many buyers; terms and visibility depend on the platform.
4. Brokers and intermediaries. They assess, prepare and introduce your data to qualified buyers, then you license directly. This is the model DataToLab follows: you keep ownership, and nothing is shared without your approval.

Whatever the route, keep control of three things: who sees samples, what the licence allows, and how many times the same data is licensed.

7. A checklist before your first conversation

• One-paragraph description: what, how much, how collected
• Confirmation of ownership and contract rights to license
• Personal-data position: anonymised, consented for this use, or neither
• Datasheet with formats, volumes, annotation and known gaps
• A representative, de-identified sample (ready, not yet shared)
• Licensing models you are open to, and what is off limits
• A price floor below which you would not license

Mistakes to avoid

• Sharing the full dataset before a signed agreement. Share a sample under an evaluation agreement instead.
• Assuming old consent covers AI training. It often does not; check the wording.
• Granting exclusivity too cheaply. Exclusivity blocks every future licence; price it accordingly.
• Overstating volume or quality. Buyers check samples carefully, and a mismatch ends the deal and the relationship.
• Forgetting what happens to trained models when the licence ends. Decide it before signing, not after.

If you hold data that might fit, the simplest next step is a short, confidential description: describe your dataset. We will tell you plainly whether there is likely demand, and what would need to be in place before an introduction.