Home / Blog / How to license real-world training data for AI
How to license real-world training data for AI
Licensing real-world training data usually takes five steps: write a short spec, check the supplier's sourcing and consent, run a small paid pilot, review the pilot against a dataset card, then sign a license for scaled delivery. Most of the risk sits in the consent and license terms, not in the data itself.
Public web data is easy to get and hard to differentiate on. The data that makes a model better at a specific job (reading a clinic's notes, transcribing a messy support call, extracting a handwritten bill) usually sits inside organisations and has never been published. Getting it means licensing it. Here is how that process works when it goes well.
1. Write a one-page spec
Suppliers can't quote or plan collection without knowing what "good" looks like. A useful spec answers:
- Task. What the model has to do: transcribe, classify, extract fields, summarise.
- Domain and source. For example, outpatient notes from general practice, or inbound calls to a delivery company's support line.
- Format. Scans, phone photos, audio sample rate, CSV exports.
- Labels. The exact fields or tags you need, and how ambiguous cases should be handled.
- Volume and diversity. Rough count, plus what must vary: speakers, handwriting styles, specialties, store types.
- Use. Training, evaluation, or both. This changes the license.
2. Check how the supplier gets the data
Ask where each record comes from and what the source agreed to. A supplier should be able to tell you, for every record, which agreement covers it and whether that agreement mentions AI training and commercial use. If a supplier can't trace records to sources, the risk passes to you. Our training data rights checklist lists the questions to ask.
3. Run a small paid pilot
A pilot of a few hundred records is cheap compared with a bad full delivery. Use it to check:
- Label accuracy on a random sample you review yourself.
- Whether the de-identification removed what it should, without destroying the signal you need.
- Whether the data actually moves your metric. Train or fine-tune on it, or run it as an eval set.
4. Read the dataset card
Every delivery should come with documentation describing how the data was collected, who labelled it, what was removed, and what is known to be missing. The format popularised by Datasheets for Datasets is a good baseline. Gaps listed honestly in a card are worth more than a card that claims none.
5. Negotiate the license
| Term | What to decide |
|---|---|
| Field of use | Training, evaluation, fine-tuning, or all of them |
| Exclusivity | Exclusive data costs more; non-exclusive is cheaper but competitors may license it too |
| Term and region | How long you can use it, and where |
| Model ownership | Whether models trained on the data are fully yours |
| Withdrawal | What happens if a source later withdraws consent |
| End of license | Whether data must be deleted, and what happens to trained models |
How pricing is usually structured
Real-world data is usually priced per unit (per document, per record, per hour of audio) with a premium for exclusivity, specialist labelling, or hard-to-reach sources. Custom collection is often quoted per project. Ask for the pilot to be priced on the same unit basis as scaled delivery, so the pilot also tests the economics.
How we run it at Datadhara
We work exactly this way: spec first, a small paid pilot with a dataset card, then scaled delivery under a written license. You can look at the format of each data type in our free synthetic samples, and read how collection and de-identification work on the methodology page.
Questions
How long does it take to license custom training data?
A pilot typically takes a few weeks, because sources have to be signed up and records collected, de-identified and labelled. Scaled delivery depends on volume and how many sources are needed.
Is exclusive training data worth paying for?
It is worth it when the data gives you an edge competitors could otherwise buy, for example a rare domain or language. For evaluation sets, non-exclusive is usually enough.
What should a training data pilot include?
A few hundred records in the final format, with the same labels and de-identification as scaled delivery, plus a dataset card describing collection, labelling and known gaps.
Tell us what your models are missing
A few lines is enough: the domain, the format, and roughly how much.
Request data →