- Start with a written data request that names the modality, domain, volume, labels and rights you need.
- Licensed data from the original holder gives you the clearest provenance and fewer surprises in diligence.
- Evaluate a representative sample against your own model before committing to volume.
- Training data disclosure rules in California and the EU make provenance records a requirement, not a nice-to-have.
Step 1: Write a data request
A clear request saves weeks. Include:
- Modality and format (text, audio, video, code, images, trajectories)
- Domain and task (what the model should learn)
- Volume and minimum quality bar
- Labels or metadata required
- Languages, regions and time period
- Rights needed: training, fine-tuning, evaluation, commercial use, territory, term
- Exclusivity preference
- Budget range and timeline
See how to write a training data request.
Step 2: Choose a sourcing route
| Route | Best for | Watch out for |
|---|---|---|
| Data marketplaces | Off-the-shelf datasets | Thin provenance, non-exclusive by default |
| Human data vendors | New labeling, RLHF, expert work | Cost, rater quality |
| Direct from holders | Unique private data | Finding holders and negotiating rights |
| Brokers | Unique data without building a sourcing team | Fee structure, independence |
We work on the direct-from-holders route: you tell us what you need, we find and vet holders, and you license directly from them. See AI data brokers vs marketplaces.
Step 3: Check provenance before samples
Ask who created the data, under what terms, and what consent or rights basis covers it. Provenance documentation is what your legal team and regulators will ask for later. Read data provenance explained.
Step 4: Evaluate a sample
Test a representative sample against your own model or evaluation set. Check duplication, label accuracy, language detection, toxicity, personal information leakage and how much overlaps with public data you already have. Read how to evaluate a data sample.
Step 5: Negotiate the license
Agree permitted uses, exclusivity, warranties on rights and de-identification, indemnities, delivery format and acceptance criteria. Read AI data license agreements.
Step 6: Record what you bought
California AB 2013 requires developers of generative AI systems made available to Californians to publish information about their training data, including whether datasets were purchased or licensed (LegiScan). The EU AI Act sets transparency duties for general-purpose AI models (European Commission). Keep a dataset card for every purchase.
Questions
Is licensed data better than scraped data?
Licensed data comes with permission and provenance, which lowers legal risk and supports disclosure duties. It also tends to include private data the web does not have.
What does it cost to source a custom dataset?
It depends on the data. See our AI data prices page for published reference points, and talk to us for a range on your request.