Resources
How to evaluate a training-data vendor
A repeatable way to judge a vendor on provenance and rights auditability, not just samples and turnaround.
Published 2026-07-22 · 7 min read
Key takeaways
- You are buying a supplier, not just a dataset. Judge repeatability.
- Provenance you can query beats aggregate assurances.
- Rights auditability means reading the licence on a specific asset.
- Ask how withdrawals, takedowns, and disputes are handled after delivery.
- A vendor who cannot repeat clean sourcing will fill gaps with risk.
A strong sample does not make a strong vendor. You are not buying a single dataset. You are buying a supplier you may return to, and whose rights work you will have to defend if a buyer, a regulator, or a court ever asks.
So evaluate the vendor the way you would evaluate any critical supplier. Can they prove what they sell. Can they repeat it to a brief. Will the paperwork hold up under pressure. Provenance and rights auditability matter more than turnaround time.
Start with provenance
Provenance is the record of where each asset came from and how it reached the vendor. Ask a direct question. Can you trace any single asset back to its source, on request, after delivery.
A vendor with real provenance answers yes without hesitation. They can show the source, the contributor, the date, and the rights that travelled with the file. A vendor who can only speak in aggregates, that it all comes from licensed creators, is describing a hope, not a record.
On fiund, provenance is available during diligence. Treat that as the standard, not a bonus.
Test rights auditability
Auditability means you can inspect the rights on a specific asset, not just accept a blanket assurance. Ask to read a sample licence. Ask how the AI-training grant is recorded per asset. Ask how voice and likeness consent is captured for identifiable people, and how you would audit it later.
The difference matters. A vendor can truthfully say every file is licensed and still be unable to show which licence covers which file. When a dispute arrives, the sentence "we license everything" is not a defence. The document is.
Understand the sourcing model
Ask how the vendor sources, in plain terms. Commissioned capture. Contributions from consenting creators. Archives with clear ownership. Each can be clean. Then ask the harder question. Can you repeat this for a custom brief without cutting corners on consent or rights.
A vendor built for repeatable, consented sourcing can scale to your brief. A vendor who found one good batch cannot, and may fill the gap with whatever is fast. See how we source for what repeatable sourcing looks like.
Check process, security, and delivery
Rights are the first question. They are not the only one. Ask what metadata and provenance ship with delivery. Ask whether the vendor holds to a written deliverable spec. Ask how they handle data security, access controls, and deletion.
Ask one more thing that separates serious vendors from the rest. How do you handle takedowns, withdrawals, and disputed assets after delivery. A vendor with an answer has done this before. A vendor without one has not.
Vendor evaluation questions
- Can you trace any single asset back to its source on request
- Is there a signed AI-training licence on every asset, and can we read a sample
- How is voice and likeness consent captured for identifiable people, and can we audit it
- What is your sourcing method, and can you repeat it for a custom brief
- What metadata and provenance ship with delivery
- How do you handle takedowns, withdrawals, and disputed assets
- What are your data-handling, access, and deletion practices
Frequently asked questions
What is the single most useful question to ask a vendor?
Pick one asset at random and ask the vendor to show its source, its AI-training licence, and its consent status. Real provenance survives that test. Marketing does not.
How is a vendor different from a marketplace?
A vendor sources and sells data. A marketplace connects many owners with buyers under a shared licence and consent standard. On a marketplace, owners keep ownership and approve buyers, and provenance is available in diligence. See the vendors hub for how to compare.
Do I need to re-evaluate a vendor for every purchase?
Re-check what changes. The rights model and provenance practice tend to hold across deals. The specific consent, spec, and exclusivity change per dataset, so audit those every time.
Related resources
Want data that clears this in diligence?
Whether you're building a model or sitting on an archive, the first conversation is short and specific.
Send a brief