How to choose AI training data
Choose AI training data by scoring provenance, rights, reasoning trail, coverage and delivery terms — in that order. A dataset that fails rights cannot be fixed by size, and a dataset without a reasoning trail teaches answers without judgement. The checklist below is the one used to describe supply on this site.
The checklist
- Provenance: who produced the records, in what systems, over what years — documented, not asserted.
- Rights: the seller owns or controls the records, client and employee obligations are mapped, and the licence is signed by the rights-holder.
- Reasoning trail: threads, revisions and outcomes are attached to final documents.
- Coverage: enough years and volume for patterns to repeat — most archive buyers publish a floor of 20+ staff and 3+ years.
- De-identification standard: named method, who performs it, who warrants the result.
- Terms: scope, term, exclusivity and deletion/retention written down before delivery, not after.
Red flags
- Customer or client data offered by someone who does not own it.
- “Exclusive” archives also listed on public marketplaces.
- Prices quoted from a buyer's published ceiling as if typical — the median public sale is about $31,600 (AIData.Broker).
- Raw data demanded before any NDA, manifest or licence exists.
Want supply checked against this checklist?
Send the specification and we will score what exists against it — including where the answer is “not available.”
Cite this entry: Best Data for AI. “How to Choose AI Training Data — A Buyer's Checklist.” Content reviewed 10 November 2026. Figures on this page are published buyer ranges with sources and check dates shown — they are not offers.