Your model does not want more data. It wants retrieval units.

Why the unit of sale for AI data is shifting from files to small, structured, provenanced records.

Most data deals still ship the way they did ten years ago: a bucket of files, a licence PDF, and a hope that someone downstream will make sense of both. For a model or an agent, that is the wrong shape.

What a retrieval unit is

A retrieval unit is the smallest piece of data that can answer a question on its own. It carries:

• the facts, as fields, not prose: entities, dates, amounts and relationships;
• provenance: where it came from, down to the section, with a hash and a timestamp;
• confidence: how sure we are that it is right, and how we checked; and
• terms: what you may do with it, and what each use costs.

A model can learn from it. An agent can cite it. A lawyer can audit it.

Why it matters when you buy

When a head of AI asks for “ten thousand clinical notes”, the real request is narrower: notes that cover these conditions, in these languages, with this much agreement between labellers, cleared for training. Buying units instead of files lets both sides score the match before money moves: coverage, accuracy, freshness and licence fit.

That is how PXYZ approaches data work. Every request gets a quality bar in writing. Every finding is scored against it. Nothing is delivered that cannot show where it came from.

The short version

Files are how data is stored. Units are how data is used. Buy the second.