Who actually buys company data for AI training?
Frontier research labs, and the intermediaries already holding contracts with them. No lab opens an account with a fifty person plumbing contractor. That is why the chain has a middle.
The labs
The companies training general models buy operating records to teach a model how work is actually done: how a quote becomes a job, how a thread becomes a decision, what a file looks like on the fourth revision. They buy through procurement and they buy in volume, which means they buy from counterparties who can warrant rights across many sellers.
The intermediaries
A layer of firms sits between, aggregating corpora and reselling them under their own paper. They publish what they pay in aggregate and almost never what they pay per deal. Their published aggregate figures put the average company payment in the low hundreds of thousands.
The evaluation shops
A separate and growing buyer: firms building agent evaluations and reinforcement learning environments out of real company workflows. They want the same records for a different reason, and they are usually happy with a narrower slice.
Why you are not talking to them directly
Nothing stops you. What stops most sellers is that the buyer wants a rights warranty, de-identification done to a standard, an export in a shape they can ingest, and a counterparty who will still exist in a year. Producing all four for one archive costs more than the deal is worth to you.