AI data companies broker human experts for AI training — but why are some trying to sell more than just labor by the hour?

This explores why companies that broker human experts for AI training are trying to become something more than labor marketplaces, and what that 'something more' could be. The clearest evidence is Mercor: its gross annualized revenue reached $2 billion in June 2026, but 60–70% of that goes to contractors. That leaves roughly $600–800 million of real net revenue, and every new dollar brings a new payout with it Can AI data companies sustain margins beyond labor payout ratios?. The market has moved from cheap labeling toward physicists and finance experts, and expert judgment now commands premium prices. A brokerage model can't escape that ratio, which is why these firms are turning to packaged products. To be direct: the collection names this pressure, but it doesn't yet catalog which products these companies are actually shipping.

Other notes in the collection suggest where those products could come from. The strongest candidate is evaluation. One analysis of 960 real occupational workflows found that AI agents win abstract contests but fail long, multi-step professional tasks. The authors argue the gap comes from how benchmarks are designed, since the field has been measuring contests rather than work Why do agent benchmarks not predict real economic value?. A company with a pool of credentialed professionals is well placed to package 'what real work looks like' as reusable test suites, not just one-off training labels. A related argument says AI can now produce knowledge faster than humans can check it, and the tools used for checking are increasingly AI-made themselves Can AI generate knowledge faster than humans can evaluate it?. If that's right, the scarce thing worth selling is verified human judgment, a product rather than hours.

A more speculative direction comes from DeepMind researchers. They propose an 'Automated Scientific Economy' in which AI-generated research ideas are licensed, validation earns royalties, and scarce lab capacity is allocated by a market Could markets allocate scarce lab resources to AI-generated research ideas?. The paper isn't about data companies, but the model is a useful lens: it treats expert validation as an asset that keeps paying, not a task paid once. That is the move a data company would need to make to break the payout ratio.

The corpus also offers a less flattering view. Harris argues that AI math firms recruit senior mathematicians, then use contracts that erase their names and their ownership of the work, all while talking about democratizing knowledge Does AI math recruitment mask the commodification of expert labor?. Read next to the Mercor numbers, this points to something you might not expect. 'Productizing' expert judgment often means turning the expert's contribution into a company-owned asset, so the margin problem gets solved partly by changing who owns what. Acemoglu, Autor and Johnson argue that firms tend to profit more from automating expertise than from creating new work for experts Why do firms build automating AI instead of pro-worker AI?. That suggests the experts feeding these products may be training their own substitutes.

If you want to go further, start with the Mercor note for the economics. Then read the benchmark-gap note for the likeliest product opportunity, and the Harris piece for the cost to the people whose judgment is being packaged.