Shifting AI Data Costs to Engineers

Proximal is trying to turn training data from a labor marketplace into a software and infrastructure product. Instead of paying thousands of people to hand write prompts, grade outputs, and check edge cases one by one, the company uses engineers to build task generators, RL loops, and test harnesses that can create and verify many coding tasks repeatedly. That can make cost scale more like software and compute than hourly annotation work.1 2

• The clearest contrast is Scale AI’s older model. Scale built a large business by routing labeling jobs to global contractor networks and layering workflow software on top. Proximal is pushing the opposite direction, where more of the spend sits in internal engineering and model driven evaluation instead of crowd operations.2 3
• This does not remove humans from the loop, it changes which humans matter. In coding post training, the scarce input is not generic click work, it is people who can spot weak capability areas, design hard tasks, and decide whether automated checks are actually measuring the right thing.1 4
• That setup also lets Proximal test its own data before selling it. Because it runs RL training environments itself, it can see whether a batch of tasks actually improves model behavior, which is closer to a closed loop manufacturing process than a plain labeling vendor handing over files.1

If this model works, data vendors in AI will look less like BPOs and more like specialized infra companies. The winners will be the teams that can build reliable automatic task generation and verification for narrow domains like coding, then prove with training runs that their data changes model behavior in measurable ways.1 2 3