Snorkel sells completed datasets and RL environments

Snorkel’s 2025 move turned data work from a software seat sale into a higher value deliverable business that sells model improvement directly. Instead of only giving enterprises a platform to upload datasets, assign annotators, and run labeling workflows, Snorkel began packaging the finished output itself, expert built post training datasets and realistic agent environments with tools, documents, and reward signals, which maps much more closely to what frontier labs actually buy when training and evaluating models.1 2 3 4 5

• The older product centered on customer managed datasets inside Snorkel’s platform. The docs show users uploading their own train, validation, and test data, while the 2025 product material shifts toward Expert Data as a Service and completed datasets for evaluation and post training, which is a different buyer promise and a more service heavy revenue model.2 6 7
• RL environments extend that shift one step further. Snorkel describes them as sandboxes with APIs, tools, state, policy documents, and structured feedback, and its enterprise environment pages package thousands of task scenarios for domains like insurance underwriting. That means Snorkel is no longer just labeling examples, it is constructing the simulated workplace where agents practice and get graded.3 4 5
• This also explains why enterprise traction helped the new business ramp. Snorkel already had deployments with customers like Wayfair, and earlier materials place Chubb and BNY Mellon on the platform side. Those relationships gave it domain access and workflow knowledge that are useful when building specialized datasets and enterprise style agent environments, unlike pure labor marketplaces that mainly sell human hours.8 9 10

Going forward, the category is moving toward whoever can deliver the full training loop, not just labels. The winners will combine enterprise distribution, domain experts, finished datasets, and hard to fake evaluation environments. That puts Snorkel on a path closer to specialized AI infrastructure for labs and large regulated enterprises than to a classic labeling software vendor.1 3 4 5 10