Shift from Annotation to Integration
Oscar Beijbom, co-founder and CTO of Nyckel, on the opportunites in the AI/ML tooling market
The main shift is that labeled data is no longer the scarce input for most AI products, product integration is. Older ML stacks were built around collecting thousands of examples, sending them to outside annotators, then wiring together training, deployment, and monitoring. As pre-trained models got better, many everyday use cases moved to a much lighter workflow where a product manager or developer can label around 100 examples, test on their own data, and ship a working classifier fast.
-
Scale got big by serving data hungry markets like autonomous driving, where every new weather condition, lighting change, and road scenario created more edge cases to label. That made outsourced annotation a core part of the product, not a setup step.
-
For narrower tasks like document classification, moderation, or image tagging inside software products, the work shifts to the domain expert. The person who knows spam, invoices, or community rules can label samples directly in the product, which removes the overhead of writing instructions for a third party workforce.
-
This does not eliminate human labor, it changes where it matters. Scale itself expanded from labeling into Nucleus, Spellbook, InstantML, and Launch, while newer human in the loop businesses increasingly focus on high judgment work like RLHF and expert evaluation rather than bulk annotation.
The market is heading toward fewer standalone annotation vendors and more bundled AI products that hide the ML stack behind an API or workflow. Human work will concentrate in frontier, regulated, and safety critical domains, while mainstream software use cases will be won by whoever makes custom AI feel like adding payments or messaging to an app.