Jason Droege: Why AI still needs someone to dig up the road

Early gen-AI systems used cheap, generalist labelers; today’s frontier models require highly skilled experts (engineers, doctors, PhDs) who spend hours on tasks like building full websites or explaining nuanced medical topics, and defining what ‘good’ looks like in each domain.

Enterprises and governments increasingly need robust eval suites that encode ‘what good looks like’ for their specific processes (e.g., medical chart review, insurance claims), because off-the-shelf benchmarks and generic RAG/fine-tuning only get you part of the way.

POCs often stall at 60–70% performance; getting to production-grade reliability requires months of iteration, domain data collection, policy and legal approvals, change management, and careful human–AI handoffs, which news headlines and demo videos tend to obscure.

Droege argues that as long as new human skills, judgments, and organizational contexts matter, models will need fresh human-generated data and judgment; ‘no-human-in-the-loop’ would imply a world where no new human knowledge is worth encoding, which he sees as very far off.

When launching Uber Eats, Droege systematically compared ideas (convenience vans, grocery, point-to-point delivery) and prioritized a marketplace with strong unit economics, clear incrementality for restaurants, and the potential for network effects and large scale.