At Waymo, an AI project isn't ready until its evals are — not when the model performs well

Waymo engineering leader Manasi Joshi said the company uses eval-centric development to decide when autonomous driving AI is ready, making testing a core engineering discipline rather than a final prelaunch step. The company applies continuous evaluations during training, after training, and in simulations to manage high-stakes real-world risk.

Waymo’s approach highlights a broader enterprise lesson: AI systems need measurable reliability before deployment, especially as companies adopt agents for customer service, coding, finance, and operations. The trend points toward evaluation infrastructure becoming as important as model performance in enterprise AI governance.