What it actually takes to ship GenAI to a live vehicle fleet

The demo is the easy part. A conversational feature that delights in a conference room is a different animal from one running on a fleet that is already on the road, in customers’ hands, under warranty. The gap between those two states is where most GenAI-in-vehicle ambitions quietly stall — not because the model was wrong, but because nobody owned the operations.

When we shipped the first production conversational GenAI feature to 900K+ vehicles in six weeks, almost none of the hard problems were about the model. They were about rollout. How do you gate activation so a bad day affects a controlled slice, not the whole fleet? How do you sequence five org boundaries — app, backend, vehicle, security, legal — so they arrive at the launch line at the same time instead of blocking each other in series? What, exactly, does “safe enough to launch” mean when “roll it back” is not a clean option at fleet scale?

The answer was entitlement and feature-flag infrastructure doing the unglamorous work: gating who gets the feature, when, and under what conditions, so activation could ramp without becoming an all-or-nothing bet. The lesson is plain. If you are weighing GenAI in the vehicle, spend less time on the model bake-off and more on the operations plan. That is where the launch is won or lost.