Summary
Amit Desai, a voice AI specialist who has worked at Alexa, Roku, and his own startups, presents a framework for smarter decision-making in voice AI assistants at AI Engineer 2026. His central argument: improving model accuracy is only one lever for user satisfaction. The second lever — underused in the field — is a system’s ability to decide whether to act, confirm, or stop when it faces uncertainty about what the user intended.
Desai formalizes this as a cost optimization problem he calls OUCH (Outcome User Cost Heuristic). Rather than guessing a confidence threshold for when to ask for clarification, he demonstrates how to quantify the relative user cost of each outcome (wrong action vs. unhelpful stop vs. correct action), plot the total cost across a distribution of real interactions, and find the mathematically optimal confidence threshold. In his worked example using a smart speaker music scenario, the naive threshold guess of 65% produced a user cost of 1.9 OUCH points per turn; the optimal threshold turned out to be 43%, yielding measurably better outcomes.
Desai then scales the framework to wearables and embodied AI robots, where the stakes of acting on a misheard command are significantly higher — a robot discarding a physical object is far costlier than playing the wrong song. He argues that system decision logic should be treated as a first-class engineering concern, not an afterthought, and that the same mathematical scaffolding applies across voice surfaces from smart speakers to autonomous physical agents.
📺 Source: AI Engineer · Published September 15, 2026
🏷️ Format: Keynote Launch
