Backseat: who in the car asked for what

Backseat is a drive-thru voice agent that knows who in the car asked for what. A car is not one customer. The driver orders, a passenger adds a drink, a child shouts for a milkshake, the next lane talks into the same speaker. Public AI drive-thru pilots broke there: McDonald's ended its IBM pilot in 2024 after viral wrong orders, one for 260 McNuggets; Taco Bell slowed its rollout after a prank order for 18,000 waters. Intouch Insight's 2025 study found AI lanes 83% accurate against an 87% average, and 95% when a person stepped in. Backseat listens with two AssemblyAI ears on one microphone. The Voice Agent API runs the conversation: Universal-3.5 Pro, far-field Voice Focus, tool calls. A second transcript could only check what was said; Backseat's second ear, streaming speech-to-text with speaker labels, says who said it. Every ticket line carries an owner (whose food) and a requester (which voice asked). Another voice's request waits for the driver's spoken yes; "hers without pickles" changes only her burger; 260 nuggets goes to a person. The page books every tool call into Backseat and into a cart that trusts it. A recorded lane replays AssemblyAI's own events, decided again in your browser, no microphone or key. Measured on the real APIs with 21 scripted cars: Backseat passes 19 and gets 91% of orders exactly right; the same audio through a single-ear agent passes 16, at 81%. Food nobody asked for: once, against three times. Both absurd orders went to a person, against neither. On 156 real customer orders from Amazon's FoodOrdering set, spoken by a synthetic voice, 96% come out exactly right. 103 tests; what still fails is published, and the README ends with ten Voice Agent API findings, each with its fix. Who pays: AI drive-thru vendors, licensing the attribution layer at $150 per drive-thru per month; the room ear costs under a cent an order. A chain pilots it in shadow mode: both carts side by side, and the gap is the business case.