When McDonald’s announced its latest AI-powered drive-thru system, one number quickly became the headline: more than 90% accuracy. The figure sounds compelling, but it also raises a larger question for an industry increasingly experimenting with voice automation: what does accuracy actually tell a restaurant operator about whether an AI system is working? Hi Auto, whose voice AI powers drive-thru ordering at brands including Bojangles, offers a useful lens for answering that question. A more meaningful measurement may not be whether an AI understands an order, but whether it can complete it without a human having to step in.
That distinction matters at fast-food scale. Even a system that performs correctly more than 90% of the time still leaves a meaningful share of interactions requiring correction, clarification, or intervention. For a restaurant processing thousands of drive-thru orders, the difference between an AI that is technically accurate and one that consistently completes transactions can translate directly into labor requirements, throughput, and customer experience.
Accuracy Is Only One Part of the Equation
An AI system can be accurate while still creating operational friction. A customer may eventually receive the correct order but have to repeat a request several times, wait while an employee takes over, or navigate an interaction that feels less convenient than speaking directly with a worker. From an operator’s perspective, the technology may technically be functioning while still failing to deliver the full operational benefit expected from automation.
That is why third-party measurements of real-world performance can be more revealing than a single accuracy figure. Intouch Insight’s 2025 Drive-Thru Study evaluated 2,265 mystery-shop orders across 13 major quick-service restaurant brands, measuring factors that included speed, accuracy, customer satisfaction, speaker clarity, customer sentiment, and employee intervention.
The broader set of measurements gets closer to the question restaurant operators actually need to answer: does the technology work inside a live drive-thru, with real customers, real menu modifications, and the unpredictable conversations that come with them?
The Metric That Changes the Conversation
Employee intervention is particularly useful because it connects AI performance to restaurant operations. In the Intouch Insight testing, Bojangles, whose drive-thru AI is powered by Hi Auto, recorded an employee-intervention rate of 3%. The other major chains tested with voice AI were around 30%, with Taco Bell at 30% and Wendy’s at 33%.
The significance is not simply that one system produced a lower intervention number. Intervention measures whether automation is actually functioning as automation. If employees regularly have to monitor conversations and take over when the AI encounters an unusual request, the restaurant has introduced another interface that employees must manage rather than genuinely removing work from the process.
For operators, that can be a more useful metric than a model-level accuracy percentage. An AI system that understands 95% of an order but requires employees to resolve a significant portion of transactions may produce a very different operational outcome from one that can independently complete almost every interaction.
From AI Accuracy to AI Completion
Hi Auto makes a similar distinction in its own reporting, separating completion from accuracy. The company defines completion as the percentage of orders sent to the POS without staff intervention, while accuracy measures whether the order contains the correct items and customizations. Hi Auto currently reports 93% completion and 96% accuracy across approximately 1,000 stores.
This distinction matters because the metrics answer different questions. Accuracy asks whether the system understood the customer’s request correctly, while completion asks whether it handled the transaction without requiring someone else to finish the work. For restaurant operators evaluating automation, both matter, but completion is much closer to measuring whether AI is actually reducing the workload at the ordering point.
Customer satisfaction adds another layer. An AI can complete an order without human intervention and still deliver an experience that customers dislike. Conversely, an interaction that feels natural and efficient may be more valuable than a system that performs well on a narrowly defined technical benchmark.
The Guest Still Gets the Final Vote
Drive-thru AI ultimately has to disappear into the customer experience. People pulling into a drive-thru are not evaluating speech-recognition models or debating AI architecture; they want to order what they want, make their modifications, receive confirmation, and move through the lane without unnecessary friction. Intouch Insight’s study reflects that broader reality by examining satisfaction alongside operational and technical measures, including the speed and accuracy of the transaction and the customer’s perception of the interaction.
That points toward a more useful framework for evaluating the next generation of drive-thru AI. Instead of focusing on accuracy alone, restaurant operators can look at three connected outcomes: how many orders the system completes independently, how frequently employees need to intervene, and whether customers finish the interaction satisfied.
Accuracy still matters, but it is only part of the equation. For a technology designed to automate one of the busiest points in a restaurant, the more consequential question is whether the AI can reliably complete the job without creating another job for someone else. Hi Auto’s results at Bojangles suggest why that distinction is becoming increasingly important as voice AI moves from pilot programs into everyday drive-thru operations.
