Salwa Al-Tahan
22/09/2026
The industry talks a lot about drive-thru accuracy. Intouch Insight 2025 study put the overall average at 87% with a team member, and Voice AI trailing at 83%. So what are we actually measuring, and why does it fail in the lane?
A Lab Isn’t a Drive-Thru:
Experience shows that the lab is nothing like a real drive-thru, so let’s be clear on that. Poor mics, noise, heavy accents, and mid-order changes all affect accuracy. AI usually nails “give me a number one with a Coke.” But it trips on the messy modifiers, the half-changes, and the order the guest actually meant.
Where Does Voice AI Fail?
Anyone who’s ever worn the headset already knows how guests order, and they definitely don’t recite the POS tree.
The reality is, every guest orders in their own unique way, “the usual but no pickle,” “make it a meal,” “wait…. can I swap the fries”, “actually two of those,” and then a kid shouting from the back seat “can I have a treat?” The menu board might look pretty, but realistically the conversation usually isn’t.
That gap is where a lot of drive-thru tech still fumbles. Systems that shine on hero SKUs quietly fall apart once the guest starts customising. The Intouch Insight 2025 drive thru study which analysed brands in the US utilising Voice AI, shares that customisation drove 62% of incorrect AI orders- that’s roughly six in ten of the ones that they tracked. True Voice AI adoption isn’t just about automated speech recognition, it’s about seamless execution. And that successful solution must be able to process complex, real-world human ordering seamlessly while eliminating operational friction for both store teams and drive-thru guests.
Sure, a few guests may not want to speak to AI or will purposely try to break the system for YouTube likes, but most just want lunch and they want their order right. They might change mid-sentence because the car behind them is waiting and they get flustered, or they realise they won’t have time for what they first asked for. They might even use brand slang that often isn’t on the menu board or be ordering for three people at once. A good team member gets it, as does a good Voice AI system. A fragile, scripted AI will either guess or reset and hand it over, but here’s the thing: guests will remember both experiences.
It’s More Than “No Lettuce”
It’s important to understand that menu complexity isn’t only “remove the lettuce.” It’s modifiers stacked on modifiers, LTOs that live in one daypart and not in the next, regional names familiar to locals, and out-of-stocks that should redirect to other options rather than confirm something unavailable. In the real world, the ordering layer needs contextual awareness, not a rigid script, and it needs to be a system that actually understands what was said and pushes that correctly to the POS.
The more we understand how guests order and know that the speaker isn’t filling in a form, the better accuracy will be. One asks if their favourite is available. Another asks what’s spicy. Some don’t know if the meal comes with a drink. Some want the total before they commit, then change one thing when they see it on the confirmation board. Accuracy isn’t “did we hear the base item.” Accuracy is “did the ticket that hit the kitchen match the meal that left the window.”
Stop measuring success on plain combos
A note to operators, measure where your misses live. If most errors sit in customisation, a clean tick on a plain combo isn’t a win. Profile accuracy by modifier count. Treat confirmation boards as part of the workflow, not just looking pretty. And when you evaluate voice AI, ask the boring questions first: does it hold multi-item context, does it validate against live POS, and does it know when to stop guessing and hand it over to a human?
Remember you’ll never make guests order like a spreadsheet or a lab script. The real win with Voice AI is in live deployments. Guests speaking normally, and the ticket still coming out right.
Curious to learn more, talk to us or follow along for the next blog in this series.