TL;DR
- Natural language processing is what turns a spoken, changing, messy order into one correct ticket for the POS.
- At a drive-thru speaker it has to hear the words, understand them in context, track changes, handle modifiers, match the live menu and confirm the result.
- Clean audio proves little. Test with tricky orders: mid-sentence changes, stacked modifiers, back-seat voices, Spanish words and engine noise.
- Hi Auto processes 100M+ orders per year and reports 93%+ completion and 96% accuracy across more than 1,000 stores.
“Can I get a number two, actually, make that a large, no onions, and add a shake, chocolate, no wait, vanilla.”
That took about eleven seconds. In that time the guest changed course three times: the size, the onions and the shake flavor. A good order taker tracks all of it, keeps the running order in mind and does not ask the guest to start again.
That is the job of natural language processing in a drive-thru: turning speech like that into one correct order. This guide walks through what it has to do at the speaker, where it gets hard, and which tricky orders to run in a pilot.
What is natural language processing, in plain words?
It is software that reads and responds to everyday human language, the way people actually talk. At a drive-thru, that means turning a spoken order with corrections and extras into a structured ticket the POS accepts.
The glossary covers the neighboring terms, including natural language understanding and conversational AI. This guide stays in the lane. Every example here is a drive-thru order, and companion guides cover the conversation model (“What Is a Large Language Model (LLM)? A Guide for QSR Technology Evaluators”) and accents (“Accent Recognition in Drive-Thru AI: Why It Matters for Diverse QSR Markets”) in their own right.
What does natural language processing do at the speaker?
Six things have to go right between the first word and a correct order on the screen.
- Hear the words. Speech recognition turns the guest’s voice into text. A drive-thru is a hard place to do it, with engines, wind and a radio playing in the car. Hi Auto uses proprietary in-house speech recognition and audio filtering to deal with that noise. The noise handling entry in the glossary goes deeper.
- Understand the words in context. A transcript is not an order. “Make that a large” refers to the last item, not the whole ticket. “No wait, vanilla” is a correction, not an addition.
- Track changes as the order evolves. People change their minds mid-sentence. That is normal speech. The final order should reflect the last thing the guest chose.
- Handle modifiers. “No onions, extra sauce, light ice” all have to attach to the right item. Modifier handling is where a lot of orders are won or lost.
- Match what was said to the live menu. Guests do not speak in item codes. At Hi Auto, the system connects to the POS, and new menu items are set up by the Hi Auto team as a managed partnership.
- Confirm the result. The order confirmation board shows items, modifiers and prices as the guest speaks, so the guest can see the order take shape.
Hi Auto supports English and Spanish, with automatic language detection and switching mid-order.
Where does it get hard?
These are ordinary lunch-hour orders, and each one tests a system.
Stacks of modifiers. Four changes to one burger, spoken in one breath, all have to land on the right item.
Corrections. A good result is a clean order that reflects where the guest ended up, without reading every change back.
Similar names. “The number four” and “the number fourteen” can sound alike at speed. Limited-time offers add more variety, since guests shorten names or describe items in their own words. Menu complexity grows with every promotion.
Requests for something the store does not sell. Off-menu requests need a graceful reply that keeps the lane moving.
A second voice from the back seat. A passenger adds an item halfway through. The order should keep its thread.
Independent data shows why these are worth testing. QSR Magazine’s 2026 Drive-Thru Report (Danny Klein, 1 October 2026) reports that, among the 15 drive-thrus shopped where voice AI could not continue the order, it could not understand a customization request in 53.3% of cases and could not answer a question in 40%. The figures belong to the report and are not Hi Auto numbers. Questions and customizations are where AI ordering gets tested, so that is where to look first.
At Hi Auto, the AI handles the order. When it cannot confidently determine what the guest wants, a trained Hi Auto supervisor resolves it instantly through the platform. Every one of those moments is used to train the model further, and the process is seamless to the guest and the restaurant.
Which tricky orders should you run in a pilot?
A demo with clean audio proves very little. Run these seven orders in a real lane, and watch the confirmation board as the guest speaks.
- A mid-sentence change. “A number three, no, make it a number five, large.” The order should show the number five, large, with no leftover number three.
- A stack of modifiers. Give an item five changes. All five should attach to the correct item and appear on the board.
- A combo swap. Change a side inside a combo. The swap should happen inside the combo and not appear as a separate full-price line.
- A request for something not on the menu. The guest should get a clear answer and an easy next step, with no dead air.
- A second voice from the back seat. Both people’s items should be captured, and the driver should not have to start over.
- A Spanish word in an English order. The order should carry on smoothly, without a clarification loop.
- An idling engine. Run the same order with a loud engine nearby. The guest’s words should come through, and the noise should not change the order.
For each order, record three things: what the guest said, what the board showed, and what reached the POS. Run every order at a quiet time and again at the lunch peak, and in more than one store. No special tooling is needed, only a script and someone to keep score.
If a system handles all seven in a real lane and not a quiet room, it is worth taking to the next stage.
What does good look like in production?
A pilot is one thing. Doing this every day is another. Hi Auto processes 100M+ orders per year and reports 93%+ completion and 96% accuracy, always together, across more than 1,000 stores. Completion is the percentage of orders sent to the POS without staff intervention. Accuracy is the percentage of orders sent to the POS with every item and customization exactly right.
Jose Armario, CEO of Bojangles, put it this way about how the technology behaves with guests:
“The beauty of this technology is that we haven’t missed a beat when it comes down to the ability for this technology to interact as friendly, as effective, as efficient as a human person.”
Jose Armario, CEO, Bojangles. QSR webinar “AI Order Taker That Works: The Bojangles Model”, July 2025. Watch the QSR webinar: Armario on friendly, efficient ordering
The tricky orders you test in a pilot are the ones that decide whether those numbers hold as you add stores.
Language understanding is proven by the messy orders
Any system handles a clean order. The real test is the guest who changes their mind twice, stacks four modifiers and gets an addition from the back seat.
When you run a pilot, look at what the confirmation board shows as the guest speaks. Look at whether the results hold in the busiest hour, and whether they hold in the tenth store the way they held in the first. That is how you find out what you are buying.
Messy orders are the job. They are what a Tuesday lunch looks like.
See it in a live drive-thru
Want to see how the conversation works in a live drive-thru deployment? Talk to the Hi Auto team about what the messy orders look like in the field, across more than 1,000 stores.