TL;DR
- A large language model (LLM) is software trained on large amounts of text so it can follow and produce natural conversation. In a drive-thru, it is the part of an order taker that follows what the guest means.
- The lane makes the conversation hard: mid-order changes, stacked modifiers, a menu full of variations and a second voice from the back seat.
- A language model does not hear through engine noise, does not know today’s menu and prices, and needs a backup when it is unsure. The system around the model has to supply all three.
- Judge the whole system by results across all live stores. Across 100M+ orders processed per year, Hi Auto reports 93%+ completion AND 96% accuracy.
“Two number threes, no wait, make one a large, and hold the pickles on the other.” A person at the speaker follows that without a second thought. For software to follow it, you need a large language model, or LLM: software trained on large amounts of text so it can follow and produce natural conversation.
This guide looks at what an LLM does inside a drive-thru order taker, where the conversation gets hard, and what the rest of the system has to provide around it. Every example stays in the lane.
What is a large language model, in plain words?
Older ordering software listened for keywords from a fixed list. A large language model follows what someone means. It can tell that “that” refers to the item just ordered, and it can answer in ordinary sentences.
That is the whole idea, and you do not need architecture terms to use it. If you want the neighboring terms, the glossary covers natural language processing and conversational AI. Our guide Natural Language Processing in QSR: How NLP Powers Accurate Drive-Thru AI covers how speech becomes a structured order, so we will not repeat that here.
What does a drive-thru lane ask of a conversation?
Most software conversations are one question and one answer. A drive-thru order runs differently. The guest changes direction mid-sentence, stacks requests and expects a reply before they have finished thinking. Here is what the lane asks of the conversation.
- Mid-order changes. “Actually, make that a large” arrives after the item is already stated. The system has to know which item “that” means.
- Stacked modifiers. “No pickles, extra sauce, and swap the fries for onion rings” is three changes to one item in one breath. Each has to land in the right place.
- A menu full of variations. Menu complexity means many base items, sizes, combos and limited-time offers that change by region and by week. The conversation has to reflect what is available today.
- Questions mid-order. “What comes on the number four?” can arrive after two items are already in the order. The order has to pick up where it left off.
- Requests the menu does not cover. Off-menu requests need a clear, friendly answer that does not derail the rest of the order.
- A second voice. A passenger calls out an addition while the driver is still talking, and both inputs belong in one order.
All of this happens outdoors, under time pressure, with guests who have no reason to adapt how they speak. Hi Auto’s average response time is 1.52 seconds, and the greeting typically comes within 3-5 seconds.
What does the model do well, and what does the rest of the system supply?
The model’s job is the conversation. It follows what was ordered, changed and confirmed, understands casual wording such as “that chicken sandwich thing”, and replies in natural language. That is the difference between an exchange that feels like talking to a person and one that feels like a phone menu.
Take the opening order. “Make one a large” points to one of the two number threes, and “the other” points to the second one. The pickles come off the second combo, not the first. Following those references is the model’s work, and it is why the order that reaches the POS matches what the guest said.
It cannot do three other jobs alone.
It does not hear through engine noise. A language model works with words, so it depends on the audio being turned into accurate words first. In a lane full of wind, idling engines and road sound, that is a separate problem. It calls for noise handling: audio filtering and speech recognition that cope with that environment. Garbled words in mean a garbled order out, however good the conversation layer is.
It does not know today’s menu. A model trained on text has no idea what is on the menu this morning or what a combo costs right now. That information has to come from a live connection to the POS, and new menu items are set up by the Hi Auto team. Without solid POS integration, the conversation works from stale information, and guests notice when a price does not match the board.
It needs a backup when it is not confident. A muffled word, an unusual request or a sentence with two meanings will come up. Hi Auto’s AI handles the order. When it cannot confidently determine what the guest wants, a trained Hi Auto supervisor resolves it instantly through the platform, and every one of those moments is used to train the model further. The process is seamless to the guest and the restaurant.
Hi Auto’s AI order taker is hybrid conversational AI. The highest-performing large language models handle the conversation, and the pieces above cover the rest.
Which questions should you ask about the conversation layer?
Ask about outcomes rather than model design. A vendor who answers everything in terms of architecture may not have production results to talk about instead. Four questions do most of the work.
- How does the system handle a change mid-order? Ask what the guest hears, what changes in the order and how the change is confirmed. Ask for a live demonstration with a realistic, multi-part order.
- How does it stay in step with the live menu? Look for a live POS connection, and ask who sets up new menu items. A periodic data export is a different thing.
- What would a pilot on our own lane show? Ask how completion and accuracy would be measured, and whether you can check both against your own POS data.
- What results do you show across all live stores? Look for completion and accuracy together, averaged across a large number of stores rather than a single pilot site. Ask which stores, which dayparts and what period the numbers cover.
If the answer to any of these turns into a talk about model confidence scores, ask again until you get a plain description of what happens in the lane.
What does good look like in production?
A pilot number is easy to show. A vendor can pick a strong location, run a controlled test and present results that look good on a slide. What matters is whether results hold across many stores, dayparts, menus and noise conditions. The mid-order change and the noisy lane are exactly the moments where a demo and a production lane tend to part ways.
Hi Auto states its two headline numbers together, because either one alone can mislead. High completion with low accuracy means orders finish but come out wrong. High accuracy with low completion can mean only the easy orders get finished. Across 100M+ orders processed per year in more than 1,000 stores, Hi Auto reports 93%+ completion AND 96% accuracy. Guests can order in English or Spanish, and the system detects the language and switches, even mid-order.
Bojangles’ CEO described how the results felt to his team.
“It’s hard to imagine there could be technology this effective if you’ve not experienced it.”
Jose Armario, CEO, Bojangles. QSR webinar AI Order Taker That Works: The Bojangles Model, July 2025. Watch Jose Armario on how the technology felt to his team
The model matters, and the system around it decides the result
A large language model makes it possible for software to follow a conversation that changes direction mid-order and to reply in natural language. That matters. It is one layer, though, in a system that also has to hear through noise, know today’s menu and prices, and recover when it is unsure.
So assess the whole system rather than the model alone. Ask for paired completion and accuracy numbers across all live stores, measured in real operating conditions, and give the demo less weight.
Next step: see the conversation layer in a live deployment
To see how the conversation layer works in a live drive-thru deployment, talk to the Hi Auto team.
