The Drive-Thru is a Design Problem

I was standing in the back of house during a QSR site visit, headset on, listening to the drive-thru line work.

The AI was handling it cleanly. A customer pulled up to the speaker, said what they wanted — no fumbling, no repeating, no "I'm sorry, could you say that again?" The order moved through. The customer pulled forward. From where I was standing, it looked like exactly what the industry has been promising: frictionless, fast, almost elegant.

Then the POS went down.

What happened next had nothing to do with artificial intelligence. The manager on duty became something closer to a circus acrobat — physically running between the front counter and the drive-thru station, manually reconstructing each order, conducting ten things simultaneously with no baton and no score. The AI had done its job. The system around it had not.

I stood there, headset still warm from the conversation the technology handled perfectly, watching a human sprint to hold together what the technology couldn't see.

Then it hit me: the drive-thru doesn't have a voice AI problem. It has a systems design problem.

The Four Systems

The drive-thru is not one system. It is five — and the industry has spent fifty years optimizing four of them.

The physical lane. The communication chain that moves a spoken order into a kitchen. The production system that turns that order into food. The technology layer — POS, KDS, AI, forecasting — that ties the others together, until it doesn't.

Every major QSR brand has people whose entire job is one of those four systems. Facilities owns the lane. IT owns the technology. Operations owns production. Vendors own the communication hardware. Decades of optimization, billions of dollars of investment, and a 25-year industry study tracking every second of service time.

The drive-thru accounts for an estimated 50–70% of revenue at major QSR chains — and with 75% of all restaurant traffic happening off premises, it is no longer a convenience feature. It is the dominant commerce format of American eating. And it is slowing down. Average service time has climbed to over four minutes — up from sub-three-minute benchmarks a decade ago — not because operations have gotten worse, but because menus have gotten more complex. 

The drive-thru ordering moment was designed around a specific set of assumptions — a driver, a static menu, a single-lane queue, a human on the other end of a speaker. None of those assumptions are universally true anymore. Mobile orders enter the lane mid-queue. Multi-passenger vehicles produce competing voices. Menus have grown from dozens of items to hundreds. And the speaker box — degraded, acoustically unengineered, unchanged in its fundamental interaction model for fifty years — is still the primary interface between a customer and a $60 billion commerce channel.

This is not a speed problem. It is a complexity problem that has been mistaken for a speed problem — which is why every technology intervention for the last decade has been aimed at throughput rather than experience. Faster AI. Faster screens. Faster timers. The lane gets instrumented. It does not get designed.

And still, the average drive-thru takes over four minutes.

The design question isn't how to make AI work at the speaker. It's what the ordering moment should actually be.

What The Drive-Thru is Missing

By the time a car has been sitting in line for three minutes without reaching the speaker, something has already shifted. I've watched it happen from the lane — the small ticks first, a head shake, fingers drumming, the phone coming out. Then one of two things: the car noses out entirely, or it stalls when the car ahead moves because the driver has already left mentally. The order isn't lost at the speaker. It's lost in the queue.

Science has a name for this phenomenon. Queue psychology research has documented for decades that perceived wait time inflates with every minute past three — and that once the five-minute mark is crossed, perceived duration drastically inflates. The drive-thru, with an average service time now exceeding 5 minutes, is operating inside that inflation zone for most of its customers, most of the time. It is not a coincidence that abandonment rates climb in the pre-speaker portion of the lane. The science is specific: pre-process waits — the time before service begins — feel measurably longer and generate more frustration than in-process waits. The minutes before the speaker are experienced as longer than the minutes after it. The queue itself is designed to feel worse than it needs to.

And then the customer reaches the speaker.

What happens at that moment is one of the most cognitively compressed interactions in everyday commerce. Time pressure. Social pressure — the car behind has been waiting too. Auditory interference from road noise, engine noise, a speaker system that has been dropped and degraded and never acoustically engineered for the environment it lives in. And a digital menu board that, in the worst cases I've seen from the customer side, presents a wall of promotional graphics where the actual menu used to be — because the marketing team pushed a campaign to the screen without anyone checking whether the base products were still visible. The result is a customer who pulls up ready to order and instead says: "Uh... can I have the... where is your..."

If an interface makes something notably worse for the user, the fault is in the design, not the customer. 

From inside the system, the failure compounds. I've seen digital menu boards running promotions that aren't connected to the POS — the screen shows one thing, the register doesn't know it exists, and a team member has to manually reconstruct a customized order in real time because nobody thought to push the same campaign to both systems. The customer experiences a delay they can't explain. The team member absorbs a pressure spike the technology created. The gap between what the screen promised and what the system can deliver lands, invisibly, on a human being.

Voice AI carries the same pattern into the ordering moment itself. Voice AI can handle 72% of orders end-to-end — but achieves 81% accuracy doing it. When a human employee steps in after AI, accuracy jumps to 95%. The technology is optimized for the happy path. The human judgment that catches what the AI missed is what makes the lane actually function.

The McDonald's-IBM pilot made this visible at scale. Between 2021 and mid-2024, McDonald's deployed IBM's Voice AI to more than 100 U.S. drive-thru locations. The system failed publicly — nine sweet teas instead of one, random items added to orders, adjacent lane orders crossed. McDonald's ended the pilot in June 2024.  

While sophisticated, algorithms tend to miss one crucial environmental reality: the drive-thru is an acoustical warzone. Engine rumble, wind pressure against the microphone, car radios bleeding, competing speech, passengers talking over each other – all this creates a cacophony of chaos for the technology. Research on automatic speech recognition has documented for decades that systems achieving over 95% accuracy in controlled environments can fall below 50% in real-world noise conditions. 

A system designed for this environment requires advanced beamforming — microphone arrays that create a spatial focus on the driver's voice, filtering everything else out. The IBM system processed every sound it could hear. That is how one car's order ended up on another car's tab.

The lane was never modeled as a design environment. The technology was built for a conversation. It was deployed into a construction site.

We have thirty years of science documenting what makes waiting tolerable and decisions easier, yet none of it seems to have permeated the average drive-thru’s design. Why? Not because of an an absence of operational expertise, but an absence of experiential design. 

The Accessibility Silence

The drive-thru has always had a user it wasn't designed for. It just never had to admit it.

I was at a BBQ chain on a Sunday — post-church rush, the busiest hour of the week. The owner, Jose (name changed)_, told me something that stopped me mid-conversation: "The older crowd hardly ever uses the drive-thru, and the ones that do — I end up repeating myself longer than it takes for them to actually order anything."

I looked out at the dining room. Most of the dine-in customers were older. Several had mobility aids. A few wore hearing devices. They had made a choice — not about food, but about which format of commerce they could actually access. 

A survey by Inclusion Solutions found that 42% of deaf respondents reported leaving a drive-thru without ordering. The speaker box — already acoustically degraded, already engineered for nothing — is the only interface available. There is no alternative input. There is no fallback. The format simply excludes people who can't use it and calls that a preference.

Voice AI alone does not solve this. Deployed without non-standard speech patterns, regional accents, or alternative input modalities in mind, it may widen the gap rather than close it. 

The McDonald's-IBM system that struggled with background noise was the same system being asked to interpret the speech of elderly customers, non-native English speakers, and anyone whose voice doesn't match the training data. Accuracy rates that look acceptable in aggregate can mask systematic failure for specific populations.

The drive-thru has treated accessibility as a compliance problem — the minimum required by law, applied at the minimum possible scale. That is a different design intent than asking: who can't use this, and what would it take to include them? The distinction determines whether you design the floor or design for the full range of human variation.

The Fifth System

Because there is a fifth system. And nobody owns it.

It has no interface. No documentation. No dashboard. It runs entirely on human judgment — the continuous, real-time, invisible work of adapting the operation to whatever reality has decided to do today. Who bumps a ticket when two orders collide? Who pulls a car forward when production falls behind? Who reallocates labor at 12:07 on a Tuesday when three things fail at once?

The industry built technology on top of an assumption: that human judgement would handle everything the technology couldn’t anticipate. 

And that assumption, that the gaps between systems would be handled by someone, means the technology works exactly as designed: with painful fissures that force operators to become the human integration layer. 

Veteran operators think about this layer constantly. They model the failure path — because that's where they live. The rest of the industry models the happy path. And then it deploys technology into it.

What Systems Design Would Actually Look Like

I want to tell you about a spatula.

I was in the back of house of a U.S. QSR location, doing what I always do in stores — watching how the system actually runs, not how it's supposed to run. There was a spatula sitting on the counter near a food station. I picked it up to put it away. The manager stopped me.

She needed it.

I watched her take the spatula and use it to press the bump button on the kitchen display system — the screen mounted above the station that sequences orders through the kitchen. The button was too high for her to reach with her hand. So the spatula became the interface. Not a workaround, not a complaint — just how that kitchen runs. Nobody had filed a ticket. Nobody had flagged it. The KDS wasn't broken. It was working exactly as designed. The design had just never accounted for the full range of people who would use it.

That spatula is the fifth system made physical. A gap filled by human ingenuity so quietly that it becomes invisible. And invisible gaps don't get fixed — they get inherited.

The drive-thru has never had a systems designer in the room. It has had operations people, technology vendors, facilities architects, and marketing teams — each optimizing their own layer, handing off to the next, assuming someone else owns the boundary. Nobody does. The result is a $60 billion commerce channel held together by spatulas, sprinting managers, and the decision-making of whoever happens to be on shift.

Systems design would start by treating the four-minute drive-thru experience as a narrative, not a transaction. What should be true at the pre-speaker moment — when queue psychology says the customer is already losing patience? What should happen at the order moment — when cognitive load is highest and the interface is worst? What is the handoff designed to feel like, not just function like?

It would apply environmental design to a lane that has never had any. Acoustic engineering for a speaker system that has been degraded and ignored for decades. Menu board design that reduces cognitive load at the exact moment decisions have to be made. Signage that gives customers a sense of progress before they reach the speaker — the single cheapest intervention the research supports, and the one nobody has made.

It would design for the full range of people who use the lane, not the median. Not as a compliance exercise. As a design standard.

And it would finally design the human-AI handoff — the moment that may matter more than anything else in the lane right now, and the moment that currently receives the least design attention of anything in it.

Every Voice AI deployment has an escalation path: the moment the AI can't handle the interaction and a human takes over. In about 21% of AI-assisted orders, employees stepped in — most often because the AI couldn't handle customization, answer a question, or process an unavailable item. That moment is currently an afterthought — a fallback, not a feature. It has no designed trigger, no designed tone, no designed experience for the customer who just watched a machine fail to understand them. It is the most sensitive moment in the AI-assisted drive-thru, and it has been left entirely to chance.

Designing that handoff — when it happens, how it's communicated, who owns it, what the customer experiences in the transition — may be more important than optimizing AI accuracy by another two percentage points. Because the handoff is where trust is either recovered or lost. And right now, nobody owns it.

The brands that figure this out will not be the ones with the best AI. They will be the ones with the best designers — who finally showed up for the lane.

At 2:40 PM

At 2:40 pm, in the same location where I'd watched a manager sprint between the front counter and the drive-thru station hours earlier, her phone alarm went off. The lane was empty. Every customer in the dining room had already ordered. She walked to the kitchen and started a drop.

I asked her what she was prepping for.

"The high school lets out soon," she said.

It took me a moment. She was dropping fried chicken at 2:40 so it would be fresh for the hold time — because she knew, from years of standing in that lane, that students would come through in about twenty minutes. No dashboard told her that. No forecast model flagged it. No AI prompted the drop. She just knew.

At 3:11 pm, two cars driven by teenagers came through the drive-thru. They knew her name. She asked about their grades.

Her judgement – dropping the chicken at 2:40 pm – was the highest-value labor decision made in that location all day. It is the fifth system at work, honed over years of building two-minute relationships through a speaker box. It is a system that doesn’t live in any database, yet is the most valuable thing operating in that lane. 

The technology that keeps arriving to fix the drive-thru isn’t built by people who have worn the headset when the POS goes down, watched a manager become the integration layer between systems that were never designed to talk to each other, or seen a spatula forced to become the KDS interface.

They will keep building building systems that work in the demo and break in the lane.

Meanwhile, the fifth system will keep running as it always has: through un-captured operator intelligence that is acting as a human integration layer. 

It deserves, finally, to be designed for. 

_____

Jessa Parette is Head of Design at Byte by Yum! and an advisor to design and product leadership teams. She works at the intersection of AI-driven systems, organizational complexity, and experience design at scale. Speaking and advisory: www.jessaparette.com

_____

REFERENCES

Intouch Insight. (2025). 2025 drive-thru study. QSR Magazine / Intouch Insight. https://www.intouchinsight.com/resources/studies/drive-thru/ and https://www.qsrmagazine.com/story/the-2025-qsr-drive-thru-report/

Katz, K.L., Larson, B.M., & Larson, R.C. (1991, Winter). Prescription for the waiting-in-line blues: Entertain, enlighten, and engage. MIT Sloan Management Review, 32(2), 44–53. https://www.researchgate.net/profile/Richard-Larson-9/publication/304582002_Prescription_for_the_Waiting_in_Line_Blues_Entertain_Enlighten_Engage/links/5774627708aeb9427e2422c4/Prescription-for-the-Waiting-in-Line-Blues-Entertain-Enlighten-Engage.pdf

Larson, R.C. (1987). OR forum — perspectives on queues: Social justice and the psychology of queueing. Operations Research, 35(6), 895–905. https://doi.org/10.1287/opre.35.6.895

Maister, D.H. (1985). The psychology of waiting lines. In J.A. Czepiel, M.R. Solomon, & C.F. Surprenant (Eds.), The service encounter: Managing employee/customer interaction in service businesses (pp. 113–123). Lexington Books. https://www.columbia.edu/~ww2040/4615S13/Psychology_of_Waiting_Lines.pdf

McDonald's ArchIQ AI drive-thru: Google-powered system explained. (2026). Abacus News. https://www.abacusnews.com/mcdonalds-archiq-google-ai-drive-thru/

McDonald's ends AI drive-thru test with IBM. (2024, June 17). CNBC. https://www.cnbc.com/2024/06/17/mcdonalds-to-end-ibm-ai-drive-thru-test.html

McDonald's quietly tests AI drive-thru ordering at 5 Chicago locations. (2026). Techlicious. https://www.techlicious.com/blog/mcdonalds-ai-drive-thru-test-chicago/

McDonald's to test AI drive-thru system at five locations. (2026, June 5). The Washington Times. https://www.washingtontimes.com/news/2026/jun/5/mcdonalds-test-ai-drive-thru-system-five-locations/

National Restaurant Association. (2025, April 16). From trend to transformation: Off-premises dining now essential for restaurant consumers, operators [Press release]. https://www.prnewswire.com/news-releases/from-trend-to-transformation-off-premises-dining-now-essential-for-restaurant-consumers-operators-302429894.html

New drive-thru for the deaf. (2012, January 23). QSR Magazine. https://www.qsrmagazine.com/news/new-drive-thru-deaf/

Gong, Y. (1995, April). Speech recognition in noisy environments: A survey. Speech Communication, 16(3), 261–291. https://dl.acm.org/doi/10.1016/0167-6393(94)00059-J