McDonald's ArchIQ and the Future of AI in Business Operations

About your AI Guides

Gary Sloper

https://www.linkedin.com/in/gsloper/

Scott Bryan

https://www.linkedin.com/in/scottjbryan/

Macro AI Website: 

https://www.macroaipodcast.com/

Macro AI LinkedIn Page:  

https://www.linkedin.com/company/macro-ai-podcast/

Gary's Free AI Readiness Assessment:

https://macronetservices.com/events/the-comprehensive-guide-to-ai-readiness

Scott's Content & Blog

https://www.macronomics.ai/blog

Transcript

00:57
 I'm Gary Sloper joined as always by my co-host, Scott Bryan. And today we're going to go somewhere most of us visit ah more than we'd probably like to admit and that's McDonald's. But we're not here to talk about the McRib or the McFish or the  other yummy things that they have.

01:26
ah Um, yeah. Uh, we're here to talk about one of the most significant enterprise AI developments,  uh, happening right now in American retail. It's called Arch IQ announced at McDonald's global worldwide convention in Las Vegas just this past month. And, and Scott, you're the one who kind of found the headline. Maybe you want to tell us what you saw. Sure. Yeah. The, uh, the headline is that, uh, McDonald's is attempting.

01:55
something that most companies  really haven't yet. ah They're not bolting an AI onto one part of their operation. They're building what they're internally calling a master brain, a single AI system that simultaneously manages customer order taking, kitchen throughput, inventory, and real-time demand routing.  And that's  really a different class of deployment from what most enterprises are doing.

02:21
And there are some pretty good lessons about it and  some ideas ah for every business leader that might be listening. Right. And to understand why they're doing this now, you have to understand the business pressure they're under. McDonald's just replaced their multi-year growth strategy, accelerating the arches with something new called McDonald's Next. And this was not a branding exercise. They had to make a bold move because if you look at the data between 2020 and 2024,

02:50
they've had, you know, with the persistent inflation that's existed, really hammered low income consumer spending and McDonald's core customer base pulled back on spending. So consumer perception of their value proposition dropped by double digits and stayed down into this coming year  to 2026 despite promotions and price adjustments. CEO Chris Kempinski has been very direct. The old playbook just isn't enough anymore.

03:20
Right, exactly. Yeah, and they,  what he outlined in the next strategy really organizes around four pillars. So they focused on better food, uh deeper fan personalization, restaurant productivity, and hospitality. And what's interesting technically is that ArchIQ really sits at the intersection of two of those,  productivity and hospitality. And the strategic argument is that if you  automate the purely transactional parts,

03:49
So just the purely transactional parts of the customer interaction, you can free up human staff for genuine hospitality. At least that's the thesis and it's pretty sound one, but the execution is really where it uh gets really challenging. Right. it's worth spending a minute on why McDonald's has to completely rethink their approach because this is not their first attempt at artificial intelligence ordering. ah If you go back around 2021, they piloted an automated voice ordering system with IBM and

04:20
At the  at more than a hundred US drive-through locations, I ah And from a business standpoint, it was a disaster the system, you know would add bizarre items to orders  You might receive something like nine sweet teas bacon on ice cream  That is not not delicious  At least not for me  And it couldn't handle real-time corrections. It went viral on tik-tok  Some of you probably saw some of those those reels

04:47
And then McDonald's pulled the plug in July, 2024, a  long time ago in AI terms, if you think about it two years ago. Right. Yeah, exactly. Yeah. And I, I don't think that the technical reason for that failure is directly relevant to understanding what makes ArchIQ different. uh The IBM system was cloud dependent. So every time a customer spoke, that audio had to travel over the internet to remote servers for processing and then come back again.

05:16
In a fast food drive-through,  know, you've got diesel engines idling, you've got passengers talking in the background, you've got rain coming down. So all that noise creates uncertainty in the audio signal and  any latency or network hiccup during that cloud round trip is compounded, uh all the errors. So the architecture was really just fundamentally wrong for the environment. And Gary, you and I do a lot of, you know, low latency types of designs for clients. And this is kind of, you know,

05:45
Kind of obvious. Yeah, completely agree. And so if you're listening, you're probably wondering what did McDonald's and Google do differently? The core decision was to reject cloud first processing for AI inference. Yeah. And that's the real time thinking and really a move to edge computing. Right. Yeah. And edge computing is starting to come up in more of our wide area network conversations. I know it is for me and Gary. I think it is for you too. And that

06:14
kind of naturally leads into AI use case discussions.  And so there's a lot going on with edge computing. I know a lot of carriers tried to roll it out years ago, but it was just kind of ahead of its time. But now you've got real use cases and AI applications. Exactly. Mine as well.  In one case,  I was thinking about every domestic McDonald's location is being outfitted with physical Google edge hardware and it's installed on site at the restaurant. So that's an interesting use case. These are

06:44
dedicated servers  with what are so, you know, are called ASICs.  Many of you that are familiar with that. uh It's application specific integrated circuit. It's not something new. ASIC's been around for years within hardware. And then plus you have, you know, NVIDIA tensor processing units built specifically to run the neural network models locally. And this minimizes that latency that you were just talking about, Scott.  So the AI brain, you know, the core function here lives in the building.

07:14
It lives on your main street McDonald's, not only in a data center somewhere, it's there locally. Right. Yeah. And I think the,  business implications of that architecture choice are huge. So on-premise processing  is  relatively zero latency so that the conversation really feels instantaneous. So it's not like talking to a phone system with a delay across a wide area network.  Um, and also it's important for a franchise operation. It means resilience.

07:44
So if the internet goes down at a restaurant or  it  gets really bad for whatever reason, and as you and I, Gary, know that happens all the time, the system keeps running.  So the AI model is cached locally, orders keep flowing into the kitchen, and that's a pretty big deal, especially for a franchisee who loses revenue every minute that the drive-through stalls. And that reliability is hugely valuable for them. Right.

08:11
So now let's talk about the actual voice interface because this is what customers experience.  So they've come out with what's called Archie  and  Archie and the engineering behind it addresses specific real world failures. So the acoustic layer does real time noise cancellation. So help with the use case that you're talking about, diesel engines and rain on the roof.  And it's filtering that engine noise and cabin audio to isolate just the customer's voice.

08:41
And that's foundational, but the  more interesting work happens at the linguistic level, and there are three specific capabilities worth understanding. And maybe you can jump in on that one, Scott. Yeah, sure. Yeah. And these are also conversations that I have with clients who are evaluating contact centers. And Gary, I'm sure you're doing the same. So  first is bilingual fluidity. So the system handles what linguists call uh code switching.

09:09
So when a speaker flips between two languages mid-sentence, so if customer says, can I get a number one?  Un cafe, por favor. So mixing English and Spanish in the same sentence.  And Archie is able to track the semantic meaning across both vocabulary simultaneously without losing its place in the conversation. And that's usually relevant to McDonald's customer demographics all over the US. ah And its capability most enterprise voice systems just simply don't have at that

09:37
level of seamlessness,  at least not yet. It's, coming all over the place and through contact centers, but they, most of them don't have it yet.  Um,  and, uh, Gary, you want to take intent buffering? That's next one. Yeah. Yeah. We talked about this before the  show. The second is contextual,  uh, intent buffering. So, you know, humans are messy orderers.  Uh, we change our minds mid sentence. say, give me a Sprite, wait, cancel that, make it a Diet Coke and a large fry.

10:05
may have somebody yelling in the background to screw you up as you're trying to ask for the order as well. A basic voice system treats each utterance  as a new command and gets confused. So Archie holds the entire conversational sequence of memory, figures out what you actually want, suppresses the retracted item, and updates the order correctly. This is  a direct technical fix to the IBM failure mode where...

10:32
Corrections got mangled into wrong items and then ultimately led to, you know, poor delivery, poor customer experience. Yeah. And that was, like we said, just a few years ago, it's,  and everybody knows how quickly things are changing with the technology. ah But yeah, that, what you just mentioned there, uh obviously causes huge pain at the drive-through. uh But then there's another one. It's the identity resolution. So  this is where the business strategy dimension comes in.

11:02
When a customer pulls up with the McDonald's loyalty app active, so either through geo-fencing or a quick QR scan, Archie accesses their order history. So when someone says, just give me my usual, the system looks at their historical ordering patterns, identifies the most probable order, and then repopulates it. And that's kind of the drive-through equivalent of a barista who knows your coffee when you walk in in the morning.

11:30
So it's technically a personalization feature, but strategically it deepens that loyalty relationship that the CEO was working on. And that goes right to that fan engagement pillar in their next framework.  I mean, that has to be a game changer because your risk for errors  placing that order with a human also is removed. So it's really interesting.  All right. So now let's get into the part of ArchIQ that

11:58
generally sets it apart from every other drive-through AI system on the market. The back of house  integration. Most automated ordering technology does one thing. Captures the order and passes it into the kitchen, right? Very simple. Arch IQ is bi-directional. It doesn't just receive information, it actively manages operations in real time. So if I give you an example,  think of it this way. ah It's a lunch  rush hour and the burger grill station falls behind, which can be very... ah

12:28
anxiety driven if you've ever worked as a short order cook, that the kitchen can't keep up with the quarter pounder demand. ArchIQ uh detects that through its connection to the kitchen display system, calculates what it's calling a real time uh friction index, and then without any human decision automatically adjusts the digital menu board and shifts its suggestive selling algorithm towards items the kitchen can actually fulfill like McCrispy or McNuggets.

12:57
customers see a featured promotion, hopefully the kitchen ah can catch up  because the throughput stabilizes. No manager has to make that call. So it's all done within the algorithm in real time using that display system. Yeah, just pushing the ad right to the display. ah Yeah,  and from a business operations standpoint, think about what that actually is. So it's live demand management at the point of customer interaction.

13:24
So the AI is reading supply signals from the kitchen  and shaping demand signals and pushing it right through the  drive-through um really simultaneously. So that's kind of a closed loop operational system. And what the industry hasn't had before is a single platform  doing both sides of that loop in real time. And when you extend that to inventory, ArchIQ is connected to inventory tracking too. So you've got a system that's essentially running continuous dynamic optimization across

13:54
the whole restaurant. Right. And so if you're listening right now, you're probably wondering, you know, how is it performing?  They've been running pilots at five undisclosed high volume US locations already. And excuse me, they've proposed over 1 million trends. They've they've processed over 1 million transactions. So the automated completion rate is approximately 90 percent, meaning 90 percent of the orders are handled entirely by AI with no human escalation.

14:22
That's a meaningful leap over the IBM attempt just 24 months ago, but the remaining 10 % is where the story gets really complicated. Yeah. Yeah. And  that's the hard part. So the 10 % that doesn't complete cleanly comes down to three failure vectors. So the first is acoustic interference. So the environmental conditions that kind of overwhelm even sophisticated noise cancellation. So we talked about, you know, diesel engine idling right at the speaker.

14:52
or heavy rain right at the roof. ah So, but when the system's confidence that it correctly decoded a sound drops below uh a certain threshold, it can't safely commit. So it either loops back with a clarifying prompt,  which obviously frustrates customers, or it just makes an error, which is obviously also frustrating.  Yeah. And the second failure vector is what they're calling high order semantic pivots, meaning the customer keeps changing their mind.

15:21
in layered complicated ways. Archie handles one or two corrections gracefully, but if someone says, actually, wait, go back, cancel everything, let me start over. ah The context window can break down. The system  has to hand off to a human crew member at that point. And that handoff itself ah is really the third failure vector.  Right. Yeah. Yeah. And the handoff design is  really an interesting user experience and operational problem. uh So when ArchIQ

15:51
decides to go and escalate, it routes the audio feed silently to a crew member's wireless headset so the human can hear it  and read a transcription at the same time. So if that transition isn't perfectly engineered, if the crew member doesn't immediately see a real-time text transcription  of what's been said so far, ah the customer then has to go and repeat themselves.  And that's kind of the worst possible outcome. It really...

16:19
breaks down the user experience that the automation was supposed to improve. So the quality of the escalation, the actual escalation itself is really important. So, so I did some number crunching and here's the business math that makes the 10 % more number, I would say more serious than it sounds. At 120 cars per hour during a peak lunch rush,

16:47
And high volume McDonald's locations  hit that uh at a 10 % failure rate, means 12 human escalations every hour. That's a nearly constant stream of interruptions to crew members who are simultaneously making food and delivering change and other components that they're required to do. It's not just a customer experience issue, it's an operational throughput issue. Right. Yeah, just 10%. And there's a  technical principle at work here that every

17:16
executive that's thinking about AI deployment should really understand. So getting a model from zero to 90 % accuracy is hard, but it's achievable with a well-structured training data set. And then getting from 90 to 99 requires exponentially more effort. So you're chasing increasingly rare, increasingly weird edge cases. And that's what's called asymptotic. You'll hear that in a lot of AI conversations. So meaning the closer you approach 100%, the more

17:45
impossibly expensive the next fraction of a point becomes. That's why the last 10 % always turns out to be so much harder than the first 90. Yeah. And that's a principle that transfers to every industry. The demo looks great at 90%. It's the production reality at scale that reveals the gap and bridging that gap is where most AI projects run out of runway or budget.  So let's get into the financial model because the technology only creates that value if the

18:14
economics work and the economics here are, I'd say, generally complicated by McDonald's franchise structure. Right. Yeah. And there's a lot of franchise structures out there, obviously. So  about 95 % of domestic McDonald's locations are independently owned by franchisees. So they operate within the McDonald's system, but they bear their own capital costs.  the  ArchIQ rollout requires each location to install

18:42
dedicated Google edge cloud hardware. So those are the on-premise servers within video accelerators that area. I you mentioned the beginning. And that's real upfront capital expenditure and franchisees are asking the obvious question, you what's the payback period and Gary, you I have seen this over, know, wide area network deployments across franchises where they just need, you know, router and switch upgrades. So imagine this expense. Right. And I'd say from what we'd looked at for high volume urban locations,

19:11
The ROI case is actually pretty compelling. If ArchIQ improves drive-through throughput by even 15 seconds per car during peak hours, and the pilot supports something in that range, multiply that number across 120 cars per hour, across multiple peak periods a day, 365 days a year, and the additional transactional volume is meaningful. A high volume location can build a rational payback case, but for

19:41
you know, lower volume, rural franchises doing say 50 cars at lunch instead of 120, that same capex produces a much longer payback horizon, which naturally will be a question by those franchisees owners. Like why should I be doing this? Exactly. Yeah. Yeah. So that, uh, that kind of that volume tier disparity is a, is a real point of tension right now between McDonald's corporate and its franchise associations. So corporate wants system wide rollout to

20:11
drive operational consistency and  uniformity. a  network of 14,000 US  locations all running the same AI platform is enormously more powerful than just a thousand of them running it. but franchisees want proof that the unit economics work for their specific location or situation. uh So I think how McDonald's resolves that, whether through tiered rollout timelines or

20:40
co-investment structures we've seen in some of the network things that we worked on, or subsidized hardware, I think is one of the key business execution questions  for the next couple years inside of  McDonald's. Yeah, and if we were to talk about labor, because no AI deployment  in fast food gets a pass on this question. McDonald's has been deliberate in framing ArchIQ as, ah you know, reallocation, not reduction,  employees.

21:08
And I think that framing is actually defensible, at least for now. It's a standard, you know, in a standard drive-through configuration, a meaningful chunk of front of house labor  hours go to a single task, know, standing at the speaker, taking orders, delivering cash, delivering food. That person is essentially a human audio interface when they're just  simply, you know, speaking and taking orders.  They're not making food. They're not interacting face to face with customers.

21:36
ah So when you automate that specific task, you create options. Right. And the operational deployment of that freed up labor actually does improve measurable outcomes. So redeploy that person to the food assembly line  and order errors go down. You put them at the pickup window  for genuine customer interaction, like they're looking to get to. So handing food out warmly, checking accuracy and satisfaction scores improve. So

22:05
Total labor hours per shift probably stay similar, but productivity yield per hour can go up. So the near term story  is generally about optimization and not displacement. And if, if McD's gains back some of those lost customers that we talked about in the very beginning, then perhaps ultimately they  can create jobs. Yeah, possibly. I think the longer term story is less certain once you've...

22:29
put in infrastructure that can handle 90 % of orders autonomously, and that number is going to keep rising to close the gap on that 10%. The structural case for a dedicated drive-through order taker erodes over time. That's a real conversation that the QSR industry hasn't fully had yet. Where that trajectory leads for employment over the next decade is an open and legitimate question for a company like McDonald's.

22:58
Yeah, definitely agree with that long-term. And I think McDonald's isn't alone in trying to navigate that.  we know about, we looked at some other cases like Wendy's has been running their fresh AI system, which is built on Google Cloud LLM since 2024,  and it's now deployed at hundreds of locations. So they're reporting 22 second reductions in average wait times.  Another example we looked at Yum Brands, the parent of Taco Bell and KFC. So they're also...

23:28
partnering with Nvidia to push voice AI across 500 or more stores. So the entire quick service restaurant industry is converging on AI driven ordering. But I think what distinguishes ArchIQ is not that McDonald's got there first because they didn't. It's really that architectural ambition that we covered in the beginning. They're really trying to get to that full  brain  next level. Yes, agreed.

23:54
Wendy's Fresh AI is primarily a conversational ordering interface. takes orders well. ArchIQ is claiming to be a full operational orchestration layer, slightly different  than what we see at Wendy's. And it's managing the customer conversation. It's managing the kitchen throughput and inventory signals and real-time demand steering ah as a single integrated system. If that holds up its scale, that's  a qualitatively different product.

24:24
The analogy I'd use is, know, Wendy's built, you know, a smart front door. McDonald's is trying to build a smart building. Yeah. Perfect. Yeah. And I think the distinction is the big lesson for business leaders outside of fast food. So the value of AI doesn't scale linearly with how smart the model is. It scales with how deeply the system is integrated across your operations if it's done successfully. So, you know,

24:50
Sophisticated AI that only touches one layer of your business produces modest gains, but a system that can read signals from multiple operational layers and act on them in real time produces  something closer to a real step change improvement, which is what I think a lot of CIOs and CEOs are trying to get to. So the question every executive should be asking is, where in my business could a closed loop AI system work?  know, one that reads both the customer side and the operation side simultaneously.

25:20
So, you know, how can we, how can we create that kind of leverage? Yeah. And I'd add that the architecture question matters enormously.  Yeah. The IBM failure wasn't a failure of intelligence. IBM is a great company. around for a long time. ah It was a failure of architecture, know, cloud dependent inference  in a noisy, latency sensitive, connectivity, unreliable physical environment was the wrong tool for the job. Right.

25:49
Yeah, I agree. You look now edge computing is increasingly the right model for any ah AI that operates in the physical world. Think of manufacturing floors, retail locations, healthcare, logistics. ah If your AI deployment lives in a place where real time response matters and network reliability isn't guaranteed, the ArchIQ architecture is a blueprint worth studying for your business. Yeah. Yeah.  And then

26:18
plan your failure mode before you plan your success case. every AI system has that 10 % problem that we talked about, or a 5 % or a 15%. But how gracefully it fails, how seamlessly it transfers to human judgment when it should, and how quickly it learns from those failures determines whether the deployment succeeds or becomes part of a TikTok lore. So I think

26:45
You McDonald's was the cautionary tale in 2024. They spent two years rebuilding the architecture and are trying again. And that's actually the right lesson. So not that AI ordering fail, but that failure is information and then use it to redesign towards success. McDonald's arch IQ is a  super instructive AI case study, edge computing in a real world physical environment, sophisticated NLP under generally difficult conditions, I think is.

27:16
easy argument.  Yeah, language processing. Yep. A compelling unit, uh economic thesis and a 10 % problem that will determine whether this becomes the new industry standard or just another footnote. We'll absolutely be following this one past  the summer and see this through the end. So it'll be interesting. If you're building or evaluating AI for your own operations, ArchIQ story is worth

27:43
your time, regardless of your industry, especially if, you know, those real world models in place, like retail and logistics have to make decisions. The architecture, integration depth and the failure mode design are universal.  Thanks for listening to the Macroad Podcast. I'm Gary Sloper.  And I'm Scott, and we'll catch you next time. Thanks for listening.