McDonald's AI Drive-Thru debacle is a warning to us all | Hacker News Reader

McDonald's AI Drive-Thru debacle is a warning to us all

creativebloq.com

20 pointsmooreds57 comments

Went into a McDonalds recently that was all touch-screen ordering. Could maybe get someone at the counter with enough patience.

It made me feel so unimportant. It took a REALLY long time for orders to be fulfilled, and it was just me and a crowd of other customers standing around awkwardly hoping the touch screen listened to any of us.

Awful experience, avoiding McDonalds for a while.

More than once, I’ve gone through the touchscreen fast food checkout only to be shown a message at the end that the credit card reader doesn’t work. You have to print out the list of what you ordered and hand it over to the clerk after standing in line.

I know when I go to McDonalds, I go because the person behind the counter makes me feel powerful and important. It's the best. I'm flying high for a week after :).

Naw, but I get that if it took a REALLY long time for orders to be fulfilled, I'd be disappointed.

You're adding words I never said. I value the person across the counter from me as well, because in human-to-human interactions there is empathy and connection that transcends the business exchange. There is a person there that I care about way more than I care about fries, and usually that person cares more about me than they do the McDonalds Corporation.

So if they mess up my order- no big deal, you're a person after all. Or if I need to run back to get my wallet I left in my car- no big deal, I'm a person after all.

Humans generally care more about the people local to them than they do the faceless business operating in the background.

But in the human-to-AI scenario, that connection is vaporized, and now it's just me and a machine whose only loyalties lie solely to McDonalds Corporation. It makes you feel unimportant in a very cold way.

That's been in the UK for ages. Like 5 years. Everything works fine, have been to a McD's dozens of times over the last few years and never seen a problem, apart from sometimes the printer is low ink.

The reason it takes longer has nothing to do with the touch screens. It's because the stores that are touch screen are now cooking to order, no picking stuff cooked already off warming shelves.

It is annoying when you want to grab something quick. But on the other hand, no more floppy, soggy, burgers. Everything is freshly cooked. Less waste.

They just have coupled the two changes, touch screens + cook to order, together. So when a store gets one, they get the other. And if you look behind the serving counter you should be able to see the difference, no warming shelves, but a slide that fresh food is coming down, which gets immediately bagged and then put straight into your hand.

Interesting. In the USA McDonald's has been made to order for decades now in general.

I don't understand why a touchscreen is necessary to make food to order?

It's not. McDonald's chose to roll the two out at the same time. They would have benefitted from an advertising campaign raising awareness that the restaurants with kiosks are now cooked to order. Maybe their franchise rules prohibited them?

there's a process-control failure here - the reason the traditional drive-through works is that the cars stay in order and if it gets crowded then they can't reach the order speaker so it's rate-limited. The touchscreens don't do this at all, it's failure mode is the crowd just gets bigger and bigger. And since the workers don't _have_ to interact with the customers, they mostly don't, so exceptions don't get noticed or handled. I'm sure GrubHub and Uber Eats deal with similar issues to some degree, but at least in those cases the customer isn't stuck standing there, plus the apps can keep tabs on order progress and you can escalate issues. the McDonald's model is just fire and forget. And it's not just McDonalds! Taco Bell has the _exact_ same design with the exact same failure mode

When the touchscreens first came out, it was a relatively decent experience. You just chose what you wanted, maybe made a few simple modifications, and then you were done.

I went more recently and the touchscreens have adopted the same adversarial behavior that plagues much of the internet. Every other screen is begging you to add something else to your order that you have to click to close, or pushing an item that you didn't want. Dark patterns galore.

I just want to upgrade my egg McMuffin meal to a large OJ, but the kiosk won't let me.

At the McDonald's local to me, I use the kiosk because it's quicker and I don't have to repeat myself 12 times because the person at the counter can't keep up or is having trouble hearing me or is barely paying attention.

Wild you blame it all on the employee

I did not. Not being able to keep up could be me or them. Not being able to hear me well is likely because of background noise. Barely paying attention could be because of a lot going on around them. Wild that you only read it one way.

I used to be that employee. I understand the issues and it's easier for me as a customer to skip the issues.

Their kiosk UI is unusually poorly designed. It takes more than a dozen clicks to buy what used to be "I'd like a number two with a Coke" at the counter.

Their old number 1/number 2/etc was incredibly efficient and swept through the entire fast food industry.

Has anyone seen a report or analysis or similar on why they abandoned that incredibly fast and effective system and replaced it with a purchase UX (both kiosk and at the counter signage) which looks pretty incompetent to the casual observer.

Can't answer your immediate question, but in my experience, the enshitification of UI is usually a result of layers and layers on top of the user <> product connection. It's everywhere

It's a certainty that they A/B tested the UI and anything that added a cent or two got the green light. You are seeing the end result of that evolution.

I don't eat fast food much, but In n Out is where I last experienced that kind of simplicity.

Similar but the touchscreen is fine

McDonald’s “food” has not been in my diet since at least my high school and early college years. Have been avoiding them for at least a decade, long before the AI testing

Wonder how much of this failure was due to the IBM specific tech issues. Especially since McD is moving their POC to GCP

> "Developed by IBM"

I wonder how many people got fired.

> Especially since McD is moving their POC to GCP

McD has been using GCP for sometime now so it's unrelated to this fiasco [0]

The IBM-McD Voice Assistant project was actually a startup that McD acquired in 2019 that was sold to IBM in 2021 [1][2], and was even a thread on HN a couple years ago [3]

[0] - https://corporate.mcdonalds.com/corpmcd/our-stories/article/...

[1] - https://www.verdictfoodservice.com/news/ibm-mcdonalds-mcd/

[2] - https://web.archive.org/web/20180902173557/https://www.appre...

[3] - https://news.ycombinator.com/item?id=18151768

I know this is being used as an "AI bad" parable... but it seems like a really poor understanding and use of the technology? I'm not a subject matter expert but this seems like it would fail immediately.

Audio in controlled scenarios is really difficult, and the drive through experience can often go wrong with audio for real live people on both ends... why would they even expect this to half work?

Because I am willing to bet execs that greenlit this project have never worked a drive thru in their entire life. It was likely tested in ideal scenarios (ie, no background noise, order articulated clearly with no hesitation, happy path only tested).

From TFA: "After a thoughtful review, McDonald's has decided to end our current partnership with IBM on AOT (automated order taking) and the technology will be shut off in all restaurants currently testing it no later than 26 July, 2024."

I'm glad they thoughtfully reviewed after not thoughtfully reviewing the tech itself

> Developed by IBM

That's all you need to know, right there

Even before the LLM AI hype, IBM talked a big AI game but rarely delivered. Their market seemed to be out-of-touch CEOs who fell for their marketing.

> Developed by IBM

Not surprising there to be honest. Probably tested the tech in a room with no background noise, a couple of easy use cases, and an “ideal” customer that speaks clearly and knows what they want.

But also not surprised that AI has failed to replace human order takers. The order taker software by IBM reminds me of Siri, maybe they picked up Apple rejects.

Frankly I wouldn't trust IBM to tie my shoes. How they keep getting these large multi-million dollar contracts is beyond me.

Like, seriously, I'd go hire a small team of MIT grads (5-ish) and I bet they deliver a higher quality product than giving IBM four times as much and twice as long.

Because they are BIG and they understand how BIG works.

It’s very hard to be small and understand it.

I'm not convinced that's true any more.

Let me get this out of the way: McDonald's is BIG. I'm not going to argue against that.

That said, BIG in this case means about 63 million customers a day[0]. Yes, they come in bursts. No, you can't divide 63M by 86400 to address their traffic. That said, 63M requests to do something per day is no longer IBM-scale volume. At a prior job we'd designed a system to handle about 30M transactions per second at peak (think "it's the Super Bowl: guess whether the offense is going to kick or run the ball in the next play! 5... 4... 3..."). McD's requests are more complex than that, but so is the technology available to work on it today.

IBM knows BIG. I think the notion of what BIG means is a lot different now though. Lots of problems that were nearly intractable a couple decades ago genuinely are in the wheelhouse of clever startups today.

[0]https://corporate.mcdonalds.com/corpmcd/our-company/who-we-a...

There's usually some way to "cheat" to make things easier. In this case, orders don't need to ask go into a single globally accessible instantaneously updated database like BigTable. A McDonald's in San Francisco doesn't care what a McDonald's in Paris' customers have ordered right that second. Or even the one down the street. So the worst you could do is give each restaurant their own copy of the stack. And that stack meds to handle maybe 3 orders a minute? The busiest McDonald's still isn't going to be all that busy in orders per second, relative to a busy website. Everything you can do to optimize and use fewer resources per restaurant is an optimization and cost reduction problem, but the base problem is inherently tractable. Chick-fil-A runs an Edge Kubernetes cluster at each of their 2800 restaurants.

Definitely. And that's the kind of thing that could certainly be managed by a "small" (at least compared to IBM) sized team. While it's not quite as simple as "Terraform it and call it good", it's probably not radically different from that, at least in concept.

One thing about big is that they take forever to pay. Months to make the sale, months to cost the deal, months more to get paid. A tiny company may not have the resources to wait that long.

You're right, but there are two versions of how big works. One is you're buying something to solve a career need, the other is you're buying something to solve a business need.

People in big companies hire external services so that the contractor is culpable when things go wrong and they can’t be personally blamed. Most of the cost and value to them has nothing to do with the actual work. Nobody has ever been fired for trusting IBM. The same reason companies will pay big money for a commercial Linux distribution even if there is a free one that is better.

> The same reason companies will pay big money for a commercial Linux distribution

To companies like… IBM!

Haha, I was thinking about RedHat when I wrote that, not even realizing they were owned by IBM.

Most of the time, this isn't even the executive's choice but a requirement of their corporate liability insurance. It would be nice if we could just find responsible executives to fix these types of issues, but they're systemic.

Large companies only talk to large companies. So it is accenture, ibm and others. But it is amazing how their absolute garbage that they produce has no effect on their business going forward. They just keep racking in the millions upon millions.

Because their failures don’t really have an impact on their clients, who also keep racking in millions upon millions.

At that scale, thought and attention are proportionate to money moved. So even if 5 interns produced a better tool, the company wouldn't notice it. There's no skin in the game unless millions are involved.

Once the millions have moved to a contractor, it can hire 5 interns to deliver the same product. And it will be used because it cost millions.

It is not just 5 interns developing a product. It is also deploying it to how many hundreds or thousands of locations over reasonable timeframe. Sorting problems in each of them and customer support for all that.

And failing too many of those deployments get really expensive. Having your drive-trough down for multiple hours or days get franchisees really annoyed...

This was predicted by SNL about a decade ago...

https://www.youtube.com/watch?v=EcwcU0ZG8nk

If I need to interact with a company for some reason, I will usually try the website first because it can be faster. If I can't do it using the website, I call the phone number. The automated phone system usually tries to send me to the website, so I yell "AGENT" into the phone until I talk to a person.

I am tired of this.

[dupe]

More discussion: https://news.ycombinator.com/item?id=40701868

Other than the egregiousness of 2,510 McNugget meal orders, I often find that human order takers are not particularly accurate either.

Y’all are legit bad at ordering it appears. You’ll learn terse syntax to code but won’t accommodate the human order taker

Who are you talking to? IBM?

Anyone feel another AI winter coming on?

Yip. Soooo want to short NVidia, but the market can remain irrational longer than you can remain solvent.

My elderly mother was just telling me how my elderly uncle is going hard on NVDA because they make the AI chips…

So, it might be a good time to place some puts.

Totally true but every bone in my body is telling me that this bubble is going to burst in the next 18 months when the current generation of AI reveals its limitations more and more when put into production. But for now, I’m long on Nvidia but dying to short it.

Not really. The trough of disillusionment is not winter.

If so, it will be much much shorter than the last one.

Someone has been saying this for every negative piece of AI news for the past four or five years. I'm going to go with "no".

It more the overshoot of expectations. Computer Vision, LLM's, etc ... are very cool but their usefulness is not as general as folks make the out to be. The idea the LLM's demonstrate emergent behavior was a huge overshoot. The bigger idea that we are near AGI and the end of the world is waaayyy overshoot.

Reply on news.ycombinator.com