Starbucks Just Killed Its AI Inventory Tool | Fast Five Shorts

Starbucks has scrapped its AI powered inventory counting tool just nine months after rolling it out across its North American stores after the system repeatedly miscounted and mislabeled products, including confusing similar milk types and missing items altogether, according to the must read retail coverage in the Huff post.

Of course I'm joking, but that tells you something about this story.

Starbucks has terminated what it called its automated counting program this week with an internal company newsletter confirming the shutdown.

The tool provided by Nomad Go was deployed in September 2025 across more than 11,000 North American locations as part of CEO Brian Nichols back to Starbucks turnaround strategy.

Nomad Go had claimed the system could count inventory up to eight times faster than manual efforts and with 99% accuracy.

In a slight bit of irony too, Laura Starbucks had previously told Reuters and as recently as February so just three months ago, that the adoption of the tool had improved product availability in stores.

The company has since deleted the original September 2025 blog post announcing the rollout.

Laura Starbucks just pulled the plug on an AI inventory tool that couldn't count milk.

What does this story tell us about the state of AI deployment in retail operations right now?

And do you think this is a one off stumble or a warning sign for the industry?

I think the simple answer is I think it's a stumble.

But with any activity in this space it's always going to give us a useful data point of some kind just in case people don't know.

Nomad goes technology.

It uses devices with spatial vision, a form of computer vision, to count what's on a shelf and see what's missing.

An associate does interact with the device.

It's not a fixed camera like so much of what we know with computer vision for inventory.

And the part about a human being involved is probably the main issue and just highlights the variability and challenge and how all of these tools work.

You know, I feel like every example of AI inventory tracking, often computer vision and the use of it as shelf has taught us that there are very specific and narrow circumstances where it does work.

You, you need things to be very predictable as the oversimplified version of it.

And a restaurant and a business like Starbucks is not predictable.

It's very high turn.

There's seasonal drinks coming in, you know, in the tools defense.

All milk, whether it's a milk product or dairy milk looks the same.

And so then you mentioned 11,000 locations that multiply that by number of associates and then you've got an associate who is looking at the screen based on Nomad goes on Videos and kind of checking it.

And so the tough thing for Nomad is that restaurants and food service are listed as its top capabilities.

And so that's unfortunate for them.

You know, if it was further down their list of capabilities, that might be better.

But, yeah, I would also have to imagine, you know, that I would have thought that there'd be someone who'd recognize some of these errors.

And so it really shows you how when you get these tools into a store and into operation, store, restaurant, what have you, you're relying on so many different factors that you can control for when you're testing them.

The other thing that I do think is important to note is that this is a type of AI solution that has existed much longer than what we think of in 2026 as AI deployment and retail operations.

Definitively.

Yeah.

And so, you know, maybe this just reminds us of how much refinement there's going to be for every type of AI in retail that we think of going into the next, you know, decade.

Who knows?

You know, going back to what I heard at the lead, though, was a lot of enthusiasm for test and learn culture being first instead of just a fast follower and then learn, you know, iterating fast from there.

And so you hope that Nomad learns.

Starbucks has shown an affinity for that.

And so I don't think we have to be worried about Starbucks on that front specifically.

So, yeah, that's my take.

I think there's a lot to learn from it.

Yeah, I think.

I think 100% it's.

I think it's definitively a stumble.

You know, I think the interesting part about this, I think there's two things I'd say off what you said.

I think one is, what is the interesting point is the egg on the face.

Like, when you're talking about this being like a significant part of your operational improvements in February, it's kind of like, oh, you got to make an about face on that.

That's not good for Starbucks.

It's definitely not good for Nomad.

Go for the reasons you said, too.

And so, you know, I think that tells me too, you can't always buy into the hype cycle.

You got to give it time and to do this effectively.

My first takeaway is, my first takeaway is you have to have continuous learning.

It's not just jumping be first to try new technologies or experiment, but you have to be continuously learning and basically coming at it with the mindset of you're never going to get this right, and so you almost don't even want to talk about these things publicly because you're kind of setting yourself up for failure when you start to do that, especially when they're just at the early stages of implementation.

The other part, the word you said predictability, Laura.

For me, the reason I think that's the number one reason it's a stumble is it's predictability.

But it comes down to also the predictability angle, I would say too, is it comes down to the predictability of who's using the technology when your technology requires an employee to use it.

With the turnover in the retail industry, that just, that just is a massive issue.

And so whenever that fact is in play, it's hard to deploy a system that works consistently the same way every time.

And that's where automation works best.

Whether it's AI, computer vision, whatever.

You want to be doing the same thing the same way every time.

That's why things like robotics, fixed position cameras like you mentioned, or robotics that always do it the same way, because it's a frickin robot, not a person that you have to train each time, or even overhead rfid, which we're going to talk about as well.

So those are where I see things going in the long run.

And that's why I think this is, it's a much harder technology to implement than people think.

Even though to your point, it's been around for a really long time.

Like self scanning with a, with the employee device or an iPad is not new, but it's really hard to do it well for the reasons we.

Yeah, I mean, and I think from the tech company perspective, maybe that it elevates the need to emphasize employee training as part of the solution.

Like we're going to help, this is how we're going to help your employees learn.

Because like you said, they have to, they have to train people quickly and you know, a new person every day.

And if that's going to be a huge part of it, and if you're going to promise that the human in the loop is a big part of the solution, then maybe employee training needs to be a higher priority.

Yeah, and when I hear you say that, I'm already like, yeah, not for me, like as a former executive, I'm like, no, not for me.

Then we're going to look for a different answer that provides an easier path to what we're trying to accomplish or what our objective is.

That's how I think about what you just said, Laura.