Starbucks tried using AI to count syrup bottles – it kept hallucinating the inventory

[Alfonso Maruccia]

AI NO: Brian Niccol became Starbucks CEO to pursue one major goal: improving the company's profits while driving sales growth. The executive has been trying to turn things around with significant technology upgrades, but AI-based solutions have so far failed to show real gains in terms of efficiency or operational standards.

After a nine-month pilot turned out to be a dud, Starbucks recently decided to retire a new AI-based tool designed to ease its inventory management pains. The coffeehouse chain tried to bring more automation into its stores' beverage inventory, but it was ultimately forced to admit that human workers are still much better at replenishing the shelves.

According to Reuters sources, Starbucks recently sent an internal newsletter announcing that its AI inventory tool, named Automated Counting (AC), was being retired. In its place, coffeehouse staffers will have to count milk and other beverage "components" the same way they count other product types.

The AC app was developed by Seattle-based firm NomadGo, and was tested for years before Niccol took Starbucks' helm. The new executive decided to deploy AC across all Starbucks locations across North America last September, announcing a rapid rollout to make inventory management faster and more efficient.

Automated Counting was designed to automate counts of syrup, milk, and other beverage products, using tablets equipped with cameras and LiDAR sensors to identify the products that needed a refill. Sources said that the app was essentially "hallucinating" actual coffee shop inventories, failing to detect something like a peppermint syrup bottle while counting other bottles that were on the same shelf.

Starbucks confirmed to Reuters that the AC program is no more because the company has now decided to "standardize how inventory is counted across coffeehouses as we continue to focus on consistency and execution at scale."

Alternative solutions to the failed AI tool apparently include more frequent shipments of beverage products to stores, plus some unspecified improvements within Starbucks' own supply chain. The company even shared screenshots with feedback from internal staffers, who welcomed the decision to abandon AI-based automation. Meanwhile, NomadGo said that it will take this useful customer feedback to further improve its app.

Niccol deployed AC with the idea of optimizing Starbucks' inventory with smart technologies and AI, but the coffee chain has tried to embrace several technological innovations over the years. Even before its new CEO arrived, the company tried the NFT avenue, 3D printing, and other high-tech solutions, including dark-pattern tricks to allegedly rake in hundreds of millions of dollars. Though we wouldn't call the latter innovation.

Permalink to story:

Starbucks tried using AI to count syrup bottles – it kept hallucinating the inventory

 

[Squid Surprise]

While there will be the usual trolling of “AI bad, humans good”, this is clearly just a case of a terrible program not being suited to the task it was given.

Computers don’t “hallucinate “ - that’s a human term from people who don’t understand how tech works!

When an AI “hallucinates”, it simply means it wasn’t programmed properly for the task it was assigned to perform.

Don’t blame the AI - blame the people who program/use it.

 

[Alfonso Maruccia]

While there will be the usual trolling of “AI bad, humans good”, this is clearly just a case of a terrible program not being suited to the task it was given.

Computers don’t “hallucinate “ - that’s a human term from people who don’t understand how tech works!

When an AI “hallucinates”, it simply means it wasn’t programmed properly for the task it was assigned to perform.

Don’t blame the AI - blame the people who program/use it.

Computers don't hallucinate. Black-box AI models do.

 

[Plutoisaplanet]

Computers don’t “hallucinate “ - that’s a human term from people who don’t understand how tech works!

Sadly TechSpot put that word in the headline and it wasn't even in the source by Reuters. And the reason Reuters left it out is clear: it didn't hallucinate. Here are the issues Reuters reported with the tool: "But the app frequently miscounts and mislabels items, such as confusing similar milk types or missing them altogether, according to 10 cafe workers and managers. For example, a video uploaded by Starbucks shows the app failing to recognize a peppermint syrup bottle on the shelf as it counts adjacent bottles."

Missing bottles isn't hallucinating and misreading something isn't either. Making up something that doesn't exist is hallucinating. There were no "fake" items counted, and Reuters didn't report that the tool was counting items that weren't there.

 

[JaredTheDragon]

Love watching the apologists cry when this "tech" fails to be technology, time after time.

 

[Theinsanegamer]

Love watching the apologists cry when this "tech" fails to be technology, time after time.

Love watching doomers cry about new tech not being perfect right out of the gate.

 

[godrilla]

While there will be the usual trolling of “AI bad, humans good”, this is clearly just a case of a terrible program not being suited to the task it was given.

Computers don’t “hallucinate “ - that’s a human term from people who don’t understand how tech works!

When an AI “hallucinates”, it simply means it wasn’t programmed properly for the task it was assigned to perform.

Don’t blame the AI - blame the people who program/use it.

I wonder how much of so called hallucinations or errors are from memory leaks and or incompetent chief software engineering or underlings? The suits don't don't know any better in what goes into proper ai execution because non of them are likely software engineers. Going forward any company that lacks a competent chief engineer which is either part the executive team will experience this level or higher of beta testing. Biggest ai growth companies have a strong engineering team from the looks of it. This is even more so due to the current trend for ai vibe coding.
The non strong software engineering companies don't even know what ai hallucinations are. This is going to be a new trend and probably a foreshadowing of things to come for most companies that are have an ai growth outlook. Some non software engineer executive comes witha half baked idea and runs with it will only be fueled by yes men that will throw more money at the situation rinse repeat until some companies go belly up!

 

[JMGuy]

People really need to start talking about AI, and its application, properly or just admit they are uneducated.

AI is not a “black box” nobody understands except for anyone who hasn’t been educated in the technology. Nobody expects anyone to know something they haven’t learned, but people should at least attempt to bother to educate themselves before deploying any tool.

Computers do not hallucinate; they do exactly as instructed with extremely high levels of accuracy. The problem here, like most computing-issues, is the user. Uneducated people that don’t bother to learn how their tools function, apply them improperly, and then wonder why their application failed to reveal results. This is a basic cognitive failure.

Trying to use a probabilistic tool to get deterministic answers is akin to asking why a guesstimate wasn’t accurate. Really? You wouldn’t wonder why your house fell down if you eyeballed its construction instead of measure it. Or, maybe you would, but nobody would sympathize, because the reasons against it are obvious.

If people in construction used tools the way so many people attempt to use computers they’d be fired on day one for absurd levels of safety incompetence. Do not use a pen as a screwdriver or pry-bar (a true story I watched unfold that resulted in ink everywhere; my tiny violin)!

Using this tech for the type of application described in the article is just bonehead. Of course it failed. Any CEO of competence should better understand the technologies they deploy. It isn’t difficult to do. There are plenty of people and resources available to anyone willing to be an autodidact. RE: the article—we’ve had computers that calculate basic math, near error-free, since the 1960s. FFS. Use the right tool for the job.

 

[Theinsanegamer]

People really need to start talking about AI, and its application, properly or just admit they are uneducated.

AI is not a “black box” nobody understands except for anyone who hasn’t been educated in the technology. Nobody expects anyone to know something they haven’t learned, but people should at least attempt to bother to educate themselves before deploying any tool.

Computers do not hallucinate; they do exactly as instructed with extremely high levels of accuracy. The problem here, like most computing-issues, is the user. Uneducated people that don’t bother to learn how their tools function, apply them improperly, and then wonder why their application failed to reveal results. This is a basic cognitive failure.

Trying to use a probabilistic tool to get deterministic answers is akin to asking why a guesstimate wasn’t accurate. Really? You wouldn’t wonder why your house fell down if you eyeballed its construction instead of measure it. Or, maybe you would, but nobody would sympathize, because the reasons against it are obvious.

If people in construction used tools the way so many people attempt to use computers they’d be fired on day one for absurd levels of safety incompetence. Do not use a pen as a screwdriver or pry-bar (a true story I watched unfold that resulted in ink everywhere; my tiny violin)!

Using this tech for the type of application described in the article is just bonehead. Of course it failed. Any CEO of competence should better understand the technologies they deploy. It isn’t difficult to do. There are plenty of people and resources available to anyone willing to be an autodidact. RE: the article—we’ve had computers that calculate basic math, near error-free, since the 1960s. FFS. Use the right tool for the job.

Crying about other people needing education while mis defining what "black box" means. Just ROFLMAO.

Buddy, unless the AI tool is open source, it's a black box. You cannot see the inner workings of proprietary software.

computers do not hallucinate, but AI sure does!

And blaming users when the tool in question was skipping product and mislabeling thins while running autonomously......bruh.

 

[Squid Surprise]

Computers don't hallucinate. Black-box AI models do.

No, they don't... there's just a programming or user mistake that is unknown...

And why is "hallucinating" in your headline? It's not in the original article you based this off of... I know AI articles get lots of clicks... but... maybe try a bit harder?

 

[bviktor]

Is this AGI yet?

 

[Theinsanegamer]

No, they don't... there's just a programming or user mistake that is unknown...

And why is "hallucinating" in your headline? It's not in the original article you based this off of... I know AI articles get lots of clicks... but... maybe try a bit harder?

So they dont hallucinate....but if they do its user error or a programming bug? Why are we using doublespeak to deny something that occurs frequently enough that simply googling it will find dozens of examples? If the AI hallucinates, arguing about the cause is a strawman argument to distract from the fact that, yes, the AI hallucinates.

When AI encounters something it doesnt fully know, it hallucinates. Hell even if it DOES know it will hallucinate and give incorrect answers. It's not hard to do. It cant even answer basic questions with public ally available documentation, like "what weight oil does the ford maverick use" without getting it wrong half the time.

 

[Alfonso Maruccia]

No, they don't... there's just a programming or user mistake that is unknown...

And why is "hallucinating" in your headline? It's not in the original article you based this off of... I know AI articles get lots of clicks... but... maybe try a bit harder?

Do you understand what a "black box" actually means?

 

[Squid Surprise]

Do you understand what a "black box" actually means?

Yes... but just because something is constantly changing, doesn't mean there aren't REASONS for these changes. When an AI "hallucinates", it's simply because something (or many somethings) unexpected occurred, that weren't accounted for by the initial program.

Kind of like how people used to believe in magic... magic simply meant something we didn't understand... Now we call it "hallucination" -- but it's still just as misguided as a belief in magic...

Maybe read the initial article from Reuters and see what ACTUALLY happened.. no hallucinations occurred.

https://www.reuters.com/business/starbucks-scraps-ai-inventory-tool-across-north-america-2026-05-21/

In case you lost it...

 

[Squid Surprise]

So they dont hallucinate....but if they do its user error or a programming bug? Why are we using doublespeak to deny something that occurs frequently enough that simply googling it will find dozens of examples? If the AI hallucinates, arguing about the cause is a strawman argument to distract from the fact that, yes, the AI hallucinates.

When AI encounters something it doesnt fully know, it hallucinates. Hell even if it DOES know it will hallucinate and give incorrect answers. It's not hard to do. It cant even answer basic questions with public ally available documentation, like "what weight oil does the ford maverick use" without getting it wrong half the time.

Because "hallucinate" is not a synonym for "unexpected result".

If I were to give you the instruction, "turn left at every intersection, and wave at every person you see on the right", you might decide... hmm... I can't really wave at people unless I turn right instead of left... You might decide... maybe I'll turn right, do my waving, then do a u-turn and do the original left turn...

I would see you turning right at that first intersection and think "WTF?!?!!?"

Is this your fault? Are you "hallucinating" and turning right instead of left?!?!?

No... it's MY fault for giving you conflicting / unclear instructions.

Now... to add upon this... if I were to tell someone else "look at this great guy I hired - he'll always turn left at every intersection and greet everyone on the right side of the road".

That person I told would see the resulting chaos and think that I hired a fool...

 

[Axeia]

Before employing AI perhaps start with the question is this a task for AI?

Sounds like the type of thing where RFID in the packaging would do a near 100% perfect job at a fraction of the cost? just call it AIS for "All identifying system" so it still has AI in the name to make green line go up.

 

[JMGuy]

Crying about other people needing education while mis defining what "black box" means. Just ROFLMAO.

Buddy, unless the AI tool is open source, it's a black box. You cannot see the inner workings of proprietary software.

computers do not hallucinate, but AI sure does!

And blaming users when the tool in question was skipping product and mislabeling thins while running autonomously......bruh.

It is indeed fascinating that you chose to lead with clearly unearned confidence while lacking basic professionalism and fundamentally misunderstanding both computer science and my original point.

You, sir, are conflating proprietary source code with foundational computer science while simultaneously resorting to internet slang and 'ROFLMAO' emojis that only serve to underscore a baseline maturity that is just as deficient as your technical literacy.

A 'black box' in software engineering refers to an architectural system where the internal data transformations are opaque, but the mathematical and architectural frameworks governing them—such as transformer networks, weights, and probabilistic tokens—are thoroughly documented academic science. One does not need to read OpenAI’s proprietary codebase to understand how a Large Language Model processes vectors.

Regarding 'hallucination': data synthesis errors are a predictable byproduct of a probabilistic engine. The system is executing its mathematical instructions flawlessly; it is maximizing probability, not verifying factual truth.

When an autonomous system mislabels or skips data, it is a failure of system architecture, deployment strategy, and programmatic guardrails. Expecting a generative engine to act as a deterministic database without integrating validation layers is, fundamentally, a failure of technical implementation by the deployment team.

Relying on marketing buzzwords like 'hallucination' to excuse poor implementation proves my original point: a baseline education in the mechanics of these tools is required before deploying them.

Feel free to correct your comprehension gaps and/or professionalism any time there…..bruh.

 

[Liljonny11]

Maybe want to mention the type of AI before fear mongering it?