Starbucks

Starbucks has reportedly retired its AI-powered “Automated Counting” inventory system across North American stores this week — less than a year after rolling it out company-wide.
The system used computer vision, 3D spatial intelligence, and AR-enabled tablets to scan shelves and count inventory like syrups, milk, and cups much faster than manual checks.
In theory, it sounded like a perfect retail AI use case.
In practice, real stores are messy.
The tool reportedly struggled with:
Similar-looking products
Partially obscured items
Shelf clutter
Inconsistent lighting
Missing or misplaced inventory
Examples included confusing milk varieties, missing bottles entirely, or failing to recognize seasonal syrups like peppermint.
Instead of improving inventory visibility, the errors sometimes created additional supply-chain friction.
Starbucks is now reverting to manual counts while continuing broader operational and supply-chain improvements under CEO Brian Niccol.
The bigger lesson here is important:
AI often performs extremely well in controlled demos and structured environments. But deployment in chaotic, real-world physical settings is much harder.
Retail stores generate endless edge cases:
Damaged packaging
Human stocking inconsistencies
Constant layout changes
Occlusions
Lighting variation
Seasonal product churn
That’s where reliability becomes more important than raw capability.
This doesn’t mean AI in retail is failing. It means the industry is learning that replacing human operational workflows requires extremely high accuracy — especially when small errors compound across thousands of stores.
Classic example of the gap between “AI can do the task” and “AI can do the task reliably at scale.”