Starbucks has ended one of the restaurant industry’s most visible experiments in automated inventory counting. The decision is not evidence that artificial intelligence has no place in restaurant operations. It is evidence that a fast demo and a dependable store process are two different things.
Restaurant Technology News reports that Starbucks deployed a computer-vision counting tool across more than 11,000 North American locations in 2025 and retired it roughly nine months later. The system used a phone or tablet to identify products in storage. Its appeal was obvious: reduce counting labor, improve stock visibility, and give employees more time for guests.
But restaurant storage is not a clean computer-vision test bench. Cartons hide behind cartons. Similar dairy products share nearly identical packaging. Cases get split, shelves move, labels face backward, and every store develops small layout differences. If employees must inspect and correct an automated count, the tool can add a verification step instead of removing work. That is the same practical filter we recommend in our guide to AI and automation in the kitchen: judge the complete workflow, not the most impressive feature.
📝 USA-RS take: Inventory accuracy is not primarily a camera problem. It is a storage, labeling, par-level, receiving, and replenishment problem that software may help manage after the physical process is stable.
The bigger story is replenishment, not counting
Starbucks still needs better availability. Restaurant Dive reported that the company is targeting 24-hour inventory replenishment as it works to keep food and beverage ingredients available. That ambition matters more than whether a phone can count a shelf in seconds.
A precise count has limited value if ordering, distribution, receiving, and put-away cannot respond. Conversely, a disciplined operation can run reliably with ordinary counts when pars are sensible, deliveries are predictable, and exceptions are visible. Operators should therefore separate five jobs that are often bundled under the vague label “inventory technology”:
- Identification: What product is this, and what unit of measure does the system use?
- Counting: How much usable stock is physically present?
- Forecasting: How much will the restaurant need before the next delivery?
- Ordering: What quantity should be purchased, from which supplier, and when?
- Replenishment: Can the supplier and distribution network actually deliver it on time?
Automating one box does not repair the other four. A restaurant can have a flawless count and still stock out because a promotion drove demand above forecast. It can also have an excellent forecast and lose availability because cases were received under the wrong item, product was stored in an overflow area, or an employee pulled from the newest case first and left older inventory to expire.
Why computer vision struggles in real restaurant storage
Computer vision performs best when objects are visible, consistently oriented, and visually distinct. Restaurant inventory routinely violates all three conditions. A walk-in cooler may contain manufacturer cases, opened cartons, repacked pans, squeeze bottles, house-made sauces, and products transferred into clear containers. The useful unit may be a case, an each, a pound, a hotel pan, or a fraction of an open container.
Those ambiguities are operational, not merely technical. A camera can recognize a carton without knowing whether it is full. It can see two pans without knowing that one is reserved for catering. It can identify an item without knowing whether it has passed its discard time. Packaging recognition also says nothing about the quality of produce or whether a damaged case should be excluded from available inventory.
The physical environment can either reduce or multiply that ambiguity. Clearly assigned locations on commercial shelving, front-facing labels, and one item per slot make both human and machine counts easier. Scattered product, mixed cases, and unmarked overflow make every system less trustworthy. For coolers, washable polymer shelving units can also simplify cleaning and create repeatable zones without asking employees to interpret a pile of mismatched cartons.
Standardized containers matter for the same reason. A defined set of commercial food storage containers makes quantities more legible than a shelf of random tubs. A known two-quart container filled to a marked line is easier to estimate than an unmarked deli cup. For a concrete example, the Cambro RFS2PP190 round container gives operators a repeatable vessel that can be labeled and assigned to a specific ingredient.
Build the physical system before buying the digital one
Before an independent restaurant pays for AI counting, it should be able to answer a simpler question: could a new manager walk into the storage area and understand where each high-value item belongs? If not, software will digitize confusion.
| Foundation | What good looks like | Why it matters |
|---|---|---|
| Locations | One primary slot per SKU, labeled and mapped | Prevents hidden and double-counted stock |
| Units | Case, each, pound, and recipe conversions documented | Stops purchasing and recipe systems from disagreeing |
| Labels | Item, received/prepared date, use-by date, and owner visible | Connects quantity to usable shelf life |
| Pars | Adjusted by daypart, season, promotion, and lead time | Turns counts into sensible orders |
| Exceptions | Waste, transfers, substitutions, and unavailable stock recorded | Explains gaps between theoretical and actual usage |
Labeling deserves special attention. A consistent program using food storage labels gives staff information that visual recognition alone cannot infer. Broader data standards from GS1 US also show why identifiers and units need consistency across suppliers and locations. The technology can change; a clean item master and disciplined labels remain useful.
Equipment placement is part of the data model, too. If milk is kept in three refrigerators and an emergency case sits behind the bar, a count must include all four locations. Consolidating high-turn inventory in an appropriately sized reach-in refrigerator can reduce search time and counting errors. A one-section unit such as the True T-23-HC reach-in refrigerator can serve as a defined station for a compact operation, while higher-volume stores need zones designed around actual turns and delivery frequency.
Test automation against net labor, not headline speed
A vendor may say its system counts eight times faster. That number is incomplete unless the pilot measures preparation, exception handling, correction, synchronization, training, and manager review. The right metric is net labor per accurate, usable count.
Run a pilot with a control group and write down the complete sequence. How long does an employee spend straightening shelves before scanning? How often does the system confuse variants? How long does correction take? Does a manager repeat the count because trust is low? Does the resulting order reduce stockouts, emergency transfers, and waste? A fast scan followed by slow reconciliation is not a win.
Accuracy should also be weighted by consequence. Missing one sleeve of cups is inconvenient. Confusing dairy and nondairy milk can create a guest-service and allergen problem. Misreading a costly protein can distort food cost. A single accuracy percentage hides those differences. Operators need category-specific thresholds and a human review step for high-risk items.
The National Restaurant Association regularly frames technology as an operational tool rather than an end in itself. That distinction is useful here. The best pilot question is not “Did the AI recognize products?” It is “Did availability, waste, labor, and order quality improve enough to justify the subscription and change-management burden?”
Where automation can still earn its place
Starbucks ending one implementation should not trigger a retreat to clipboards everywhere. Computer vision may work well in tightly controlled beverage stations, packaged grab-and-go cases, or fixed shelf sets where products are visually distinct. Barcode scans can outperform vision for receiving. Scales can provide better estimates for bulk ingredients. POS-linked forecasting may create more value than automating the count itself.
The winning system may therefore be a stack of modest tools rather than one magical camera:
- Receiving scans that confirm item and quantity at the back door.
- Digital labels that make dates and locations visible.
- POS forecasts that adjust pars for day of week and promotions.
- Temperature monitoring that identifies inventory at risk before spoilage.
- Manual exception counts for expensive, fast-moving, or visually ambiguous items.
- Supplier integrations that shorten the time between order and replenishment.
Cold storage reliability remains a prerequisite. Inventory software cannot preserve product during equipment failure. Operators evaluating layout and capacity should start with our reach-in refrigerator guide and our comparison of walk-ins versus reach-ins. The goal is enough organized capacity to support delivery rhythm without burying inventory.
A practical 30-day inventory reset
Operators do not need an enterprise AI budget to learn from this story. A 30-day reset can reveal whether the real constraint is counting, forecasting, ordering, or storage.
Week 1: Map and clean
List every inventory location, including line backups, bar coolers, catering storage, and manager-held supplies. Remove obsolete items, assign primary slots, and mark overflow. Evaluate whether ingredient bins, shelf dividers, or additional storage capacity would remove ambiguity.
Week 2: Standardize
Define one count unit for each item and document case conversions. Align recipe, purchasing, and count names. Establish a labeling rule and make first-in, first-out physically obvious. Do not automate until two employees can count the same area and produce similar results.
Week 3: Measure exceptions
Track stockouts, emergency buys, waste, substitutions, and receiving discrepancies. Those events identify the broken link. If counts are accurate but deliveries miss, counting software is not the priority. If the item master is full of duplicate names, fix that before asking a model to recognize products.
Week 4: Pilot one narrow workflow
Choose one station or category with a measurable pain point. Compare the new method with the current process. Set success criteria in advance: labor minutes, count variance, stockouts, waste dollars, and manager corrections. Keep the pilot only if the combined result improves.
✅ Buying rule: Do not buy inventory technology to compensate for inadequate shelves, refrigeration, containers, labels, or receiving discipline. Fix the physical workflow first; then automate the stable parts.
Why we are watching this
Restaurant AI is entering a more useful phase. Operators are becoming willing to retire systems that create impressive presentations but weak store-level economics. That is healthy. A failed or discontinued pilot can save years of labor friction if leaders listen to frontline employees and distinguish a promising category from an unsuitable implementation.
Starbucks’ next test may be more important than the one it ended. Faster replenishment connects forecasts to actual availability, but it also changes storage needs. More frequent deliveries can reduce safety stock and crowded shelves. They can also increase receiving touches, delivery-window pressure, and dependence on supplier execution. Each operator should model those tradeoffs before copying a national chain’s cadence.
For independents, the durable advantage is not possessing the newest AI. It is knowing where inventory errors originate and matching the simplest reliable tool to that point. Sometimes that will be software. Sometimes it will be a scanner, a scale, a label, a better shelf, or a refrigerator sized around the delivery schedule.
If you are planning around inventory accuracy
Start by walking the operation from receiving door to storage to prep line. Note every place product becomes hard to identify, count, rotate, or replenish. USA Restaurant Suppliers can help you compare storage and shelving, refrigeration, containers, and prep equipment that make the process easier to control. Contact our team for help matching capacity and layout to your delivery rhythm, or browse the full catalog.