Starbucks Ties Tech Bonuses to AI After the Store Tool Failed

Starbucks is tying 25% of most technology workers’ bonuses to department goals that include AI adoption, according to an internal document. Software developers must use an AI assistant multiple times a week to stay in compliance.

The pay rule was set down in May 2026. That is the same month baristas were told to peel QR codes off backroom shelves and count milk by hand again.

The 25% AI Slice of the Tech Bonus

The document covers most tech workers, not a small pilot group. Starbucks declined to comment. People familiar with the matter, who were not authorized to speak publicly, said developers are scored on whether they open an AI assistant several times a week.

The company is also counting how many priority projects in its Back to Starbucks plan are powered by AI. That is a headcount of tagged work, which is easier to audit than a faster latte or a cleaner inventory sheet.

THE TECH BONUS SPLIT

The 25% department slice is not AI alone. It also folds in other tech priorities, including how the order-and-pay function on the Starbucks app is doing. Still, AI is now written into the same bonus math as the company’s own sales.

Managers who want a clean dashboard will get one. Weekly assistant use is a binary that a system can log. Whether those prompts cut ticket times, reduce outages, or improve the mobile queue is a different score, and it is not the one named in the compliance rule.

The Inventory Scanner Died in May

Automated Counting was the store-floor proof of the same efficiency pitch. Built with Redmond startup NomadGo, it put computer vision, 3D spatial tools, and augmented reality on handheld tablets so cafes could scan beverage components instead of ticking them off by hand. Starbucks finished the North America rollout by the end of September 2025, putting automated counting at 11,000 stores and then some, all company-operated.

At launch, NomadGo said the system could count up to eight times faster than a manual tally, with 99% accuracy in controlled tests. Then chief technology officer Deb Hall Lefevre told stores the scans would free partners from the back room. She resigned on September 29, 2025, the same month the rollout finished.

THE STORE AI CALENDAR

1. End of September 2025: Automated Counting is live across company-operated North American cafes.
2. April 3, 2026: Starbucks tells NomadGo the tool is finished. The 30-person vendor cuts a large share of staff, including the technical team on the account.
3. May 18, 2026: Baristas get the memo. QR tracking codes come off the shelves. Milk and other beverage components go back to hand counts.
4. May 2026: An internal document pegs 25% of most tech bonuses to department goals that include AI adoption.

People who worked on the tool described physical misses, not a lab demo that never left Seattle. A shift supervisor near the city aimed an iPad at a steel fridge; the camera caught the reflection and five oat milk cartons registered as ten. Other stores saw milks marked as the wrong type, syrups swapped, and a bin counted as food. In Graham, Texas, patchy Wi-Fi wiped a count mid-scan after managers had already ruled hand counts out, which left the store with no usable number.

WHAT BROKE ON THE SHELF

• Reflections: Shiny fridge doors doubled what the camera thought it saw.
• Lookalikes: Similar cartons and bottles were labeled as the wrong product.
• Dead air: A dropped connection erased a count that staff were not allowed to replace by hand.
• Moving SKUs: Seasonal cups and limited-time packaging could need weeks of retraining, and developers sometimes learned about a new item only once it was already on the shelf.

NomadGo chief executive David Greschler said computer vision struggles when the inventory itself keeps changing. People on the project also pointed at Starbucks’ older IBM AS/400 backend, which made live store data hard to move. Insiders estimated the program cost more than $10 million over several years. That figure is an estimate, not a filed total.

There’s nothing you can do when leadership and strategy change.

David Greschler, chief executive, NomadGo

He called the April notice a complete surprise. Starbucks told staff it was standardizing how inventory is counted so stores could focus on consistency and execution at scale. A spokesperson later said the company uses technology to support human connection, that it had put $500 million into extra coffeehouse staffing, and that it changed course when the tool fell short.

That store failure is the other half of the AI inventory tool Starbucks retired. HQ still wanted an AI number it could put on a bonus sheet. The scanner that had to work in a walk-in cooler did not survive. The prompt log did.

Why Weekly Prompts Count as Compliance

Using an assistant multiple times a week is a volume rule. It does not ask whether the code merged, whether the app crashed less, or whether a store got oat milk on time. Companies that want proof of AI spend like volume because it shows up fast.

Research on pay design is blunt about that trap. A Wharton note on incentives argues that value versus volume in AI metrics is the split that matters, and that scores should track business outcomes, not logins. Microsoft, in that same roundup, ties a portion of team OKRs and bonuses to efficiency gains and customer impact from AI workflows. Accenture has gone the other way for promotions, with HR tracking weekly log-ins on internal AI tools for senior managers who want leadership jobs.

Starbucks is closer to the login model than to the outcome model. A developer who never opens the assistant can miss the compliance bar even if the work is clean. A developer who pastes every function into a chat window can clear it even if the product does not move. Short-term waste is likely, because people will feed the meter. The intended message is still the one managers keep repeating elsewhere: skipping the tools is now a performance problem, because the output gap between people who use them and people who do not is obvious to anyone reading a sprint board.

The metric also collides with cost. If the same group is later asked to write software that replaces licensed systems, every prompt has a token bill. A cheap-looking usage target can raise the inference tab that the software cut was supposed to lower.

Mobile Order-and-Pay Shares the Same Scorecard

The 25% department bucket is a bundle. Beside AI adoption, the document names the performance of order-and-pay in the Starbucks app. That is the line customers already fight with: mobile tickets stacking up, drinks waiting on the bar, cafe guests watching a bagging shelf fill.

So a Seattle engineer can be paid, in the same slice, for using a coding assistant and for whether the app’s order pipe holds up at 8 a.m. One of those is a personal habit. The other is a system that also depends on staffing, store layout, and how Smart Queue sequences drinks.

Store AI did not all leave with the scanner. Green Dot Assist on store iPads is the barista-facing companion Starbucks first piloted in 35 coffeehouses, built to answer recipe and machine questions in the flow of a shift. Order sequencing tools remain part of the Back to Starbucks operations kit. The inventory camera is the piece that went back to paper.

Green Apron partners, the people on those shifts, got a different cash hook. At the close of the third quarter, Starbucks launched a Best of Starbucks Reward that lets eligible store partners earn up to $300 a quarter for coffeehouse goals across sales, operations, and customer service. HQ tech is being graded on prompts. Stores are being graded on the floor.

Building Replacements for Microsoft and IBM

In July 2026, an internal presentation showed where the prompt habit is supposed to pay off. Starbucks is building in-house replacements for Microsoft and IBM tools it currently buys, including a Microsoft system that tracks inventory and an IBM tool that manages maintenance. Some of that homegrown software could roll out by the end of next year, if tests hold. The company has also been working for years on a point-of-sale stack to replace Oracle Simphony.

Chief technology officer Anand Varadarajan told workers in an internal forum that Starbucks spends about $400 million a year on software and that there are “clear opportunities to reduce the spend in software.” AI-assisted coding was central to the IBM replacement, according to the presentation. The same presentation said the enterprise technology team is on track to cut its budget by about $30 million in the fiscal year ending in late September.

THE TECHNOLOGY BUDGET CUT

• Software bill: About $400 million a year, per Varadarajan.
• This year’s cut: About $30 million off the enterprise technology budget by late September.
• Licenses: About $10 million of that from software spending.
• Contractors: Another $13 million, mostly by cutting professional-services contractors and filling some roles with staff, including new offices in Nashville and India.

Those pieces do not add to $30 million on their own; the rest of the cut was not broken out. The work sits inside a wider push to take $2 billion out of costs and to review every technology contract and service. The bet is that assistants make internal software cheap enough to walk away from vendors. The risk, which showed up as soon as the presentation leaked, is that token spend, testing, and a second inventory system could cost more than the licenses they replace, especially after the last inventory AI had to be unwound by hand.

Other companies with large software bills will try a version of this. If a coffee chain can train engineers to ship maintenance software, a bank or a grocer can too. That is why the bonus rule is not a side HR note. It is the staffing lever for a build-versus-buy shift that Starbucks has already put on a calendar.

Half the Payout Still Hinges on Company Results

Whatever the prompt logs say, 50% of the tech bonus still follows company results. That half does not care how many times Copilot opened. It cares whether guests came back.

On the July 29, 2026 earnings call, covering the quarter ended June 28, chairman and chief executive Brian Niccol reported global comparable store sales of 7.9%, driven mainly by a 4.2% rise in comparable transactions and a 3.5% rise in average ticket. North America comps were up 8.1%. U.S. comps were up 7.9%, on a 4.2% rise in transactions and a 3.6% rise in ticket. Consolidated net revenues were $9.3 billion, down 1%, reflecting the China joint-venture shift. Non-GAAP operating margin expanded 430 basis points to 14.4%, and non-GAAP earnings per share grew 70% to $0.85.

Our Back to Starbucks plan was built on the belief that an extraordinary cup of coffee, human connection and customer experience win the day, every day. Our third quarter results are proof they do.

Brian Niccol, chairman and chief executive officer, Q3 fiscal 2026 earnings call

Niccol took the job in September 2024 after a run of falling comparable sales. He has leaned on Green Apron Service, extra labor, Smart Queue, and store “uplifts” more than on backroom cameras. Food availability is close to 99%, about 10 points better than a year earlier, he said on the same call. That gain landed after Automated Counting was already gone, which is a reminder that the sales recovery and the AI bonus are running on parallel tracks, not as cause and effect.

Full-year guidance now looks for U.S. comparable store sales slightly greater than 6% and global comps nearing 6%. The fourth-quarter call is tentatively set for October 29, 2026. Before that, the fiscal year closes in late September, which is when the technology group is supposed to show the $30 million budget cut and when those weekly assistant logs become part of a live bonus score rather than a memo.