Operations

AI's Enterprise Reckoning: When the Budget Runs Out and the Milk Goes Uncounted

From Starbucks' scrapped inventory tool to Uber burning through its annual AI token budget in four months, the gap between AI hype and operational reality is getting harder to paper over.

Written by Marcus Wren · Bureau Business · Jun 2, 2026

The tools are in the stores. The results aren't.

Starbucks doesn't usually make news for what it quietly stops doing. But last week Reuters reported that the company had discontinued an AI-powered inventory counting system built by NomadGo — less than a year after rollout. The reason was straightforward and unglamorous: the tool kept getting it wrong. Store employees flagged persistent inaccuracies on basic tasks, including miscounting milk carton volumes and failing to reliably track back-of-house beverage syrups.

This is the kind of failure that doesn't show up in earnings calls. It shows up when a shift manager has to reconcile what the system says with what's actually on the shelf at 6am.

Token budgets don't survive contact with developers

Uber's situation is different in kind but similar in structure. The company gave thousands of software engineers access to Anthropic's Claude Code — one of the more capable AI coding tools on the market — and watched its entire annual AI token budget disappear in four months. Individual engineers were running up monthly bills between $500 and $2,000.

Uber is now describing that spend as increasingly hard to justify and says it will rethink its budgeting approach. That's a reasonable response. It's also a signal that enterprise AI rollouts need cost controls that most companies haven't built yet. Giving developers access to a powerful tool without usage guardrails is the operational equivalent of handing out corporate cards with no spending limits and being surprised when the bill comes in.

The ROI math is brutal for almost everyone

The financial picture behind these deployments is not encouraging. Investment bank Panmure Liberum modeled the returns on AI infrastructure spending across major tech companies and found that, under best-case assumptions, Microsoft's AI initiatives are returning -9% on investment. Google sits at -15%, Meta at -28%, and Oracle at -35%. Amazon is the only one in positive territory, barely.

These are best-case numbers. The actual figures, if the deployments continue to underperform, will be worse.

Regulation is arriving before the returns do

Illinois passed SB315, making it the first state to require independent third-party safety audits, risk disclosures, and incident reporting for large frontier AI developers. Industry groups flagged compliance cost concerns. That's a predictable response, but the law reflects something real: public and legislative patience with self-regulation is running thin.

In Pennsylvania, Republican lawmakers introduced bills to repeal tax breaks for AI data centers and give municipalities authority to impose 18-month construction moratoriums. A mid-May Gallup poll found more than two-thirds of adults oppose new AI data center construction — a majority said they'd rather have a nuclear plant nearby.

The political coalition that made large-scale AI infrastructure buildout easy is fracturing. That matters for capital planning, site selection, and the timeline on which any of these investments could realistically pay off.

What operators should watch

None of this means AI tools don't work. Some do, in specific contexts, with the right implementation. But the current moment is clarifying: the cost of a bad deployment is no longer just a write-off. It's a story. Anthropic has filed a draft S-1 with the SEC, which means the scrutiny on enterprise AI performance is about to get significantly more intense.

For operators, the practical question isn't whether to use AI tools — it's whether the deployment has clear success metrics, a defined budget ceiling, and someone accountable when the milk count is wrong.

Key takeaways

Starbucks quietly killed an AI-powered inventory counting tool less than a year after rollout after store employees reported it miscounted milk cartons and failed to track back-of-house syrups accurately.
Uber burned through its entire annual AI token budget in four months after giving thousands of developers access to Anthropic's Claude Code, with some engineers running up $500–$2,000 monthly bills.
Profitability modeling from Panmure Liberum shows most major AI investors are deeply underwater: Microsoft at -9% ROI, Google at -15%, Meta at -28%, and Oracle at -35% under best-case scenarios.
Illinois became the first state to mandate independent third-party safety audits, risk disclosures, and incident reporting for large frontier AI developers, adding compliance cost pressure to an already strained investment case.
Pennsylvania lawmakers — including Republicans — introduced bills to repeal data center tax breaks and allow municipalities to impose 18-month construction moratoriums, a sign that political support for AI infrastructure is softening.

FAQ

Why did Starbucks discontinue its AI inventory tool?

Store employees reported persistent inaccuracies on basic tracking tasks, including miscounting milk carton volumes and failing to accurately track back-of-house beverage syrups. The system, developed by NomadGo, was discontinued less than a year after rollout.

How did Uber overspend on AI coding tools?

Uber gave thousands of developers access to Anthropic's Claude Code without sufficient usage controls. The company burned through its entire annual AI token budget in four months, with some engineers running up individual monthly bills between $500 and $2,000.

What does Illinois's new AI law actually require?

Illinois SB315 mandates independent third-party safety audits, risk disclosures, and incident reporting for large frontier AI developers. It is the first state law in the country to impose these requirements.

Are any major tech companies making money on AI infrastructure?

According to profitability modeling from investment bank Panmure Liberum, Amazon is the only major tech company showing a slightly positive return on AI investment under best-case scenarios. Microsoft, Google, Meta, and Oracle all show significantly negative returns.

What is driving political opposition to AI data centers?

A combination of factors: community concerns about energy use and local impact, skepticism about tax break deals, and broader public anxiety about AI's effects on jobs and society. A mid-May Gallup poll found more than two-thirds of adults oppose new AI data center construction.

Citations

AI's Reality Check Has Finally ArrivedStarbucks discontinued its NomadGo AI inventory counting system after employees reported persistent inaccuracies including miscounting milk cartons and failing to track beverage syrups.
AI's Reality Check Has Finally ArrivedUber burned through its entire annual AI token budget in four months after giving thousands of developers access to Anthropic's Claude Code, with individual engineers spending $500–$2,000 per month.
AI's Reality Check Has Finally ArrivedPanmure Liberum modeling shows Microsoft AI ROI at -9%, Google at -15%, Meta at -28%, and Oracle at -35% under best-case scenarios, with only Amazon in positive territory.
AI's Reality Check Has Finally ArrivedIllinois passed SB315, the first state law to mandate independent third-party safety audits, risk disclosures, and incident reporting for large frontier AI developers.
AI's Reality Check Has Finally ArrivedA mid-May Gallup poll found more than two-thirds of adults oppose AI data center construction, with a majority saying they would prefer a nuclear power plant nearby instead.