Three AI Defeats in One Week: Lessons for the Mid-Market

Key Takeaways

• AI still needs a lot of oversight. After a nine-month pilot, Starbucks shut down an inventory counting system in over 11.000 stores because employees had to manually verify every count. 99 percent accuracy and eight times faster inventory were promised.
• The bill skyrockets with usage. In mid-May 2026, Microsoft canceled most internal licenses for Anthropic’s AI coding assistant and brought its developers back to its own tool. Reason given in the internal email: too expensive because too successful.
• Starbucks: when AI can’t tell milk from milk

  In September 2025, Starbucks rolled out an inventory counting system in its North American stores, based on an AI model from the provider NomadGo. Employees scan the shelves with a tablet; the system automatically counts using LIDAR and camera data and transfers the result to the merchandise management system. The advertising promise: 99 percent accuracy, eight times faster than manual inventory.

  On May 19,

  Uber: Billions Spent, ROI Not Measurable

  According to Fortune, perhaps the most openly communicated case comes from Uber. The company rolled out AI coding tools to approximately 5.000 engineers, with the adoption rate rising from 32 to 84 percent between February and April 2026. By April, 70 percent of all code commits were co-written by AI tools. Sounds like a textbook example of adoption for any conference stage.

  The other side of the numbers, also documented in Fortune’s reporting on May 26, 2026: Uber exhausted its AI budget for 2026 in four months. Per developer, the monthly bill ranges between approximately 460 and 1.840 Euro, depending on the intensity of use. COO Andrew Macdonald stated in the earnings call, in essence, that as long as they cannot clearly draw the line between AI usage and delivered product improvements, the investment will become increasingly difficult to justify.

  Key Figures Uber Budget

  • Rollout: approximately 5.000 engineers equipped with AI coding tools
  • Adoption: 32 to 84 percent between February and April 2026
  • Output: 70 percent of code commits with AI contribution in April 2026
  • Costs: approximately 460 to 1.840 Euro per engineer per month
  • Budget Status: 2026 annual budget consumed after four months
  • COO’s ROI Statement: Connection to product improvements not clearly demonstrable

  What’s remarkable is not the high consumption, but the transparency. Uber is thus demonstrating live what is currently causing headaches for many mid-sized company CFOs: high investment, high usage, but no clearly attributable impact on business results. If a publicly listed tech company with its engineering depth cannot draw this line, a German mechanical engineering company with 200 employees certainly won’t be able to do it off the cuff.

  What SMEs Should Take Away From This

  The three cases reveal three different vulnerabilities: quality, cost curve, and impact measurement. From these, concrete consequences can be derived for investment decisions in SMEs, which should be included in every template for 2026.

  First: Make verification effort a mandatory disclosure. Every AI investment requires an estimated figure for how many human verification minutes per 100 AI outputs are necessary to safely operate the process. This figure is not a cosmetic flaw; it determines the business case. If the verification effort is 50 percent or more, the tool is not an automation but expensive duplicate work.

  Second: Realistically calculate token models. For every tool with consumption-based billing, a scenario must be included in the spreadsheet where usage per user increases three- to fivefold. Exactly that happened with Microsoft and Uber. Anyone who only calculates for the pilot group and forgets about scaling actively creates a cost trap.

  Third: Firmly establish impact measurement before launch. A rough, honest key metric before project start is worth more than precise reporting afterward. Which lead time, which error rate, which sales metric should change by how much? Anyone who cannot answer that question on day one will not be able to answer it after twelve months either and will then find themselves in an earnings call with their own shareholders, just like Uber.

  Fourth: Mitigate vendor lock-in with license clauses. Microsoft implemented its change within weeks because the tool was interchangeable. Anyone who deeply embeds an AI tool into their own workflows without agreeing on an exit period and a data extraction mechanism in the contract will be taken advantage of by the vendor at the first price change.

  Fifth: Orient pilot sizes toward learning objectives, not hype. A pilot with 20 carefully selected users provides more reliable data than a broad rollout to everyone. Starbucks equipped 11,000 branches before it was clear that employees had to double-check every count. This order is reversible and even imperative for SMEs.

  Frequently Asked Questions

  Do these three cases mean that AI makes no sense for SMEs in 2026?

  No, they mean that the standard sales slides are too optimistic. AI brings measurable benefits in narrowly defined use cases, such as in text processing, case prioritization, or structured evaluations. What doesn’t work is the uncritical adoption of promises like 99 percent accuracy or ROI in six months. Those who start with a clearly defined pilot and realistically estimate verification effort and cost curve will benefit. Those who roll out indiscriminately across the board will lose money.

  How high should the AI budget of a mid-sized company be in 2026?

  There is no flat figure, but a general rule: Plan for double the initial offer and define a hard cap, beyond which the investment will be stopped. Uber used up its budget in four months because usage increased faster than planned. For SMEs, amounts in the low five-figure range are usually sufficient for a serious pilot. The cap is important, not the size.

  Which contracts with AI providers are particularly risky?

  Token- or consumption-based contracts without a hard monthly cap are the biggest operational risk. With popular tools, they can deplete an entire fiscal year’s budget within a few weeks. Anyone signing such contracts should negotiate at least a monthly hard cap, an automatic usage report, and a 30-day notice period. Flat-rate licenses per user are more expensive upfront but more predictable.

  How do you measure the impact of AI if Uber can’t?

  With a single key metric per use case, defined before the pilot starts. In sales, for example, the quote turnaround time; in accounting, the rate of documents posted without intervention; in customer service, the average handling time. A rough figure that everyone accepts is worth more than a precise dashboard that no one maintains. Uber measures too much at once, which is why it lacks a clear direction.

  What about German AI providers, are they better?

  The logic is identical. Whether the model comes from a US or DACH provider changes neither the verification effort nor the token mathematics. What German providers often do better: data protection contracts in German, shorter response times, and a more realistic pricing model without pure consumption-based billing. For the risks described here, origin is secondary, contract design is primary.

  Editor’s Reading Recommendations

  • Stanford AI Index 2026: Inaccuracy Beats Cybersecurity as Top Risk
  • When AI Tools Suddenly Eat into Margins
  • When AI Can Do Everything: Why Trust Becomes the Hardest Currency Again

  More from the MBF Media Network

  digital-chiefs

  Microsoft’s AI Bet Devours CIO Budgets

  cloudmagazin

  The Inference Bill Nobody Budgeted For

  securitytoday

  Self-Replication: AI Agents from 6 to 81 Percent

  Cover Image Source: Pexels

  Also available in

  FrançaisEspañolDeutsch