Procurement Theater Comes Due

Starbucks didn't fail at AI. It failed at procurement, nine months ago, before the tool ever touched a store.

Reuters reported Wednesday that Starbucks terminated its worker-facing AI inventory program across North America, citing an internal newsletter and two people with direct knowledge. Nine months from deployment to scrap. The framing in the coverage is product failure. The framing is wrong.

This is the third or fourth time in eighteen months we've watched the same shape. McDonald's killed its AI drive-through pilot with IBM in 2024. Klarna walked back the customer-service AI it had spent a year bragging about. IBM Watson Health is the canonical version: a billion dollars in, no clinical workflow that stuck. The failure pattern is consistent enough now to name: procurement theater. A tool gets bought to signal modernization to a board, a press cycle, or an internal stakeholder, and the problem definition is either skipped entirely or compressed into a vendor pitch that doubles as the procurement document. Success criteria never get written down in falsifiable form. Nine to eighteen months later, the tool quietly comes out.

Reuters didn't surface whether Starbucks ran pre-deployment validation. That gap matters. It's possible this was a planned pilot that hit a defined kill switch, in which case "termination" is too dramatic a frame. But the nine-month timeline and the across-North-America rollout don't read like a contained pilot. They read like a full deployment that didn't stick. Either way, the question I want answered isn't "did the model work" — it's what the success metric looked like on day one, and who owned it.

Which brings us to the other story this week. The Wall Street Journal ran a piece in which two figures who helped launch the agentic-AI tooling boom are now warning that the code their products generate is dangerous enough to "catch up to us." Read the incentive on that honestly. People who shipped the first wave of vibe-coding tools have a clean differentiation play in becoming the responsible-AI voices for the second wave. The warning may also be sincere. Both can be true. The signal worth keeping after you discount the incentive is the engineering one: AI-generated code is entering production codebases without the review gates engineering organizations spent twenty years building, and the people who shipped the generators are telling you the output quality is uneven.

That's the same failure as Starbucks, one layer down the stack. A tool gets deployed. Problem-definition work (in the developer case, what does code review look like when 40% of commits are AI-drafted) gets skipped. The failure surfaces later, but it surfaces as technical debt rather than a Reuters scoop. No press release announces the termination, because there's nothing to terminate. The code just sits there until something breaks in production and the post-mortem points at a function nobody on the team actually wrote.

The trajectory question is why both stories landed this week. My read: we're entering the part of the cycle where the first wave of unforced AI deployments comes due. The 2024 procurements signed under "we need an AI story for the board" framing are hitting their renewal windows. Vendors are being asked to justify spend against outcomes that were never specified. The deployments where someone did the problem-definition work upstream will renew. The procurement-theater deployments won't.

What would make me wrong: if Starbucks replaces this tool with another inventory AI from a different vendor inside ninety days, the failure was about the product, not the procurement. If they replace it with a non-AI process improvement, or with nothing, the procurement read holds.

I think it'll be nothing.