Starbucks' AI-built replacements for Microsoft and IBM software move the 4:30 a.m. page in-house

A SaaS contract buys an escalation path. The dev.to essay "Coase's Coffee" makes that case, and its author wrote that "a SaaS contract is a counterparty with skin in the game" ^[13]. The essay then walks through an incident: "When the inventory system is wrong today, Starbucks calls Microsoft, and an army of people whose careers depend on that phone call gets paged. When the in-house replacement is wrong at 4:30 in the morning before the rush, Starbucks calls Starbucks." ^[6]

Insourcing takes the vendor's support staff out of that path and puts Starbucks engineers in it. The author does not claim the bill goes away. According to the author, part of it turns into engineers and pager rotations, and into the company slowly learning that saying the model wrote the code is not acceptable in a Sev-1 postmortem ^[8].

The published figures leave little margin. About $10 million of expected software savings against roughly $400 million a year in software spend is 2.5 percent ^[1]. That saving is booked against this fiscal year. The first replacements could roll out by the end of 2027, pending testing ^[3][4]. For the in-house code to pay back, the yearly cost of the engineers who own it and the rotation that answers its pages has to come in under the license fees Starbucks stops paying Microsoft and IBM ^[1]. The essay, relaying Bloomberg, does not say what those fees are.

Automated Counting is the essay's strongest-looking evidence. In May, Starbucks retired that vendor-built AI system for counting milk and syrups after it kept confusing products that looked alike, and stores went back to counting by hand ^[5]. I think the essay is right about where the maintenance cost lands. I think it leans on this example harder than the example can hold. The failing system came from a vendor, and the remedy was still retirement and a return to hand counting ^[5]. It also failed as a model misreading products at runtime. The replacements Bloomberg described are software built with AI tools ^[1]. That code fails the way other software fails. The difference is that someone at Starbucks has to debug code a model wrote, at the hour the essay describes.

The insourcing case itself is sound engineering. The essay frames it with Ronald Coase's 1937 paper The Nature of the Firm: a company brings work inside when coordinating it internally is cheaper than buying it ^[12]. Starbucks' problems are narrow. The essay lists inventory across roughly 40,000 stores and espresso machines that have to keep running. It adds that the data comes off Starbucks' own machines and registers, and that the tools being replaced are general-purpose products ^[11]. A narrow domain with first-party data is where a small internal team can beat a product built for everyone.

By the essay's account, the cost AI moved is the cost of producing working software, cut "by some large and still-unmeasured factor" ^[9]. "The honest version of the math has three terms, and AI only changed one of them," the author wrote ^[10]. The cost of keeping that software running after it ships is outside that one term. "Maintenance is where AI-built software goes to get quietly expensive," the author wrote ^[7].