The AI Transformation Sequence

Gartner expects worldwide AI spending to reach $2.59 trillion in 2026, up 47 percent. McKinsey's 2026 survey of 1,719 business leaders found the share of organizations attributing any EBIT impact to AI flat at 37 percent, unchanged from the previous year, with 6 percent qualifying as high performers. Spending grew by almost half. The share of companies who can find the result in their earnings did not move.

The usual diagnosis is immature measurement. The evidence from 2026 points somewhere more specific and more fixable: programs are failing in a predictable order, and the order is the problem.

This paper is built on two cases with unusually good documentation. Meta's Project OT, reported by Reuters in August after a review of internal documents, set out to make the company AI native, explored cutting some teams by 60 percent, and planned two rounds of layoffs. The first round ran in May. The second was canceled after internal figures showed code changes to internal platforms up 220 percent year over year while changes that reached users rose 36 percent, major incidents up 40 percent, and time spent resolving them up as much as 70 percent. The second case is a direct-to-consumer business the author worked with from pre-MVP through launch, which produced a 10,000 item catalog in six weeks with two people, and then discovered its pipeline could generate 500 items a day while review could approve 200.

Both organizations accelerated production. Both found that the constraint had moved downstream rather than disappeared. One restructured as though it had disappeared and lost a program and a round of layoffs it did not need. The other treated the new constraint as an engineering problem and funded it.

Two further engagements sharpen the method. A global manufacturer wired automatic ticket creation into its project tracking system, made intake free, left triage unfunded, and lost the ability to tell an urgent request from a passing remark. A multinational consumer products company found internal communication degrading as employees drafted memos with AI, where no queue formed anywhere and the artifact simply got worse as it got longer. The second case is the one the constraint framework does not cover, and it yields a test: between the generation step and the person who bears the cost of its output, can you place a filter? A catalog can be reviewed before a customer sees it. A memo cannot, because the person bearing the cost is the person it was addressed to.

The paper turns all of this into a method. It gives a screening procedure for locating the binding constraint before funding an acceleration, a distinction between substitution and expansion investments that carry different economics and different gates, the filter test and the categories of work where generation should be withdrawn, a full cost model for the evaluation layer built from real figures where inference was 2.4 percent of the true total, and a five gate sequence with a specific evidentiary test at each gate. It closes with a one page brief a steering committee can approve or reject on the evidence rather than the narrative.

The argument in one line: the return on an AI investment is set by what you do about the constraint the acceleration exposes, not by the acceleration itself.