MIT's finding moved markets the week it published: roughly 95% of enterprise AI pilots produced no measurable P&L return (2025). PwC followed with 56% of CEOs reporting no financial benefit (2026), and IBM found only 25% of AI initiatives delivered the ROI expected (2025). If you have concluded that AI does not work for businesses like yours, you are reading real numbers.
You are also reading them wrong, because the same studies name the cause. The failures cluster around unchanged workflows and unready data. And inside MIT's own findings sits the tell: solutions purchased or built with partners succeeded roughly twice as often as tools companies built alone. The 95% is not a verdict on the technology. It is a map of the common approach: bolted on, unmeasured, built solo.
What the failing pilots have in common
The typical failed pilot looks like this. A team buys a copilot or builds a chatbot. It sits on top of existing work; nothing about the workflow changes. Nobody sets a baseline, so nobody can prove anything changed. The pilot impresses in a demo, drifts for two quarters, and dies quietly in a budget review. Multiply by twenty tools and you get the pattern CEOs describe in the surveys: real spend, no financial benefit, and a growing suspicion that the whole category is theater.
Notice what never happened in that story. No workflow was redesigned. No number was owned. No one was accountable for a P&L result. The pilot was never structured to produce a return, so it did not.
The five disciplines of the 5%
The companies that clear the bar run a different program. McKinsey's high performers, the 6% reporting significant earnings-material value, are nearly three times as likely to have fundamentally redesigned workflows around AI (2025). Underneath the studies, five disciplines repeat:
- Diagnose before buying. They pick the workflow where value is provably trapped, instead of piloting whatever the vendor demoed.
- Foundation first. They fix the data the workflow runs on before scaling the workflow.
- One workflow end to end. Depth beats breadth: AI does the operating work, people hold the judgment seams, and the redesign is finished before the next one starts.
- Measurement from day zero. Baseline before the build, then one composite number, quarterly. No baseline, no proof; no proof, no program.
- Phased gates with a stop rule. Each phase states the evidence required to continue, and the program can be stopped. Paradoxically, the stop rule is what makes boards fund it.
The question to take to your team
Do not ask whether your AI pilots are impressive. Ask which of the five disciplines your current program runs, and be strict about the count. Zero or one is the 95% pattern with better branding. The full sequence, including the diagnostics that pick the first workflow and the board pack that funds it, is in The 2026 AI Strategy, free. The failure rate has a cause. Causes have fixes.
