Every failed pilot in MIT's 95% shares one trait.

It isn't the model. It isn't the sector. It isn't the budget. It's the layer.

The finding itself, 95% of enterprise GenAI pilots with no measurable P&L impact, gets quoted in every steering committee in the country. Usually as a warning. Rarely with an explanation. So look down any list of failed pilots and check what kind of project each one is. Nearly every one is an Execution project: automate this task, accelerate this step, draft this document.

Execution, Execution, Execution.

And Execution deployed on its own fails twice, by design.

Execution without Perception is blind. The pilot automates a task, but nobody has verified the task is still the right one. The process was designed for market conditions that may no longer exist, and the pilot inherits that judgement unexamined. An agent doing the wrong work with precision isn't progress.

Execution without Design is fragile. A probabilistic component gets wired into a deterministic process that was never built for it: no confidence routing, no structured data, nothing to absorb the mismatch. The pilot impresses in the demo, then meets reality's edge cases with no architecture behind it.

At an international bank, an AI pilot cut retail loan processing time by 50%. In production, that meant one thing: form-filling time on the customer journey fell from 1 minute 40 seconds to 50 seconds.

Half the process. Fifty saved seconds. That was the whole prize.

The layers have a sequence: Perception first, Design second, Execution third. Run it in that order and agents deploy into a structure built to channel them. Start at the end, as 95% do, and MIT's number stops being a mystery.

It's the failure rate of building in reverse.

The full argument is in "The Ferrari Engine in the Horse Cart". The architecture itself is Chapter 12 of Architecture of Intellect.

Which layer is your current AI portfolio actually invested in? Count honestly.