Why most AI rollouts return nothing

If you've rolled out AI and can't point to what it changed, you're in the majority. This lesson explains why that happens even when people genuinely use the tools, and why the fix starts somewhere most rollouts never look.

People are faster. The business isn't.

MIT's 2025 study of generative AI in business found that 95% of organisations get no measurable return from their pilots. It's an early study and the number is blunt, but every larger survey since points the same way. McKinsey's 2026 survey of 1,719 companies shows where the value goes missing.

AI isn't failing to work. Individual people really are faster. The gap is between faster people and a better business, and it has a cause you can see in any piece of work that passes through more than one pair of hands.

Every piece of work has one slowest step

Work moves through steps. A request comes in, someone drafts, someone reviews, someone approves, it goes out. Each step can handle a certain amount a week. The step that can handle the least sets how much comes out the other end. That step is the bottleneck.

Speed up any other step and nothing changes at the end. The extra work just waits longer in front of the bottleneck. Try it below. First give AI to every step except review, then try fixing review on its own.

Goldratt built a whole theory of management on that sentence, the Theory of Constraints, in a novel about a failing factory. Your organisation isn't a factory. The arithmetic is the same.

Why AI can make things worse, not just no better

AI doesn't only speed up the wrong steps. It floods the right one. When engineering teams adopted AI heavily, Faros AI found they merged 98% more pull requests, and the time spent reviewing them rose 91%. The step that checks the work became the bottleneck, and AI fed it faster than it could cope.

The same thing happens outside engineering. Researchers at BetterUp and Stanford have a name for AI output that looks finished but shifts the real work to whoever receives it: workslop. It's what speed without a bottleneck in mind produces.

Fix the bottleneck, and the next one appears

When you fix the slowest step, another step becomes the slowest. That isn't a failure. It's the job. You saw it in the simulator: fix review, and drafting becomes what everything waits on.

That's why the Deployment Framework is a loop rather than a project. You find the bottleneck, improve it, build only what improving can't fix, give what runs an owner, and then find the next one. This course takes you around that loop once, in 90 days.

A bottleneck only exists relative to a goal

Which step is slowest depends on what you're trying to produce. If the goal is more signed contracts, the bottleneck might be legal review. If the goal is a faster month-end close, it's somewhere else entirely. Before you can find the bottleneck, you need to say what you're aiming at.

That's why this course starts with the aim, and not with tools or training.