Pilot purgatory: why most AI pilots never reach production — and the one question that gets you out

Your organization probably has more AI pilots than AI products. That ratio is the problem.

The most quoted number of the past year — from MIT’s widely-cited research on the state of AI in business — is that around 95% of enterprise GenAI pilots deliver no measurable P&L impact. Even if you distrust the exact figure (I always do; verify before you quote), every survey points the same direction: only about 29% of organizations report significant ROI from generative AI, while a majority spend over a million per year on it.

So companies aren’t failing to start. They’re failing to cross. Welcome to pilot purgatory: enough activity to feel like progress, not enough consequence to be progress.

Why pilots stall — by design

Here’s the uncomfortable mechanic: most pilots are designed to avoid the organization, not to change it.

Think about what a pilot is allowed to be. Safe scope, side team, sandboxed data, no impact on the core process, no threat to anyone’s role. Every one of those properties makes the pilot easy to launch — and each one removes a reason for the organization to ever adopt the result. A pilot engineered to be harmless is also engineered to be inconsequential. The purgatory isn’t an accident. It’s the design working as specified.

Three specific gaps keep pilots suspended:

No owner of the Monday morning. The pilot has a project lead, but nobody owns the workday that’s supposed to change. When the pilot ends, the result lands on a desk that never asked for it.

No verifiable success criterion. “Promising results” is not a criterion. AI automates fastest where output can be verified — and pilots that never define what “worked” means can never be declared to have worked. They can only be extended.

No path into the workflow. The pilot ran beside the process. Moving it inside the process means touching roles, responsibilities, and systems — everything the pilot was scoped to avoid. That bill was never budgeted, so it’s never paid.

What crossing actually looks like

I’ve spent 26 years helping organizations get things out of pilot mode — most of that time before we called it that. The clearest counterexample I’ve lived is IRISnet: a telecom operator drowning in back-ends, parallel interfaces and semi-manual processes. We could have run a pilot beside that mess. Instead, the first thing we built was alignment — management and shareholders convinced of one architecture as the cornerstone of the company’s future — and then a real product, in the real workflow, for real customers, with success criteria nobody could argue with. Within a year of launch, over 50% of customers had onboarded; NPS rose 40 points; the company made it the flagship of its strategy.

The difference wasn’t talent or budget. The difference was that the work was designed to be consequential from day one — visible direction, named owners, verifiable outcomes, and the political work done up front instead of postponed.

The one question

If you want to know whether a pilot will ever leave purgatory, ask this in the kickoff meeting:

“If this works, who changes their Monday morning — and have they agreed?”

A name and a yes: the pilot has a path. Silence or “we’ll see”: you’re about to fund an experiment whose main deliverable will be a slide, and eighteen months from now it will still be “showing promise.”

Purgatory, in the original sense, is where you wait until someone decides your fate. Your pilots are waiting for the decision your organization is avoiding: whose work changes. Make that decision first, and most of the purgatory empties out.

Sources

• Writer, Enterprise AI adoption 2026
• Firecrawl, 13 Agentic AI Trends