They start without an owner, without an agreed measure of success, and without the people expected to use them seeing the tool work on their own tasks. The model is rarely the problem. The NoW of Work calls this pattern Pilot Purgatory: experiments everywhere, connected to nothing, with no measurable impact.
How often do AI pilots fail?
Often enough that a stalled pilot is the normal result. In The GenAI Divide: State of AI in Business 2025 (July 2025), MIT's Project NANDA looked at enterprise-grade AI systems, custom-built or bought from a vendor. Sixty percent of organizations evaluated one. Twenty percent reached the pilot stage. Five percent reached production. The report's summary is blunt: "95% of organizations are getting zero return."
- 60%evaluated one
- 20%reached the pilot stage
- 5%reached production
Grey: organizations that stopped at that stage
Source: MIT Project NANDA, The GenAI Divide: State of AI in Business 2025 (July 2025)
The report's explanation matters more than its headline. It puts the failures down to "brittle workflows, lack of contextual learning, and misalignment with day-to-day operations." Every item on that list is about how work is organized.
What does Pilot Purgatory look like from the inside?
It looks busy. The licences get bought, a few enthusiasts run experiments, a vendor demo goes well and a pilot launches in one department. Months later nobody can say what changed, because nobody agreed what change would count.
Meanwhile, people use AI anyway, just not the pilot. The same MIT report found that while only 40% of companies said they had bought an official LLM subscription, workers from over 90% of the companies surveyed reported regular use of personal AI tools for work tasks. That's Shadow AI: real use, outside the pilot, with nobody owning the risk or the results.
The pilot
40%said they had bought an official LLM subscriptionThe real use
90%+had workers using personal AI tools for work tasksSource: MIT Project NANDA, The GenAI Divide: State of AI in Business 2025 (July 2025)
So the pilot stalls while AI use grows around it. What's missing is direction and ownership.
Why does the team matter more than the tool?
People use AI regularly once they've seen it work on their own tasks. The NoW of Work surveyed 333 staff in organizations working with NoW. Among people who agreed that their team has practical examples of how AI can improve its work, 61% said AI had become a regular part of how they work. Among people who disagreed, 10% did. (Agree means 7 or higher on a 0 to 10 scale; disagree means 0 to 3.)
The same staff show where effort tends to go. 85% understand when AI output needs human review, and 75% know what information not to put into AI tools. Only 18% say they've received practical AI training relevant to their work, and 31% say their team has practical examples. The rules reached people. The hands-on help didn't.
The rules
- 85%understand when AI output needs human review
- 75%know what information not to put into AI tools
The hands-on help
- 18%have received practical AI training relevant to their work
- 31%say their team has practical examples
Source: NoW of Work staff survey, 333 people in organizations working with NoW
A pilot owned by IT, or by one enthusiast in one corner, never produces those team examples. It can work perfectly and still change nothing, like a great practice that never makes it into a game.
What do pilots that get out of purgatory have in common?
They come from the team, and they're judged against a measure agreed before they start. The NoW of Work's method has four steps, and it doesn't begin by picking a use case for you.
Build the AI Backlog from across the team
Collect ideas from the people doing the work, as well as from leadership.
Vet the ideas
Decide which are worth testing and who would own each one.
Pilot one AI tool on real work
One owner, one workflow, actual tasks.
Validate whether it adds value
Measure against the result you agreed at the start. We call that result the AI Dividend: named by you, measured your way.
The order matters. A pilot chosen by a vendor, or by whoever argues hardest, skips the first two steps and reaches the fourth with nothing to measure against.
How can you tell if your pilot is stuck?
Ask four questions of any pilot you're running now.
- Can someone name the person who owns it?
- Was the measure of success written down before it started?
- Has anyone outside the pilot team seen it work on their own tasks?
- Do you know what your people already use AI for outside the pilot?
Any "no" shows you what's missing, and that's where to start.
Questions about failed AI pilots
Is the technology usually why an AI pilot fails?
Rarely on its own. The reasons the MIT report gives are about workflows and fit with day-to-day operations. Ownership, an agreed measure and examples from the team tend to decide the result.
Should we stop running AI pilots?
Run fewer, and run them differently. Choose each one from your own backlog, give it an owner, and agree the measure before it starts.
Who should own an AI pilot?
A named person close to the work, with a sponsor on the leadership team. In The NoW of Work's view, AI belongs to leadership, not to IT alone.
How do we choose which pilot to run first?
From your team's backlog, vetted against what matters to the business. We don't pick a use case for you, and you should be wary of anyone who does before they've seen your work.

