Pilot Purgatory — The Pattern Killing Your AI Initiatives

So, leadership approves the expansion. And then slowly, with no single bad decision to point out, it stops. The pilot is still a pilot 18 months later. It is then either quietly discontinued or it survives on paper, while the people who build it have moved on to other priorities. This pattern has a name, and I call it a Pilot Purgatory. Right now, it's not a rare failure. It is a typical outcome for AI initiatives. This episode is about understanding Three research groups have measured the same problem from different angles.

The first one, Gartner, tracked generative AI proof-of-concept projects throughout 2025. By the end of that year, 30% had been abandoned after the initial proof-of-concept. Not canceled because the technology failed, but simply not continued. Another one is a subdivision of MIT, MIT Nanda, looked specifically at embedded AI tools, the kind that actually change how people do their jobs rather than stand-alone demonstrations. Their funnel is instructive. 80% of organizations invested into a tool, 50% piloted one, and only 5% successfully implemented in practice. McKinsey took the broadest view.

Their analysis concludes that only 39% of companies could attribute any measurable EBIT impact to their AI investment. And here, we are speaking not about significant impact, but just about any impact. So as you see, three different research groups, three different methodologies, three different ways of measuring. And they all converge on the same picture. Most AI investments right now are not reaching the point where they generate measurable value. The question worth asking is not whether this is happening or not, it is happening. But it is where things actually go wrong, because the failure modes are distinct.

Knowing which one is operating in your organization tells you what specifically to address. This is why I want to spend some time The first one is technology failure. Here, the model output is not sufficient. Accuracy is too low for the decision to be trusted, or the required capability does not exist for the quality use case requires. This is the failure mode most organizations assume is dominant when things move slowly. It is also the least common. The pace of model improvement means that many use cases which genuinely could not be solved reliably a couple of years ago are viable now.

Technology failure has a clean diagnostic question. Does this capability exist in any available tool at any price? If yes, technology is not your constraint. The second reason is integration failure. Here the AI works, the output is useful, and it lives somewhere the workflow does not reach. It may be a separate application, a separate tab, a step outside the place where the work is actually being done. Employees have to go looking for the output rather than encounter it naturally. This failure mode is more common than technology failure.

Normally, a demo environment is always designed around the tool. The real workflow is not. Asking busy people to add an extra step, even a valuable step, to every task they perform is a significant request. Most will not do it consistently. As a result, the tool first becomes optional and then it changes its state to dormant, and nobody uses it. The third place where AI breaks is people and organization failure. Here the technology works. Integration is technically feasible. At the same time, adoption is assumed rather than designed.

Nobody has a clear mandate to drive behavior change. Change management was not planned from the start. It was expected to happen naturally once the tool becomes available. Leadership attention that was present during the pilot has moved to the next initiative. Employees who would benefit most from the tool were not part of designing it and feel no ownership over it. Procurement took nine months, and as a result, the tool arrived feeling already too old. More than 1,200 companies across 68 countries looking at what actually separates organizations generating real AI value from those stuck in pilot purgatory.

Here is their finding on what success depends on. 70% is determined by people, process and organizational change. 20% by infrastructure and integration. And only 10% by the model or algorithm itself. And Deloitte put a concrete number on what inverted that ratio cost. Normally, a team builds something really useful. For example, a tool that cuts a significant manual process from hours to minutes. The technical work is solid. The business case is clear. A pilot with a small group of people produces results that are hard to argue with.

And finally, approval to scale comes through. Then the approvals begin. First, security review. Four months. Then integration sign off from the central architecture team. Two months. After that, a data protection assessment because the tool processes personal data. Another six weeks. By coincidence, a leadership reorganization happens mid-process and the original executive sponsor moves to a different role. And the incumbent leader has different priorities and no personal investment into the initiative. By the time the tool clears all the gates, a different team in a different part of the company is piloting a different platform that covers similar functionality.

And nobody wants to have conversation about which one to continue. The first project quietly stops. It is important to know that nobody in that sequence acted in bad faith. Every approval step existed for a legitimate reason. All of these factors at the same time resulted into a cumulative 12-month delay that the project momentum could not survive. This is not a dramatic failure and there is no single decision that caused it. The process worked exactly as designed. But the result is unfortunately a pilot which has never been implemented in practice.

There is also a second pattern worth mentioning here and which industry practitioners are sharing more and more on the AI conferences which I attended in the last three months. MIT research found that employees from more than 90% of surveyed companies use AI tools regularly on their own. While only 40% of those same organizations have an official AI subscription. The people using AI most actively in their personal lives are frequently the same people who encounter the widest gap when they come to work. When the gap is persistent enough and visible enough, a specific thing happens.

Those employees just stop advocating for better internal tools. They adapt around the official environment instead. They use personal accounts for what they cannot do officially, which creates exactly the data governance exposure the careful approval process was designed to prevent. The response to slow governance is not slow governance. The response is governance that moves at the speed the environment now requires. for this episode is Pilot Purgatory Diagnostic. It is free, the link is available on Unreasonable AI platform and in the description or in pinned comment depending on where you are watching this video.

There are four sections which will take about one hour to complete. You can do it alone or with a small group of colleagues. The first one is portfolio inventory. This is the list containing every active AI initiative in the company with name, status, owner, information about current stage. Writing this down often reveals something immediately. The portfolio is more fragmented than leadership realizes. There are initiatives running in parallel that nobody had compared. If portfolio is too long, you should agree on the list of main initiatives which are currently running and analyze them.

The second one is the failure mode assessment. For each initiative, there are eight questions across the three failure mode dimensions. The goal is not to score each of them. It is to surface where the real constraint is sitting. When technology scores are high and people scores are low, improving the model is not the intervention that matters here. The third is the governance gap audit. Here you should list five things employees in your company can do with AI personally today that they cannot do through official company approved channels.

For each item, estimate how long the standard approval process would take. This number will tell you the pace at which your governance system currently operates, relative to the pace at which AI capabilities are moving in the real world. And the final one is the output. This is the result of the analysis and should contain just one initiative to examine differently and just one governance process worth accelerating. This is enough to make the next steering meeting a different kind of conversation. All the materials mentioned in the episode are available on the Unreasonable AI platform.

The link is also provided in the pinned comment or in the description, depending on the platform on which you are watching it. The failure modes gives you a diagnostic language. They tell you where to look when something is not progressing. But they do not fully explain why the dominant one, people and organization, is so consistently under invested in across so many companies and many industries. Underneath, it is a structural pressure that most companies have not named precisely. A gap between what people can do with AI personally and what their companies officially enable them to do.

And why that gap tends to widen rather than close on its own.