Why AI Pilots fail, and what works afterward

There are two standard explanations for this. Either the technology isn’t up to par, or the employees aren’t on board. In practice, however, you usually see something else: things go wrong in the way that first step is set up, which prevents adoption from ever taking off afterward.

An average AI pilot goes something like this: a few accounts are set up, an enthusiastic group gets to work with it, and after a few weeks, there’s a demo that impresses everyone.

There’s nothing wrong with that. The technology does indeed deliver on its promises. It’s just that this is rarely the issue that causes such a project to stall later on.

The questions that matter afterward are different. Does the work surrounding it change as well? Is a colleague who doesn’t feel like learning to use prompts any better off? Does someone trying it again three weeks later still know what it was useful for? And does anyone notice when usage declines?

These questions are rarely addressed in a pilot, because a pilot is designed to demonstrate that something is possible. So if usage declines afterward, it’s not the technology’s fault. It’s simply that no one has set up the framework needed to sustain that usage.

Five Things That Are Often Missing
Organizations that have moved past this phase consistently identify the same five points. None of them are about technology; they’re all about adoption. And they’re all rather uninspiring, which probably explains why they’re overlooked.

1. No one is responsible, with dedicated time
In many organizations, someone handles AI on the side—usually someone from IT and sometimes a member of executive management. Without dedicated time and without a mandate to reorganize the work, such a person gets no further than answering questions. So the question isn’t whether someone has been assigned to the task, but whether that person has been given the time and decision-making authority to do it.

2. There’s nothing ready that people recognize
Access to an AI tool is not the same as an actual solution. For most people, a blank text box is primarily a barrier. What does work are ready-made applications that already know what kind of work someone does: the template for your own reports, the structure of your own quotes, the questions your customers ask. That saves the user from having to do the step they like the least—namely, figuring out for themselves what this tool is good for.

3. Guidance Only Reaches Those Who Already Know How
In our baseline assessments, 62% of employees rate their own AI knowledge as basic or below, while 82% actually want to do more with it. So there’s no lack of willingness, just a lack of guidance. Early adopters will figure it out on their own anyway. The group behind them needs someone to demonstrate how to apply it once in their own work; general training on language models is of little use to them.

4. The number of licenses issued is counted
That’s the most common measurement error, because a distributed license doesn’t necessarily mean adoption. You want to know how many people are actually using it each week, broken down by department. Without that figure, you’re making decisions based on gut feeling—and that feeling is almost always more optimistic than reality.

Our baseline measurements show just how big the difference can be: employees use the free ChatGPT about six times as often as the paid license that’s already available to them. This is rarely due to a lack of willingness. People gravitate toward what they know and what’s within easy reach.

5. There’s no moment when anyone steps in to course-correct
As soon as attention wanes, usage declines. That’s to be expected—it’s simply how behavior works. In organizations where adoption takes hold, AI is on the agenda of regular meetings, and something new is added every quarter. Without that rhythm, every rollout remains a one-time event.

What you can do on Monday
Three steps that cost nothing but attention.

First,measure
Survey the entire organization to find out how many people use AI for their work each week, and which tools they use for that purpose. Do this anonymously and not through department heads; otherwise, you’ll get biased responses. The results almost always differ from what you expected—and usually in two ways at once: AI is used more than you thought, and less of what you’ve purchased yourself.

Start with the work
Visit two or three teams and ask which repetitive tasks take up the most time and provide the least satisfaction. You’ll almost always find the same thing: experienced professionals who spend a significant portion of their week typing, retyping, and fleshing out documents. Reports, minutes, summaries—endless. That’s not what they were hired to do, and that’s where they add the least value. That’s usually where your first application lies—not with the most technically impressive idea.

Assign someone who has the time to work on it
One person, with hours per week and the authority to adjust the process. That doesn’t have to be a technical role, and often it isn’t the best choice. What carries more weight is whether someone has authority on the work floor: whether colleagues listen to this person, whether he or she can build support, and whether they can actually get a change in the way things are done implemented. Technical knowledge can be learned, but you don’t build that kind of trust in a single quarter. Someone who’s part of the day-to-day operations often already has it—and, moreover, knows exactly where the bottlenecks are.

The layer beneath the pilot
The first steps organizations take are not a waste. They demonstrate that the need exists and where the value lies. What’s often missing afterward is the layer beneath it: ownership, something that’s ready to go, guidance for the broader group, real data from a real request, and an opportunity to make adjustments. Together, these elements are what make adoption possible and give an organization control over what happens with AI.

That’s primarily organizational work, change management, and people-centered work. And that’s where those four out of five projects get stuck—on something that has very little to do with AI and technology.

This blog post is a contribution from AIStudio. AIStudio is a single, secure AI work environment for your entire organization. Want to learn more? Visit AIStudio at booth 92 during Data Expo on September 9 and 10.