Your AI pilot rarely fails on the model. It fails because no one introduces it. And introduction is a craft, not a prompt.
An AI rollout and a classic rollout across 80,000 users in 16 countries have more in common than the model debate suggests. Both are multi-tier. You take a mid-sized, dynamic pilot, look at it closely, check what holds, and then scale up. You create the organization, processes and tools that later carry the mass. Once the first thousand users are cleanly live, the remaining 79,000 are execution. The craft behind it is the same one I have been doing for years, including the works councils.
Why pilots fail at introduction
I see three patterns. First, unstructured: AI is made accessible, perhaps even promoted, but not consistently introduced. „Have a look, feel free to try it“ leaves eighty percent of the potential lying with chronically overloaded employees. It is the old joke about the woodcutter who has no time to sharpen the axe because he has to fell trees. Second, unmotivated: outside narrow applications the benefits are often not even clear. Third, unregulated: where AI produces mass, the editorial review becomes decisive, because the results are usable only once they meet corporate standards.
The mistake is always the same: you start with the tool. What has proven itself is the reverse order. First the organization with clear responsibilities and subject-matter experts in the areas, then the processes with sign-offs, checklists and test descriptions, and only then the tool: which connection, which provider, which artifacts. And the effect has to be measured. It is measured by time saved and better quality, not by token consumption, which is admittedly the harder task.
What a change practitioner brings that a data scientist does not is the experience of how value is not just identified but realized. Between the gap and the solution lies much more than a good model evaluation. So test your next AI pilot against one question: is it introduced, or is it merely made available?
