TL;DR
- An AI working group usually inherits a capability that sits in two or three individuals' personal habits. Meetings can describe that capability. They cannot move it.
- The stall is a transfer problem. What a champion does well is mostly undocumented judgment: what to ask, what to distrust, what to check. A group that shares opinions leaves that judgment where it was.
- Organizations that move past the ceiling tend to treat the champion's working method as an artifact to be extracted, written down, and made reviewable. They stop treating it as a talent to be admired.
- A useful diagnostic: if the strongest AI user left tomorrow, would the organization's results fall back to where they were? If so, the capability was never the organization's.
The pattern: a group formed, and little changed
A familiar sequence plays out in mid-sized organizations. Leadership sees that AI tools are being used unevenly. A working group forms, with representatives from several departments. It meets regularly and compares experiences. Someone circulates a list of promising use cases, and a policy draft appears. After a few months, the group still exists and is still busy, yet the way work gets done on an ordinary Tuesday looks much as it did before.
The usual explanations are insufficient time, insufficient budget, or insufficient executive attention. Those can be real. But the pattern recurs even where all three are present, which suggests a structural cause. The working group is the wrong instrument for the job it was given.
A working group is built to coordinate opinion. The problem it was asked to solve is that capability is distributed unevenly, and capability is not an opinion.
What a champion actually has
Nearly every organization that adopts AI informally develops a champion: someone whose results are visibly better than their peers'. The common explanation is enthusiasm, or technical aptitude.
Closer observation points elsewhere. A champion's advantage is usually a set of small, unrecorded judgments:
- how much context to supply before asking for anything;
- which kinds of output to accept on sight and which to verify;
- when to abandon a thread and begin again;
- how to phrase a request so the failure modes surface early;
- which tasks to keep away from the tool entirely.
None of this appears in a policy document. Most of it is not articulated even by the champion, who experiences it as "just knowing how to use it." This is tacit knowledge in the classic sense, and it has the classic property: it transfers through shared work and shared artifacts, not through discussion.
When a working group asks the champion to "share what works," the champion offers examples of good prompts and impressive outputs. The room reads these as the capability. The prompts are the visible tip of it. The judgment that produced them, and the judgment that rejected the weaker attempts, stays private.
Why the ceiling forms where it does
This is the point at which many organizations stall between the early and middle stages of what WBA has described as a three-stage maturity pattern (see The Three Stages of AI Maturity). Early adoption is individual. Each person builds a private relationship with the tool. The next stage requires something different in kind: a shared layer of context, conventions, and checks that makes one person's competence reproducible by another.
The champion ceiling is the gap between those two conditions. Three mechanisms keep it in place.
Admiration substitutes for extraction. Organizations tend to celebrate the champion publicly. Celebration feels like progress, and it relieves pressure to do the harder work of documenting how the results were produced. The champion becomes a resource consulted informally, which concentrates load on them and leaves everyone else's practice unchanged.
The group has no mandate over artifacts. A coordinating body that can issue guidance but cannot own the shared materials (reference documents, standing instructions, review steps) has nothing to change. Guidance that lives in a slide deck competes with habit, and habit wins.
Nobody is assigned the unglamorous middle. Turning a champion's method into something others can use means interviewing, observing, writing, testing with a second person, and revising when the second person gets a different result. This is closer to technical writing and process design than to innovation, and it rarely appears in anyone's role. An adjacent pattern, where problems persist because no one owns their cause, is examined in The Correction Bottleneck.
A generic illustration
Consider a regional distributor whose operations team formed an AI working group. One analyst in purchasing had been using AI tools to reconcile supplier discrepancies, with consistently good outcomes. The group invited the analyst to present, and the presentation was well received. Several colleagues tried the same approach and reported mixed results. Some outputs were wrong in ways that were not caught until later.
The difference was not the tool or the request wording. The analyst habitually pasted in the supplier's original terms before asking anything, checked every total independently, and discarded any answer that cited a clause the analyst could not find in the source. Those three habits were absent from the presentation because the analyst did not think of them as the method. They were simply how the work was done.
The group's next step was to schedule a second presentation. A different step would have been to write the three habits down as a short standing procedure, attach it to the task, and observe a colleague following it. One of those steps changes the next Tuesday. The other does not.
This case is a constructed illustration, not a measurement. The pattern it shows is that visible outputs travel and invisible checks do not.
What separates organizations that move past it
Organizations that pass this point tend to change how they treat the champion's knowledge. Four features recur in this pattern, offered as a working description rather than a measured result.
- The method is treated as an artifact. It is written down, versioned, and attached to the work it governs. It has an owner, and it can be changed.
- Transfer is tested. A method counts as captured only when a second person, without the champion in the room, produces comparable results. Failures are read as defects in the document rather than in the person.
- The checks are explicit. What the champion verifies instinctively becomes a named step. Verification is the part of expertise that disappears first when work is delegated.
- The group's role shifts. It moves from sharing impressions to maintaining the shared layer: deciding what is added, retired, and reviewed. This resembles what the earlier Cognitive Infrastructure Gap analysis describes as infrastructure, as distinct from tooling.
None of these requires new technology. They require deciding that the champion's value lies in what can be extracted from them, and assigning someone to do the extracting.
The diagnostic and its limits
The test proposed above is deliberately blunt: remove the strongest user and ask what remains. An organization whose results would revert has an individual capability and an organizational aspiration. Neither is a failure. It is simply an accurate description, and the accuracy matters because the remedy differs. More meetings address a coordination problem, while extraction addresses a transfer problem, and the stalled groups described here have the second kind.
There are limits to the claim. Some tacit judgment resists documentation, and a written procedure will capture only part of it. Some champions are reluctant to be extracted, reasonably, if their value to the organization has not been recognized in other ways. And a method that works for one task may not generalize. These caveats argue for modest scope: capture one method, for one task, and test it before widening.
The broader point stands. When an AI working group is not delivering, the most useful question is not how often it meets or who sits on it. It is whether anything it produces is owned by the organization, in a form that someone else can run without the person who invented it.
Research questions on this pattern are welcome through Inquiries.