The public conversation about AI discusses, almost exclusively, its external governance: regulation, model biases, systemic risks, alignment. My work operates on a different level. I coach executives who are deploying AI in their organizations right now, and what I see is that the critical decision is not how AI is regulated from the outside. It is in who deploys it from the inside. Because their premises speak to who they are, and that determines what systems they build.
What I see in my practice takes three distinct forms.
I see a highly controlling leader deploy monitoring agents that perform intrasystem logging and continuously analyze output. The leader's signature becomes inscribed in an infrastructure that their team can no longer negotiate.
I see an ambitious leader discover what AI makes it possible to produce and raise the bar for everyone, without any intervening conversation about the human cost this has downstream.
I see a critical leader ask the model for damning analyses of their teams' work and deliver them directly, without mediation, treating others as mediocre through an output that seems objective because it comes from the machine.
And I say this from personal experience as well. I have designed workshops with a clear idea of my own, iterated on them with a model to improve them, and ended up delivering something that sounded good but no longer sounded like me. I have seen it in myself, not just in my clients.
Something has been removed from the leader's environment, and it is worth naming. For decades, the leader's reactive stage had a natural ceiling. The controller wore themselves out controlling. The ambitious one in overdrive would burn out. The critic paid a relational price. That personal cost was finite and, in some cases, was the engine that drove development: the pain of one's own pattern made it unsustainable to maintain. At the same time, the leader existed within external frictions that regulated their stage without their knowledge. Teams resisted, complained, gave corrective feedback—even if poorly and at the wrong time. Peers provided a counterpoint. Seeking validation for an intuition required exposing oneself, waiting, managing disagreement.
AI removes these three frictions at once. It reduces the personal cost of the reactive pattern. It eliminates the relational friction of seeking counterpoints. It replaces the human collaborator who would correct them with agents that do not. What the system used to do—contain, regulate, provide a signal back—the leader must now do from within, or it does not get done at all. This is why internal governance becomes the critical decision. Not because it is morally superior, but because the external governance that unwittingly substituted for it has disappeared.
The first thesis has to do with amplification. AI does not add new dimensions to the reactive stage. It removes the ceiling that contained it. Control can be total and permanent via agents. Overdrive can be scaled to the entire organization via automation. The critic can operate on every output in real time. And all this at a much lower personal cost for the one executing it: they do not get worn out, they do not expose themselves, they do not negotiate. The signal of pain that historically pushed for vertical development is anesthetized.
What the reactive leader designs today crystallizes into infrastructure. Metrics, dashboards, committees, vendors, budgets. When two years from now they or their successor sees the damage, dismantling it will not be a matter of changing one's mind: it will be a matter of dismantling a system. The reactive signature becomes path-dependent.
And there is one more layer, which I offer as a reasoned risk rather than an assertion: the reactive human feedback that goes into fine-tuning models in enterprise deployment becomes embedded in shared layers. I have no direct technical evidence, but in the absence of any specific guardrails we know of, the logical extension is that the reactive signature does not remain within the perimeter of the person who generates it.
The second thesis has to do with validation. The leader who had overcome their dependence on external validation had a system built for that kind of "other": bosses, mentors, group norms, the gaze of their peers. In the face of the model, this architecture is not activated. The model is not an identity-forming other in the Keganite sense. There is no belonging to negotiate, no status to defend. And precisely because it does not feel like an identity-based dependency, the filters installed by classic vertical development are not triggered.
The mechanism, in my own experience, is this: I have an intuition, I articulate it so the model can help me improve it, and I end up absorbing a formulation that sounds good, sounds rigorous, sounds articulate, but no longer sounds like me. Differentiating an enhancer from a validator is not so easy. It is a more complex game than it seems, and the specific trap is that this regression is camouflaged as mature analysis.
This is especially important in professions where one's own voice is the asset. If what I deliver sounds good but is not me, I have ceded authorship precisely where authorship was the value. I formulate it thus, and I leave open what I do not yet know: Kegan's model, built on human identity-based dependencies, perhaps does not anticipate this pathway. I need to continue working on it.
The third thesis has to do with agents. An agent is a prompt, and a prompt is the written signature of its designer's stage. A transactional leader who designs an agent to analyze meetings, if their prompt does not account for implicit ideas or latent needs, will anchor everything the agent processes in the transactional. What escaped them as a human now escapes them systematically and at scale. Their personal blind spot becomes an organizational blind spot, replicated in every meeting, without appearing as such on any dashboard.
Unlike a human collaborator, the agent does not point out what is not in its prompt. It does not complain, it does not rebel, it does not correct. The agent's ceiling is the ceiling of the stage of the person who writes it.
This works just as well in reverse. The creative leader can design prompts that integrate layers they themselves did not see, contrast them with the more limited prompts from their team, and use the exercise as a diagnosis of a shared stage. AI amplifies in both directions. The difference is not in the tool.
And here lies the frontier. The agents of three years from now will not be the same as today's. The design practices we are cementing today will guide the next generation. The urgency is not just about the immediate impact: it is because the ecosystem is being calibrated right now.
And here the connecting framework comes into play: mirage or lever. The same phenomenon produces opposite effects depending on the stage of the person experiencing it. For the leader without internal governance, AI appears as an enhancer and operates as a mirage: it amplifies, anesthetizes, crystallizes, erodes, and replicates blind spots. For the leader with their own judgment, the same AI is a real lever: it extends judgment, returns information about one's own stage, and scales one's reading of situations. The difference between mirage and lever is not in the tool. It is in who is looking. And, crucially, the leader cannot know from the inside which side they are on without deliberate work on their vertical development, because the signals that would have previously warned them have been removed.
There are three positions I do not subscribe to. I reject the external governance of AI as the sole answer. It is necessary, but it is not sufficient, because it does not touch upon what is happening inside each organization right now. I reject leadership coaching that operates on skills, styles, or behaviors without a structural diagnosis of the leader's stage. In the current context, intervening without reading the stage is just noise. I reject the accelerated adoption of AI in organizations without parallel and explicit work on the vertical development of those deploying it. The speed of adoption without development is the condition that makes the problem critical.
What I do subscribe to, and what I commit to doing. I affirm that the leader's internal governance is, now, the critical decision. That vertical development, which was already a competitive advantage, is also a condition for containment. And that in a context where leading includes designing agents, independent judgment is becoming an entry requirement, not a desirable aspiration. As a practitioner, I commit to three things. To not intervene with clients who are deploying AI without a prior structural diagnosis of the stage from which they are deploying it. To work explicitly on the leader's signature in the systems and prompts they design, not only in their conversations. And to hold the uncomfortable question with my clients about what is being sedimented in their organization as adoption accelerates.
There are questions that remain open for me. Is the early adoption of AI in organizations biased towards reactive leadership profiles, because AI offers validation, visibility, and measurable control that resonate with that configuration? I have no data. The question seems relevant to me. Does Kegan's model have calibrated filters for this form of ceding authorship, or are we facing a pathway that classic vertical development does not anticipate because the model is not an identity-forming other? I need to delve deeper. What happens, technically, when the human feedback from reactive leaders is fed into models in enterprise deployment? I have no technical answer. I have reason to ask. And a question of method, for myself and for anyone reading this: how can one distinguish, from within the interaction with a model itself, when it is being a lever and when it is being a mirage? If I knew a definitive answer, I would not be framing it as a question. But I suspect the answer is not a technique. It is a stage.