A Question I Have Stopped Asking
When I started running coaching engagements with AI-augmented teams two years ago, I would open the first session by asking what people were worried about. The answers were uniformly vague. "Disruption." "The future of work." "Whether we are using it right." The conversations did not get far because the question did not get far.
I have stopped asking it. What I ask now, depending on the role of the person sitting in front of me, is one of three more specific questions. The answers are no longer vague. They are remarkably consistent across industries and remarkably different by role — to the point where I have come to think of organizational AI anxiety not as a single condition but as a three-layer stack, in which each layer has its own diagnostic profile, its own behavioural signature, and its own coaching response.
This piece is the synthesis of where the layers come from, how they interact, and what — concretely — to do about each. It draws on the empirical literature of the last three years and on the patterns I have seen across our coaching practice. It is the longest piece in our AI-and-EQ series, and it is the one I would start with if you are reading only one.
Layer 1 — The Engineer's Fear: Relevance Loss
If you sit with an engineer who has been working in an AI-augmented environment for twelve to eighteen months and ask them, privately, what they are most afraid of, what comes back is almost never "losing my job." That fear has receded as the technology has matured and as it has become clear that most engineering roles are not being eliminated. What has replaced it is more diffuse and more corrosive: I am afraid that what I am still doing is no longer the part that matters.
I call this relevance loss. It is a specific fear — not of unemployment, but of value-irrelevance. The engineer can still write the code, draw the diagram, draft the FMEA. They are also aware, with increasing clarity, that the AI can do the bulk of those things, and that what is being asked of them now is the harder, fuzzier, less-rewardable work of judgment over the AI's output. They do not yet trust their judgment to be worth what their hands used to be worth.
The empirical evidence that this fear is well-founded — at least in some sub-populations — is now substantial. Anthropic's randomized controlled trial of 52 professional developers found AI-assisted workers scored 17% lower on comprehension assessments of code they had just written, with the largest gap in debugging (Cohen's d = 0.738). A 2024 study of endoscopists who used AI polyp-detection tools found their independent adenoma detection rates dropped from 28% with the AI to 22% without it after routine use — below the pre-AI baseline. A PNAS study of approximately 1,000 students found those given ChatGPT-4 access during math practice improved 48% with the tool, but performed 17% worse than peers who never had access once the tool was removed.
The pattern is recursive: the more the engineer leans on the AI, the more their independent capacity atrophies, the more they rely on the AI, the more relevant the fear becomes. Engineers feel this, even if they cannot name it. The fear is not paranoid. It is calibrated.
The behavioural trajectories I described in our piece on the engineer's EQ gap — Amplified Craftsman, Cognitive Outsourcer, Audit Adversary, Quiet Resister, Orchestrator — are best understood as five different responses to this single underlying fear. The Outsourcer leans into the fear by surrendering. The Resister refuses the premise. The Adversary optimizes for defensibility. The Craftsman and the Orchestrator are the two trajectories that metabolize the fear into something productive, and the difference between them is whether the engineer reframes their identity around judgment (Craftsman) or around orchestration of judgment (Orchestrator).
The coaching response at Layer 1 is identity work. The engineer's professional identity has to migrate — from the person who produces the output to the person who decides what gets produced and validates that it is correct — without experiencing the migration as a demotion. This is not a slogan; it is the work. It is uneven across individuals, it takes six to eighteen months, and it depends substantially on whether the team's reward system makes the new identity visible and rewarded. None of this is technical. All of it is coachable.
Layer 2 — The Manager's Fear: Attribution Loss
The manager's fear is structurally different. They are not personally being substituted by the AI, at least not in the short term. What they are losing, slowly, is the ability to see what is happening on their team. And because management is fundamentally a job of seeing, that loss feels — correctly — existential to the role.
I call this attribution loss: the gradual decay of the manager's ability to know who actually did what, how, and how well. It is the fear behind the pattern I described in our piece on managers in the middle, where teams are using AI more than the dashboards say and the manager suspects it but cannot name it.
The empirical foundation here is the Wharton/UC San Diego experiment series with 3,346 participants, which found that managers reduced bonuses by approximately 50% for workers who had used AI to produce identical, quality-verified output. The mechanism is attribution drift: when an AI is in the work, the manager credits the AI and discounts the human. Workers respond to this rationally — by concealing AI use, by selectively disclosing only the boring uses, by optimizing the recorded trail for defensibility rather than accuracy. The manager's epistemic position quietly degrades. The team's relationship with the manager quietly degrades alongside it.
The OECD's 2025 study of approximately 6,800 workers across seven countries found that algorithmic management was associated with reduced autonomy, lower trust, higher stress, and reduced satisfaction — except in workplaces where workers had substantial voice in how the systems were configured. In those workplaces, the negative effects did not merely soften; they disappeared. The variable that decided whether the manager-worker relationship was harmed or strengthened was worker voice in configuration. Not training. Not policy. Voice.
The manager's fear, in other words, is not paranoid either. It is the felt sense of a real epistemic deficit, in a role whose value is constituted by epistemic access.
The behavioural responses I see in managers under this fear are predictable and dysfunctional. Some become surveillance-oriented — trying to close the visibility gap by tightening the recording infrastructure, which makes concealment more rational and the deficit worse. Some become permissively disengaged — abandoning the attempt to evaluate the team accurately and falling back on output metrics, which protects them from the discomfort but hollows out the developmental relationship. A small minority do the harder thing: they restructure the relationship with their team around honest disclosure, accept the short-term cost of admitting they cannot see everything, and build the trust under which the team voluntarily makes itself legible.
The coaching response at Layer 2 is relational, not informational. No dashboard, no policy, and no monitoring tool will recover the attribution that has been lost. What recovers it is a different manager — one with the capacities I named in the managers-in-the-middle piece: attribution discipline, disclosure-safe inquiry, calibrated trust, reward-system advocacy, relational repair. These are also coachable. They are also unevenly distributed across the manager population of any organization, and the unevenness now decides more than it used to.
Layer 3 — The Leader's Fear: Verification Loss
The executive layer's fear is rarely articulated cleanly, because executives are not in the habit of articulating fear cleanly. But the question that surfaces in nearly every senior leadership conversation I have, once the small talk is over, is some variant of: How would I know if it had already gone wrong?
This is verification loss. It is the fear of the leader who understands, perhaps better than the layers below them, that the organization is producing more output than ever, that the metrics look good, and that — beneath the metrics — the organization may have already lost the capability to detect when its AI-augmented work is subtly broken.
The evidence that this fear is structurally justified is the most disturbing in the literature. MIT's State of AI in Business 2025 reported that approximately 95% of enterprise AI pilots fail to scale, and the dominant mode of failure is what the research calls organizational amnesia — work produced inside stateless AI sessions whose reasoning was never captured, by an organization that increasingly cannot reconstruct how it arrived at the work it shipped. Fortune 500 firms are estimated to lose roughly $31.5 billion annually to this dynamic.
Beneath the amnesia is the Oversight Paradox, which I covered in our piece on the engineer's EQ gap but which is most properly understood as a leader's problem. The Decision Spine of any AI-augmented organization depends on the senior expert at the apex retaining verification capacity — the integrated, partly tacit ability to recognize when an AI's plausible output is actually wrong. The empirical literature is unanimous that this capacity erodes with routine reliance. Within eighteen months, the most senior people in the organization may have lost the very capability whose preservation is the load-bearing assumption of the entire architecture. The failure is invisible while it develops and catastrophic when it surfaces.
A Singapore Management University analysis of 302,600 AI-authored commits across 6,299 GitHub repositories identified 484,366 distinct issues introduced by AI assistants, with 22.7% surviving in the latest repository version — meaning roughly one in five AI-introduced problems was never caught downstream. At the scale of an enterprise codebase, this is a quietly accumulating verification debt. At the scale of a regulated industry, it is a recall or an audit failure waiting to be timed by chance.
The leader who has internalized any of this is operating with a specific anxiety: the anxiety of someone responsible for a system whose visible state and actual state are diverging on a timeline they cannot fully see. Their fear is also calibrated.
The coaching response at Layer 3 is governance work. The leader's job in the AI-augmented organization is not to verify the work themselves — they were never going to scale to that — but to ensure the verification function is being preserved across the layers beneath them. This means investing in the diagnostic capacity to detect erosion before it manifests as incidents. It means building reward structures that price the capture of reasoning as highly as the production of output. It means deliberately preserving the formation pathways through which junior workers develop the judgment that the AI will later need a human to exercise. None of this looks like the strategic work senior leaders are trained for. All of it is the new work.
How the Layers Interact
The three layers do not sit independently. They feed each other in a closed loop that makes the whole stack worse if any layer is left unaddressed.
The leader's verification anxiety becomes pressure on the manager — show me you are tracking AI use, show me the productivity, show me the controls. The manager's attribution anxiety becomes pressure on the engineer — I need to see what you are doing, I need to know how you are using these tools. The engineer's relevance anxiety becomes concealment — I will use the AI quietly and present the work as my own, because the alternative is being credited less for the same output. The concealment makes the manager's attribution worse. The worse attribution makes the leader's verification worse. The leader applies more pressure. The loop tightens.
This is not a hypothetical. In the organizations where I am working most intensively right now, the loop is the dominant dynamic. It is also the reason that single-layer interventions almost always fail. Coaching the engineers on identity work, while leaving the managers' attribution problem and the leader's verification problem untouched, simply restocks the bottom of the loop with workers who handle their identity better but face the same upstream pressure. Restructuring the manager's incentives, while leaving the leader's anxiety unaddressed, produces managers who get squeezed twice as hard from above. The leader who tries to govern verification without rebuilding the manager and engineer layers underneath them is, in effect, trying to verify a system whose participants have a structural incentive to mislead them.
The intervention has to operate across all three layers, or it does not operate at all. This is the single most important strategic implication of the AI Anxiety Stack: the fear is recursive, and the response has to be coordinated.
What a Coordinated Response Looks Like
In practical terms, a coordinated response is built on three commitments that the organization has to make explicitly and in sequence.
First, diagnose all three layers before intervening at any of them. The mistake organizations make most consistently is to assume they know which layer is in worst shape. They do not. The correlation between executive estimates of "where the problem is" and what our diagnostics actually surface is, in our client data, roughly zero. We have seen organizations whose engineers were thriving and whose managers had collapsed, and vice versa, and the executive impression was uniformly wrong in both directions. The diagnostic is not a step you skip because it slows you down. It is the only thing that prevents you from solving the wrong problem.
Second, sequence the interventions to break the loop, not to fix any single layer. In practice this almost always means starting at Layer 2 (the manager), because the manager is the joint where engineer concealment and leader pressure meet, and is the layer where intervention has the largest second-order effect on the other two. Coaching the manager to develop attribution discipline reduces the engineer's incentive to conceal, which restores the leader's verification signal, which reduces the pressure the manager is under. The loop reverses.
Third, build the diagnostic capability into the organization, not the consultant. This is the part of the work that vendors and external coaches have the strongest incentive not to recommend. But the AI Anxiety Stack is not a one-time intervention; it is a continuous dynamic, and the organization's ability to monitor and re-diagnose itself is the difference between a coaching engagement that ends and a capability that compounds.
Where to Start
The diagnostic tools we use are layer-specific because the fear is layer-specific.
- For engineers and individual contributors: EQ Signal identifies which behavioural trajectory the worker is on and surfaces the specific identity-anchor work that the next twelve months require.
- For team and middle-management dynamics: EQ Pulse measures the gap between surface engagement signals and the underlying attribution and disclosure patterns that the Wharton finding predicts will be most distorted.
- For leadership and organizational AI-readiness: EQ Frontier measures the emotional and cognitive dimensions of readiness at the leader and organizational level, including early indicators of verification erosion that conventional readiness audits miss.
- For coaching once the diagnostic is in: EQ Catalyst is our neuroscience-grounded coaching method for the identity, attribution, and governance work that the diagnostic surfaces. The diagnostic without the coaching produces awareness without change; the coaching without the diagnostic produces change in the wrong direction.
If you are reading this and you do not yet know which layer of your organization is in the worst shape, that is itself diagnostic — and is the place to start.
The conversations I have with leaders, managers, and engineers have stopped being vague. The fear is specific, the literature is robust, and the work — at all three layers — is now well-defined. The organizations that will be in the best shape three years from now are not the ones with the most AI in the most places. They are the ones that have done the unglamorous, layer-specific human work to keep the people inside the AI capable of being trusted, seen, and verified. The AI is the loud part. The Anxiety Stack is what the next decade will actually be decided on.
Each layer has a diagnostic. Start with the one that fits your role: EQ Signal for engineers, EQ Pulse for managers, EQ Frontier for leaders. Coaching follows via EQ Catalyst.