The Spending That Brought Silence
Over the past two years, organisations have spent unprecedented sums to drive AI adoption. The assumption is familiar: provide the tools, resource the rollout, measure the uptake. Returns will follow.
For most, they have not. Executives report difficulty connecting AI investment to firm-wide productivity. Usage metrics climb whilst the behaviours that convert adoption into value—flagging flawed outputs, questioning whether a task should be automated at all, proposing better approaches—remain conspicuously absent.
The pattern is not a technology problem. It is a leadership one.
AI does not reduce the need for human judgement. It concentrates that need into a narrower set of moments—moments when someone must choose to say “I do not trust this output.” To admit “I do not understand this tool well enough.” To propose “there is a better way to use this.” Each is an act of voice or help-seeking under uncertainty. Each is interpersonally risky. And each becomes more consequential, not less, as automation spreads.
Whether those moments produce voice or silence depends on the conditions leaders create. Psychological safety—the shared belief that interpersonal risk-taking will not result in punishment or embarrassment—shapes whether people engage with uncertain technology at all. But safety is not simply declared. It is enacted through observable leadership practices: structured access when things go wrong, non-punitive reception when doubt is voiced, visible follow-through when concerns are raised.
That is the reversal this article examines. Automation is sold as a way to depend on people less. In practice, it raises the stakes on what leaders must relationally provide. Three threads structure the argument: what automation stops surfacing, why adoption is not the same as engagement, and what secure-base leadership must deliver when teams navigate transformative change.
What Automation Stops Surfacing
Before automation, problems surfaced through friction. Manual checks caught errors. Handoffs between colleagues created second pairs of eyes. The ordinary pauses of human workflow—reviewing, querying, double-checking—generated countless small moments where someone noticed something was off and said so.
None of these moments required courage. They were built into the work itself.
Automation removes that friction deliberately. That is its purpose. But the friction was never only inefficiency. It was also the organisation’s distributed detection system—the means by which weak signals surfaced before they escalated into failures.
When processes run manually, error detection is ambient. When they run autonomously, it becomes discretionary.
Once the friction is gone, what remains is choice. The only remaining detector of a flawed automated decision is a person willing to flag it. The only remaining check on whether a task should be automated is a person willing to question it. The only remaining source of a better approach is a person willing to propose it.
Each of these acts carries interpersonal cost. Research on team learning has long established the mechanism: when people anticipate negative consequences for raising concerns, they raise fewer of them. Voicing doubt, disclosing errors, seeking help—these are precisely the behaviours that psychological safety enables. They are also precisely the behaviours that automation makes newly critical.
This is not solely about catching mistakes. The generative side matters equally. Teams adopting AI must continuously judge where the tools help and where they harm, which outputs to trust and which to interrogate, which workflows to redesign and which to protect. These judgements cannot be automated. They depend on people contributing what they observe, doubt, and imagine—and they require a climate in which doing so does not invite penalty.
Under automation, the cost of silence compounds. An unflagged flaw in a manual process affects one task. An unflagged flaw in an automated process replicates at machine speed.
The question is no longer whether teams surface problems—it is whether the conditions exist for them to do so before the problems scale.
Adoption Is Not Engagement
The dominant response to AI investment has been to track consumption. Some organisations went further—publishing internal leaderboards of token usage, treating volume as a proxy for productivity. The practice acquired a name: tokenmaxxing. It did not work. Employees at major technology firms reportedly ran meaningless tasks through AI agents to inflate their statistics. Code churn rose sharply—work generated, then discarded. At least one global firm exhausted its annual AI budget within four months. The leaderboards have since come down.
The failure is instructive, but not for the reason most commentary suggests. The problem was not that organisations measured the wrong metric. It was that they confused adoption with engagement.
Recent evidence makes the distinction precise. A study of more than 2,200 employees across a global consulting firm examined whether psychological safety was associated with AI engagement. The findings were notable in their asymmetry. Psychological safety was reliably associated with whether employees adopted AI tools at all—each unit increase in perceived safety corresponded to a meaningful increase in the odds of adoption, and the pattern held across seniority, experience, and geography. Yet once adoption had occurred, psychological safety bore no association with how often or how long employees used the tools.
Crossing the adoption threshold is a single act of interpersonal risk: trying something unfamiliar in front of colleagues. A climate of safety lowers the anticipated cost of that act, and adoption follows readily enough. But productive engagement is not a single threshold. It is a sustained stream of judgement, challenge, and disclosure—behaviours that must happen continuously, under pressure, and in front of colleagues. These are the behaviours the tokenmaxxing era failed to produce. They are also the behaviours no usage dashboard can detect.
A climate of safety gets people through the door. What sustains contribution on the other side is less clear—and the question matters, because the door is not where the value lives.
What sustains exploratory behaviour under uncertainty is not a climate alone. It is a relational anchor.
The Anchor That Does Not Depend on the Answer
Attachment science offers a useful frame. Decades of research demonstrate that people explore most willingly—take risks, venture into the unknown, persist through setbacks—when they have access to a secure base. Not a safety net. A relationship. One that provides availability when needed, reassurance without judgement, and support without interference.
The logic extends directly to leadership. Leaders function as secure bases when their behaviour signals three things: that they can be reached, that approaching them will not be punished, and that what is raised will be acted upon. The secure-base leadership literature demonstrates the mechanism: when these conditions are present, team members report greater self-efficacy, stronger autonomous motivation, and higher levels of initiative. They disclose more, persist longer, and take more ownership of change—not because they are told to, but because the relational conditions make it possible.
The AI transition makes this provision newly consequential.
In previous waves of change, leaders could reassure partly through expertise. They understood the new terrain better than their teams. They could answer the question, resolve the ambiguity, assess the risk. In this transition, frequently they cannot. The tools are evolving faster than any individual can master. Fluency is often distributed unevenly across hierarchies—and sometimes inverted, with junior team members understanding the technology better than those leading them.
This redistribution of expertise removes the basis on which many leaders have historically provided security. It does not remove the need for security itself.
A secure base does not depend on having the answer. It depends on being reachable when the answer is unclear, receptive when doubt is voiced, and responsive when concerns are raised. These functions survive the redistribution of expertise. They may represent the leader’s most durable contribution when nobody fully understands the tools.
Yet here the field has stalled. Existing frameworks describe secure-base leadership functionally—providing a secure base, enabling exploration—without specifying what these abstractions look like under sustained operational pressure. “Be a secure base” is an aspiration, not a practice. The question is what it actually looks like—observably, behaviourally, under real constraint.
What a Secure Base Looks Like: From Aspiration to Practice
Recent research within a high-tempo technology function examined exactly this: how leaders enact secure-base practices when operational demands compress relational bandwidth. Three distinct practices emerged, forming the Available–Receptive– Responsive framework (ARR Framework™).
Available practices signal that the leader can be reached when it matters. This is not constant presence but structured access— predictable micro-windows that lower the cost of approaching. For AI-adopting teams, availability shapes what happens in the moment a tool fails, confuses, or invites a better idea. A team member weighing whether to raise “I think the model got this wrong” is making a rapid calculation about access. Explicit response commitments, clear escalation routes, brief but reliable windows—these tip that calculation toward voice.
Receptive practices signal that vulnerability will not trigger punishment or embarrassment. In the original research, this proved the foundational practice under sustained pressure—accounting for nearly half of all coded episodes and appearing most frequently around error disclosure and voice. For AI adoption, receptiveness governs whether doubt surfaces early or goes underground—into covert use, quiet workarounds, and the performative compliance that tokenmaxxing made visible at scale.
Responsive practices demonstrate that raising something leads somewhere. Visible follow-through converts concerns into action and signals that voice is not futile. For AI-adopting teams, responsiveness governs whether the first flagged failure becomes the last. A team member who surfaces a flawed output and watches nothing happen learns a precise lesson about the cost of speaking up again.
The research suggests a further pattern. Under sustained pressure, these practices appear to operate hierarchically rather than co-equally.
Receptiveness comes first. Without evaluative safety, access goes unused—team members may know a leader is available yet remain unwilling to approach if they anticipate judgement. This is the mechanism that explains why organisations increase leader visibility during transformation—more town halls, more open doors, more check-ins—and still encounter silence. Access without reception is structurally inert.
Availability comes second, offering structured contact once evaluative safety is established. Responsiveness faces a different constraint entirely. Follow-through unfolds over time and competes with immediate operational load. Under sustained pressure, demands accumulate faster than closure cycles complete—generating what the research terms “completion debt.” Sustaining visible follow-through may require system-level supports beyond individual leader capacity.
This hierarchy matters practically. Leaders navigating AI transformation cannot simply do more of everything. Under constraint, sequencing matters: establish non-punitive reception first, then signal bounded access, then build the scaffolds that make follow-through visible.
The Diagnostic Reframe for Leaders
The patterns described above carry practical implications—not as prescriptive checklists, but as orientation shifts for leaders steering teams through AI adoption.
Start with reception, not access. The instinct during transformation is to increase visibility. The hierarchy suggests a different starting point. The prior question is whether admitting confusion about AI, or challenging an automated output, has ever been met with embarrassment or sanction in the team’s recent memory. If it has, no amount of availability will compensate. Evaluative safety comes first. Everything else depends on it.
Reinterpret quiet. During AI rollouts, leaders typically read silence as acceptance—the absence of complaint as the presence of adoption. The evidence suggests a different reading. Quiet teams may be using the tools performatively, working around them covertly, or withholding the judgement calls that would make adoption productive. Healthy AI engagement is audible. It sounds like questions, challenges, flagged failures, and proposed alternatives. A silent rollout is not a success metric. It is a warning signal worth interrogating.
Build pluralistic pathways. Team members differ in how they prefer to raise things—and research on attachment orientations suggests these differences are systematic rather than random. Some will use an open forum. Others need asynchronous channels, private conversations, or structured prompts. The principle is optionality: multiple routes to voice that allow people to self-select the pathway that fits how they process relational risk. Leaders cannot diagnose individual orientations, nor should they attempt to. But they can build sufficient variety that team members find routes that work for them.
Scaffold the follow-through. Responsiveness fails structurally, not wilfully. Visible action trackers, distributed ownership of raised concerns, and transparent closure on AI-related issues sustain the responsive function when individual leader bandwidth cannot. Without these scaffolds, the pattern is predictable: a team member raises something, nothing visibly changes, and the lesson is learned.
A Measured Claim
These mechanisms warrant explicit qualification. The ARR framework emerged from a single technology function operating under chronic time scarcity, tight interdependence, and low tolerance for error. Its sample leaned secure on attachment dimensions, with high baseline psychological safety. The adoption study is cross-sectional and self-reported, conducted within a single consulting organisation. The tokenmaxxing accounts are journalistic; they illustrate the pattern rather than establish it. The findings across all three strands constitute associations and preliminary design knowledge, not confirmed causal claims.
What can be said is this. Two independent strands of evidence, from different contexts and different methodologies, point in the same direction. Safety shapes whether people engage with uncertain technology at all. The relational practices that sustain voice under pressure are specifiable, observable, and unevenly enacted. And the gap between adoption and productive engagement—the gap the tokenmaxxing era exposed—is not a technology problem. It is a leadership one.
The patterns are offered as lenses for interpretation, not protocols for implementation.
Conclusion
The organisations that capture value from AI will not be those that spend the most on tokens or track the most usage. The tokenmaxxing era demonstrated that conclusively: consumption can be incentivised, measured, and maximised whilst the behaviours that actually produce value go missing.
Those behaviours are relational. They require the judgement no tool can replicate—and they occur when leaders provide the conditions that make voice possible. Not expertise. Not enthusiasm for the technology. A secure base.
This reframes the leadership task of the AI transition. The question is not how to drive adoption. Adoption, the evidence suggests, is associated with safety readily enough. The question is what sustains the continuous, audible, judgement-rich engagement that converts adoption into value. The answer lies not in mastering the technology but in what leaders observably provide when others are trying to.
The question for leaders shifts accordingly. Not “how much are my people using AI?” but “what are they telling me about it—and what does the silence contain?” That question has no dashboard. But asking it may be where the returns begin.