Middle Managers Will Determine the Success or Failure of AI Adoption
This article is from *Harvard Business Review* and was written by Gleb Tsipursky, CEO of the AI consultancy Disaster Avoidance Experts. He previously worked as a consultant for Fortune 500 companies and spent over 15 years in academia as a behavioral scientist at UNC-Chapel Hill and Ohio State University. He is the author of several works; his most recent book, published by Georgetown University Press, is titled *The Psychology of AI Adoption at Work: From Resistance to Results*.
Summary. Generative AI initiatives often stall—not because of the board of directors, vendors, or training programs, but because middle managers determine how AI translates into daily work. Five distinct profiles influence how managers perceive AI and launch new initiatives with their teams: AI skeptics, wait-and-see traditionalists, cautious implementers, enthusiastic experimenters, and AI catalysts. Each reacts differently to risk, testing, incentives, and support. Effective leaders must tailor their approach to each profile, combining clear governance, appropriate incentives, and targeted support. The next phase of AI-driven competitive advantage will not belong to companies with the loudest mandates, but to those that understand the types of managers they are dealing with and provide each one with exactly what they need to move forward.
The most common point of failure for generative AI lies neither in the boardroom nor in vendor selection or training platforms. It resides at the management level responsible for transforming executive ambition into everyday behaviors. That is why generative AI adoption continues to stall, even in organizations where senior leadership has approved the budget, selected the platform, and announced the mandate. Middle managers decide whether AI becomes part of the workflow, a side experiment, a compliance headache, or a silent casualty of organizational resistance.
Middle managers do more than just implement executive decisions; they interpret what those decisions mean for employees. Frontline teams look to their managers to determine whether AI represents a useful tool, a passing fad, a surveillance mechanism, or a warning sign of future workforce reductions. An executive might promise that AI will reduce tedious tasks, but employees only believe that promise when their managers adjust workloads, review practices, performance expectations, and decision-making authority accordingly. This turns the adoption of generative AI into an operating model challenge rather than a simple software implementation. Middle managers stand at the intersection where strategic ambition collides with employee anxiety, ambiguous policies, existing workflows, and accountability for flawed outcomes.
Senior leaders often misdiagnose the problem.
They assume adoption hinges on three steps: articulating the strategy, selecting the technology, and training the workforce. While necessary, these steps are rarely sufficient. AI-driven transformation has outpaced most operating models.
McKinsey’s *State of AI* report for 2025 revealed that, while AI usage is widespread,
many organizations have not yet integrated it deeply enough into their workflows and processes to generate significant business value. Furthermore, the 2024 *Work Trend Index* from Microsoft and LinkedIn found that employees were already using AI at work, yet many leaders lacked a clear path to translate individual productivity into business value—highlighting the gap between enthusiasm and execution. The missing element is the translation work performed by managers: they are the ones who decide which workflows to modify, how to verify quality, what employees can test, and what happens when AI-assisted work does not go as planned.
Consider what happens when a leadership team announces that all departments must start using generative AI. One manager authorizes employees to experiment, establishes review criteria, and identifies a workflow ripe for redesign. Another manager remains largely silent, waits for clearer guidelines, and continues to evaluate staff based on traditional processes. Within months, the first team has developed a replicable practice, while the second has fractured into two groups: employees who avoid AI and those who use it discreetly, without any oversight.
This divergence leads to more than just uneven adoption.
It results in inconsistent quality, hidden data-related risks, redundant experiments, employee frustration, and technology investments that never yield a significant return. As generative AI shifts from optional experimentation to integration into core business processes, organizations have increasingly less room to leave these decisions to managerial improvisation.
Five managerial mindsets that shape adoption
My research points to a more precise diagnosis. While developing my book *The Psychology of AI Adoption at Work*—drawing on over 35 focus groups with more than 250 middle managers and 22 in-depth interviews across sectors such as insurance, manufacturing, legal services, retail, education, and technology—I discovered that middle managers do not constitute a homogeneous block that is simply “resistant to change.” They fall into five distinct psychographic profiles. Each profile reacts differently to pressure, risk, evidence, incentives, and governance. Treating them all the same may yield some impressive pilot projects, but it rarely achieves consistent adoption across departments.
The AI Skeptic
The first profile—the AI skeptic—is the easiest to caricature and the easiest to manage poorly. Skeptics are not merely stubborn; They often harbor legitimate concerns regarding accountability, job losses, data privacy, quality, and the human cost of automation. They have seen past “transformations” arrive with catchy slogans only to leave fractured processes in their wake. When an AI-generated compliance summary contains an error, they know they are the ones who will have to answer for it. When executives speak of efficiency, they perceive a risk to the workforce.
In their day-to-day work, skeptical managers may delay the approval of tools, limit AI use to optional, individual experimentation, or require employees to justify every single use case. For instance, during a leadership training session I once conducted, a skeptic repeatedly insisted that he did not trust AI tools to handle his team’s data—even though the company’s IT department had vetted and approved the software’s terms of service, which explicitly stated that subscriber data was not used to train the models. In educational institutions and non-profits, this hostility manifests as a reluctance to use AI due to its perceived excessive consumption of electricity and water.
Teams comprising AI skeptics tend to react in one of two ways: they either avoid AI entirely or use it discreetly, without sharing prompts, results, or errors. The first reaction stifles learning, while the second leads to “shadow AI” usage—operating outside established quality or security standards.
The wrong approach to dealing with skeptics is trying to win them over with enthusiastic rhetoric. These individuals require evidence accompanied by a clear definition of responsibilities. Leaders should propose low-risk, limited-scope pilot projects featuring defined metrics, human oversight, and explicit rules regarding accountability for AI-assisted errors. The AI pilots that truly matter are not flashy demonstrations but carefully selected workflow tests that compare cycle time, quality, rework, employee experience, and risk incidents against the current process. Skeptics need to see that the organization is not asking them to shoulder unmanaged risks simply to support someone else’s productivity narrative.
The Cautious Traditionalist
The second profile creates a quieter bottleneck. These managers do not block AI; rather, they delay it. They ask for more evidence, insights into lessons learned by other departments, additional policy memos, or another quarter of observation. Their resistance often appears reasonable, as it is framed in the language of prudence. However, this passivity becomes a strategic drag when teams wait for one another to make the first move. This is where AI change management must go beyond generic training and generate positive peer pressure through credible internal examples.
“Wait-and-see” traditionalists respond better to comparison than to coercion. Pairing them with respected managers who have successfully executed pilot projects in related areas proves effective. It is helpful to show them how a marketing team reduced the time required to produce initial drafts while maintaining human editorial control, or how a finance team used AI to summarize budget variance explanations without automating the final decision. It is best to assign them a limited adoption commitment tied to their own workflow, rather than imposing a sweeping corporate objective. Their fundamental question is not “Is AI good?” but rather “Why should I devote my limited management time to this right now?” Leaders must provide an answer that takes into account the specific constraints of their department.
The Cautious Implementer
The cautious implementer is often the organization’s most valuable manager. These managers are open to AI but rigorous about controls. They demand criteria for selecting use cases, review protocols, escalation paths, and measurable outcomes. They are the ones most likely to turn initial enthusiasm into scalable practices. However, organizations often frustrate them by offering vague permissions instead of a useful operational infrastructure. A cautious implementer does not need another inspirational briefing; what they require is a library of prompts (AI instructions), approved data-use standards, pilot project templates, quality control mechanisms, and a forum where they can point out failures without fear of reprisal.
When leaders leave these requirements undefined, cautious implementers limit the scope of pilot projects, postpone wider deployment, or subject every new use case to multiple rounds of review. Their teams may use AI responsibly, but only in isolated pockets that never evolve into standard practice. What strikes executives as excessive caution may actually reflect the absence of an operating system for responsible adoption.
This is where AI strategy becomes operational. BCG’s *AI at Work 2026* study argues that strategic clarity is more important than access to tools, given that jobs evolve faster than companies can redesign their operations. Cautious implementers act as a bridge between these two realities. They are able to take an executive objective—such as “using AI to improve customer responsiveness”—and translate it into a validated process: identifying suitable customer inquiries, permitted data sources, and AI-generated content, as well as determining what humans verify and which metrics are monitored post-launch.
Leaders must grant these cautious implementers the authority to design the review process for a specific, well-defined use case, while also providing timely access to experts in legal, information security, human resources, and compliance. Furthermore, they should require them to document workflows, quality controls, issues, and results in a format that other teams can reuse. In this way, their caution transforms into organizational infrastructure rather than becoming a permanent brake on adoption.
The Enthusiastic Experimenter
The fourth type of manager brings dynamism.
Enthusiastic experimenters test new tools, share prompts (AI instructions), organize informal demonstrations, and make AI feel practical rather than abstract. They are able to turn apprehension into curiosity by showing colleagues real results instead of mere theoretical presentations. In many organizations, they are the first to demonstrate that AI adoption can move beyond individual productivity to become a team practice. Wharton’s *AI Adoption Report 2025* highlights the importance of people- and process-related levers in turning widespread technology use into sustainable return on investment (ROI); often, it is through the activities of these enthusiastic experimenters that such levers first emerge. Their risk lies in prioritizing speed over structure. They may adopt tools before legal, compliance, information security, or HR departments have defined the boundaries. They can inadvertently normalize unsafe practices, as early successes are highly compelling. Leaders should not stifle them with bureaucracy but rather channel their energy by establishing clear guidelines. It is advisable to provide them with approved tools, shared prompt libraries, review standards, and a visible channel for raising questions about established policies. Recognize them for achieving secure results, not just for moving quickly.
The AI Catalyst
The final profile, the AI catalyst, represents both an asset and a test for governance. Catalysts view generative AI as an engine for comprehensive transformation. They seek to redesign cross-team processes, influence colleagues, build alliances, and urge executives to pick up the pace. In the right environment, they become intrapreneurs who help the organization escape “pilot project purgatory.” In the wrong environment, they move faster than trust allows, underestimate regulatory compliance, and trigger pushback from teams that feel like subjects of an experiment.
For catalysts, AI governance must act as an enabler, not merely a formal hurdle. The generative AI profile within the NIST AI Risk Management Framework identifies risks such as data privacy, information security, collusion, harmful bias, and intellectual property issues. These risks are not reasons to slow catalysts down, but rather design requirements for responsible acceleration. Ask them to conduct “pre-mortem” analyses before launching high-impact pilot projects, define scaling thresholds, document human oversight points, and communicate lessons learned in a format that other teams can reuse.
Tailor the intervention to the manager’s profile
Identifying these five profiles is just the first step. Most organizations harbor all five, often within the same function or leadership team. Therefore, senior executives need a systematic method to diagnose what is holding each manager back and to align support, incentives, and accountability accordingly.
A practical diagnosis begins with three questions:
What does this manager fear losing?
What evidence would change their mind?
What support would enable them to act responsibly?
Skeptics fear accountability and job displacement. “Wait-and-see” traditionalists fear wasted effort and political exposure. Cautious implementers fear unmanaged risks. Enthusiastic experimenters fear missed opportunities, and catalysts fear organizational timidity. These differences matter because each profile requires a distinct leadership response.
Leaders must also redesign incentives. Too many organizations publicly reward enthusiasm for AI while privately punishing AI-related errors. This drives managers into a defensive posture. If a manager is urged to experiment but blamed for any imperfection, the rational choice is to wait. If speed is celebrated without regard for accuracy, privacy, or trust, the rational choice is to overreach. A better system rewards disciplined learning: clear hypotheses, responsible pilot tests, measurable workflow impact, transparent error reporting, and evidence-based scaling.
For reluctant managers, the intervention involves providing evidence and psychological safety. This entails having transparent conversations about the impact on jobs—especially given that the World Economic Forum’s *Future of Jobs* report identifies AI and information processing as key forces that will transform work and skills between now and 2030. Leaders should not pretend that there will be no disruption. They must explain what will change and what will not, which skills will be relevant, and how managers who learn to leverage AI will be better positioned than those who avoid it.
For managers who support the initiative, the intervention involves providing the necessary means and establishing boundaries. Offer cautious implementers the operational discipline they need, and provide enthusiastic experimenters with a safe testing environment and a mechanism to share useful practices. Give catalysts a mandate tied to risk controls. The classic “pre-mortem” exercise is particularly useful here, as it legitimizes dissent before a pilot is launched. Asking, “Imagine this AI pilot fails spectacularly: what happened?” helps teams identify risks related to legal issues, quality, trust, workflows, and adoption while there is still time to redesign the initiative. The true test of enterprise AI is manager trust. Executives should stop asking, “Why aren’t employees adopting AI?” and instead ask, “Which middle-management profiles are driving adoption in each functional area, and have we tailored our intervention to those profiles?” This approach transforms a vague cultural issue into a practical management agenda. It helps leaders avoid wasting months on repetitive pilots, reduce anxiety before it turns into resistance, and channel the energy of early adopters without letting them get ahead of governance frameworks.
The bottleneck is real, but it is solvable. Middle managers can block, dilute, or distort generative AI—or they can turn it into quantifiable business value. Senior leaders seeking practical guidance on responsible AI adoption should start by viewing managers a
