The company's AI blueprint is very impressive, but its success or failure hinges on these five types of middle managers.
Middle managers can block the implementation of generative AI, weaken its effectiveness, and distort its direction; they can also translate generative AI into measurable business value. Senior leaders who want to obtain practical guidelines for responsible AI implementation should regard middle managers as the operating system of transformation, rather than as inconspicuous supporting pipelines under the strategy. The competitive advantage in AI in the next stage will not belong to the enterprises that shout the loudest slogans, but to those that understand their own managers and provide everyone with the resources needed to promote change.
The most common failure point of generative AI implementation is not in the boardroom, the supplier shortlist, or the online training platform. The problem lies at the level of middle managers — they need to translate the strategic vision of executives into the daily work behaviors of employees. That is why even when senior leaders have approved budgets, selected platforms, and issued implementation instructions, the implementation of generative AI still repeatedly stalls. Middle managers determine whether generative AI will eventually become a normalized workflow, a superficial trial, a compliance problem, or a project quietly shelved in the organization.
Middle managers do not only execute top-level decisions; they also interpret what these decisions mean to employees. Front-line teams observe their direct managers to judge whether generative AI is a practical tool, a passing trend, a monitoring method, or an early warning of future layoffs. Executives may promise that generative AI can reduce repetitive drudgery, but employees will only believe this promise when managers adjust workloads, review methods, performance standards, and decision-making authority accordingly. This means that the implementation of generative AI is essentially an operational model transformation, not an ordinary software launch. Middle managers are at the intersection of strategic vision, employee anxiety, ambiguous policies, existing workflows, and accountability for errors in generative AI outputs.
Senior leaders often misjudge the root cause of the problem. They think that implementation only requires three things: clearly articulating the strategy, selecting the technology, and training employees. These steps are essential, but far from sufficient. The advancement speed of generative AI transformation has far exceeded the iteration speed of most enterprises' operating models. McKinsey's *2025 State of AI Report* points out that the use of AI is already very common, but many enterprises still fail to deeply embed it into business processes to obtain tangible corporate value. In addition, Microsoft and LinkedIn's *2024 Workplace Trends Index* shows that employees have begun to use AI at work, while many managers still have not found a clear path to translate individual productivity gains into business value — this is exactly the gap between enthusiasm and implementation.
And the missing key link is the managers' transformation work: who will determine which workflows need to be transformed, how to verify quality, which scenarios employees can try, and how to deal with errors after generative AI-assisted work goes wrong.
Just imagine what will happen after the executive team announces the full rollout of generative AI in all departments: some managers allow employees to experiment, formulate review standards, and select a set of workflows that can be reconstructed; other managers choose to wait and see for more clear guidance, and continue to evaluate employees according to the old standards. Within just a few months, the team led by the former has precipitated a set of reusable practical methods; the latter's team is divided into two groups: one group completely avoids generative AI, and the other group uses it privately without supervision.
The consequences of this difference are not only uneven implementation of generative AI, but also unstable output quality, hidden data risks, repeated trials, and reduced employee morale, which ultimately leads to no substantial return on technology investment. As generative AI moves from optional pilots to core business processes, enterprises can no longer leave relevant decisions entirely to managers to figure out on their own.
Five Managerial Mindsets That Affect Generative AI Implementation
My research provides a more accurate diagnosis. To write the book *The Psychological Logic of AI Implementation in the Workplace*, I conducted 35 focus group surveys, interviewed more than 250 middle managers, and completed 22 in-depth interviews in industries including insurance, manufacturing, legal services, retail, education, and technology. The study found that middle managers are not a monolithic "change-resistant group" and can be divided into five psychological profiles. Each group has different reactions to pressure, risks, empirical information, incentive mechanisms, and governance rules. If the same approach is applied to everyone, a few eye-catching pilot projects may be produced, but it will be difficult to achieve stable implementation across the entire department.
1. The AI Skeptic
The first category is generative AI skeptics, who are the most easily labeled and the most easily mismanaged. Skeptics are not simply stubborn; their concerns are often reasonable: accountability, job reduction, data privacy, output quality, and the human cost brought by automation. They have seen too many transformation projects that start with loud slogans but end up leaving a bunch of broken processes. Once a compliance summary generated by generative AI has an error, they know they are likely to be the one held responsible. When executives talk about improving efficiency, it sounds like layoff risks to them.
In daily work, skeptical managers will postpone tool approval, limit generative AI to voluntary individual attempts by employees, or require employees to explain each usage scenario. I once met a skeptic in a leadership training session. Even though the company's IT department had reviewed and approved the software's terms of service and made it clear that it would not use the data of registered users for model training, he still repeatedly emphasized that he did not dare to let his team's data come into contact with generative AI tools. In educational institutions and non-profit organizations, this kind of resistance also stems from a concern: AI will consume huge amounts of water and electricity resources.
Teams led by skeptics usually end up with two results: either completely abandon generative AI, or use it quietly without sharing prompts, results, or reporting errors. The former prevents the organization from accumulating experience; the latter forms "shadow AI usage" without unified quality and safety specifications.
When dealing with skeptics, you cannot just preach the benefits of generative AI. They need empirical cases and a clear division of powers and responsibilities. Leaders can delineate limited, low-risk pilot projects for them, set clear indicators, retain manual review, and explicitly stipulate the accountability for errors in generative AI-assisted content. A truly valuable generative AI pilot is not a fancy demo, but a carefully selected process test: compare the cycle, quality, rework volume, employee experience, and risk events of the old and new processes. Skeptics need to be assured that the enterprise will not let them bear the unknown risks alone just to fulfill the productivity story in other people's mouths.
2. The Wait-and-See Conservative
The second type of manager will create a more hidden implementation bottleneck. They will not stop generative AI directly, but choose to postpone its promotion. They always wait for more supporting evidence, experience from other departments, new policy memos, and observe for another quarter. Their resistance sounds very reasonable, full of prudence. But if all teams wait for others to test the waters first, this passive waiting will drag down the overall strategy. This requires AI change management to jump out of general training and form positive peer-driven momentum based on credible internal success cases.
To persuade wait-and-see conservatives, case comparison is suitable rather than forced pressure. You can pair them with recognized managers in adjacent business departments who have already run successful pilots. Show them how the marketing team shortens the time for drafting while retaining manual editing and review; or how the finance team uses generative AI to interpret the causes of differences, with humans still making the final judgments. Set lightweight implementation indicators that match their own business processes, rather than grand overall group goals. Their core question is not "Is generative AI good?", but "Why should I invest my limited management energy in this matter right now?" Leaders' answers must fit the practical constraints of their departments.
3. The Prudence-Focused Implementer
Prudence-focused implementers are often the most valuable type of managers in the entire system. They embrace generative AI but have strict requirements for control mechanisms. They will proactively think about scenario screening, review processes, reporting mechanisms, and measurable results. It is this group of people who have the ability to turn enthusiasm into scalable practices. But enterprises often frustrate them: they only give general permission for use, but do not provide supporting implementation support. Prudence-focused implementers do not need to attend another inspiring all-hands meeting; they need prompt libraries, approved data usage rules, pilot templates, quality verification standards, and a channel where they can report defects without punishment.
If senior leaders do not clarify these supporting rules, prudence-focused implementers will limit the pilot to a very small scope, postpone full promotion, or repeatedly go through the review process for each new scenario. Although their teams can use generative AI responsibly, they can only be limited to scattered small scopes and cannot form standard practices. What seems overly cautious in the eyes of executives is essentially the lack of an operating system for responsible implementation.
AI strategy is put into practice at this link. Boston Consulting Group's *2026 Workplace AI Research* proposes that strategic clarity is more important than tool access permissions, because the speed of job changes has exceeded the speed of enterprise operation reconstruction. Prudence-focused implementers are exactly the bridge connecting the two. Executives put forward goals such as "improving customer response speed with the help of AI", which are translated by them into verified processes: which customer consultations are applicable, which data sources are allowed to be called, what content AI is responsible for generating, which links are manually reviewed, and which indicators are tracked after launch.
Leaders should grant prudence-focused implementers the authority to design review processes for limited scenarios, and allow them to connect with legal, information security, human resources or compliance teams in a timely manner. Then ask them to organize the processes, quality inspection items, defects and results into documents for reuse by other teams. In this way, their prudence will become organizational infrastructure, rather than a brake that permanently hinders implementation.
4. The Active Experimenter
The fourth type of manager is the driving force for project advancement. Active experimenters are willing to try new tools, share prompts, and organize informal demonstrations, turning generative AI from an abstract concept into a practical tool that can be used. They can turn employees' concerns into curiosity and produce real outputs, not just PPT plans. In many enterprises, it is they who first prove that generative AI can be upgraded from individual efficiency improvement to team-level practice. The Wharton School's *2025 AI Implementation Report* emphasizes that to translate popular usage into sustainable return on investment, people and process leverage are crucial, and active experimenters are often the starting point of this leverage.
But their risk lies in prioritizing speed and lacking frameworks. Active experimenters may use tools in advance before legal, compliance, information security, and human resources define boundaries. Their early results are very convincing, which inadvertently makes unsafe operations the norm. Leaders do not need to slow them down with cumbersome processes, but should set guardrails to guide them: provide approved tools, share prompt libraries, clarify review standards, and establish channels for reporting policy doubts. Reward their safe implementation results, not just the speed of advancement.
5. The AI Catalyst
The last category is generative AI catalysts, which are not only valuable assets of the organization, but also put governance capabilities to the test. Catalysts regard generative AI as the engine of comprehensive transformation, hoping to restructure processes across teams, influence peers, build collaborative networks, and push executives to accelerate the pace. When the environment is suitable, they will become internal entrepreneurs and lead the enterprise out of the pilot quagmire — countless pilot projects have been stuck in the experimental stage and cannot be promoted. But if the supporting environment is insufficient, they will exceed the trust boundary, underestimate compliance risks, and cause resistance from teams that are treated as test subjects.
For catalysts, generative AI governance should empower rather than be a mere formality. The U.S. National Institute of Standards and Technology's *Generative AI-Enabled AI Risk Management Framework* lists risks such as data privacy, information security, AI hallucinations, harmful bias, and intellectual property. These risks are not reasons to stop catalysts, but design requirements for responsible and accelerated implementation. Catalysts can be required to carry out pre-drill reviews before high-impact pilots, set reporting thresholds, record manual review nodes, and sort out experience for reuse by other teams.
Implement Targeted Promotion Plans Based on Individual Profiles
Identifying these five profiles is only the first step. Most enterprises have five types of managers coexisting internally, even concentrated in the same business department or management team. Senior leaders need a standardized diagnostic method to find the root cause that hinders each manager's actions, and match corresponding support, incentive and accountability mechanisms.
A practical diagnosis starts with three questions:
· What is this manager afraid of losing?
· What kind of evidence can change his mind?
· What support is needed for him to promote the work responsibly?
Skeptics are afraid of taking responsibility and job replacement; wait-and-see conservatives worry about wasting effort and taking political risks; prudence-focused implementers fear uncontrollable risks; active experimenters are afraid of missing opportunities; catalysts are disgusted by the organization's conservative and timid attitude. These differences are crucial, and each type of profile requires a different leadership style.
Leaders also need to redesign the incentive mechanism. Too many enterprises publicly recognize the enthusiasm for embracing AI, but privately hold people accountable for AI-related mistakes. This will make managers adopt a defensive attitude: on the one hand, they require experiments, and on the other hand, they blame all flaws. The rational choice is to wait and see; if you only blindly praise speed and ignore accuracy, privacy and trust, it will lead to reckless behavior. A better incentive system rewards rigorous exploration: clear preset goals, responsible pilots, measurable process improvements, transparent error reporting, and evidence-based large-scale promotion.
For resistant managers, the solution is to present empirical evidence and create psychological safety. We need to communicate frankly about the impact of AI on jobs. The World Economic Forum's *Future of Jobs Report* points out that until 2030, AI and information processing technologies will be the main forces reshaping work patterns and professional skills. Leaders do not have to pretend that transformation has no impact, but should explain what will change, what will remain the same, which skills are more important, and that managers who learn to use AI to enhance their work will have more advantages than those who avoid AI.
For supportive managers, the countermeasure is empowerment + delineation of boundaries. Give prudence-focused implementers the required operating specifications; give active experimenters a safe test field and channels to promote effective practices; give catalysts authorization tied to risk control. The classic pre-project rehearsal review is particularly applicable here, which allows objections to be raised before the pilot starts. Put forward the hypothesis: "If this generative AI pilot fails completely, what will be the reason?" Help the team discover legal, quality, trust, process and implementation risks in advance, and there is still enough time to optimize the plan.
The Real Test for Enterprise AI: Trust at the Manager Level
Executives should stop asking: "Why are employees reluctant to use generative AI?" Instead, think: "Which type of middle manager in each department is leading the generative AI implementation? Do we have a corresponding promotion plan in place?" This question can turn the vague cultural problem into a manageable agenda that can be implemented. It helps leaders avoid spending months on repeated pilots, resolve conflicts before anxiety solidifies into resistance, while absorbing the vitality of early users without letting them cross the governance bottom line.
This bottleneck is real, but it can be solved. Middle managers can block the implementation of generative AI, weaken its effectiveness, and distort its direction; they can also translate generative AI into measurable business value. Senior leaders who want to obtain practical guidelines for responsible AI implementation should regard middle managers as the operating system of transformation, rather than as inconspicuous supporting pipelines under the strategy. The competitive advantage in AI in the next stage will not belong to the enterprises that shout the loudest slogans, but to those that understand their own managers and provide everyone with the resources needed to promote change.
Keywords: #AI
Gleb Tsipursky | Text
Gleb Tsipursky is the CEO of Disaster Avoidance Experts, an AI consulting firm. He has provided consulting for Fortune 500 companies and has engaged in academic research on behavioral science at the University of North Carolina at Chapel Hill and The Ohio State University for more than 15 years.
Zhou Qiang | Editor