Imagine your company deploys an AI tool, whether for writing proposals, analyzing contracts, or preparing reports. In the pilot, it works flawlessly, and people praise it during the management presentation. Three months later, almost nobody clicks on it. The tool hasn't slowed down, hasn't started hallucinating, nor has its output quality degraded. People simply stopped using it.
This scenario repeats itself so frequently that researchers have begun studying it. Das Narayandas and Shunyuan Zhang from Harvard Business School found that at least 30% of generative AI projects in companies end up quietly abandoned—not because of a technical failure, but because employees quietly stop using them¹. Nobody announces it at a meeting. Nobody complains. Adoption simply fades away gradually.
This article is the first in a four-part series examining what a company needs to resolve to make AI truly work—not just technically, but in day-to-day operations. We start with the role, because without it, the other three aspects (compensation, processes, and leadership) have nothing to lean on.
Why People Reject a Tool That Works
The key finding from the research is that employees do not reject AI because it works poorly. They reject it because it threatens their professional identity. This is a different problem than what companies usually address when planning a new technology deployment, which is why standard solutions (training, manuals, presentations at town halls) fail to work.
Narayandas and Zhang describe three specific mechanisms through which this threat occurs:
Loss of interesting work. When AI takes over the part of the work that required judgment and expertise, the employee is left with the less interesting, lower-status part. The authors describe this as a situation where "the machine does all the interesting stuff"¹. A senior analyst who previously spent the day on deep data analysis and presenting conclusions now merely checks and approves what the tool generated in a few minutes. The work has accelerated, but the part that provided a sense of expertise has vanished.
Shift of control. Decision-making authority shifts to the algorithm, even if it formally remains with the human. The job title doesn't change, but the sense of authorship over the decision does. A manager who previously compiled recommendations for a client independently now merely clicks to approve what the AI suggested. They still hold the responsibility, but the feeling that they are the decision-maker is gone.
Erosion of influence. Managers lose part of their influence over people and processes they previously controlled. The authors describe this metaphorically as becoming a "kingdom without a throne": formally still leading the team, but real influence over how the work gets done is taken over by the system.
None of these three mechanisms are solved by telling people how to operate the tool. They are solved by adjusting the role itself.
What People Actually Want from AI
Interestingly, this resistance is not tied to people rejecting AI across the board. In its State of AI in Business 2025 research, MIT found that for routine tasks—such as emails, summaries, or basic analysis—70% of users prefer AI over a human². For work that is complex or carries a high risk of error, the ratio flips: 90% of people prefer a human over current AI systems².
This is precisely the boundary that an adjusted role must respect. If a role shifts routine tasks to AI (which people welcome) while retaining complex tasks with the human—where people trust themselves more than the tool—the role remains valuable and no one feels threatened. However, if the boundary is drawn incorrectly—giving AI responsibility for the complex, judgment-dependent part while leaving the human with only checking and approving—the exact loss of interesting work described by Narayandas and Zhang occurs.
What Adjusting a Role Means (and What It Doesn't)
Adjusting a role does not mean writing a new job description in the HR system. It means answering a specific question: What does this role now do differently, and why is it still valuable work that makes no sense to replace?
Research points to five mechanisms companies can use, which work best together rather than individually¹:
Redefine accountability. Instead of keeping the role identical with the caveat that "now you also use AI," it is necessary to articulate specifically what has changed and where the human's new added value lies. For a sales representative, this could mean shifting from drafting proposals (now handled by the tool) to leading more complex negotiations where the tool cannot reach.
Preserve the right to override the tool. If the AI provides recommendations and the human merely rubber-stamps them, they lose their sense of authorship. If they have a clearly defined right and responsibility to challenge and alter recommendations, decision-making remains theirs, even when relying on AI input.
Add a layer that enhances judgment, rather than replacing it. Where the tool serves as a second perspective on a problem rather than a replacement for decisions, people maintain their role as meaningful.
Offer a realistic upskilling pathway. Not a declaration like "we invest in people's development," but a concrete plan detailing what the person will learn and how it will benefit them when part of their previous work is automated.
Ensure the change is publicly supported by leadership. This is often underestimated. If a manager does not state publicly that the new role makes sense and that they back it, employees interpret it as a signal that this is merely a temporary solution until a way is found to replace them entirely.
What It Looks Like in Practice
A good example—though from a slightly different area than analyst productivity—is an experience from our recent project, where we developed an AI tool for a client designed to give sales representatives feedback after every customer meeting. Across 47 advisors and one manager, this fundamentally changed the role of the manager. Previously, they physically had time to personally coach only a fraction of the team, leaving others untouched. Feedback was given to everyone only once a year during annual reviews. Now, the manager's role has shifted: they do not coach all 47 people personally, but instead decide where the AI-suggested feedback requires their involvement and guidance, spending time on what AI cannot see—the client relationship, the context of a specific situation, and decisions under pressure. Their role did not shrink. It shifted toward what the manager does best and previously lacked the capacity to do.
This is the exact difference between "AI replaced part of the work" and "AI enabled the role to perform different, higher-value work." The first version leads to rejection. The second to adoption.
A similar pattern can be found outside sales. In recruitment, AI is now commonly used for the initial screening of CVs. Companies where this works do not give AI the final say; the recruiter still conducts the interview, decides on cultural fit, and takes responsibility for the final recommendation. Conversely, companies where recruiters quietly bypass the tool allowed AI to decide on matters recruiters previously considered the core of their expertise—namely, who even gets a chance at an interview. The difference is not in the quality of the algorithm. It lies in where the company drew the line between what the machine does and what remains with the human.
How to Tell if a Role Is Properly Adjusted
Before a company claims it has "adjusted a role for the AI era," it is worth running through four specific questions. If the company cannot answer most of them specifically, the role has likely changed only on paper:
Can the person in this role clearly describe what they do differently compared to a year ago, and why it is more valuable, not just different?
Do they have both formal and informal rights to reject AI recommendations without it triggering extra paperwork or suspicion that they are doing something wrong?
Do they know precisely what new skill they need to acquire, and do they see how it will practically help them—rather than a vague "it will be useful in the future"?
Have they heard someone from leadership state publicly and specifically why the new form of their role makes sense, rather than just receiving general encouragement to "try AI"?
What Happens When Roles Are Not Adjusted
If a company deploys a tool and leaves roles unchanged, employees will not use the tool to its full potential. They will use it only to the minimum extent necessary to keep it unnoticed, and after six months, the company will wonder why the AI investment failed to deliver expected returns. The answer is rarely that "the tool was bad." Far more often, it is that "nobody changed what is expected of the people working alongside the tool."
The technology works in most failed deployments. What is missing is the work on what the people around it do.
Why This Is Not Just a Question for Sales or IT
Role adjustments around AI are most frequently addressed where the impact is most visible—in sales or technical teams. However, the same logic applies where it is discussed less, such as in leadership development itself. If AI provides a manager with ongoing feedback on how they lead people, the manager's role changes just like the sales representative's role in the example above: previously, they received feedback once a year during a formal review; now, it is available continuously for immediate application. For this to work, this role must also go through the same four questions: what the manager does differently, whether they can challenge recommendations, what they need to learn, and whether someone above them publicly supports it. Adjusting a role is not a one-off task tied to a specific department. It is a question that every role incorporating AI must ask.
What Comes Next
An adjusted role is the first step, but not the last. Even a well-designed role hits a limit if the company continues to pay and reward employees according to old rules—that is, based on metrics that fail to account for how the work has changed. That will be the focus of the second article in this series: why compensation is one of the most underrated aspects of AI adoption, and what specifically to do about it.
This text is the first in a series on what a company needs to resolve for AI to truly work: roles, compensation, processes, and leadership. It is based on publicly available research and Kogi's practice.
Sources
Harvard Business School Working Knowledge, "Why Employees Resist AI, and How Companies Can Win Them Over," research by Das Narayandas and Shunyuan Zhang. https://www.library.hbs.edu/working-knowledge/why-employees-resist-ai-and-how-companies-can-win-them-over
MIT NANDA, "State of AI in Business 2025" (The GenAI Divide), data on human preferences regarding AI in routine vs. complex tasks. https://mlq.ai/media/quarterly_decks/v0.1_State_of_AI_in_Business_2025_Report.pdf