AI Doesn’t Sign Off: Why The Most Skilled Professionals Hesitate To Work With Generative AI

An MBA consultancy project started with a simple question and ended with an unexpected answer about people, trust, and generative AI

by Anton Vasile, EMBA RO15

Generative AI is now part of almost every conversation about work, and many organizations are running into the same puzzle: interest is high, yet everyday use lags far behind. When I started my MBA consultancy project, the obvious explanation was that people simply did not know enough yet. The data pointed somewhere else. The question that mattered was not “How much do people know?” but “Who is accountable when the AI gets it wrong?

What we thought we knew

 Many digital transformation programs start from the same line of reasoning: people lack the skills, so they resist, so the answer is training. The assumption is intuitive, and it is easy to budget for. A course catalog can be measured in hours and headcount without much effort.
 I wanted to see whether that reasoning would hold in an organization in Romania. The study followed a sequential mixed-methods design. It began with a survey of 47 respondents, made up of a core team and a comparison group from adjacent functions. It continued with 22 semi-structured interviews, analyzed using the Gioia methodology (Gioia et al., 2013). Instead of asking people about theoretical constructs, the interviews stayed close to concrete situations from their own work, in the spirit of the critical incident technique.

The Paradox: skill does not predict resistance

 The first result went against my starting hypothesis. In this sample, I found no relationship between technical competence and resistance to change; the correlation was close to zero (r = −.038). Put plainly, you could not tell who would hold back from AI by looking at how well they understood the technology.

What a model-based cluster analysis did show was four distinct behavioral profiles:

• Silent Adopters, the majority of the team: skills slightly below average, real openness, but mostly passive use.

• Expert Skeptics: among the most competent people in the team and, at the same time, among the most resistant.

• The Innovation Champion: rare, highly skilled, and very open.

• Critical Resistance: limited skills and strong resistance.

 The second group is the one worth a closer look. Expert Skeptics are not afraid of the technology. The core team was clear that learning AI did not feel hard and that they did not see it as a threat to their expertise. The comparison group, whose technical skills were weaker, felt that threat more strongly. The experts’ reluctance has a different source. Because they understand the technology well, they can see exactly what is at stake when the AI gets something wrong.

The Barrier Triad

 The interviews made sense of the paradox. Resistance did not grow out of a single cause. It came from three barriers that keep each other going, which I call the barrier triad.

 1. Individual accountability. Almost every respondent (45 of 47) believes that final responsibility for an error lies with the person who validated the AI output. Hardly anyone placed it with the technology vendor or with the IT and governance functions. One participant summed it up: “The responsibility is ours, not the AI’s.” The AI does not sign off on anything; people do. As long as the organization leaves open who is accountable, and under which conditions, going back to manual work is a rational choice rather than a conservative one. At the individual level, this is the ground Gartner covers under the label AI TRiSM, short for AI trust, risk, and security management.

 2. Lack of psychological safety to experiment. Trying out generative AI means asking beginner questions, producing rough drafts, and making mistakes in front of others. Where mistakes are punished, people stop experimenting. Respondents pointed to three main obstacles: no training tied to their actual tasks, security rules that felt either too strict or too vague, and simply not enough time. The format they wanted most was a protected space to experiment, followed by learning from colleagues. Theory-based courses came last.

 3. The verification burden. Even when the AI gets the result right, many professionals redo the work step by step, because they cannot trace how the answer was reached. “It takes longer to validate the output than to do it manually,” another participant told me. Research helps explain this. People lose trust in an algorithm faster after a single error than they lose trust in a colleague who makes the same mistake (Dietvorst et al., 2015). Generative AI also has what Dell’Acqua et al. (2023) call a jagged frontier: it does remarkably well on some tasks and surprisingly poorly on others that look much the same. If you cannot tell in advance where it will slip, you end up checking everything. It is telling that specialists felt confident about traditional checks on results but far less sure they could catch errors specific to AI, such as hallucinations.

 Why they work together. The three barriers form a vicious circle. Unclear accountability breeds fear, and fear shuts down experimentation. Without experimentation, familiarity never builds up. Without familiarity, you cannot tell where the tool deserves trust, so you check everything. Checking everything makes AI look like more trouble than it is worth, and the fear grows. This is an interpretive model drawn from the interviews, not a proven causal chain. Its practical lesson is clear all the same: an intervention aimed at only one barrier will be undone by the other two.

 One further finding is encouraging, and it concerns leaders. Managers and specialists turned out to be equally open to innovation, and managers approached AI with more confidence and less anxiety. The group of managers was small, so this is only an indication. Even so, it suggests that leaders are an asset rather than a brake, and that they can be the first to experiment in plain view. The challenge is to turn their confidence into the practical conditions their teams need, since the day-to-day risk sits mostly with the people doing the hands-on work.

What managers can do

 If resistance is not rooted in missing skills, more courses will not solve it. The response has to deal with all three barriers at once and allow for the different profiles within a team.

1. Settle accountability before asking for adoption. Decide who is responsible and under what conditions, what has to be documented, and when an output should be labeled as AI-assisted. Clear rules turn a vague fear into a risk that can be managed.

2. Make room, and time, for experimentation. Teams need a protected environment with real but non-critical tasks and anonymized data, where a mistake counts as a lesson and not as a breach. Specialists described a few hours a month, built into the regular schedule, as a realistic pace.

3. Agree on a handful of simple rules (Sull & Eisenhardt, 2015). Boundary rules state what AI should and should not be used for. Priority rules state which tasks to automate first. Exit rules state when to fall back on purely human analysis.

4. Match the intervention to the profile. Silent Adopters need hands-on practice in short sprints. For Expert Skeptics, one more technical course is money spent in the wrong place; coaching and trust-calibration exercises will do more. Champions can step into mentoring roles. The Critical Resistance group needs the fundamentals and one-to-one support.

5. Help capture tacit knowledge. A good many working rules exist only in the heads of senior specialists, and what has never been written down cannot be handed over to an AI system.

6. Experiment where people can see it. When leaders show their own attempts with AI, failed ones included, the team gets the clearest signal of psychological safety it could ask for.

What specialists can do

 Change does not have to come only from the top. The people who use AI every day can also break the vicious circle from within.

1. Calibrate trust instead of checking everything. Check a sample, run AI alongside your manual approach for a limited time only, and keep a record of where it goes wrong. In a few weeks you will have your own map of where the tool can be relied on.

2. Learn to spot AI-specific errors. Noticing that a figure does not add up is not enough, because hallucinations often sound perfectly reasonable. Knowing how generative models differ from traditional ones is no longer a technical detail; it has become part of professional competence.

3. Use self-change heuristics (Atanasiu, 2021; Atanasiu et al., 2025). A “Ulysses contract” keeps you from sliding back into an old habit, for example: “I will no longer redo by hand what I can generate and then check on a sample.” A “stepping stone” supports the new habit, for example: “Before I hand in an analysis, I ask myself which part AI could have sped up.”

4. Be open about it. Ask for clarity on who is accountable, and say plainly which parts of your work were done with AI. Openness protects you better than avoidance does.

5. Learn from colleagues, then pass it on. Specialists rated peer learning far above theory-based courses. Successes and mistakes that get shared shorten the learning curve for everyone.

6. Rethink your role. The value no longer lies only in producing the result but in interpreting it and explaining it. One participant from the business side said what they needed from specialists was someone to turn the “black box” into something they could actually base a decision on. That part of the job is not one AI can take over.

Limitations and Conclusion

 These findings come from a single case study with a small sample and self-reported competence data, all collected at one point in time. The profiles and their relative sizes should therefore be read as indicative. Whether the barrier triad shows up in other organizations and industries is a question for further research.

 The main conclusion, though, is easy to put to the test anywhere. When people feel answerable for whatever AI helps them produce, adopting it is less a training problem than a question of how trust is built into the way work gets done. That means clear accountability rules, permission to make mistakes safely, and verification routines that cost less than the work they are supposed to save. Organizations that invest only in courses may end up with well-trained people who keep doing things by hand.