Fear of AI Agents: Why Teams Hesitate and How to Move Forward
Introduction
Most people who feel fear of AI agents are not afraid of technology itself. Instead, they fear two specific things. First, they do not trust the work an agent produces. Second, they believe the agent will eventually take their job. Both fears are reasonable. However, neither is a good reason to stand still.
We see this hesitation in nearly every team we talk to. For example, Developers use AI chat tools privately but resist giving an agent a real story. Managers approve a pilot but quietly hope it stalls. Meanwhile, team members ask, sometimes out loud and sometimes not, whether helping the agent succeed means training their own replacement.
In Frank Herbert’s novel Dune, the young Paul Atreides repeats a litany to stay in control under pressure. Its most famous line is “Fear is the mind-killer.” That line describes what we see in teams today. Fear does not protect people from AI. Rather, it stops them from thinking clearly about it.
In this blog, we name that fear and look at what the research actually says about it. We then lay out a practical, Scrum based way for teams to move from reluctance to confident adoption. Our conclusion comes first: teams should treat fear of AI agents as an impediment to make transparent, inspect, and address. It is not a personality trait to argue away.
The fear and its object
Two layers are easy to conflate, so we keep them apart:
- The fear itself concerns trust, control, and identity. It lives in how people feel about handing part of their work to something they do not fully understand.
- The object of the fear is a real change in how work gets done. Agents will take over some tasks, and some roles will change shape.
Reducing the fear does not stop the change. Likewise, ignoring the change does not reduce the fear. Therefore, teams need to work on both.
The Three Fears Behind the Reluctance
When we listen closely, the reluctance almost always breaks down into three distinct fears. Each one needs a different answer.
1. “I don’t trust what it produces”
Mistrust is the most common fear and the most practical one. People do not simply worry that AI might be wrong. Rather, they distrust it because it is often wrong in a way that looks right. An agent can invent a statistic, cite a paper that does not exist, or write code that compiles and still fails. Worse, it presents all of this with the same calm confidence it uses for correct answers.
People have seen the consequences in public. In 2023, for example, a US federal court sanctioned two lawyers who filed a brief citing court cases that ChatGPT had invented (Mata v. Avianca). Developers report the same problem at a smaller scale every day. In the 2025 Stack Overflow Developer Survey, the most common frustration with AI tools was output that is “almost right, but not quite.”
As a result, people who answer for the final result feel a real risk. If the agent fails, they carry the cost, not the agent. Consequently, many conclude it is safer to do the work themselves.
This mistrust often runs strongest among experienced professionals. They know exactly what good work looks like, so they notice every flaw. They also carry the reputational cost when something slips through.
2. “It will take my job”
This fear concerns the future, not any single task. People worry that every task they hand to an agent proves their role matters less. Some respond by hiding how much they use AI. Others refuse to use it at all, hoping the threat goes away if the agent does not succeed.
This fear cuts especially deep for people early in their careers. After all, their work often consists of exactly the tasks agents now handle well.
3. “I’ll lose control of my work”
The third fear is quieter. It concerns identity and craft. People who take pride in their work worry that they will become reviewers of someone else’s output instead of makers. They also worry about losing visibility. If the agent does part of the work, who understands the whole system? And who answers for it when it breaks?
In the end, all three fears share one root. We ask people to trust something before they have seen it work, inside a process they did not design, with consequences that feel personal.
Fear Is the Mind-Killer: A Lesson from Dune
Frank Herbert published Dune in 1965, and its warning about fear still applies to the arrival of AI. The book’s Litany Against Fear opens with the line “Fear is the mind-killer.” The litany teaches its speaker to face fear, let it pass, and keep thinking clearly.
Frozen by the machine
Early in the novel, an assassin sends a hunter-seeker after Paul: a small flying weapon that homes in on movement. A panicked person would run, and the device would strike. Paul, however, stays completely still while it hovers near him. Then, at the right moment, he grabs it and destroys it. He survives because he controls his fear, not because he avoids the machine.
Many professionals face AI agents in a similar way today. Some freeze in the wrong sense: they stop learning, stop experimenting, and hope the technology passes them by. Others flail and hand everything to the agent without checking it. Neither response works. Instead, the right response is calm attention followed by a decisive, well judged move.
The cost of banning thinking machines
Dune also describes a universe that, long before the story begins, fought a war against thinking machines and then outlawed them entirely. Herbert’s world shows that fear of machines can shape a whole civilization. In our view, the lesson for teams is clear. Neither banning AI from your work nor surrendering to it is a strategy. Mastering it is.
What the Research Says
The research confirms three things. The fear is widespread, it is partly justified, and teams that use AI agents deliberately get measurable gains. To reach that conclusion, we reviewed worker surveys, trust in automation research, controlled productivity studies, MIT research on AI use, and labor market analyses published between 1997 and 2025.
The fear is common and often hidden
- A Pew Research Center survey of US workers (published February 2025) found that 52% feel worried about the future impact of AI in the workplace. By contrast, only 36% feel hopeful. About a third expect fewer job opportunities for themselves in the long run.
- Microsoft and LinkedIn’s 2024 Work Trend Index found that 75% of knowledge workers already used generative AI at work. Yet 52% felt reluctant to admit using it for their most important tasks, and more than half worried it would make them look replaceable.
The pattern is clear. People use AI, but privately and cautiously. In other words, the fear does not stop adoption. Instead, it drives adoption underground, where nobody can learn from it.
Mistrust runs deeper than use
- The KPMG and University of Melbourne global study of trust in AI (2025, about 48,000 people in 47 countries) found that only 46% of people feel willing to trust AI systems. Notably, a majority of the same people use AI regularly.
- The 2025 Stack Overflow Developer Survey found that 84% of developers use or plan to use AI tools. Even so, more developers distrust the accuracy of AI output (46%) than trust it (33%).
These numbers show a gap between using AI and trusting it. People keep using the tools because they help. At the same time, they do not believe the output without checking it. That instinct is healthy, as long as it turns into a method rather than a refusal.
Trust in automation is a calibration problem
Researchers have studied human trust in automation for decades. Parasuraman and Riley (1997) described two failure modes. In misuse, people over rely on automation. In disuse, people reject automation that would help them. Later, Lee and See (2004) argued that the goal is neither more trust nor less trust. Rather, the goal is calibrated trust: trust that matches what the system can actually do.
Similarly, Dietvorst, Simmons, and Massey (2015) identified algorithm aversion. People stop trusting an algorithm after they see it make a single mistake, even when it outperforms humans overall. By contrast, they forgive human error far more easily. Logg, Minson, and Moore (2019) found the reverse effect in laypeople, who often prefer algorithmic advice. However, they noted that experts tend to discount it.
This research explains a lot. Experienced professionals see one hallucination and conclude that nobody can trust the agent. That reaction is human, but it is not calibrated.
AI agents help most on clearly defined tasks
| Study | Who | What AI did | Finding |
|---|---|---|---|
| Dell’Acqua et al., Harvard Business School and BCG (2023) | 758 consultants | Assisted realistic consulting tasks | Inside AI’s capability frontier: 12.2% more tasks, 25.1% faster, over 40% higher quality. Outside it: 19 percentage points less likely to be correct |
| Brynjolfsson, Li, and Raymond (2023, published 2025) | 5,179 customer support agents | Real time AI assistant | 14% more issues resolved per hour on average, about 34% for novice workers |
| Peng et al., GitHub Copilot experiment (2023) | 95 developers | Coding assistant | Task completed 55.8% faster |
| METR (2025) | 16 experienced open source developers, 246 tasks | AI tools on their own mature codebases | 19% slower, although developers believed they were about 20% faster |
Two lessons stand out. First, the gains are real and sometimes large. Second, they depend on where and how people use the agent. The Harvard and BCG authors called this the jagged frontier: AI excels at some tasks and fails at others that look similar. Moreover, the METR result shows that even experts misjudge where they sit on that frontier. So the evidence supports neither blind trust nor blanket rejection. It supports measurement.
What MIT research shows about who succeeds
Two MIT studies from 2025 point in the same direction: the people and organizations that succeed with AI are the ones who stay actively engaged with it.
- MIT NANDA, The GenAI Divide: State of AI in Business 2025. Based on interviews, surveys, and a review of public deployments, this report found that about 95% of organizations saw no measurable return from their generative AI pilots. The small group that succeeded did not simply deploy tools. Instead, they integrated AI into real workflows, gave it feedback, and held it to measurable business outcomes.
- MIT Media Lab, Your Brain on ChatGPT (Kosmyna et al., 2025). In this early, small study (54 participants, not yet peer reviewed), people who relied on an AI assistant to write essays showed the lowest brain engagement, and many struggled to quote their own work. By contrast, participants who did their own thinking first and then used the AI showed stronger engagement.
A 2025 Microsoft Research and Carnegie Mellon survey of 319 knowledge workers (Lee et al.) adds an important detail. Workers with more confidence in AI applied less critical thinking. Meanwhile, workers with more confidence in their own skills applied more. The researchers also found that critical thinking shifts toward verifying, integrating, and supervising AI output.
Taken together, these findings point to one conclusion. AI does not make people better or worse on its own. What matters is whether the human stays in charge, checks the work, and keeps thinking.
Job impact is real, but it is mostly a shift
- The World Economic Forum’s Future of Jobs Report 2025 projects about 170 million jobs created and 92 million displaced by 2030, a net gain of about 78 million. It also expects around 39% of workers’ core skills to change in that period.
- A 2025 Stanford Digital Economy Lab study by Brynjolfsson, Chandar, and Chen found a relative employment decline of about 13% for workers aged 22 to 25 in the occupations most exposed to AI. However, more experienced workers in the same occupations stayed stable or grew. The declines clustered where AI automates tasks rather than augments people.
The honest reading is this: the job fear is not imaginary, particularly for entry level work. Still, the data points to a clear protective factor. People whose work AI augments, who direct it, review it, and combine it with judgment and context, do better than people whose tasks AI simply automates. Therefore, learning to work with an agent is the response the evidence supports.
Verify Every Output: How Checking AI Makes Humans Smarter
Trust in automation requires proof. For that reason, our rule is simple: a human verifies every output an AI agent produces, from start to finish, before it counts as done. Verification is not a sign of distrust. Rather, it is how trust gets earned.
Once humans can verify they will realize that the human in the loop slows things down. Then they start training an independent agent to verify even better than a human does. Then will then spot check that independent agent so they master the control of agent output without slowing it down.
Why verification is the answer to mistrust
Mistrust of AI is justified when people accept output blindly. However, it becomes unnecessary when people check output systematically. The difference lies in the method, not in the tool. A team that knows how to verify agent work no longer needs to fear it, because the team catches errors before they reach a customer.
Strategies for checking an AI agent
Effective teams do not check agent work at random. Instead, they build deliberate strategies, such as these:
- Define acceptance criteria before the agent starts. If you cannot say what correct looks like, you cannot check it. So write clear criteria into the story first.
- Check facts against the source. Open every cited paper, link, and statistic. If the agent cannot point to a real source, treat the claim as unverified.
- Test, don’t read. For code, run the tests and add new ones for edge cases. Code that looks right proves nothing until it passes.
- Ask the agent to explain its reasoning. Then challenge the weak points. Often, an agent’s explanation reveals a gap that its answer hid.
- Spot check large outputs. For bulk work, such as data cleanup or document updates, review a sample in detail and look for patterns of error.
- Compare against a known answer. Periodically give the agent a task the team has already solved, and compare the results.
- Keep an error log. Record what the agent gets wrong. Over time, the log shows where the agent is reliable and where it needs close supervision.
How verification makes people smarter
Verification does more than catch mistakes. In fact, it sharpens the person doing it. To check an agent’s work, you must understand the problem well enough to know what a good answer looks like. You must also ask better questions, spot weak reasoning, and test assumptions.
The research supports this. As the MIT Media Lab study suggests, people who hand their thinking to AI engage less. In contrast, people who think first and then use AI stay sharp. Likewise, the Microsoft Research and Carnegie Mellon survey shows that critical thinking does not disappear with AI. Instead, it moves to verifying, integrating, and supervising.
So the choice is not between using AI and staying skilled. People who verify AI output end up building new skills: they learn to direct, question, and quality check a powerful collaborator. As a result, the human becomes more capable, not less necessary.
The Aikido Approach: Open Hand, Then the Right Move
Aikido offers a useful model for working with AI agents: welcome the energy coming at you, then direct it where you want it to go. Instead of meeting force with force, the practitioner opens the hand, engages, and turns the incoming energy into control.
A short history of Aikido
Morihei Ueshiba (1883 to 1969) developed Aikido in Japan during the 1920s and 1930s. He drew on older Japanese martial arts, especially Daito-ryu Aiki-jujutsu, as well as his own spiritual practice. The name combines three ideas: ai (harmony), ki (energy or spirit), and do (the way). Together, they mean “the way of harmonizing energy.”
Unlike many martial arts, Aikido does not aim to defeat an opponent through strength. Instead, the practitioner blends with the attack, redirects its momentum, and gains control without needing to overpower anyone. For this reason, traditional Aikido has no competitions. Ueshiba summed up the goal in the phrase masakatsu agatsu: true victory is victory over oneself.
Applying Aikido to AI agents
Aikido is not passive, however. The practitioner steps toward the attack, engages fully, and acts decisively. That is exactly how we recommend teams approach AI agents.
- Start with an open hand. Welcome the agent instead of resisting it. Give it real work, not a token task designed to fail.
- Step in and engage. In Aikido, you move toward the energy, not away from it. Likewise, learn how the agent works, where it is strong, and where it fails.
- Blend with its energy. Use the agent for what it does well, such as speed, volume, and first drafts. Don’t fight it on tasks where it is weak.
- Redirect toward your goal. Clear stories, acceptance criteria, and a Definition of Done steer the agent’s output toward the result you actually want.
- Take control. Verify every output and make the final decision yourself. The agent provides the energy; you provide the direction and the judgment.
Victory over yourself
Most importantly, Aikido teaches that the hardest opponent is your own fear. A practitioner who panics loses balance, whereas one who stays calm can handle almost any attack. The same holds for AI. Teams that master their fear and master the agent at the same time get the results. In practice, both happen together: the more skillfully you direct and verify the agent, the less you fear it.
Why the Fear Is Rational, and Where It Misleads
The fear is rational because it points at real risks. AI can be confidently wrong, agents will take over some tasks, and accountability can blur when people share work with a machine. Pretending otherwise damages credibility. In fact, it makes people more resistant, not less.
Nevertheless, the fear misleads in three ways.
- It judges the agent on its worst output. Because of algorithm aversion, a single visible mistake outweighs hundreds of correct results. Nobody judges a human teammate this way, so we should not judge an agent this way either. Instead, we should judge both on evidence collected over time.
- It treats avoidance as protection. Refusing to work with an agent does not stop the shift in how work gets done. It only means the change happens without you and without your judgment shaping it. As the research shows, the workers who do best are the ones who learn to direct and verify AI.
- It assumes the agent works alone. Most fear imagines an agent operating unsupervised. In a well run team, however, an agent works inside the same process as everyone else: clear stories, a shared Definition of Done, human verification, and inspection at every Sprint.
In other words, the problem rarely lies with the agent itself. Rather, it lies in adding an agent without a structure that makes its work transparent. Fortunately, a Scrum Team already knows how to solve that problem.
The Fear Map
Each fear has a true core, a distortion, and a Scrum response that addresses it directly.
| Fear | What is true | Where it distorts | Response |
|---|---|---|---|
| “I don’t trust its work” | AI can be confidently wrong, especially outside its capability frontier | One visible error becomes proof that the agent is never reliable | A human verifies every output; agent work meets the same Definition of Done as any Increment |
| “It will take my job” | Agents will take over some tasks; entry level roles face the most exposure | Avoiding the agent feels like protection | Developers decide which work the agent takes; people move toward directing, verifying, and integrating work |
| “I’ll lose control” | Accountability can blur when people share work with an agent | Assumes the agent operates outside the team’s process | Every agent task is a visible Product Backlog item with a named human accountable for acceptance |
| “I’ll lose my skills” | Passive reliance on AI reduces engagement | Assumes every use of AI is passive | Active verification builds new skills in questioning, testing, and judgment |
| “It will create more work than it saves” | Poorly scoped tasks can slow experienced people down | Assumes results from one setting apply everywhere | Measure against a pre agent baseline and inspect the result at each Sprint Retrospective |
A Scrum Approach to Overcoming the Fear
Scrum rests on empiricism: transparency, inspection, and adaptation. Fear of AI agents lacks exactly these three things. Fear grows where work stays invisible and people assume results. Conversely, it shrinks where work is visible and people measure results.
The approach below comes from our own work integrating an AI agent into a Scrum Team’s Jira workflow. It follows the Aikido pattern: open the hand, engage, then make the right move.
Phase 1: Make the fear visible
- Name the fear as an impediment. Raise it openly in a Sprint Retrospective, and ask each team member which fear they feel most strongly. Once someone speaks a fear, the team can address it. A hidden fear, by contrast, turns into quiet resistance.
- Start with one Scrum Team. Do not roll an agent out across the whole organization. Instead, choose one team of fewer than ten people that is willing to experiment, and treat the effort as a joint experiment rather than a mandate.
Phase 2: Build the evidence
- Measure a baseline first. Record the team’s velocity and quality before the agent joins. Without a baseline, every opinion about the agent remains just a feeling. With one, the team can see what actually changed.
- Give the agent the right first stories. Begin with work that is well defined, repetitive, and easy to verify, such as documentation, test cases, data cleanup, and first drafts. This keeps the agent inside its capability frontier while the team builds trust. Writing stories an agent can actually execute is a skill in its own right.
- Verify every output. A named Developer checks each result from start to finish, using the strategies above. No agent output reaches the Increment unchecked.
Phase 3: Keep humans in control
- Apply the same Definition of Done. Agent work is not done until it meets the same quality standard as human work. In this way, concrete checks replace vague distrust.
- Keep a human accountable for every item. The Developers decide which work goes to the agent, and a named Developer accepts each result. The agent contributes to the Increment, but people remain accountable for it.
Phase 4: Inspect and adapt
- Inspect results at every Sprint Review. Show what the agent delivered alongside the rest of the Increment. That way, stakeholders and skeptical team members see real output instead of imagining it.
- Adapt in the Sprint Retrospective. Ask what the agent did well, where it failed, and what work to give it next. Then expand its scope only as far as the evidence supports. This is calibrated trust, built Sprint by Sprint.
- Reinvest the time saved. Make it explicit that freed time goes to higher value work, such as design, customer conversations, complex problem solving, and learning. This step answers the job fear most directly, because it shows people moving up, not out.
Ultimately, nobody hands trust to an AI agent at the start. The agent earns it through verified work, the same way any new team member does.
How Each Accountability Helps
Overcoming fear takes the whole team, and each Scrum accountability plays a specific part.
| Accountability | What they do to reduce fear |
|---|---|
| Developers | Decide which work the agent takes, verify every output against the Definition of Done, and own the quality of the Increment |
| Product Owner | Orders the Product Backlog so agent ready items are clear and valuable, and communicates results honestly to stakeholders |
| Scrum Master | Makes the fear visible as an impediment, protects psychological safety, and coaches the team through the experiment |
| Organizational leaders | State clearly how the team will use freed time, avoid using the agent as a headcount argument, and support teams that report problems openly |
Psychological safety comes first
Psychological safety matters here more than any tool. Amy Edmondson’s research on team learning shows that people improve faster when they can admit mistakes and raise concerns without penalty. Consequently, a team that can say “the agent got this wrong” or “I’m worried about my role” will adopt AI faster and more safely than a team that cannot.
Conclusion: From Fear to Evidence
Fear of AI agents is real, common, and partly justified. AI can be wrong, work is changing, and accountability needs protecting. Even so, the research makes it just as clear that avoidance is not a strategy. Teams that bring an agent into a transparent process, verify every output, and keep people accountable for the outcome gain capacity and keep control.
The shift we recommend is simple to state. Stop asking whether you can trust AI, and start building the evidence that tells you where you can. Scrum gives every team the tools to do that, one Sprint at a time.
In the spirit of Aikido, open your hand to the agent, engage with it fully, and then make the right move. And as Dune reminds us, fear is the mind-killer. The teams that thrive will not resist this change, nor will they hand everything over blindly. Instead, they will learn to work with an agent as a real teammate: directing it, checking it, and growing sharper through the process.
If your team is ready to run that experiment, we are working with a small number of Scrum Teams to integrate an AI agent into their existing Jira and Scrum workflow. Contact us through ScrumAI.org to learn more.
Selected References
- Pew Research Center. U.S. Workers Are More Worried Than Hopeful About Future AI Use in the Workplace. February 2025.
- Gillespie N, Lockey S, Ward T, Macdade A, Hassed G. Trust, Attitudes and Use of Artificial Intelligence: A Global Study 2025. University of Melbourne and KPMG. 2025.
- Microsoft and LinkedIn. 2024 Work Trend Index Annual Report: AI at Work Is Here. Now Comes the Hard Part. May 2024.
- Stack Overflow. 2025 Developer Survey: AI. 2025.
- Mata v. Avianca, Inc., No. 22-cv-1461 (S.D.N.Y. June 22, 2023), order imposing sanctions.
- Parasuraman R, Riley V. Humans and automation: use, misuse, disuse, abuse. Human Factors. 1997;39(2):230 to 253. doi:10.1518/001872097778543886.
- Lee JD, See KA. Trust in automation: designing for appropriate reliance. Human Factors. 2004;46(1):50 to 80. doi:10.1518/hfes.46.1.50_30392.
- Dietvorst BJ, Simmons JP, Massey C. Algorithm aversion: people erroneously avoid algorithms after seeing them err. Journal of Experimental Psychology: General. 2015;144(1):114 to 126. doi:10.1037/xge0000033.
- Logg JM, Minson JA, Moore DA. Algorithm appreciation: people prefer algorithmic to human judgment. Organizational Behavior and Human Decision Processes. 2019;151:90 to 103. doi:10.1016/j.obhdp.2018.12.005.
- Dell’Acqua F, McFowland E, Mollick E, et al. Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality. Harvard Business School Working Paper 24-013. 2023.
- Brynjolfsson E, Li D, Raymond L. Generative AI at Work. The Quarterly Journal of Economics. 2025. doi:10.1093/qje/qjae044.
- Peng S, Kalliamvakou E, Cihon P, Demirer M. The Impact of AI on Developer Productivity: Evidence from GitHub Copilot. arXiv:2302.06590. 2023.
- METR. Measuring the Impact of Early 2025 AI on Experienced Open Source Developer Productivity. arXiv:2507.09089. July 2025.
- Challapally A, Pease C, Raskar R, Chari P. The GenAI Divide: State of AI in Business 2025. MIT NANDA. 2025.
- Kosmyna N, et al. Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task. MIT Media Lab. arXiv:2506.08872. 2025.
- Lee HP, Sarkar A, Tankelevitch L, et al. The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers. Proceedings of the CHI Conference on Human Factors in Computing Systems. 2025.
- World Economic Forum. The Future of Jobs Report 2025. January 2025.
- Brynjolfsson E, Chandar B, Chen R. Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence. Stanford Digital Economy Lab. August 2025.
- Edmondson A. Psychological safety and learning behavior in work teams. Administrative Science Quarterly. 1999;44(2):350 to 383. doi:10.2307/2666999.
- Herbert F. Dune. Chilton Books. 1965.
- Ueshiba M. The Art of Peace. Translated by John Stevens. Shambhala. 1992.
- Schwaber K, Sutherland J. The Scrum Guide. November 2020.