People Don't Fear AI. They Fear Losing Agency.

The customer-support chatbot may make a promise your company then has to honor. An efficient assistant may become evidence that someone’s job is no longer needed. The search result, fluent and certain, may invent an answer. AI products can infer, generate, recommend, remember, and act.

Once software can do those things on behalf of a person or an institution, trust stops being a usability issue and becomes a question of authority. KPMG and the University of Melbourne found that two-thirds of people use AI day to day, and most expect it to help society. Yet only 23% trust these systems and 70% want them regulated.[1]

The common fear is that AI is shrinking us: making workers easier to replace, turning expertise into supervision, automating meaningful interactions, and reducing our role to just correcting whatever AI produces.

Product managers should take this fear seriously. Treating AI resistance as a communications problem leads to shallow fixes: better launch copy, more AI literacy, a disclosure banner, another warning that the system may be wrong.

These fixes are insufficient when the user is asking a harder question: does this product expand my agency, or does it take agency away from me?

The fear of job displacement

U.S. workers are already more worried than hopeful about the future of AI in the workplace.[2] Gallup found that nearly a quarter of U.S. employees at organizations using AI believe their job could disappear within five years.[3] A 2026 Reuters/Ipsos poll found that half of Americans fear AI could put someone in their household out of work.[4]

Executive rhetoric makes the fear worse when it frames AI mostly as labor reduction: cut headcount, automate support, do more with less. That message may appeal to investors, but it teaches employees and customers to see the product as a substitute for human judgment rather than a tool that strengthens it.

Klarna framed AI as replacement. IKEA framed it as redeployment.

Klarna announced that its AI assistant was doing the work of 700 customer-service agents, then faced backlash and had to reassure customers that they could still reach a person.[5] The concern was that the company seemed to be replacing meaningful human contact with AI, all while pushing people out of work.

As IKEA's AI assistant Billie handled routine customer queries, the company trained 8,500 call-centre workers to become interior-design advisers rather than making the story about headcount reduction. Its AI customer service assistant handled 47% of call-centre customer queries over two years, while sales through Ingka’s remote interior-design channel accounted for €1.3 billion in sales.

In IKEA's case, automation helped convert support capacity into a revenue-producing channel by moving workers toward higher-value customer advice, sales, and design judgment.[6]

That is the product standard AI teams should aim for. Do not only ask what the system can automate. Ask what new human work it makes possible.

The best AI products should leave behind stronger human roles, not thinner ones.

The fear of lost agency

AI pushes users to confront questions of control and consent that older software rarely made urgent.

Every time an AI system enters a workflow, users start testing the boundaries, often before they can fully name what bothers them.

• Did I really consent to the system being involved at this point?
• What information did the AI access, process, or store about me and my work?
• When something fails, can I recover or get meaningful help from a human?

More than half of U.S. adults want more control over how AI is used in their lives. Yet only 14% of the public say they have a great deal or quite a bit of control over whether AI is used in their lives.[7] People are not only worried that AI will make mistakes. They are worried that it will absorb their data, alter their work, and shape decisions about them without a clear moment of consent or a reliable path to regain control.

Microsoft Recall showed how quickly this fear surfaces. The intended benefit was obvious: help users find what they had seen on their PC. But users spotted the risk immediately. The feature depended on screenshots of their activity, which made the product feel less like memory and more like observation.

After the backlash, Microsoft changed the control model: Recall became opt-in, snapshots stayed local, authentication was required, and users could pause, delete, filter, or review what Recall kept.[8]

That is the product lesson. Designing for agency has to happen before the build. Product teams should ask where users want AI to automate, where they want AI to assist, where human judgment must remain, and what controls need to be visible at the moment AI enters the workflow.[9]

The hardest fear to see: losing purpose

Some repetitive tasks deserve automation. Others, though repetitive, remain meaningful because they rely on judgment, craft, care, or the chance to exercise taste and identity. Product teams can often erase this line and treat all repetition as the same kind of drudgery.

A 2026 paper, Are We Automating the Joy Out of Work?, found that AI is most likely to touch the tasks that contribute to a sense of agency and happiness.[10] The default instinct is to make the assistant more capable, more fluent, and more agentic. That is not always what users want. They want the system to stay useful, bounded, and easy to overrule.

Relieving people of true busywork—formatting, transcription—tends to land well. As AI moves beyond that, into roles that require judgment or taste or authorship, resistance grows.

Users rarely call it lost purpose. They say the product feels generic, pushy, or irrelevant, then work around it or use it with resentment.

Then comes the hidden labor. Glean's Work AI Index 2026 reported digital workers now spend 6.4 hours a week "botsitting": feeding the AI more context, fixing outputs, rerunning prompts, cleaning up. 87% say AI saves them time, but only 13% say it makes a real difference for their organization.[11]

AI removes friction in one place and creates supervision work somewhere else.

The user becomes editor, auditor, prompt engineer, exception handler, and the one left holding the blame. Ignore that hidden labor and you misstate productivity gains. The ROI case weakens when invisible work piles up.

The fear of missing accountability

AI failure unnerves because the mistake arrives dressed as authority.
Traditional software bugs give themselves away—a crash, a blank field, a visible error state. A generative AI mistake lands as a finished response. Polished, self-assured, and often wrong.

Interface design can help—the product can surface uncertainty, mark contradictions, and show what is speculative or fabricated.[12] But a well-placed tooltip or label does not shift the verification burden off the user. “AI can make mistakes” is a thin shield if the product still presents answers as final.

When Air Canada’s chatbot gave a customer false information about bereavement fares, the company tried distancing itself—arguing the customer should have checked elsewhere and the chatbot was accountable for its advice. The tribunal disagreed. Air Canada, not its chatbot, was held responsible for what the customer was told.[13]

Cheap disclaimers are starting to fail the moment AI speaks with company authority.

A generic warning that “the AI may be wrong” does not show what is certain, what is inferred, where doubt remains, or what decisions turn on that uncertainty. It does not give recourse or make clear who bears responsibility. Too often, it asks the user to shoulder the risk while the company keeps the automation upside.
A trustworthy AI product must go further: show its evidence, declare its changes, surface uncertainties, and offer users a real path to recover or appeal when the AI is wrong.

What people really want from AI

People are not asking AI products to feel harmless. They are asking for control, proof, and recourse when the stakes are real.

• Control means the user can decide when AI enters the workflow, what it can see, what it can remember, what it can change, and when it must stop.
• Proof means the product shows the evidence behind a summary, recommendation, or action before asking the user to trust it.
• Recourse means the user can undo, correct, appeal, escalate, or recover when the AI gets something wrong.

A caution about control. The reflex to answer every fear with one more toggle, setting, or consent screen usually backfires. A wall of permissions is its own tax, and it quietly signals that the team does not trust its own product. Most people never open the settings panel; they live on the defaults. The defaults are the real trust decision.

A disclosure banner and a stack of opt-outs are cheap. A sane default, a reversible action, and a company that owns the liability when the system is wrong are expensive. And yet, they are the ones who actually earn trust.

Turning trust into product behavior

Before the feature moves into build, a product team should be able to state what the user might reasonably fear, whose work changes if the feature succeeds, which parts of the workflow still require human judgment, and where the user can step in once the AI is acting.

A trusted AI system keeps human authority close to the parts of the work where judgment, relationship, identity, and accountability matter.

Product teams should not confuse adoption with trust. High usage can come from defaults, lock-in, or the absence of real alternatives, but not earned confidence. A spike in usage or workflow coverage says little about whether users believe the AI, understand its limits, or feel protected from its errors.

Trust is not measured by engagement alone. It shows up in signals a team can instrument: how often users reverse or undo what the AI did, how long it takes to reach a human after an escalation, what share of appeals get resolved, and how much of a wrong action the product can reverse.

If a team wants to earn trust, it should not stop at driving usage. It should watch reversals, escalations, appeals, unresolved errors, and recovery paths, then make those numbers part of the release decision.

The real test of trust is not whether the AI works. It is whether the product protects the user when it does not.

References

[1]: KPMG and the University of Melbourne, "Trust, Attitudes and Use of Artificial Intelligence: A Global Study 2025." https://kpmg.com/xx/en/our-insights/ai-and-technology/trust-attitudes-and-use-of-ai.html

[2]: Pew Research Center, "On Future AI Use in Workplace, U.S. Workers More Worried Than Hopeful," February 25, 2025. https://www.pewresearch.org/social-trends/2025/02/25/u-s-workers-are-more-worried-than-hopeful-about-future-ai-use-in-the-workplace/

[3]: Gallup, "Rising AI Adoption Spurs Workforce Changes," April 2026. https://www.gallup.com/workplace/704225/rising-adoption-spurs-workforce-changes.aspx

[4]: Reuters, "Half of Americans Fear AI Could Put Someone in Their Household Out of Work, Reuters/Ipsos Poll Finds," June 10, 2026. https://www.reuters.com/business/world-at-work/half-americans-fear-ai-could-put-someone-their-household-out-work-reutersipsos-2026-06-10/

[5]: Customer Experience Dive, "Klarna Changes Its AI Tune and Again Recruits Humans for Customer Service," May 9, 2025. https://www.customerexperiencedive.com/news/klarna-reinvests-human-talent-customer-service-AI-chatbot/747586/

[6]: Helen Reid, "IKEA Bets on Remote Interior Design as AI Changes Sales Strategy," Reuters, June 13, 2023. https://www.reuters.com/technology/ikea-bets-remote-interior-design-ai-changes-sales-strategy-2023-06-13/

[7]: Pew Research Center, "Artificial Intelligence in Daily Life: Views and Experiences," April 3, 2025. https://www.pewresearch.org/internet/2025/04/03/artificial-intelligence-in-daily-life-views-and-experiences/

[8]: Microsoft Support, "Privacy and Control Over Your Recall Experience." https://support.microsoft.com/en-us/windows/privacy-and-control-over-your-recall-experience-d404f672-7647-41e5-886c-a3c59680af15

[9]: Yijia Shao, Humishka Zope, Yucheng Jiang, Jiaxin Pei, David Nguyen, Erik Brynjolfsson, and Diyi Yang, "Future of Work with AI Agents: Auditing Automation and Augmentation Potential Across the U.S. Workforce," arXiv, 2025. https://arxiv.org/abs/2506.06576

[10]: Jaspreet Ranjit, Ke Zhou, Swabha Swayamdipta, and Daniele Quercia, "Are We Automating the Joy Out of Work? Designing AI to Augment Work, Not Meaning," arXiv, 2026. https://arxiv.org/abs/2603.14963

[11]: Glean Work AI Institute, "The Work AI Index 2026." https://www.glean.com/work-ai-institute/reports/work-ai-index-report

[12]: Nielsen Norman Group, "AI Hallucinations: What Designers Need to Know," February 7, 2025. https://www.nngroup.com/articles/ai-hallucinations/

[13]: The Guardian, "Air Canada Ordered to Pay Customer Who Was Misled by Airline's Chatbot," February 16, 2024. https://www.theguardian.com/world/2024/feb/16/air-canada-chatbot-lawsuit