Why Employees Resist AI Tools: The Personal Costs That Keep Employees From Using AI

Two in 3 office professionals at companies with at least $500 million in annual revenue have used an AI tool at work that they believed their employer did not allow, according to a 2026 survey of 1,250 workers run by Wakefield Research for PagerDuty. Employees resist AI tools, or hide their use of them, when using them openly carries personal costs: uncertainty about what is allowed, a reputational penalty for visible use, lost trust after a visible error, and friction with how they already work. More training helps with some of this, but it cannot remove these costs on its own.

At The Decision Lab, we see these barriers throughout our AI adoption work. A Fortune 500 HR technology company we worked with had deployed AI tools across its 20,000-person workforce, backed by training and internal communications. Employees knew the tools existed, and adoption still stayed low because the main obstacles were behavioral. We now measure those obstacles with our SPROUT framework, validated with data from more than 20,000 employees.

This guide covers four barriers, three of which also appeared in our Fortune 500 diagnostic. The exception is the reputational penalty for visible AI use, and it comes from peer-reviewed research. The diagnostic also found two further barriers, time constraints and perceived complexity, which are covered in our guide to behavioral barriers to enterprise AI adoption.

Most resistance to workplace AI shows up as silence about using it rather than outright refusal. In the PagerDuty survey, close to 4 in 10 respondents said they would rather keep their AI use to themselves than disclose it and risk being told to stop, and that share rose to nearly half at companies above $1 billion in revenue. A separate 1 in 3 said they would hide AI use to avoid scrutiny from managers.

Unclear usage guidelines were one of the barriers our Fortune 500 diagnostic identified. A rollout plan should track how much of its adoption is visible, since reported usage already misses the people who will never disclose it.

A social evaluation penalty is a drop in how capable or motivated colleagues judge someone to be, based on the method they used rather than the quality of their results.

Jessica Reif, Richard Larrick, and Jack Soll, at Duke University's Fuqua School of Business, documented one for AI in the Proceedings of the National Academy of Sciences in May 2025. Across four preregistered experiments with 4,439 participants, people who used AI at work both expected and received negative evaluations of their competence and motivation, and the penalty carried into assessments of job candidates. The researchers also found that evaluators who rarely use AI themselves were the ones who saw AI users as lazy, and that the penalty shrank for tasks where AI was clearly useful.

We treat this as the barrier training is least able to reach, because a well-trained employee faces the same reputational risk as an untrained one.

Algorithm aversion is the tendency to stop relying on an automated tool after seeing it make an error, even when it remains more accurate than the human alternative.

Berkeley Dietvorst, Joseph Simmons, and Cade Massey established the effect in the Journal of Experimental Psychology: General in 2015. Across five studies, participants who watched an algorithm make forecasts were less confident in it than participants who had not seen it perform. They were also less likely to choose it over a less accurate human forecaster. This held even for people who had watched the algorithm outperform the human.

Low trust in AI accuracy was one of the barriers in our Fortune 500 diagnostic. A skeptical segment of employees distrusted AI's value and stalled at first contact, so an employee's first weeks with a new tool carry disproportionate weight. A team that hides errors during a pilot turns a small credibility problem into a larger one at full rollout.

Workflow integration was the fourth barrier our diagnostic found, and it shaped the programmes we built for the client. They ranged from hands-on tasks embedded in real workflows to a mission-based challenge series run in Slack.

Fit also depends on control. In a 2018 follow-up in Management Science, Dietvorst, Simmons, and Massey found that people were considerably more likely to use an imperfect algorithm when they could modify its forecasts, even when the permitted changes were severely restricted. Most enterprise deployments remove editability in the name of consistency, but it's a small price to pay for meaningful adoption.

Our client had already delivered training and internal communications before we started, and usage stayed low. Its workforce also split into two groups with opposite failure modes. Enthusiasts were already experimenting but kept running into friction, while skeptics stalled at first contact. A single training track cannot serve both groups at once.

We designed 20 candidate programmes, each aimed at a specific barrier for a specific segment, and piloted eight of them for one month with more than 100 employees against a control group, using before-and-after measures of ease and confidence. Organisation-wide AI adoption rose by about a third, and use of AI in employees' core responsibilities rose by nearly a quarter. Confidence in exploring AI tools rose by about two-fifths in the pilot group and fell by about a quarter in the control group.

These are results from a single proprietary pilot at one company. The published case study does not report baseline adoption rates or statistical significance, so the figures show what moved in that organisation rather than a benchmark for others.

• State in performance criteria that AI-assisted work is judged on output quality, and publish plain guidance on which tools are permitted. Repeat an anonymous disclosure question after 90 days and check whether the share of people who would stay quiet has fallen.
• Ship approved tools with an accept-and-edit step at the point of output, and compare weekly active use against a team given a locked result.
• Demonstrate the tool on a task it fails at, in front of the team, before the wider rollout. Compare 30-day retained use against a team shown only a clean demonstration.
• Have managers document their own AI use before asking anyone else to, then compare disclosure rates among their reports against teams whose manager said nothing.

The people doing the hiding and the people doing the judging are the same population. Reif and colleagues captured both halves in the same experiments, with participants expecting a penalty for their own AI use and then imposing it on someone else's. A team could agree unanimously to stop marking colleagues down for AI use and still do it the following week, because the judgment forms before anyone gets a chance to decide anything.

Why do employees resist AI tools?

Employees resist AI tools when open use costs them something personally. The most common costs are unclear rules about what is permitted, colleagues judging AI users as less capable, trust that drops after a visible mistake, and tools that sit outside existing workflows. Many employees respond by using AI quietly rather than refusing it.

Is lack of training the main reason employees do not use AI?

Usually it is not. Our Fortune 500 client had delivered training and internal communications, and employees knew the tools existed, yet usage stayed low. Training helps employees who find AI complicated, but it leaves the reputational and trust barriers in place.

How can leaders build trust in workplace AI?

Let employees adjust AI output rather than accept or reject it whole, since research on algorithm aversion shows even limited control increases use. Show the tool's limits early instead of hiding errors during a pilot, and design a separate first experience for skeptical employees.

How can companies encourage employees to disclose AI use?

Publish clear guidance on permitted tools, judge AI-assisted work on output quality, and have managers disclose their own use first. An anonymous question about willingness to disclose, repeated after 90 days, shows whether those changes are working.

What should organizations do when employees fear AI will replace them?

Measure the fear before responding to it. Fear of replacement is one of the factors in the Psychological Readiness pillar of our SPROUT framework. The free AI Adoption Diagnostic scores it anonymously at team level, so leaders can see how widespread the concern is before deciding how to address it.