Staff Do Not Trust AI. What Do You Do About It?
Last Updated: September 2026
A team where staff do not trust AI is usually reacting to one event, not to AI as a whole. Someone saw a tool get a client detail wrong. Someone lost a task they were good at, or was handed the job of checking output they never agreed to own. That memory sits under the shrugs and the unused logins.
AI Smart Ventures has guided growing businesses through AI adoption for more than a decade, and the same scene repeats. Leaders bring a plan to win people over, while the team brings a list of things that went wrong. The second list is the more useful document, and it is shorter than anyone expects.
Skip that list and you pay twice. You fund another training day that answers a question no one asked, while the original defect stays in the workflow, generating the next reason to say no.
Key Takeaways
- Distrust is a finding, not a mood: ask when the tool last let someone down, and you get a short list of defects that sit inside one or two workflows.
- The complaint is written down now: AI mentions in US staff reviews rose 240% in a year, and the positive share fell from 81% in 2019 to 43%, Glassdoor data shows.
- One error you can see does the damage: across five studies, people lost confidence in a model faster than in a person making the same mistake.
- The loudest no sits on serious work: a 2023 study found 49.3% turned down a stronger model when the stakes were grave, against 29.2% when they were slight.
- Trust points at you, not the tool: only 27% of workers fully trust their employer to use AI well, in a poll of more than 1,000 working adults.
- Give the doubter the edit: people take up an imperfect model far more readily when they can change its output, so the sceptic is your best reviewer.
Read together, they turn the usual order around. Persuasion comes first, the asking never comes at all, and the thing that caused the distrust outlives every workshop you run. Flip it, and the workshop has something real to teach.
What is staff distrust of AI actually telling you?
It tells you where a tool lost touch with the actual work. Distrust clings to an event: a figure that went out wrong, a summary that dropped the one line that mattered, a new process that removed a step someone was proud of. Ask three people and you get three incidents, and they usually sit inside the same two workflows rather than across the whole business.
The mood can be counted now, which is new. Glassdoor data reported in September 2026 shows AI mentions in US staff reviews rose 240% in a year. The positive share fell from 81% in 2019 to 43%, and more than half were negative. Those reviews name things: fear of replacement was the top theme, in 20% of critical comments, and 41% of the positive ones named real help such as tools or training.
Why does one bad AI answer stick so hard?
Because people drop a machine faster than they drop a colleague for the same mistake. Berkeley Dietvorst, Joseph Simmons, and Cade Massey called this algorithm aversion in five studies. People who watched a model forecast lost confidence in it faster than in a human making the same errors, then tied their money to the weaker human. Seeing it work was the trigger.
The reaction lands hardest where the stakes are real. In a 2023 study of 143 people, 49.3% chose a human expert over a stronger model in grave scenarios such as medical scans. Only 29.2% did so in light ones such as recipes. So the loudest no in your business usually sits on your riskiest workflow, which reads as a map of the step that needs a person on the output.
How do you find the incident behind the distrust?
Ask about the last time, never about the general view. “What do you think of AI?” gets you a stance, while “When did the tool last let you down, and what were you working on?” gets you a date, a file, and a name. Then ask what they did instead, because the workaround shows what the tool failed to do. Ten minutes each, one person at a time.
Write their words down and ask to see the file. A draft with the wrong figure still in it beats any account of the meeting, and it often shows the fault was smaller than the story around it. Two people will describe one incident from two angles, which is how you spot the one that spread. Note the date as well, since a March case no one closed is a different problem from last week’s.
| Ask this | What it surfaces |
|---|---|
| When did it last let you down? | A date and a real file |
| What were you working on? | The workflow that carries the risk |
| What did you do instead? | The gap the tool never filled |
| What would have to change? | The fix, in their words |
Which complaints are accurate, and which are not?
Most are right about the event and too broad about the tool. The wrong client name was real, but “AI cannot be trusted with client work” is the part that ran past the evidence. Keep the first half and retire the second, which you can only do once the first half is on paper. Sort what you gather into four kinds, because the fix differs in each case, and only some of them involve skills at all.
Sorting also protects the people who spoke up. A complaint that gets named and answered reads as respect, while one met with a pep talk reads as change management done to them. That difference decides whether anyone tells you about the next incident, which matters more than any single repair. The fault you never hear about is the one that shapes how your team quietly works around the tools all year.
- A defect: the tool produced something wrong that reached real work, so add a check at that step or take it off that task.
- A missing check: no one agreed who reads the output, so whoever noticed became the unpaid last line.
- Lost work: the rollout took a task someone valued, which is a job design question rather than a tool question.
- A borrowed story: it happened at an old employer or in the news, so show what your own setup does differently.
What do you fix once you have found it?
Fix the narrow thing, then say so out loud. If the summary tool mangled figures, it stops touching numbers and a person owns that column. If no one was checking, name the checker in the workflow rather than in a policy nobody reads. Then go back to the person who raised it, in their own words, because trust returns through repair that people can see.
Then hand over the edit. In a 2018 study, the same three researchers found people were considerably more likely to use an imperfect model when they could change its output, and they performed better as a result. The pull held even when the change allowed was tiny. So give the doubter authority over the last step, because whoever found the error is often the sharpest reader you have, and the role is real work rather than a gesture.
AI Smart Ventures offers AI Training built on the workflows your team already runs, so a session starts from the defect you found and not a tour of tools. Ask for practical AI workshops tied to the step that broke.
When is more AI training the wrong answer?
When the finding is a defect, not a skill gap. The stock answer to a doubtful team is a campaign: build buy-in, beat the resistance, appoint champions. That treats an accurate report as a mood to be managed, and it fails for a plain reason, because nothing about the tool changed between the complaint and the workshop. People are not holding back enthusiasm; they are holding evidence.
Training earns its place once you know which finding you have, and never before it. A missing check needs a written review standard that somebody then has to be taught. Lost work needs a redesigned role, and a borrowed story needs a walk through your own setup. Tool-first AI vendors sell one course into all of them, which is why a second day of AI enablement lands flat. AI Smart Ventures observes that teams recover fastest when one named defect is closed first.
Frequently Asked Questions
Are employees sabotaging AI?
Real sabotage is rare. What looks like it is usually quiet non-use: opening the tool for a demo, then doing the job the old way. Motive matters more than the act. Fear of replacement was the top complaint in Glassdoor’s 2026 review data, appearing in 20% of critical AI comments, and people guarding a job rarely offer improvements. Treat avoidance as a report, then check what it is about.
Is there a reason to not trust AI?
Yes, and it is narrow rather than general. Any model can give a confident answer that is wrong in a way only someone who knows the subject will catch. So trust belongs to a task and not to a tool. That is why refusal lands hardest on serious work: close to half the people in one 2023 study picked a human expert there. Match the checking to what the output touches.
What causes lack of trust in the workplace?
Broken expectations that no one names afterwards. Something goes wrong, the people hit by it notice, and nobody explains what happened or what changed. AI adds a second layer, because staff judge the employer as much as the software. In an SHL poll of more than 1,000 working adults, only 27% fully trust their employer to use AI well. Silence after a bad week sets that.
How do you find out what an AI tool got wrong?
Start from the file, not the memory. Get the draft, email or sheet, then find what went in: the source, the request, the date, and which version of the tool was running. Compare the output with what a competent person would have produced from the same inputs. Nine times in ten the fault is small, such as a missing source or an unwritten rule, and it points at one step.
Should you tell the team when an AI tool fails?
Yes, and quickly. A failure that gets named, fixed and reported back reads as competence, while one absorbed in silence reads as a cover-up. Keep it factual and short: what happened, what it affected, and what changed as a result. Two or three of those, handled in the open, do more for AI adoption than any launch note, because people see that speaking up produces a change.
Can someone opt out of a tool they think is unsafe?
Treat the objection as evidence before you treat it as a request. Ask what they think the tool would get wrong and on which task, then route that task away from it while you check. Often they are right about one workflow and wrong about the other five. Set a review date, so the exception stays open to change rather than becoming permanent by drift.
How long does it take to rebuild trust in a tool?
Faster than most leaders expect, once one named defect is closed. A visible fix on the workflow that caused the complaint tends to move behaviour within a few weeks, because people test it themselves rather than take your word for it. Trust rebuilt through messages alone takes far longer, and often never arrives. The unit of repair is one incident, one owner, one change.
How do we start after an AI rollout went badly?
Run the conversations first, before you buy anything else. Speak to six people, gather the incidents, sort them into defects, missing checks, lost work and borrowed stories, then fix the one that shows up most. Scope drives what comes next, and it shrinks sharply when you start from named faults instead of a broad relaunch. Schedule a consultation to work through what your team reported.
Executive Summary
Staff distrust of AI is usually right about one event and too broad about the tool, so ask before you persuade. Find out when the tool last let each person down, what they were working on, and what they did instead. Get the file rather than the story. Sort the findings into defects, missing checks, lost work and borrowed stories, since only some of those are skill gaps. Fix one named defect, tell the team what changed, and give the checking step to whoever found the error. AI upskilling built on your own workflows makes it hold.
What Should You Do Next?
Book six short conversations this week with the people using the tool least. Ask about the last time it let them down, rather than what they think of AI. Write each incident down with a date and the file it produced. Then pick the defect that comes up most and close it in the open.
AI Smart Ventures offers AI Training for growing businesses that want capability building tied to their own work, with clarity and confidence. Schedule a consultation to turn what your team reported into a workable AI strategy.
People Also Read
- AI Vendor RFPs for Owner-Operators: A Practical Template
- AI for the GM in an Owner-Operated Business: A Daily Rhythm
About the Author
Nicole A. Donnelly is the Founder of AI Smart Ventures and an AI Adoption Specialist with 20 years of experience as a founder and CEO and over a decade leading AI adoption initiatives. She helps businesses integrate artificial intelligence with clarity and confidence, driving innovation and sustainable growth. Nicole has trained over 20,217 professionals in Applied AI, delivered 624 workshops, and worked with close to 1,000 organizations across diverse industries.
Expertise: AI Transformation, AI Strategy, AI Implementation, AI Adoption, Applied AI, Marketing, Business Operations
Disclaimer: This content is for informational purposes only and does not constitute professional business or technology advice. Results vary based on industry, existing systems and implementation commitment. Contact AI Smart Ventures for a consultation regarding your specific situation.
