Most guidance on AI at work says the same thing. Do not trust it blindly. Check its work. Keep a human in the loop.
That advice is sound, but it is incomplete. It protects against one failure and says nothing about the other.
Failure one: following when you should not
The first failure is the familiar one. People accept AI advice that is wrong.
Research by Klingbeil, Grützner and Schreck, published in 2024, shows how strong this pull can be. People followed advice labelled as AI even when it contradicted the context. They followed it even when it contradicted their own view.
Read that again. The people in the study had information that pointed the other way. They had their own opinion. The label "AI" was enough to move them.
At work, this looks ordinary. A credit analyst in Kuala Lumpur accepts a risk summary that missed a late payment. A planner in Manila accepts a demand forecast that ignored a public holiday. Nobody was careless on purpose. The advice looked finished, so it was treated as finished.
Failure two: ignoring when you should not
The second failure gets less attention. People reject AI advice that is right.
This happens for understandable reasons. Some people distrust the tool after one bad answer. Some feel that accepting the AI makes their own role look smaller. Some prefer their usual method because they can defend it if things go wrong.
In September 2026, Harvard Business Review previewed an article by Das Narayandas of Harvard Business School. It argues that employees often override proven AI systems out of self-protection. The override is not about the evidence. It is about cover.
Both are failures of calibration
John Lee and Katrina See described this pattern long before generative AI. Their point was simple. Appropriate reliance depends on calibrated trust: trust that matches what the system can do. Over-trust leads to misuse. Distrust leads to disuse. Both are failures.
The word that matters is "matches". Good judgment with AI is not maximum trust or minimum trust. It is trust that moves with the evidence. You accept sound advice quickly. You check in proportion to the stakes. You challenge advice that does not fit what you know.
Why "always check" is not enough
"Always check" sounds like calibration, but it is not. If people check everything with the same effort, two things happen. They slow down on the many easy cases where the AI is right. Then, tired of checking, they start to skip checks, often on the hard cases.
Calibrated checking is uneven by design. It is light where the AI is strong and the stakes are low. It is heavy where the AI is weak or the stakes are high.
Questions for your team
- Where is our AI tool usually right? Do we still slow down there?
- Where is it often wrong? Do we check harder there?
- When did someone last overrule the AI and turn out to be wrong?
- When did someone last accept the AI and turn out to be wrong?
If your team can answer only the last question, you are watching one failure and missing the other.
Sources
- Klingbeil, Grützner and Schreck (2024). Computers in Human Behavior 160, 108352.
- Lee and See (2004). Human Factors 46(1):50–80.
- Ignatius, A. (25 September 2026). Why Your Employees Override AI. HBR Executive Agenda, previewing an article by Das Narayandas. HBR.
