What happens when people have to defend a decision they do not understand
Harvard Business Review published research this week that is ostensibly about AI. I think it is about something older and more useful than that.
Researchers followed a German bank for six years. The bank had automated its loan approvals. Loan officers could no longer override the system, but they were still the ones sitting across the table from customers, explaining why an application had been rejected.
The system gave them a few bullet points. Often those bullet points made no sense against what the officer could see in front of them. A customer with steady income and no outstanding debt would be flagged for an unstable financial situation.
The officers did not say they were confused. They were afraid that admitting confusion would cost them the expert status their customers expected. So they invented explanations instead. Weak credit history. Tighter income thresholds because of inflation. Reasons that sounded like the kind of thing a banking expert would say, and that had nothing to do with why the decision had actually been made.
Customers noticed. Not the invented reasons, but the uncertainty underneath them. Some of them left for other banks.
The pattern underneath
Strip out the technology and what remains is a pattern I see constantly.
Someone is held responsible for something they do not fully understand. Saying so feels professionally dangerous. So they perform understanding instead. And the performance is worse than the admission would have been, because everyone in the room can feel that something is off even if they cannot name what.
This shows up in leadership teams far more often than anyone admits.
The executive who nods through a strategy presentation full of terms they could not define if asked. The manager who cascades a decision from above using language they have not examined. The team lead who is asked why the roadmap changed and produces a confident answer that they do not actually believe.
None of these people are being dishonest. They are protecting a role. The role requires competence, competence is assumed to mean having answers, and so not having an answer becomes something to conceal rather than something to say.
Why the concealment costs more than the confusion
The bank's loan officers were trying to preserve trust. What they actually did was destroy it.
Customers could not articulate what was wrong, but they registered the mismatch between the confident delivery and the hesitation underneath. They asked follow-up questions and watched the officer struggle. They concluded, reasonably, that this person did not know what they were talking about.
The same thing happens in teams. People are remarkably good at detecting when someone is performing certainty. They usually cannot point to the specific tell, but they come away with a sense that something does not add up. And what they do with that sense is quietly adjust. They stop asking real questions. They stop expecting real answers. The conversation moves to the corridor.
The cost is not that one person did not understand something. The cost is that the team learns understanding is not the point. Appearing to understand is the point. Once that norm is set, it applies to everyone, about everything.
What the research found actually works
The most interesting case in the study was a biotechnology company where the response was completely different.
Seed experts there also lost their decision-making role to an AI system. But they were given access to the underlying data, time to examine it, and regular contact with the people building the system. When they did not understand a classification, they could investigate rather than improvise.
Over time they built genuinely new expertise. They found visual patterns the model was detecting that they had never used themselves. They learned from supply chain managers which explanations were actually useful downstream. Their role shifted from making decisions to making decisions meaningful.
What made this possible was not better technology. The company built the conditions: a lab next to the sorting floor, weekly sessions with developers, funded research when experts spotted something the model had never seen. Not knowing was treated as the start of an investigation rather than a failure to be hidden.
The bank had none of that. Its officers had a screen, a customer waiting, and no route to understanding. Given those conditions, inventing an explanation was the rational choice.
The question this leaves for leadership teams
The researchers make a point worth sitting with. Employees in all three organisations faced systems that were hard to interpret. What differed was not the technology but whether the organisation made it possible to say so.
The question for any leadership team is the same, with or without AI in the picture.
When someone does not understand a decision they are expected to carry, what happens? Is there a route to understanding it, or only pressure to represent it convincingly?
And more uncomfortably: when was the last time someone on your team said, in the room, that they did not understand something? If you cannot remember, that is not evidence that everyone understands everything. It is evidence about what your team believes it is safe to say.
The bank's loan officers were not weak. They read their environment accurately and behaved accordingly. That is what people do. The behaviour was a symptom. The environment was the cause.
Which is where the work is.
The research by Anne-Sophie Mayer, Elmira van den Broek and Tomislav Karacic was published in Harvard Business Review in July 2026.
Mees Loman is the founder of Loman Leadership, a leadership coaching practice for founders and leadership teams of fast-growing companies in Amsterdam and beyond. lomanleadership.com