Organizations are rapidly embedding AI into decisions that were once considered the domain of human experts—from hiring and lending to healthcare and public administration. But what is often overlooked is what happens at the employee level, when workers are expected to communicate, justify, and defend the AI-generated decisions they did not make and often do not fully understand.

Our research explores that under-examined aspect of AI implementation. Drawing on a multi-year field study across banking, recruitment, and biotechnology, we find that employees rarely share AI outputs with customers, managers, or colleagues as-is. Instead, they reinterpret, reshape, and sometimes even hide them to protect their professional role and credibility. How they respond depends largely on who they are accountable to and whom they must convince.

Understanding these responses is critical because they shape whether AI decisions become accepted, contested, or quietly undermined inside organizations. Our research identifies diverging ways employees respond when they are asked to explain AI-generated decisions and points toward better ways companies can navigate the risks.

When Experts Mask AI

In a large German bank we followed over six years (2019 to 2025), loan officers faced a predictive AI system that fully automated loan approval decisions. Loan officers could not override the system’s decisions, yet they remained responsible for communicating and justifying those decisions to customers.

To help them do so, the AI system generated basic, technical explanations, typically a handful of bullet points with rules for a rejection. Yet many loan officers struggled to make sense of these explanations as they strongly departed from their expert standards. For example, the system might state an “unstable financial situation” even when a customer had no open loans and a steady income.

Loan officers were hesitant to openly admit their confusion as they feared this would undermine their expert role customers expected from them. They decided to mask the role of AI and hold onto familiar expert standards and terminology for explaining loan decisions even though decisions were no longer grounded in them.

For example, officers would cite common expert standards in banking such as a weak credit history or stricter income thresholds due to inflation, as reasons for loan rejections, even when they did not reflect the AI’s actual rules. They hoped that customers would find those explanations plausible and continue to place trust in loan officers.

But masking came at a cost. We observed how customers often left loan consultations confused when loan officers’ explanations did not match their individual financial situation. In conversations, several customers reported that they noticed officers’ confusion and uncertainty when asking follow-up questions, doubting whether they truly understood the decision and deserved their services. Consequently, some customers decided to move to a competitor bank. Rather than strengthening trust in AI decisions, masking created distance between the organization, its employees, and its customers.

When Experts Embrace AI

A very different response emerged at a global consumer goods company we studied between 2018 and 2022. The company’s internal recruiters enthusiastically adopted a new AI-based selection tool that predicted job candidates’ suitability by identifying patterns in personality and performance data from past employees.

Read the full article here.