Summary  

Most organizations now rely on AI to inform credit, hiring, or investment decisions. Yet many employees quietly wonder, “Should I trust the algorithm or my own judgment?”  

Our research published in the Journal of Management Studies followed that question across six longitudinal studies in corporate credit-rating tasks. We found that across three decision sessions, a clear pattern emerged: with experience, people’s initial estimates grew more similar to the AI’s suggestions and overall accuracy improved. But trust was guarded—anchoring to one’s own first impression rarely disappeared. 

The takeaway? Decision makers don’t instantly trust AI; they learn to calibrate between human intuition and machine guidance. For organizations, designing AI tools that explain their reasoning and highlight learning over time can bridge that trust gap and unlock real performance gains. 

The Experiment in a nutshell 

Participants worked with two types of AI: one purely data-driven (“neutral AI”) and another that learned from users’ own patterns (“user-resembling AI”). Over 30 decision rounds, they made initial credit estimates, received AI advice, and choose whether to adjust. Sometimes, they also received extra information — like a prior year’s credit score. This design let us observe how people’s reliance on AI evolved as they gained experience. 

What actually happened as users gained experience 

Early in the process, participants treated the AI as a second opinion. Many viewed the algorithm’s advice, nodded, and changed their rating only slightly. The data confirm this: when the gap between a person’s first estimate and the AI’s advice was large, they were more likely to make an adjustment—but their final decision still leaned toward their own original view. 

Over time, however, familiarity worked quietly in the background. As people saw feedback on how accurate they and the AI were, their first guesses began to inch closer to the AI’s predictions. Session-level mean accuracy increased steadily, and relative dependence on AI advice rose. The trajectory suggests that repeated exposure and performance feedback help users internalize how the AI “thinks.” 

Still, total deference never occurred. Most participants positioned their answer somewhere between human and machine—evidence of what psychologists call anchoring. Once an anchor is set, people adjust, but rarely enough. 

When more information changes the picture 

We next introduced a second anchor: last year’s credit score. This extra cue reduced initial similarity between user and AI estimates overall. Yet something interesting happened when that prior-year figure happened to be close to the AI’s recommendation. Under those circumstances—what the study defines as higher relative proximity to AI—participants were much more likely to follow the algorithm. 

So, extra data did not automatically increase reliance; it helped only when it pointed in the same direction as the AI. When the two sources conflicted, confidence fractured, and people reverted to intuition. In organizational life, this means that simply piling on additional metrics may confuse users rather than reassure them. 

Why reliance matters 

Across all models, the relationship between AI reliance and decision accuracy was consistently positive. Participants who adjusted further toward the AI achieved smaller prediction errors. Even partial trust improved performance. Supplementary studies confirmed the robustness of these findings. When a human rather than an AI offered advice, the same biases appeared—people still anchored to their first estimates—but the egocentric tendency was slightly stronger with human advisors. Replications using full-time employees produced broadly similar results, suggesting that professional experience helps, but does not eliminate, anchoring and selective trust. 

Practical takeaways for organizations 

1. Build transparent feedback loops. To foster effective human-AI interaction, organizations must build transparent feedback loops. Our research suggests that providing real-time, comparative dashboards enables employees to accurately calibrate their confidence. A good example is a logistics coordinator who monitors a dashboard that contrasts her delivery predictions with the AI’s and the actual outcomes. This data allows her to systematically learn which contexts require overriding the system (like weather issues) and which call for full reliance. 

2. Make reasoning visible. Enhancing the visibility of reasoning processes enables users to place greater trust in the systems they utilize. Providing transparency transforms opaque recommendations into informed dialogues by elucidating the primary factors influencing an AI’s suggestions—such as critical variables or outcomes of analogous previous cases. For example, a loan officer receives an AI rejection and the system instantly displays the top three weighting factors: “High Debt-to-Income Ratio (40%),” “Recent Delinquency on Auto Loan,” and “Less than 1 Year at Current Employer.” This visibility empowers the officer to ask targeted questions during the follow-up call. 

3. Manage multiple anchors. Employees often face several cues—AI predictions, historical data, and peer opinions. You should highlight agreement among sources, since this immediate convergence builds essential trust. For example, if the AI, maintenance logs, and a peer comment all signal machine failure, confidence is high. Conversely, when indicators diverge—such as an AI failure warning combined with normal log readings—organizations must prompt a structured discussion rather than letting users quietly default to intuition.  

4. Train for calibration, not blind faith. Our results show that trust in AI evolves through repeated, feedback-rich experience. Therefore, managers must design training that lets people experience both when AI excels and when human judgment adds value. Training for customer service agents, for example, should include simulated high-stakes calls. In some simulations, the AI’s script is perfect for a standard refund issue (building confidence); in others, the AI fails spectacularly on a complex, emotional escalation, demonstrating the critical value of the agent’s empathy and unique override.