Hard to Halt: Automation Bias in Agent-Driven Sequencing Prior Authorization Workflows

The study examined the ability of large language models (Gemini 3 Pro, Gemini 3 Flash, and Claude Opus 4.5) to automate the prior authorization process for DNA sequencing tests. Researchers created a simulated environment with 836 patient records, some containing errors or deficiencies. Larger models achieved high task completion rates (Gemini 3 Pro 95.45%, Claude Opus 4.5 93.67%), but nearly all failed to withhold deficient requests. Gemini 3 Flash paradoxically showed better withholding performance (17.33%) despite lower completion rates (56.05%). When models were tested without web interface interaction, Gemini 3 Pro correctly identified 91% of issues, indicating that withholding failure is caused by form interaction rather than reasoning limitations. The research demonstrates a systematic bias of these agents toward form submission, which poses risks for clinical practice.