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

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Genomic Medicine

Background:

• Prior authorization (PA) for exome/genome sequencing delays rare disease diagnosis.
• Large language model (LLM)-based browser agents could automate PA workflows.
• Clinical reliability of these agents for PA is uncharacterized.

Purpose of the Study:

• Evaluate LLM-based browser agents for automating prior authorization for exome/genome sequencing.
• Assess clinical reliability in terms of task completion, accuracy, and appropriate withholding of deficient profiles.

Main Methods:

• Developed a simulated payer portal and synthetic EHR with 836 patient records.
• Evaluated Gemini 3 Pro, Gemini 3 Flash, and Claude Opus 4.5 on PA submission tasks.
• Assessed task completion rate, form accuracy, and withholding of deficient profiles.

Main Results:

• Larger models (Gemini 3 Pro, Claude Opus 4.5) had high task completion but failed to withhold deficient profiles.
• Gemini 3 Flash showed better withholding (17.33%) but lower task completion (56.05%).
• Gemini 3 Pro identified 91% of issues in a non-agentic setting, suggesting interaction bias, not reasoning limits.

Conclusions:

• Current LLM browser agents have a bias towards form submission, posing risks in PA workflows.
• A modular, multi-agent architecture with human supervision is crucial for safe clinical deployment.
• Addressing systematic biases is essential for reliable automation of healthcare administrative tasks.