Artificial intelligence in prior authorization and coverage decisions: A systematic review of methods, evidence gaps, and future implications for patient access.

BACKGROUND: Prior authorization (PA) is intended to support appropriate use and spending of services and medications, yet 1 in 6 insured adults report PA-related problems linked to delayed care, worse access, and higher financial burden. As artificial intelligence (AI) expands in utilization management and other pharmacy benefit operations, evidence guiding responsible AI use in PA remains fragmented. OBJECTIVE: To systematically map AI applications across PA workflow and evaluate model development, validation, equity, and implementation outcomes relevant to managed care pharmacy and patient access, while identifying AI adoption priorities. METHODS: A systematic review was conducted in Embase, PubMed, Scopus, and Google Scholar from October 1, 1988, to January 27, 2026. Eligible studies applied AI (rule-based systems, classical machine learning, deep learning, large language models, or hybrid approaches) to real-world PA workflows. Data were extracted on workflow stage (initiation, submission, payer review, appeals, patient engagement), model type, dataset characteristics, outcomes, and bias assessment. Risk of bias for prediction models was evaluated using PROBAST-AI. RESULTS: Of 3,417 records, 16 studies met eligibility criteria. AI use was concentrated in payer review/decision-making (62.5%), followed by post-denial appeals (18.8%), initial PA submission (12.5%), and early PA initiation (6.3%), with limited focus on patient engagement. Supervised classical machine learning predominated (37.5%), followed by hybrid multimodal models (31.3%) and deep learning (25.0%). Administrative PA datasets (25.0%) and synthetic datasets (18.8%) were the most common training data. Most studies reported technical metrics (F1, precision, accuracy, recall, area under the receiver operating characteristic curve), while few demonstrated operational gains, including faster decisions, lower denial rates, and reduced cost. Most studies had high analysis bias, limited generalizability, inadequate validation, and limited subgroup reporting. Only 1 study assessed demographic bias and found significant disparities. Patient- and provider-centered outcomes and long-term system impact were rarely evaluated. CONCLUSIONS: This systematic review synthesized AI literature into a PA workflow-based map and identified critical gaps in equity, stakeholder engagement, and validated real-world outcomes. Although AI demonstrates acceptable technical performance and some operational gains, its impact on medication access, financial burden, and trust remains uncertain. Managed care pharmacy should adopt a risk-tiered hybrid AI-human model spanning workforce augmentation, workflow automation, and business process innovation, with human oversight for complex decisions. This should be guided by governance, vendor accountability, transparency, subgroup monitoring, external validation, and workforce training. Future research should quantify the value of AI-enabled PA by evaluating return on investment and its impact on appropriate, equitable, patient-centered access while reducing administrative burden.