AI Prescribing Blog: Innovation, Risk, and the Need for Guardrails

Updated: March 26, 2026

What Do We Mean by AI Prescribing?

Artificial intelligence is rapidly moving from back-office support to direct participation in clinical decisions. Traditionally, AI in health care has been used for administrative tasks such as scheduling, documentation, or administering digital questionnaires. Increasingly, however, AI systems are being integrated into clinical decision-making, including evaluating symptoms, recommending treatments, and potentially capability to issue prescriptions.

"AI prescribing" refers to situations where an AI system participates in, or autonomously performs, decisions about medication prescribing. This can range from AI-generated recommendations that physicians review to systems that independently renew prescriptions for patients. Supporters argue that these tools could improve efficiency and help address physician shortages. The APA’s AI 2024 position statement emphasizes that these technologies should be understood as augmented intelligence that supports, rather than replaces, clinician judgment. But allowing automated systems to make medication decisions raises serious safety concerns. Prescribing medications is one of the most consequential actions in clinical care, and errors can have immediate and lasting consequences for patients. As AI moves closer to the point of care, it raises new policy and safety questions.

The Current Policy Landscape

The policy environment around AI prescribing is still evolving, with states beginning to experiment cautiously. A notable example is legislation and regulatory experimentation in Utah that allows certain AI systems to renew prescription medications under specific conditions. Currently, these programs are limited in scope. They focus on routine refills of medications that have already been prescribed and exclude drugs such as controlled substances. Critics, however, warn that even routine prescribing decisions like refills require clinical judgment and contextual understanding that may be difficult for automated systems to replicate.

Across the country, lawmakers, healthcare providers, patient advocates and technology developers are watching these early experiments closely. Some states are exploring similar approaches, while federal policymakers continue to debate how AI systems should be evaluated, approved, and monitored when they play a role in medical decision-making. Part of the debate is that there should be a limit on AI prescribing authority to carefully designed, regulated pilot programs with close oversight and rigorous safety evaluation. Part of the current debate is that there should be a limit on AI prescribing authority to carefully designed, regulated pilot programs with close oversight and rigorous safety evaluation.

Unique Concerns in Mental Health

AI prescribing raises concerns across medicine, but mental health presents unique challenges. Psychiatric prescribing often depends on nuanced assessments of mood, cognition, and social context. These factors can be difficult to capture through structured questionnaires or algorithmic decision trees. For example, assessing treatment progress in patients with mixed features of bipolar disorder, managing comorbid substance use and polypharmacy, and differentiating between bipolar disorder, borderline personality disorder, PTSD, or their overlap all require nuanced clinical judgment, longitudinal observation, and a strong therapeutic relationship, elements that are difficult for automated systems to replicate.

Mental health treatment also relies heavily on ongoing monitoring and therapeutic relationships. Medication decisions for conditions such as depression, anxiety, or bipolar disorder often depend on subtle changes in behavior, side effects, life circumstances, and patient preferences. These signals can be difficult for automated systems to interpret.

There are also concerns about bias and data limitations. If training data do not adequately represent diverse populations, AI prescribing tools could worsen disparities in mental health care. In addition, many patients seeking mental health treatment benefit from human interaction and trust, which is more difficult to replicate through automated systems. This emphasizes the need to include ongoing bias audits of any AI prescribing tool to mitigate potential bias in data or output.

The Path Forward: Innovation with Guardrails

AI prescribing is unlikely to disappear. The pressures driving its development are significant, including provider shortages, rising health care costs, and growing demand for services. The central policy challenge is not whether AI will be used in clinical care, but how it will be governed.

A balanced path forward should focus on innovation with guardrails. Policymakers can encourage responsible development while protecting patient safety through several steps:

• Requiring human oversight for higher-risk prescribing decisions
• Conducting independent clinical validation before widespread deployment
• Increasing transparency around training data and model performance
• Clarifying liability when AI systems contribute to medical errors
• Monitoring systems continuously after deployment in real-world settings

Carefully designed pilot programs can play an important role in this process. They allow policymakers and clinicians to test how AI tools function in practice while maintaining safeguards for patients. With thoughtful oversight, AI could help expand access to care and improve efficiency without displacing the clinical judgment that remains essential to good medicine.