JMIR Publications

ABSTRACT

The rapid evolution of Large Language Models has accelerated the development of agentic AI systems capable of pursuing autonomous goals, creating an urgent need for structural frameworks in psychiatry and psychotherapy. While existing classifications often draw parallels to autonomous driving, this paper argues that the mental health domain requires a distinct theoretical grounding due to specific differences in error tolerance, fat-tail risks (low-probability, high-consequence events), and the reliance on common therapeutic factors. To guide clinicians and researchers through these developments, we propose a five-step taxonomy for language-based AI systems, differentiating technical functionality from clinical effectiveness. The taxonomy progresses from Level 1 (Knowledge Level), where systems perform static benchmark tasks; to Level 2 (Elementary Level), characterized by dynamic engagement in specific therapeutic micro-skills; and Level 3 (Integration Level), where systems achieve case-level conceptualization suitable for blended therapy under human oversight. Level 4 (Saturation Level) describes human-in-the-loop systems capable of autonomous function with minimal supervision, while Level 5 (Mastery) represents fully autonomous agents that theoretically outperform human error rates across the entire care pathway. We conclude by discussing the necessity of shifting benchmarking from static knowledge tests to dynamic evaluations of therapeutic capabilities to safely navigate the transition toward autonomous care.