Abstract The adoption of artificial intelligence (AI) into clinical practice is accelerating, outpacing the development of organizational guidance, training, and governance. A recent study indicated that two-thirds of hospitalists are using AI, particularly large language model (LLM)–based platforms, in their clinical work. However, as with prior disruptive health technologies, adoption alone does not ensure meaningful improvement in care. Drawing on lessons from electronic health record implementation, we argue that AI’s ultimate impact will be determined not by use rates, but by implementation quality and fit. Poorly implemented digital tools have been shown to increase clinician workload and burnout, despite their intended benefits. Early evidence on LLM-based diagnostic AI further underscores this risk: clinical-decision making supported by AI may be suboptimal when integration, training, and workflow design are inadequate. To provide value, AI tools must be thoughtfully embedded into clinical reasoning processes through evidence-informed training, intentional workflow design, and supportive organizational culture. As AI technologies are rapidly adopted, three priorities come into focus: training clinicians on AI inputs and interpreting outputs, applying implementation science frameworks for AI deployment in clinical environments, and establishing strategies for ongoing evaluation of the impact of AI tools over time. Implementation science frameworks offer practical guidance to assess workflow integration, training needs, infrastructure, and potential unintended consequences that can then inform adaptation of implementation strategies to enhance contextual fit. In parallel, learning health system infrastructure can enable continuous monitoring and iterative adaptation using routinely collected clinical and workflow data that reflect the value of the intervention across the quintuple aim of clinical outcomes, health equity, cost, and patient and clinician experience. AI adoption in hospital medicine is likely inevitable. Its ability to advance the quintuple aim will depend on how effectively these tools are implemented, supported, evaluated, and adapted in practice.
Maw et al. (Mon,) studied this question.