Ask, Audit, Apply: A New Rulebook for Doctors Thinking with AI

Artificial intelligence has quietly slipped into the doctor’s black bag. Tools powered by large language models now sit inside electronic health records, generating differential diagnoses, summarizing evidence, and proposing management plans in fluent, confident prose. But unlike the stethoscope, which medical students learn to use under careful supervision, these AI-enabled clinical decision support systems have arrived with remarkably little guidance on how clinicians should actually reason with them. A new perspective published in the Journal of General Internal Medicine argues that this gap is not a minor oversight but a genuine safety hazard, and it proposes a deceptively simple three-step framework—Ask, Audit, Apply—to fill it.

The paper, authored by Andrew Jenkins of Brigham and Women’s Hospital, Katherine Eisenberg of the University of Rochester Medical Center and Dyna AI, and Roy Ziegelstein of Johns Hopkins University School of Medicine, is a conceptual rather than empirical work. That distinction matters, and the authors are candid about it: the framework has not yet been validated in clinical trials or workflow studies. Yet its premise is grounded in well-established research. Diagnostic error remains a leading cause of patient harm, frequently driven by cognitive biases such as anchoring and premature closure. Decision support systems, decades of human factors research show, introduce their own hazard: automation bias, the tendency of users to over-rely on automated output and reduce independent verification. Systematic reviews in health informatics indicate that automation bias can meaningfully increase both the frequency and severity of errors, particularly when recommendations appear authoritative.

Existing frameworks, the authors contend, were built for a world that no longer exists. Evidence-based medicine has long emphasized integrating the best available evidence with clinical expertise and patient values, and tools like PICO sharpen question formulation while GRADE structures the rating of evidence certainty. But these systems predate a setting in which AI synthesizes heterogeneous sources and presents recommendations directly, often obscuring the distinction between rigorously curated evidence and unvetted text. Meanwhile, AI governance frameworks from the World Health Organization and the U.S. National Institute of Standards and Technology articulate transparency, accountability, and human oversight at the organizational or policy level—principles that rarely translate into explicit cognitive behaviors a clinician can enact at the bedside. Ask–Audit–Apply is designed to bridge exactly that gap, operating not as an institutional oversight model but as a decision-level reasoning loop for individual clinicians.

The first step, Ask, may sound trivial but is anything but. It requires deliberate specification of the clinical question, the relevant patient context, and the purpose of the AI interaction before the tool is engaged. Ambiguous prompts, the authors note, produce overly broad or clinically misaligned outputs, and when a response fails to answer the query, careful evaluation of the prompt is the first corrective step. More profoundly, explicit question formulation functions as a cognitive forcing strategy—a technique drawn from the clinical reasoning literature that helps counteract anchoring and premature closure, the twin biases behind many diagnostic errors. In this framework, asking is not mere preparation; it is an accountable act that shapes the downstream validity of everything the AI produces.

Audit is the step the authors identify as the framework’s distinguishing move. Clinicians must first determine whether an AI response is purely informational or contains an actual recommendation, then interrogate the evidentiary basis, recommendation strength, and transparency of what the system presents. The technical details matter here. The authors urge users to look for statements about the methods behind a response—whether the tool uses retrieval-augmented generation, for instance, and whether it draws exclusively from evidence-based clinical content or more broadly from non-curated material. They also stress that AI-generated responses are not individually reviewed by human experts before presentation, a fact that shifts the burden of appraisal squarely onto the clinician. When information about evidentiary basis is absent, the framework demands heavier reliance on independent verification and external sources.

Crucially, Audit is framed not as reflexive distrust of AI but as calibrated trust. Fluent, well-structured explanations can inflate perceived trustworthiness, yet research on explainability shows that a plausible rationale does not guarantee correctness or appropriateness. Clinicians must actively discriminate among levels of evidence and recognize when AI output extends beyond validated indications or the relevant patient population. The authors single out trainees and early-career clinicians as the group for whom this discipline is most critical, since they may lack the experience needed to independently calibrate the reliability of AI outputs against the accumulated pattern recognition of seasoned practice.

The final step, Apply, delivers the framework’s sharpest message: AI does not make decisions—clinicians do. After weighing AI input, the clinician integrates it with the patient’s values, competing risks, local context, and clinical judgment, commits to a final plan, and documents the reasoning in their own words. This step operationalizes accountability and directly counters the diffusion of responsibility that occurs when decisions are perceived as AI-driven. It aligns with governance frameworks that insist on human responsibility for AI-assisted decisions, making explicit that liability for clinical outcomes cannot be delegated to an algorithm, even when its recommendations are followed. The authors also frame a positive vision: by delivering trustworthy, evidence-based information rapidly at the point of care, AI could free clinicians to know their patients as individuals and tailor recommendations accordingly.

Why does this matter now? Because the failure modes AAA targets are neither hypothetical nor distant. Diagnostic error already harms patients at scale, automation bias is among the best-documented risks of decision support technology, and generative AI’s fluency amplifies both. The authors argue that by making verification and evidence discrimination explicit, the framework helps preserve critical appraisal skills and may mitigate the de-skilling—or as they put it, the never-skilling—associated with over-reliance on AI output. When paired with deliberate training, AI-enabled decision support could even strengthen clinical reasoning, helping learners identify gaps in their thinking and refine their illness scripts rather than atrophy them.

For medical education, the framework converts the abstract aspiration of AI competence into observable behaviors that can be taught, reflected upon, and eventually assessed. Rather than teaching students what AI is, AAA emphasizes how clinicians must reason when AI is present, complementing supervisory guidance by clarifying what supervisors should expect of learners who interact with these tools. The authors recommend that trainees maintain a low threshold for tracing citations and consulting additional sources while they develop the experience to calibrate trust in AI-generated outputs. At the institutional level, the framework aligns with patient safety priorities embedded in accreditation standards for undergraduate and graduate medical education, including expectations around learning environments, supervision, and accountability.

The authors are careful to acknowledge the limits of their proposal. As a conceptual framework, AAA awaits empirical validation: future research must test whether it can be operationalized in real clinical workflows and whether it actually improves verification behavior, trust calibration, and downstream safety outcomes. But they note that conceptual clarity is a necessary precursor to meaningful evaluation, and that many foundational frameworks in medical education and clinical reasoning were articulated long before formal validation caught up. As AI systems increasingly participate in clinical reasoning, the paper concludes, medicine needs more than technical safeguards and ethical principles—it needs a clear articulation of how clinicians should think and act when AI enters the decision. Ask–Audit–Apply offers exactly that: a workflow-embedded loop that turns the vague mandate of human oversight into three concrete obligations, one question, one audit, and one owned decision at a time.

Subject of Research: A clinician-facing framework for safe clinical reasoning with AI-enabled decision support tools

Article Title: Ask–Audit–Apply: A Framework for Clinical Reasoning with Artificial Intelligence

Article References: Jenkins, A., Eisenberg, K., & Ziegelstein, R. C. (2026). Ask–Audit–Apply: A Framework for Clinical Reasoning with Artificial Intelligence. Journal of General Internal Medicine. https://doi.org/10.1007/s11606-026-10741-8

Image Credits: AI Generated

DOI: 10.1007/s11606-026-10741-8

Keywords: artificial intelligence, clinical decision support, clinical reasoning, diagnostic error, automation bias, evidence-based medicine, medical education, AI governance, patient safety, large language models, clinical training, physician accountability