Intraoperative adverse events cause preventable harm to surgical patients worldwide [1]. Recent advances in artificial intelligence (AI) have enabled the development of real-time decision support systems that recognize instruments, delineate anatomy, classify surgical phases, and identify unsafe zones of dissection during ongoing procedures [2–4]. These technologies aim to augment surgeons' perception at critical moments, with particular promise for resource-limited settings where specialist support is unavailable [1, 5]. Despite impressive technical performance, translation into routine practice has lagged substantially. For most intraoperative AI tools, rigorous implementation studies remain sparse, and few have progressed beyond pilot testing to become evidence-based standards of care [6]. This disconnect between algorithm development and clinical deployment represents a critical barrier to realizing the patient safety benefits these technologies promise.

Successful deployment requires alignment with operating room (OR) workflows, human-factors-informed interface design, role-specific training, mechanisms for performance monitoring, and generalizability across sites with varying equipment and staffing [7–10]. While established implementation science frameworks provide valuable theoretical guidance [11–13], they operate at a level of abstraction requiring substantial interpretation for the unique demands of the OR: real-time decision-making under pressure, strict sterility requirements, complex multi-professional teams, and zero tolerance for failures that could compromise patient safety. Context-specific translation of these frameworks into actionable guidance for intraoperative AI remains an unmet need. The GoNoGoNet algorithm, a validated computer vision system for identifying safe versus unsafe zones of dissection during laparoscopic cholecystectomy, exemplifies this challenge [14]. Despite demonstrated potential to reduce bile duct injuries, GoNoGoNet’s clinical impact remains unrealized without effective implementation strategies [15].

To address the critical gap between intraoperative AI development and clinical deployment, this study aims to qualitatively identify barriers, facilitators, and implementation requirements for intraoperative AI decision support. By engaging diverse stakeholders, we developed the first implementation framework and actionable checklist for deploying AI in the OR, offering concrete guidance for future implementation efforts.

Phase I: We employed purposive maximum variation sampling to recruit OR personnel (surgeons, trainees, nurses, biomedical engineers) at one academic institution, with diverse experience levels, and attitudes toward AI adoption, supplemented by snowball sampling. Thematic saturation was operationalized as no new themes across three consecutive interviews [19]. Semi-structured interviews were conducted via Microsoft Teams (Microsoft Corporation, Redmond, WA) using an interview guide (Supplementary Material, Section 1), structured around CFIR domains. The interviewer followed the participant lead to explore emerging topics and perspectives. Interviews were audio-recorded, de-identified, transcribed verbatim, verified for accuracy.

Phase II: Patients were recruited through convenience sampling from a research volunteer database, regardless of prior surgical experience. Two focus groups with four participants each were conducted using semi-structured discussion guides (Supplementary Material, Section 2). Each 60-minute session began with standardized education about intraoperative AI, followed by moderated discussion of patient perspectives on benefits, risks, disclosure preferences, and trust considerations. Focus groups were audio recorded, de-identified, transcribed, and verified.

Identified themes were synthesized into a three-phase Systematic Methodology for AI Readiness and Translation in the Operating Room (SMART-OR) implementation framework with actionable checklists. The preliminary framework underwent member-checking, who confirmed relevance and suggested refinements that were incorporated to produce the final stakeholder-informed framework.

Phase I enrolled 20 OR personnel (91% recruitment rate) including 10 surgeons, 6 nurses, 2 surgical trainees, and 2 biomedical engineers to the study. Participants represented practice settings across Canada and the United States (Table 1). Phase II enrolled six patient participants (recruitment rate: 6%), comprising three males and three females who participated in two focus groups. In Phase I, each interview lasted a mean of 29 minutes (range, 18–45 minutes). Five major themes emerged from the analysis that defined requirements for implementing intraoperative AI decision support: intuitive design, adequate training, maximizing adaptability, ongoing support, and fostering buy-in. These themes span pre-implementation, implementation, and post-implementation phases and address technical, educational, organizational, and cultural dimensions of technology adoption (Fig. 1).

Surgeons, non-medical staff, and patients described a wariness toward the technology due to the potential for overreliance, leading to poor decision making. One surgeon questioned, “would you be falsely reassured by this technology that you're in the Go Zone despite not actually being there and then causing a bile duct injury?” Surgeons emphasized the importance of positioning AI as an adjunct to surgical judgment rather than a replacement. One surgeon described the importance of appropriate framing: "it's not telling you how to do an operation, and it's not making a judgment for you. It's an added layer of safety." Participants highlighted the importance of surgeons learning the algorithm's limitations and recognizing situations when clinical judgment should override AI recommendations.

Participants also expressed concern about cognitive overload from introducing additional technology into the OR. One surgeon stated, "the goal is not to distract the surgeon, because they still bear the primary responsibility for the patient’s outcome." Another surgeon noted that if the technology presents excessive clinically irrelevant information, "it's actually not helping and could be perceived as a distraction more often than not." Additionally, concerns about medicolegal implications of AI use were also raised. Views on liability varied, with some believing AI use would provide protection in litigation and others fearing it could increase liability in the event of adverse outcomes.

Participants emphasized that interface design and ease of setup directly influence technology adoption, with well-designed systems reducing training burden and facilitating integration into existing workflows. One surgeon stated, “if the integration is not smooth, barriers to adoption increase significantly.” A biomedical engineer explained: “if you can make it as easy as plug ‘A’ into ‘A’ and ‘B’ into ‘B’ and everything works, then everything will go smoothly.” Participants suggested that interface elements such as labeling, color-coding, and information density should reinforce appropriate use of AI as a decision-making adjunct rather than a directive authority. Participants also emphasized that achieving intuitive design requires iterative usability testing in both simulation and clinical environments. One surgeon stated, “usability studies analyzing the workflow, studying the environment and how these cases progress are essential to optimize product design and increase adoption.”

While AI for intraoperative decision support has advanced rapidly, translation into routine clinical practice remains limited. Existing reporting and evaluation frameworks include the IDEAL-Robotics framework (Idea, Development, Exploration, Assessment, Long-term study for surgical robotics) which provides stage-gated evaluation of innovation [21]; DECIDE-AI (Developmental and Exploratory Clinical Investigations of Decision-support systems driven by AI) which standardizes early-phase evaluation [22]; CPI-AI (Clinical Practice Integration for Artificial Intelligence) which emphasizes safe clinical integration [23]; and Consolidated Standards of Reporting Trials – Artificial Intelligence extension (CONSORT-AI) which extend reporting standards for AI-driven trials [24, 25]. Collectively, these frameworks clarify whether a surgical AI system works safely and effectively. However, they provide limited guidance on how to implement AI systems sustainably in complex, high-stakes surgical environments.

In contrast, the broader field of digital health implementation has matured over the past two decades, producing robust, theory-driven frameworks such as the CFIR, the Non-Adoption, Abandonment, Scale-up, Spread, and Sustainability (NASSS) framework [26], and the Expert Recommendations for Implementing Change (ERIC) taxonomy [13]. However, they require extensive interpretation for intraoperative use, where decision-making occurs in seconds, sterility constraints limit physical interaction with devices, and coordination among multidisciplinary teams is both dynamic and hierarchical. The SMART-OR framework directly closes the translational gap between high-level implementation science and the realities of intraoperative AI adoption. Its novelty lies in merging AI-specific safeguards (e.g., explainability, automation-bias mitigation, post-deployment drift monitoring) with OR constraints such as limited physical space, latency, and team coordination. A comparison with existing frameworks and their limitations is presented in Table 2. Additionally, the checklist is intended to be used as a practical guide when transitioning from research to intraoperative implementation.

The patterns we identified build on three decades of electronic medical record (EMR) implementation. Kaiser Permanente's successful $4 billion HealthConnect deployment demonstrated that phased rollout, physician leadership, extensive training, and workflow integration enabled large-scale health IT adoption [27]. Conversely, the United Kingdom's £10 billion National Programme for Information Technology (IT) failed due to top-down mandates, insufficient end-user engagement, one-size-fits-all design, and underestimated complexity, precisely the barriers our stakeholders identified as critical to avoid [28]. Three lessons apply directly. First, complexity across multiple domains predicts failure. Our identification of diverse user preferences signals inherent complexity requiring adaptable systems rather than standardized solutions. Second, physician engagement from inception is non-negotiable. Our emphasis on identifying surgeon champions early and framing AI as professional enhancement rather than remediation aligns with this principle [13, 29–31]. A consistent facilitator was the explicit positioning of intraoperative AI as an adjunct for surgical decision-making, reinforcing team trust. Third, training investment determines adoption. The consistent theme across stakeholder groups highlighting the importance of comprehensive training in technology uptake, mirrors findings from successful EMR implementations. These parallels suggest that surgical AI faces fundamentally similar organizational challenges to other health IT tools, with the additional complexity of real-time intraoperative use.

Our findings also highlight barriers to intraoperative AI adoption. Overreliance on AI emerged as a central concern. Surgeons identified increased risk through automation bias and deskilling unless design and training explicitly teach critical appraisal of AI outputs [32]. Unlike AI in radiology, intraoperative tools must deliver actionable, non-disruptive guidance in real time, imposing stringent requirements for latency, accuracy, and interface ergonomics [33]. Implementation must also account for tightly interdependent OR teams (surgeons, anesthesiologists, nurses, technicians) necessitating coordinated, role-specific training and clear accountability for AI-related issues. Finally, medico-legal ambiguity remains a barrier; aligning policy and guidance to recognize AI as a decision-support adjunct will be crucial to responsible adoption.

There are several limitations to this study. First, all participants were drawn from a single North American academic institution, which may limit the transferability of our findings to other practice settings. Future studies should include global stakeholders. Second, several sub-themes from this study require further exploration, including legislation governing intraoperative AI use, impacts on medical litigation, AI-introduced distractions, and optimal methods for usability testing. Lastly, purposive sampling may have introduced selection bias toward individuals with greater AI exposure, despite efforts to include both AI enthusiasts and skeptics.

This study describes the first stakeholder-informed SMART-OR implementation framework specifically designed to bridge the gap between development and implementation of intraoperative AI decision support. By systematically eliciting perspectives from surgeons, surgical trainees, nurses, biomedical engineers, and patients we identified five core implementation requirements and synthesized these into an actionable checklist. While validation through implementation is essential and ongoing, this formative work establishes a roadmap for researchers, clinicians, and healthcare organizations seeking to responsibly translate intraoperative AI from development to clinical practice.