Abstract
Artificial intelligence (AI) tools have moved from novelty to measurable clinical impact in surgical practice. This review organizes the literature and practical guidance around five problems surgeons face: documentation burden, information overload, administrative accumulation, practice intelligence, and implementation. Each section follows a consistent structure: the problem, the evidence for addressing it, the type of tool that fits, and what failure looks like. The review draws on the best available evidence across study designs and addresses compliance, implementation, and measurement throughout. The paper’s central argument is that AI’s near-term role in surgery is not to replace surgical judgment, but to reduce the cognitive and administrative overhead that surrounds it. Clinical judgment, technical skill, and operative intuition remain irreducibly human. A final section addresses the educational implications of training surgeons in an AI-enabled environment.
Keywords
Introduction
The American surgeon operates in an environment of compounding cognitive demand. Administrative work now rivals clinical care in time and effort. Burnout, driven primarily by documentation burden and bureaucratic overhead, affects 41.9% of physicians overall as of 2025 and remains above the national average in surgical subspecialties. 1 Among trauma surgeons, a 2025 meta-analysis of 19 studies reported pooled burnout prevalence of 60%. 2 Simultaneously, the AI landscape has transformed. Eighty-one percent of physicians now report using AI in practice, and the FDA has authorized more than 1247 AI-enabled medical devices as of mid-2025. 3 The evidence base has matured substantially in the past 18 months: randomized controlled trials now confirm that ambient AI scribes reduce documentation time and burnout, and a new category of agentic AI tools capable of executing multi-step tasks autonomously is beginning to reshape administrative workflows. Nevertheless, most surgeons lack a systematic framework for deciding where to start and how to measure progress. What follows is a practical field report, written by trainees at multiple stages alongside an attending surgeon, organized around the problems surgeons face and how AI can address them. Each section follows a consistent structure: the problem (with supporting evidence), the type of tool that fits, and what failure looks like. We conclude with summative guidance on implementation and the educational implications for a generation of residents training in an AI-enabled environment.
We must acknowledge specific limitations in scope: the AI tool landscape evolves rapidly, and this review reflects evidence and platform availability as of May 2026. Inclusion of specific platforms does not constitute endorsement. Institutional compliance policies supersede all vendor claims. Tool capabilities, pricing, and HIPAA status may have changed since this writing. Evidence quality also varies substantially across the tools discussed. Some have undergone prospective external validation, while others rest on vendor-reported figures alone, and these distinctions are clarified throughout.
The Non-Negotiable Foundation: Safety and Compliance
One rule governs every AI interaction in clinical practice: never enter protected health information (PHI) into a consumer-tier AI tool. Free consumer-tier AI tools carry no Business Associate Agreement (BAA), and HIPAA violations carry fines up to $1.5 million per violation category. A 2026 JAMA viewpoint on responsible AI use in medicine identified five requirements: data access policies consistent with HIPAA and state law; patient disclosure when AI affects care; encryption in transit and at rest; regular security audits; and an incident response plan. 4 All major healthcare-specific ambient AI scribes carry BAAs and are HIPAA-compliant. Enterprise-tier versions of general-purpose large language models are HIPAA-eligible through signed BAAs, but consumer-tier versions of the same products are not. Agentic tools capable of multi-step task automation have a less developed compliance landscape. HIPAA-compatible options exist but require enterprise licensing, and fewer products currently meet that bar. The practical framework: institutional tools first, enterprise-tier tools with BAAs second, consumer tools only for workflows that never touch patient data. Regardless of the tool, AI output is always a draft. The surgeon verifies, edits, and signs.
Problem 1: Documentation Burden
Clinical documentation now rivals direct patient care in the time it demands from many physicians. Primary care physicians spend 84 minutes daily on electronic inbox management, including 22 minutes outside working hours. 5 Surgeons face a different but overlapping burden: high volumes of prior authorization requests, referral coordination, scheduling correspondence, and equipment vendor communications. Documentation burden is the single most consistent predictor of burnout across specialties, and the downstream costs of surgeon turnover are substantial enough that even modest improvements justify significant institutional investment. 6
What the Evidence Supports
Ambient AI scribes, which listen to natural clinical conversation and generate structured notes within 90 seconds, now have the strongest evidence base of any AI application in clinical practice. Two randomized controlled trials published simultaneously in NEJM AI in November 2025 established the foundation. Lukac et al at UCLA Health randomized 238 outpatient physicians across 14 specialties to DAX Copilot, Nabla, or usual care over approximately 72 000 encounters; Nabla reduced time-in-note by 9.5% (P = .02). 7 Afshar et al at UW Health conducted a 24-week stepped-wedge trial of Abridge across 66 practitioners and 71 487 notes, demonstrating a 0.44-point reduction in work exhaustion on the Stanford Professional Fulfillment Index and approximately 22 fewer minutes of daily documentation. 8 Beyond the trials, Kaiser Permanente’s real-world deployment across 7260 physicians and 2 576 627 encounters estimated 15 791 documentation hours saved, with 84% reporting positive experiences. 9 A multicenter quality-improvement study across six U.S. health systems found burnout prevalence fell from 51.9% to 38.8% after 30 days of ambient AI scribe use. 10
The only published surgical-specific evidence comes from Ghanem et al at Cooper University, who evaluated DAX Copilot across 25 simulated inpatient surgical scenarios. Note quality was acceptable (modified PDQI-9: 46.91/50), and the authors concluded the tool may reduce surgical resident documentation burden. 11 The existing evidence base remains predominantly ambulatory and primary care, and additional surgical-specific trials are needed.
Practical Guidance and Failure Modes
For private practice surgeons, several clinical-grade ambient scribes offer HIPAA-compliant, low-setup options, including free tiers tied to existing clinical platforms. Academic surgeons with an enterprise EHR should explore whether AI documentation tools are already available natively within that system before adding a third-party subscription. The most capable ambient scribes have recently expanded beyond note generation to produce lab orders, referral letters, and billing codes, representing an early shift from digital scribe toward autonomous workflow agent.
The most common failure mode is adoption without verification. Ambient scribes occasionally generate plausible but incorrect information such as wrong medication doses, inaccurate procedure descriptions, and hallucinated patient complaints. The correct model is that the AI produces a high-quality first draft requiring critical review before the note enters the chart. The second failure mode is more subtle; abandoning the tool before it works. Most ambient scribes require two to four weeks to calibrate to a clinician’s vocabulary and note style, and the first week is rarely representative of what the tool becomes.
Problem 2: Information Overload
The surgeon seeking current evidence on prophylactic mesh, perioperative anticoagulation, or enhanced recovery pathways faces a literature landscape that grows faster than any individual can track. Given competing demands and current publication rates, surgeons lack time to meaningfully keep pace with relevant literature. The consequences are structural: even when high-quality evidence exists, it reaches surgical practice slowly and incompletely. U.S. surgical society guidelines carry a measurable publication-to-guideline delay, and practice often fails to change in the direction trials indicate even after guidelines are updated.12,13 AI-assisted tools for evidence synthesis address this gap—not by replacing clinical judgment, but by compressing the time between a clinical question and a reliable, evidence-based answer.
What the Evidence Supports
At the point-of-care level, AI-powered clinical decision support tools trained specifically on medical literature can provide rapid, cited summaries in response to clinical questions, offering a meaningful improvement over general-purpose search. For more structured literature work, AI-assisted systematic review tools can search tens of millions of academic papers and extract trial data into structured tables. A 2025 study in Social Science Computer Review evaluating one such tool found its data extraction accuracy approaching that of human reviewers (81.4% vs 86.7%). 14
For patient-level clinical decision support tools, validated risk calculators represent one of the most established AI applications in surgery. The POTTER calculator estimates mortality and complication risk in emergency surgery patients. A prospective bi-institutional validation in 361 emergency laparotomy patients demonstrated a mortality c-statistic of 0.90. 15 MySurgeryRisk (University of Florida) predicts eight postoperative complications with AUCs of 0.82 to 0.94, outperforming unaided physician estimates. 16 A 2025 Nature Medicine study by Rosen et al implemented an AI prediction model prospectively at a single center after validation in 18 403 patients from Danish national registries, finding that AI-guided perioperative pathways were associated with a reduction in major complications from 28.0% to 19.1%. 17 Of 1247 FDA-authorized AI devices through mid-2025, radiology accounts for 77% while general and plastic surgery hold fewer than 1%, a gap that signals both the immaturity of surgical AI and its growth trajectory. 18
Limitations and Failure Modes
A cardinal risk in AI-assisted evidence synthesis is hallucination—the generation of confident, plausible answers that are factually incorrect. Tools that return verifiable citations are more reliable than those that summarize without attribution. Exercise caution when using AI-generated clinical summaries for decisions where evidence quality, including study design, sample size, and external validity, matters. When evaluating any AI decision support tool, ask whether performance metrics derive from internal or independent external datasets, whether the validation population resembles your patient population, and whether the tool has been studied prospectively. The evidence that these tools perform well in practice remains substantially thinner than the evidence that they perform well on paper.
Problem 3: Administrative Accumulation
Prior authorization alone consumes an average of 40 requests per physician per week, with physicians and their staff spending 13 hours weekly on PA paperwork; 40% of physicians have staff working exclusively on prior authorizations. More than one in four physicians report that prior authorization has led to a serious adverse event for a patient in their care, and 95% report that it delays access to necessary care. 19 Beyond prior authorization, surgical practice generates persistent administrative accumulation: inbox triage, referral coordination, scheduling correspondence, meeting follow-through, and committee work. These tasks require accurate prose but rarely require the surgeon’s clinical judgment, making them ideal targets for AI assistance.
Prior Authorization: Enterprise-Scale Automation
AI automation of the prior authorization lifecycle has moved beyond pilot programs. MUSC Health has deployed AI agents to execute 40% of prior authorizations without human involvement, reducing a process that previously took 30 minutes of manual work to approximately one minute. 20 The regulatory environment is evolving simultaneously. CMS’s WISeR program, launched in January 2026 across six states, applies AI-assisted prior authorization to Medicare services such as skin and tissue substitutions, nerve stimulator implants, and knee arthroscopy, with coverage decisions expected within 72 hours. 21 However, the program has drawn significant resistance. Physician groups and members of Congress have raised concerns about provider workload and patient access and multiple states have enacted or proposed legislation prohibiting AI as the sole basis for a medical necessity denial.22,23 AI tools that assist with submission, such as extracting clinical documentation or pre-filling payer forms, represent sound investments for high-PA-volume practices. Tools that make autonomous denial decisions require careful scrutiny of your state’s regulatory environment.
Inbox Burden and Meeting Follow-Through
Emails and meetings are where surgical administrative work can easily accumulate and go unresolved. For email and inbox workflows involving patient-related communications, only HIPAA-compliant tools with BAAs are appropriate. Before seeking standalone tools, surgeons should check whether their institution’s existing enterprise software agreements already include HIPAA-eligible AI assistance for inbox and administrative workflows. Where institutional coverage is absent, purpose-built AI administrative platforms with BAAs are available but still warrant careful institutional vetting.
Surgeons spend substantial time in non-clinical meetings including department conferences, committee work, research coordination, and faculty meetings where decisions are made and action items are assigned, but follow-through depends entirely on individual memory. AI meeting tools address this directly: they generate structured summaries, action items, and follow-up drafts from recorded discussions, turning a conversation into a documented, actionable record.
Knowing When to Automate
Not all administrative automation carries the same risk. Workflows involving structured data, clear rules, and reversible actions are safe to automate with confidence. Workflows that require judgment applied to ambiguous inputs or trigger irreversible actions warrant more caution. Drafting a prior authorization narrative from structured EHR data is durable. Autonomously sending an email based on AI interpretation of a lab value is fragile. Agentic tools should be confined to durable workflows until reliability in your specific context has been verified through deliberate testing with reversible tasks. The most common failure mode in administrative automation is scope creep: a workflow that begins as a reliable draft-and-review process gradually becomes a send-without-review process. A second failure is building automation on top of a broken manual workflow: if the prior authorization process is chaotic before automation, AI will replicate that chaos at scale. Fix the workflow first, then automate it.
Problem 4: Practice Intelligence
Every surgeon accumulates years of scheduling records, OR utilization logs, and revenue cycle data, yet most of it sits unexplored in EHR exports and billing spreadsheets. These insights are not visible by default: which procedures have the largest gaps between scheduled and actual durations, whether first-case on-time starts have improved, and which block days have consistently low utilization. Answering these questions previously required hiring an analyst or purchasing an enterprise business intelligence platform.
Operational Metrics Worth Tracking
Key OR metrics to monitor monthly include block utilization, first-case on-time start rate, turnover time, and case duration accuracy. However, the challenge is rarely knowing what to measure; it is making the data accessible enough to act on. AI-powered data analysis platforms now accept uploaded spreadsheets and respond to plain-English queries, returning statistical analysis, visualizations, and written summaries without requiring any coding ability. Patient identifiers must be removed before uploading to any such tool, and BAA availability should be confirmed before using scheduling or operational data with any patient-adjacent information.
AI enables surgeons to develop custom analytics without any formal programming training. Large language models can convert plain-language descriptions into functional R or Python code executable in cloud-based environments like RStudio or Google Colab without any local installation required. This approach, often called “vibe coding,” is an iterative process in which users describe desired functionality, AI generates code, and outputs are refined through feedback. The two most common pitfalls are code that yields incorrect output, and code that yields correct output through incorrect calculations. The latter can be significantly harder to identify. Thus, for greatest fidelity it is critical that users are able to explain the rationale behind any generated analysis. For publication-quality analyses, generated code requires review by a statistician or senior researcher. In short, never trust and always verify.
Failure Modes
The most common failure in practice intelligence is collecting data without acting on it. Uploading OR utilization data and generating a visualization is not the same as using it. The analysis is only useful if it changes a decision. A second failure is over-interpreting small samples: a single quarter of OR data is rarely enough to draw reliable conclusions about block utilization patterns. Track trends over at least six months before making operational decisions based on AI-assisted analysis.
Problem 5: Implementation
The limiting factor in AI adoption is not capability but implementation. Tools fail because they are deployed without measurement, used without adjustment, or abandoned before enough exposure has occurred for the tool to calibrate to the specific workflow. The pattern of failed AI adoption closely resembles the pattern of failed surgical skill development: insufficient deliberate practice, inadequate feedback, and premature abandonment after initial difficulty.
A Tiered Approach to Adoption
Measurement and Failing Gracefully
The AMA’s 2025 AI Governance Toolkit provides an eight-step institutional framework including model AI policies, vendor evaluation criteria, and CME-eligible training resources. 24 Before deploying any tool, establish a baseline measurement of the target task: time yourself completing it three times under the current workflow, then measure again after two weeks of AI assistance. If you cannot demonstrate time savings or quality improvement in a specific, measurable task, that tool is not the right fit. The American College of Surgeons Informatics and AI Committee and SAGES are both actively engaged with AI integration in surgical practice, though formal position statements on the administrative and workflow applications described in this review remain in development as of this writing.
Plan explicitly for tool failure. Maintain the manual workflow capability in parallel until the AI tool has been verified over a sufficient number of cases. Keep your existing dictation system active during the ambient scribe pilot. Review AI-generated notes against your own mental summary before signing. Spot-check PA submissions for completeness during the first month of automation. When a tool is not working after deliberate testing, switching is the right call.
Implications for Surgical Education and Training
Residents and fellows entering training today will graduate into a practice environment where AI-assisted documentation, evidence synthesis, and workflow automation are standard infrastructure. A 2025 scoping review mapping AI applications in U.S. surgical training found expanding use of AI for simulation-based skill assessment and adaptive feedback, but identified a significant gap in guidance on developing AI literacy as a clinical competency. 25 A 2026 review in The American Surgeon identified three domains where AI is reshaping surgical education: technical skill acquisition through AI-guided simulation, competency assessment through automated performance metrics, and the growing administrative burden on residents that AI documentation tools are beginning to address. 26
Insufficient AI exposure leaves trainees unprepared for a practice environment where these tools are standard. Excessive reliance during training introduces a subtler risk—what Ke et al term “never-skilling”—the failure to develop foundational clinical reasoning competencies in the first place, not through loss of previously acquired skills, but through failure to acquire them at all. 27 A trainee who accepts ambient scribe notes without review, or uses LLM-generated summaries without verifying citations, may accumulate experience hours without the cognitive investment those hours are designed to produce. The risk is not that trainees use these tools, but that they use them without doing the cognitive work the tools are supposed to support.
The practical implication for program directors is structural, not motivational. Telling trainees to verify AI output is insufficient if the training environment does not require them to demonstrate reasoning independently from it. The goal is not to restrict these tools, but to ensure trainees develop independent competency that allows them to use these tools safely. Programs that train residents to use AI deliberately, recognize its failure modes, and interpret its outputs critically will produce surgeons better equipped for the practice environment they are entering than those who ignore these tools entirely.
Conclusions
AI’s near-term role in surgery is not the replacement of surgical judgment—it is the reduction of the cognitive and administrative overhead that surrounds it. Judgment, technical skill, and operative intuition remain irreducibly human. What AI can do is give surgeons more time and mental bandwidth to exercise those capacities.
The evidence base has matured enough to support specific conclusions. Randomized controlled trials confirm that ambient scribes reduce documentation burden and burnout, and prospective studies demonstrate that validated risk calculators outperform unaided clinical estimation for surgical mortality and major complications. Prior authorization automation has reached enterprise scale. The implementation evidence, while primarily from primary care and ambulatory settings, points clearly enough in one direction to justify action.
For trainees, the message is simple: you will practice in an environment where these tools are infrastructure, not novelty. The competency to adopt deliberately, verify consistently, and recognize when a tool is not working is part of what surgical readiness looks like in 2026. The limiting factor in AI adoption is not capability but willingness to start. Pick one problem, measure the baseline, deploy one tool, and measure again. Everything follows from that.
Footnotes
Funding
The authors received no financial support for the research, authorship, and/or publication of this article.
Declaration of Conflicting Interests
The authors declared the following potential conflicts of interest with respect to the research, authorship, and/or publication of this article: Kevin W. Sexton has equity in Schoolme, LLC; Biometrica, Inc; Arbizal, Inc; hDrop Technologies, Inc; Yejix.AI; and Qventus, Inc. Kevin W. Sexton has licensed intellectual property owned by Vanderbilt University Medical Center and the University of Arkansas for Medical Sciences. Carly Eckert has equity in Avante AI and SchoolMe, LLC. Carly Eckert has licensed intellectual property owned by Vanderbilt University Medical Center.
The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health. Any specific tools named reflect products evaluated in the cited literature and are provided for educational purposes only. No commercial endorsement is implied.
Declaration of Generative AI and AI-Assisted Technologies in the Writing Process
During manuscript preparation, the authors used generative AI tools including Claude (Anthropic) and ChatGPT (OpenAI) for drafting assistance, language editing, formatting suggestions, and organizational refinement. The authors critically reviewed and edited all generated content and take full responsibility for the accuracy and integrity of the manuscript.