From Tool to Teammate: The Rise of Agentic Artificial Intelligence in Orthopaedic Practice

Mohamed A. Imam, MD, MSc, DSportMed, ELD (Oxon), PhD, FRCS, UNITED KINGDOM

ISAKOS eNewsletters   Current Perspective 2026   rating

Introduction

Something fundamental has changed in the way in which artificial intelligence (AI) interacts with medicine. For years, AI in orthopaedics meant image classifiers that could identify a fracture or grade cartilage loss on MRI—sophisticated, certainly, but passive. The clinician asked a question; the algorithm answered it. The relationship was unambiguous: tool and user.

We are now entering a different era. Agentic AI—systems capable of planning multi-step tasks, reasoning across domains, integrating heterogeneous data streams, and taking autonomous actions—is moving from research into everyday clinical use. The most tangible early example is the ambient AI scribe: a system that listens to clinical encounters, generates structured notes in real time, and adapts without explicit instruction at each step. Randomised and cohort studies across ambulatory settings have documented meaningful reductions in documentation burden, lower burnout rates, and freed clinical capacity; early evidence in surgical outpatient contexts is similarly promising.

This shift, from tool to teammate, may represent the most consequential transformation in surgical practice since the introduction of arthroscopy. As members of the ISAKOS Technology Committee, it is both our privilege and our obligation to understand it clearly and to lead its responsible integration into our specialty.

What is Agentic AI and Why Does it Matter to Us?

Conventional AI models respond to a single input: an image, a query, a dataset. Agentic AI systems, by contrast, pursue goals. They can decompose a complex clinical problem into sub-tasks, execute sequential reasoning chains, call upon external tools or databases, and synthesise findings into actionable outputs—all with minimal human prompting at each intermediate step.

Consider the orthopaedic consultation of the near future. A patient with recurrent shoulder instability arrives with 5 years of wearable biomechanical data, an MRI, a history of three physiotherapy episodes, and a family history of connective tissue disorder. An agentic system integrates all of this information—cross-referencing imaging findings with kinematic patterns, flagging prognostic risk factors from the literature, modelling predicted outcomes of operative versus nonoperative pathways, and drafting a personalised surgical plan—before the surgeon has finished reviewing the referral letter. The surgeon then does what only a surgeon can do: exercise clinical judgement, establish therapeutic alliance, and decide.

This is not science fiction. Comprehensive reviews have confirmed that large language models and multimodal AI architectures are already demonstrating early clinical feasibility across the spectrum of orthopaedic practice, from fracture classification and surgical decision support to patient-facing communication and postgraduate examination preparation^2,3.

Four Domains Primed for Transformation

Preoperative Planning, Decision Support, and Outcome Prediction

Surgical planning in joint reconstruction and ligament surgery involves a complex synthesis of anatomical, biomechanical, and patient-specific variables. AI systems trained on large multimodal surgical datasets now propose implant sizing, graft selection, and tunnel positioning parameters with accuracy rivalling that of experienced trainees¹. Real-world implementation has demonstrated that the application of machine learning to large administrative and clinical datasets can improve prediction of complications, readmission, and cost, delivering on the promise of value-based surgical care¹.

Beyond planning, AI is being applied to patient-reported outcome measures(PROMs)—the most-studied application of AI in any surgical specialty³—to predict whether individual patients will achieve a clinically meaningful improvement. Machine learning models incorporating preoperative PROMs alongside demographic and clinical variables have achieved area-under-the-curve values ranging from 0.63 to 0.95 in best-performing studies after arthroplasty and sports procedures. The performance gap between models reflects differences in outcome measures, procedure types, and validation approaches. Even at the conservative end of this range, personalised MCID prediction transforms the preoperative consultation: instead of quoting population-level success rates, surgeons can offer each patient an individualised probability—a profound shift in shared decision-making that patients increasingly expect.

Intraoperative Guidance and Robotic Integration

Computer vision systems embedded in the operative environment—fed by arthroscopic or open-field video—now provide real-time tissue identification, instrument tracking, and step-level procedural guidance. In arthroplasty, robotic-assisted systems already close the loop between preoperative plan and intraoperative execution. The emerging frontier is robotic-assisted arthroscopy: AI-enabled robotic arms with computer vision-driven tissue recognition are beginning to address the fundamental limitations of conventional arthroscopy—constrained visualisation, manual fluid management, and the operator-dependency of image interpretation. Early systems promise more-consistent joint access, reduced iatrogenic risk, and richer intraoperative data capture for downstream AI learning. The surgeon remains in command; technology provides continuous situational intelligence.

Personalised Rehabilitation and Continuous Recovery Monitoring

Wearable inertial measurement units, pressure-sensitive insoles, and continuous biometric monitoring now generate dense longitudinal datasets for patients recovering from ACL reconstruction, rotator cuff repair, or total shoulder arthroplasty. AI systems trained on these streams detect early deviations from expected recovery trajectories—altered movement patterns, asymmetric loading, reduced compliance—and dynamically adjust rehabilitation protocols in response. This transforms rehabilitation from a scheduled, therapist-delivered intervention into an intelligent, continuous, patient-responsive process³.

The implications extend well beyond clinical efficiency. For orthopaedic care in low-resource settings and remote communities, AI-guided rehabilitation becomes a vehicle for health equity—delivering expert-level physiotherapy monitoring to patients who would otherwise receive none. ISAKOS’s global membership is uniquely positioned to champion this application where its impact is greatest.

Surgical Education, Skill Assessment, and the AI Study Partner

Large language models have been tested against orthopaedic postgraduate examinations worldwide. GPT-4 has achieved 58% to 74% accuracy on the American Orthopaedic In-Training Examination—performance comparable with that of PGY-3 to PGY-5 residents, depending on whether image-based questions are included—while falling short of the standard required for the UK FRCS Orth Part A³. Raw pass rates, however, are secondary to what these systems enable: generating practice questions on demand, providing immediate explanations referenced to current literature, identifying gaps in a trainee’s knowledge, and adapting to individual learning trajectories. As autonomous study agents, they are already augmenting residency programmes globally.

Beyond examination preparation, AI-driven surgical skill assessment is maturing rapidly. Computer vision systems analysing arthroscopic and laparoscopic video can classify technical performance, characterise learning trajectories, and identify specific operative behaviours—path length, instrument economy, motion regularity—that distinguish novice from expert without manual annotation2. Validated arthroscopic assessment frameworks are ready for AI automation. For ISAKOS, this represents an extraordinary opportunity: objective, scalable, bias-free competency evaluation across its global training community.

The Digital Twin: Simulating the Patient Before Operating on Them

One of the most exciting emerging applications in our specialty is the patient-specific digital twin—a dynamic computational model of an individual’s musculoskeletal anatomy, biomechanics, and physiology, continuously updated with real-world data. Digital twins leverage multibody dynamics, finite element analysis, and machine learning to allow surgeons to simulate operative interventions before committing to them in the theatre⁴. Which graft will restore native ACL stiffness in this specific knee? How will this rotator cuff repair behave under load in a 55-year-old recreational tennis player with borderline tissue quality? These are questions digital twins are beginning to answer—not at the population level, but for each individual patient.

For ISAKOS, whose community spans the full breadth of arthroscopy and joint preservation, the digital twin represents an extraordinary opportunity—to design a trial inside a simulation before risking patient harm, to train the next generation of surgeons on patient-specific virtual anatomies, and to personalise implant and technique selection in a way that aggregate trial data simply cannot achieve. The most powerful digital twins of the next decade will be built on multicentre, international datasets: precisely the kind of data that a global society like ours is uniquely placed to generate.

Regulation, Responsibility, and the Human Imperative

It would be a profound error to conflate augmentation with replacement. The qualities that define excellent surgical care—moral reasoning under uncertainty, nuanced communication with frightened patients, the courage not to operate when the evidence is marginal, the wisdom to individualise beyond what any algorithm was trained on—are not computational properties. They are human ones. An AI system trained on population-level outcome data will, by definition, optimise for the population average. The patient in front of us is not the average.

AI systems inherit the biases of their training data. If our datasets under-represent women, older athletes, or patients from low-resource settings—and they invariably do—then our algorithms will serve those populations less well5. Understanding these limitations is not timidity; it is the precondition for safe and effective deployment².

The regulatory landscape is shifting rapidly, and surgeons must be conversant with it. The EU AI Act (Regulation 2024/1689), which entered into force on August 1, 2024, classifies AI systems embedded in regulated medical devices as high-risk. As a result, these systems are subject to requirements related to transparency, bias mitigation, human oversight, and algorithmic explainability—with full compliance for medical device AI mandated by August 2027. In parallel, the FDA finalised its Predetermined Change Control Plan guidance in December 2024, enabling AI/ML device modifications without additional marketing submissions, provided that the changes fall within a pre-specified, pre-authorised plan submitted at the time of original regulatory approval. These are not abstract policy matters. Orthopaedic surgeons who deploy or endorse AI tools bear a professional responsibility to understand these frameworks, advocate for their patients within them, and demand that developers meet their obligations.

A Call to Action for the ISAKOS Community

The ISAKOS Technology Committee has a unique mandate: not merely to observe innovation, but to shape its application within our specialty. Four priorities stand out.

• First, we must champion federated learning as the architecture for global orthopaedic AI. Federated learning enables AI models to train across multiple institutions without sharing raw patient data—each site trains locally and shares only model updates, such as weights or gradients—producing algorithms trained on genuinely diverse populations and compliant with GDPR, HIPAA, and emerging national frameworks. ISAKOS is uniquely placed to coordinate such a network.
• Second, we must integrate AI literacy into surgical education at every level. Understanding algorithmic outputs—their confidence intervals, failure modes, and biases—is as essential to the modern orthopaedic surgeon as reading an MRI. This includes training surgeons to interrogate, challenge, and override AI recommendations when clinical judgement demands it.
• Third, we must advocate for responsible AI governance frameworks specific to our specialty—ones that align with the EU AI Act and FDA guidance, embed human oversight, mandate prospective outcome monitoring, and protect the therapeutic relationship at the heart of surgical care. Regulators are writing these rules now; our voice belongs in that conversation.
• Fourth, we must invest in AI-powered surgical assessment platforms. Objective, scalable, bias-free competency evaluation—applied consistently across our global training community—would represent one of the most important advances in orthopaedic education since the introduction of simulation-based training.

Conclusion and Future Perspectives

The era of agentic AI in orthopaedic surgery is not approaching—it has arrived. It is dictating our clinic letters, predicting our patients’ outcomes, guiding our rehabilitation pathways, coaching our trainees, and simulating our operations before we make a single incision. The question is no longer whether AI will change our practice, but whether we will be the architects or the subjects of that change. The history of our specialty is one of bold innovation tempered by rigorous science—from the first arthroscope to the first shoulder arthroplasty to the first ACL reconstruction. This is the next chapter. Let us write it intentionally.

References

1. Ramkumar PN, Haeberle HS, Bloomfield MR, Schaffer JL, Kamath AF, Patterson BM, Krebs VE. Artificial Intelligence and Arthroplasty at a Single Institution: Real-World Applications of Machine Learning to Big Data, Value-Based Care, Mobile Health, and Remote Patient Monitoring. J Arthroplasty. 2019;34(10):2204–2209. doi:10.1016/j.arth.2019.06.018. PMID: 31280916.
2. Oettl FC, Zsidai B, Oeding JF, Hirschmann MT, Feldt R, Fendrich D, et al. Artificial intelligence-assisted analysis of musculoskeletal imaging—A narrative review of the current state of machine learning models. Knee Surg Sports Traumatol Arthrosc. 2025;33(8):3032–3038. doi:10.1002/ksa.12702. PMID: 40450562.
3. Khoriati AA, Shahid Z, Fok M, Frank RM, Voss A, D’Hooghe P, Imam MA. Artificial intelligence and the orthopaedic surgeon: A review of the literature and potential applications for future practice: Current concepts. J ISAKOS. 2024;9(2):227–233. doi:10.1016/j.jisako.2023.10.015. PMID: 37949113.
4. Diniz P, Grimm B, Garcia F, Fayad J, Ley C, Mouton C, et al. Digital twin systems for musculoskeletal applications: A current concepts review. Knee Surg Sports Traumatol Arthrosc. 2025;33(5):1892–1910. doi:10.1002/ksa.12627. PMID: 39989345.
5. Obermeyer Z, Emanuel EJ. Predicting the Future — Big Data, Machine Learning, and Clinical Medicine. N Engl J Med. 2016;375(13):1216–1219. doi:10.1056/NEJMp1606181. PMID: 27682033.

Please note: ISAKOS Newsletter Current Perspectives are not peer-reviewed articles. For peer-reviewed articles, please visit the Journal of ISAKOS at jisakos.com.