Purpose/Aim Artificial intelligence (AI) auto-contouring tools are increasingly available in radiation therapy planning; however, translating these technologies into safe, sustainable clinical practice remains challenging. This practice innovation initiative aimed to move beyond pilot evaluation and integrate AI auto-contouring into routine treatment planning through a clinician-led, evidence-informed workflow, with emphasis on performance variability, user resistance, and adaptive quality assurance (QA). Methods/Process A departmental QA pilot evaluated AI-generated contours across pelvic disease sites. Evaluation included comparison of contouring time between manual and AI-assisted workflows, qualitative rating of contoured structures, clinician willingness to adopt AI-assisted contouring, documentation of clinical experience, and structured user feedback. Rather than emphasizing formal accuracy validation, the pilot focused on identifying systematic error patterns, workflow disruption points, and usability considerations. Findings informed iterative workflow refinements, including earlier AI integration at CT simulation, development of site-specific structure templates, and scripting tools to reduce editing burden for nodal target volumes. Conservative QA strategies were initially introduced; however, ongoing evaluation demonstrated that some steps increased workflow burden and reduced compliance. QA processes were therefore adapted to align with existing professional review responsibilities, supported by targeted training and continuous feedback. Systemic AI contouring errors identified during implementation were communicated to users and escalated for vendor feedback. Results or Benefits/Challenges Benefits included improved planning efficiency in selected contexts, particularly for structures with higher image contrast and less complex anatomy, as well as increased standardization of contour quality through reduced inter-user variability. Adoption and perceived benefit were greater among less experienced clinicians, suggesting a role of AI tools in supporting workforce variability and training. Team awareness of AI behaviour and limitations also improved. Challenges included increased contouring time for complex target volumes, particularly nodal chains requiring extensive slice-by-slice editing, and user resistance to added workflow steps such as additional image transfers and explicit QA labelling of AI-generated structures. QA workload was redistributed rather than eliminated, reinforcing the need to mitigate automation bias through clinician education, awareness of AI limitations, and ongoing professional review. Leadership support required sustained real-world evidence and iterative refinement prior to broader expansion. Conclusions/Impact This initiative demonstrated that successful AI implementation depends on clinician-led governance, workflow-aware design, and adaptive QA processes that evolve with performance trends and user behaviour rather than static procedural additions. Aligning QA with existing professional review practices supported safe and sustainable integration without increasing workflow burden. Beyond efficiency gains, AI use contributed to improved contour standardization while preserving clinician oversight. Greater acceptance among less experienced clinicians highlighted the potential role of AI tools in supporting workforce variability, onboarding, and skill development. These insights offer transferable lessons for centres integrating AI into routine radiation therapy planning.
