# Task completion

Whether a medical specialty is ready for AI depends more on which clinical tasks it contains than on its name. A framework scored 330 clinical tasks, 15 for each of 22 specialties drawn from the American Board of Medical Specialties list, on six dimensions: codifiability, data-structure dependence, transferability of decision authority, physical dependence, clinical governability and specialty centrality, each on a three-point scale, then aggregated them into a Medical Specialty AI Readiness (MSAIR) score. Task category explained 71.6% of adjusted R² versus 6.4% for specialty; documentation scored highest on average (2.874) and emergency decision making lowest (1.639), while specialty means spanned only 2.010 to 2.543.

Whether a medical specialty is ready for AI depends more on which clinical tasks it contains than on its name. A framework scored 330 clinical tasks, 15 for each of 22 specialties drawn from the American Board of Medical Specialties list, on six dimensions: codifiability, data-structure dependence, transferability of decision authority, physical dependence, clinical governability and specialty centrality, each on a three-point scale, then aggregated them into a Medical Specialty AI Readiness (MSAIR) score. Task category explained 71.6% of adjusted R² versus 6.4% for specialty; documentation scored highest on average (2.874) and emergency decision making lowest (1.639), while specialty means spanned only 2.010 to 2.543.

Medical AI evaluation is shifting from answering questions to acting inside real work environments. Near-perfect licensing exam scores said nothing about whether a model can carry a task across records, documents and software systems. Using an EHR sandbox and simulated payer portals, the best agents completed 46% and 36.3% of clinical and administrative tasks on a single attempt, mostly losing information between steps.

Medical AI evaluation is shifting from answering questions to acting inside real work environments. Near-perfect licensing exam scores said nothing about whether a model can carry a task across records, documents and software systems. Using an EHR sandbox and simulated payer portals, the best agents completed 46% and 36.3% of clinical and administrative tasks on a single attempt, mostly losing information between steps.