A new research system called CareConnect can book, change, and cancel healthcare appointments using a large language model — and its authors spent considerable effort making sure it won't try to do anything beyond that.
Built by researchers and detailed in a preprint, CareConnect chains together eight purpose-built tools using LLM function calling and retrieval-augmented generation to handle scheduling logistics. It refuses to give medical advice through what the paper calls "deterministic short-circuit mechanisms" — hard-coded rules, not model judgment. Tested across 680 scenarios, it hit a 91.8% task completion rate, a median response time of 2.2 seconds, and 96% compliance on safety-critical test cases. The average cost per appointment came out to roughly $0.03, which the authors compare favorably to manual scheduling overhead.
The cost figure is the number worth paying attention to. Healthcare administrative work is estimated to consume a substantial share of clinical operating budgets, and scheduling is one of the most friction-heavy parts of that. A system that handles routine booking at pennies per transaction — and refuses to stray into clinical territory — sidesteps the liability minefield that has stalled broader AI adoption in healthcare.
Still, 91.8% task completion means roughly one in twelve attempts fails, which is a different kind of problem when the thing being scheduled is a medical appointment. The 4% safety non-compliance rate on the hardest test cases is similarly worth scrutinizing before anyone deploys this in production. The code is public on GitHub, so scrutiny, at least, is available.