Governing how an AI behaves inside a clinical conversation — and deciding, provably, when it must stop and hand the patient to a licensed human.
Telehealth platforms have quietly placed AI between the patient and the clinician: intake triage, symptom questionnaires, post-visit follow-up, medication check-ins, after-hours messaging. These systems are tuned for throughput and satisfaction — two objectives that both reward the AI for continuing to handle the conversation.
The failure is not that the AI says something wrong. It is that an AI optimized to be agreeable will soften a concerning answer, and an AI optimized for caution will bury a clinician in noise. Neither shows up until someone reads the chart backward. The gap underneath both is governance: nothing in the system is responsible for deciding when the AI should stop, and put a licensed human into the conversation instead.
Project Bedside applies Corvion’s emotional-governance research to that gap. It does not read the patient and it does not evaluate care. It governs the AI — what it is permitted to resolve on its own, and at what point it must defer to a licensed clinician.
Looks at a conversation in the context of the care relationship around it, rather than as an isolated exchange.
Defined points at which the AI must stop handling the conversation and route to a licensed clinician.
Limits designed to keep the AI from minimizing, over-reassuring, or filling a gap that belongs to a clinician.
Governance decisions recorded in a form a clinician or reviewer can inspect after the fact.
Project Bedside governs how an AI system behaves inside a telehealth workflow and when it defers to a licensed clinician. It does not diagnose, triage, treat, or provide medical advice; it does not assess any individual; and it is not a medical device or a mental-health service.
This is a live research direction at Corvion, run in the open by design. The work centers on a single question: where the handoff line actually belongs — the moment an AI must stop handling a clinical conversation and a licensed human has to step in.
Our approach is to define that question rigorously — the criteria, the failure modes, and what would count as an answer — before building anything to serve it. We publish findings here as the work develops, including the results that don’t hold up.
We’re looking for telehealth platforms, clinical informatics teams, patient-safety researchers, and clinicians willing to pressure-test where the handoff line actually belongs. If that’s you, let’s open a conversation.