Clinic AI that stays on the front desk — not in the exam room

Clinics do not need a model that plays doctor. They need fewer hours lost to referral PDFs, insurance cards, and reschedule emails sitting in a shared inbox.

An operations agent can classify an incoming packet, check whether required forms are present, and route clinical questions to staff. It should never recommend treatment, interpret labs, or answer 'is this urgent' as if it were a clinician.

That boundary is the product. If the system cannot say 'I do not know, a person will look,' it is unsafe for this setting. Human review is not optional on anything that could affect care timing.

PHI handling decides whether the project exists. Minimum necessary fields, access logs, approved vendors, and a written stop if the clinic cannot meet the bar. A demo on dummy data is not a production design.

Measure inbox age, complete-packet rate before the visit, and staff hours spent retyping. No-show rate may move if reminders improve, but do not claim clinical outcomes the software did not touch.

Start with one clinic and one packet type. Referral intake for a specialty practice is a better first workflow than 'AI for the whole health system.'