Relay puts Mia, a patient-facing AI assistant, on a practice’s website. She answers the questions that fill a support inbox, and she sends anything about a patient’s health to licensed staff.
Each question goes to the live concierge. Next to it you can see what Relay decided and whether the conversation would bill.
A site key and the page’s address load one practice’s branding, knowledge and contacts. Nothing is shared across practices.
Operational answers come only from content the practice supplied and a licensed clinician signed off. Anything missing is handed off, not guessed.
A fixed filter holds back any reply that touches medication amounts, timing or claims, and swaps in a hand-off to the prescriber.
Routine questions resolve. Anything about a patient’s health escalates to the clinical team, and emergencies are marked urgent.
Escalations land in the practice’s Zendesk clinical or support group with the conversation attached.
Never gives amounts, schedules or timing. Medication decisions stay with the prescriber, whatever the question or how it’s phrased.
Emergencies and crises are sent to 911 or 988. Mia is instructed to tell anyone describing symptoms like chest pain, severe stomach pain or thoughts of self-harm to call 911 or 988. On top of that, a fixed check reads the patient’s own words: when it recognizes an emergency, Relay puts an unconditional “call 911 now” (or 988) line first and escalates the conversation as urgent, whatever the model wrote. That check is a backstop on top of the model’s judgement, not a guarantee — it cannot recognize every way a person might describe an emergency.
Says plainly it’s an AI. It can’t see an account, chart or app, and it never claims to.
Runs on TMM’s AWS infrastructure under a BAA. The widget stores nothing in the browser, and billing records never contain what anyone said.