The problem
A large share of the calls a healthcare provider receives are repetitive. Patients want to know how to prepare before a procedure, what is normal afterwards, and how to reschedule. The questions matter a great deal to the people asking them, but answering the same questions again and again takes clinical staff away from work that only they can do.
Traditional phone trees do not solve this. They route callers through menus, but they cannot actually answer a question, and patients often abandon them or press zero to reach a person anyway.
What we built
Megan is an AI front line for those calls. It answers by voice, can move a patient into a real-time AI video consult for pre- and post-procedure guidance, and handles the administrative tail automatically: every interaction is logged and appointments are scheduled without anyone re-entering information.
How it works
- Voice. Retell powers the conversational voice agent. Unlike a phone tree, it understands the caller in natural language, so a patient can describe what they need in their own words.
- Telephony. Twilio connects the agent to the provider’s phone line, so patients call the number they already know.
- Video consult. When a question benefits from being seen rather than heard, the conversation can move into a real-time AI video consult for guidance.
- Backend. A FastAPI service holds the logic for guidance, logging and scheduling, and connects the conversation to the systems that act on it.
- Infrastructure. The system runs on AWS.
The most important design decision: what Megan does not do
In healthcare, the boundary of an AI system matters more than its capabilities. Megan is designed to handle the repetitive, well-defined layer of patient communication: preparation instructions, post-procedure guidance, logistics and scheduling. Anything that needs clinical judgement is routed to a clinician.
That boundary is not a limitation to be engineered away later. It is the reason the system is safe to deploy. A voice agent that confidently answers a question it should have escalated is a liability, however fluent it sounds. So escalation is a first-class path in the design, not a fallback for when the model fails.
Why automatic logging matters
Every call that reaches staff still carries administrative work: recording what was discussed, updating the record, booking the follow-up. Much of the time saved by answering calls automatically would be lost if a person still had to document them. Logging and scheduling are handled automatically so the saving is real rather than shifted.
What the results mean
Megan takes a good share of the calls away from staff and gives each provider hours back. Deflection here means calls that were fully resolved without a person, not calls that were merely delayed or dropped. The hours saved are the practical measure: time returned to clinical work.
Where it can go wrong
- Escalation that is too timid. If the threshold for handing off is too high, patients get answers they should have got from a clinician.
- Escalation that is too eager. If everything is handed off, nothing is deflected and the system adds a step rather than removing one.
- Accessibility. Voice systems must work for callers who speak slowly, with accents, or in noisy environments. That needs testing with real callers, not only clean recordings.
When this approach fits
Voice AI of this kind suits organisations with a high volume of repetitive, well-defined calls and a clear line between routine questions and those needing expertise. It is the wrong tool where most calls genuinely need judgement, because the escalation rate will erase the benefit.
Questions to ask before building one
If you are considering a system like this, these are the questions we would want answered first.
- Which calls are genuinely routine? Review a sample of real calls and sort them. The share that is truly repetitive sets the ceiling on what automation can achieve.
- Where exactly is the clinical line? Agree with clinicians, in writing, which questions must always reach a person.
- What systems must it connect to? Scheduling and records integration is usually where most of the effort goes.
- How will you monitor escalations? Track both missed escalations and unnecessary ones, because each costs you differently.