NP0%
← Essays

What I Learned Building an AI Healthcare IVR System

Building an AI-powered healthcare IVR taught me that most voice bots fail before they even speak.

Not because of bad models.
Because of bad product decisions.

I built an AI-driven IVR system designed to call patients for appointment reminders, use RAG to fetch contextual patient data, handle dynamic responses (reschedule, confirm, decline), and reduce human telecaller workload.

Tech stack included FastAPI, Twilio voice calls, RAG with contextual retrieval, and structured fallback handling.

The goal wasn't just automation. It was reducing no-shows while maintaining patient trust.

01

Trust > Intelligence

Users judge the system in the first 5 seconds.

If the tone feels robotic or overly scripted, they disengage. Conversational warmth matters more than model sophistication.

Product Insight

Voice personality design is not cosmetic. It's retention strategy.

02

Latency Kills Engagement

Even a 1.5–2 second delay in response feels broken in voice interactions.

  • The system froze
  • The call glitched
  • It's spam
Product Insight

In real-time AI systems, latency matters more than model quality.

03

Personalization > General Intelligence

Fetching patient-specific context using RAG dramatically increased engagement.

“Hi, this is a reminder for your cardiology appointment on Tuesday” works better than “You have an upcoming appointment.”

Specificity builds credibility.

04

Fallback Design Is the Real Product

Most conversations don't follow the happy path. Users:

  • Speak unclearly
  • Switch languages
  • Ask unrelated questions
  • Interrupt

The fallback logic determined success more than the primary flow.

Product Insight

Design for confusion, not perfection.

  • Multi-lingual adaptive routing
  • Sentiment detection for escalation
  • Smart retry scheduling based on engagement probability
  • A/B testing conversational tone

If you're building AI voice products, design for:

  • Trust
  • Speed
  • Context
  • Failure handling

Not just model accuracy.