Twilio AI Voice Bot -
40% Faster Call Resolution
A US healthcare company's touch-tone IVR had 6 menu levels. Callers were abandoning at level 3. Average handle time was 8 minutes including IVR navigation. Replaced the entire system with a Twilio Programmable Voice bot powered by OpenAI GPT-4o callers speak naturally, the bot understands intent, resolves what it can and transfers to a human agent with full context when it can't. Average handle time dropped to 4.8 minutes. Call abandonment fell 52%.
The Brief
A US healthcare company handling appointment scheduling, prescription refill requests and insurance queries was running a touch-tone IVR that had been built in 2018 and extended several times since. Six menu levels. Callers navigating to appointment scheduling had to listen to five options at each level before finding the right one. The abandonment rate at level 3 was 34%. Average handle time from first ring to resolution was 8.2 minutes.
The internal team had tried simplifying the IVR menus twice. It helped briefly, then new services were added and the menus grew again. The fundamental problem was the architecture: touch-tone IVR forces callers to learn the system. The brief was to flip this build something where the system learns to understand the caller.
The solution: replace the IVR entirely with a Twilio Programmable Voice bot connected to OpenAI GPT-4o. Callers state what they need in plain English. The bot understands intent, asks clarifying questions when necessary, and either resolves the call or transfers to the correct agent with a full summary of what the caller needs.
"Callers are hanging up before they get to an agent. Our IVR has six levels and it's only going to get more complex as we add services. We need something that actually works the way callers expect it to work."
Project Phases
Analysed three months of call recordings (with consent) to map the 23 distinct caller intents appointment booking, cancellation, rescheduling, prescription refill, results enquiry, billing, insurance, general information, and 15 more. Identified that 68% of all calls fell into just 6 intents. Documented the resolution path for each intent: what information the bot needs to collect, what it can resolve autonomously, and what requires human agent transfer. Agreed HIPAA requirements: no PHI stored in OpenAI, conversation context cleared after each call, all transcripts encrypted at rest in AWS.
Built the voice bot on Twilio Programmable Voice with real-time OpenAI GPT-4o integration via streaming WebSockets. Twilio receives the call, converts speech to text in real-time, sends to GPT-4o with the system prompt and call context, receives the response, converts to natural speech via Twilio TTS and plays it back full round-trip under 800ms. System prompt engineered to: stay focused on healthcare call handling, ask only one clarifying question at a time, not hallucinate information it doesn't have, and trigger a clean transfer when the intent requires a human. Call context (caller intent, information collected, conversation summary) attached to the Twilio transfer so agents receive it before picking up.
No PHI sent to OpenAI caller identity verified by last 4 digits of date of birth against the internal system, but that verification result (yes/no) is what flows to GPT-4o, not the data itself. Conversation transcripts stored in HIPAA-compliant AWS infrastructure with encryption at rest. Call recording consent flow built into the bot's opening. Fallback logic: if the bot fails to understand intent after 2 attempts, it transfers immediately to a human with no friction. Load tested at 200 concurrent calls. Penetration tested. BAA reviewed with all third-party services.
Parallel running for one week new bot on 20% of traffic, legacy IVR on 80%. Metrics compared daily. Bot performing better on every metric after day 3. Full cutover on day 8. Post-launch monitoring dashboard: call resolution rate by intent, average handle time by day, transfer rate, and bot confidence scores flagged for review. Three weeks post-launch: average handle time 4.8 minutes (down from 8.2), abandonment rate 52% lower, 31% of calls resolved without agent involvement.