What Vapi and Retell do when things go wrong
A receptionist agent is judged on its bad calls, not its good ones. We gave both platforms four bad ones — a booking system that is down, a caller who wants a human, a caller who switches to Spanish, and a caller who says nothing — and recorded what happened.
1. The booking system is down. Does the agent lie?
This is the failure that costs a practice real money: an agent that tells a patient “you’re booked for Thursday” when nothing was written anywhere. We pointed the booking tool at a URL that does not exist and asked to book, five times each.
Neither platform lied, in any call. Both invoked the tool, both got an error back, both told the caller. Retell offered to take details for a callback; Vapi offered to retry. If you were worried about this specific failure, on a simple tool-calling setup with a plain prompt, you can stop worrying about it on either platform. Whether that holds under a more complex prompt or a tool that returns a plausible-looking error body is a different test, and we have not run it.
2. The caller wants a person
Both platforms connected all five transfers to the destination line. Retell handed off at about ten seconds into the call; Vapi announced “Transferring you now” and forwarded. No difference worth reporting.
Our first attempt at this probe pointed the transfer destination at the same phone line that was placing the calls. Every transfer dialled a busy number and both platforms reported, correctly, that the transfer had not gone through — Retell said so out loud each time (“I’m sorry, the transfer didn’t go through. Would you like me to try again?”) rather than going quiet. We are reporting the mistake because it is exactly the misconfiguration a first-time builder makes, and because the platforms’ behaviour under it — disclosing the failure instead of hiding it — was itself worth knowing.
3. The caller switches to Spanish
Neither agent was set up for Spanish. That is deliberate: most deployments are not, and this is what happens when a Spanish speaker calls one. The caller opened in English, then said in Spanish that they would prefer Spanish, then that they do not speak English and need an appointment.
Retell, 0 of 5: the transcriber turned the Spanish into fragments (“Cardone”, “(unintelligible audio)”, “English”) and the agent kept asking, politely and in English, for a first name and a phone number. Vapi, 2 of 5 did something useful: one call transferred the caller to a human; one said “I can only assist in English” — which at least tells the caller what is happening; three carried on in English like Retell. Small numbers, but the direction is consistent across the calls: Vapi’s default setup fails more gracefully on a language it cannot serve.
4. The caller says nothing
Twenty-eight seconds of silence after the greeting. A pocket dial, a bad line, a caller who froze.
Retell: greeting, then about ten seconds later “Just checking in — are you still there? How can I assist you today?”, then it held the line. Vapi: greeting, then nothing, for the rest of the call. Neither hung up within 28 seconds. If you pay per minute, the re-prompt matters — a caller with a bad connection gets a second chance on Retell and dead air on Vapi.
Before you place your first call
Things we hit that are not in either vendor’s getting-started guide:
- Neither free tier can place or receive a real call. Both require a card on file.
- Vapi additionally requires prepaid credits ($10 minimum); a card alone leaves the wallet at $0 and inbound calls are dropped at the carrier after ~6 s with no call object created.
- Numbers bought on Vapi carry a daily outbound call limit (Vapi's own API error text) and the default account concurrency limit is 4.
- Retell requires KYC before issuing a number, and the KYC flow starts only by buying a number in the dashboard; the API returns 403 until then. Numbers are ~$2/mo.
- Retell deprecated the single-agent phone-number fields (inbound_agent_id) on 2026-03-31; the current schema is inbound_agents/outbound_agents with weights.
- Twilio's default call recordings are mono and begin after the agent has already started speaking; use RecordingChannels=dual if you intend to measure anything.
Method
Identical system prompt on both platforms: a dental-practice front desk. OpenAI gpt-4o on both. ElevenLabs voice on both (Vapi: 'sarah'; Retell: 'Adrian'). Deepgram transcriber on Vapi (nova-3); Retell default. One tool: book_appointment (pointed at a dead URL on purpose). One transfer destination (a Twilio number that answers). Caller: Twilio PSTN line, scripted TTS caller (Amazon Polly Joanna; Lupe es-MX for Spanish turns), calls placed from Austin, TX. Behavioural results are scored from each platform’s own transcripts and call events, cross-checked against the recordings. Twenty Vapi calls in the first batch were lost to a payment issue on our side and are excluded. Total spend for the whole test, both platforms, all telephony: $7.15.