What has to work for a caller to hear an answer?
One service turns speech into words, another decides what to say and a voice service turns that reply back into audio. Vapi coordinates these steps. Choose a web or phone connection and decide what the app should do if it cannot understand the caller.
Why can two equally long calls cost different amounts?
Vapi’s usage-only pricing adds a per-minute hosting fee to the charges for transcription, answer model, voice and phone transport. Different components can have different units and rates. A support package is separate from usage. Check the parts of the bill for the workflow you plan to run.
What might be saved when the call ends?
Vapi says recording, transcription and detailed call logs are enabled by default and can be configured separately. Turning off detailed call logs still leaves a basic call record. That switch does not control webhook or API logs. Check what each chosen component and configured destination receives, and confirm who can retrieve any saved recording or transcript.
Before using a real call, decide what the caller should be told and how they can continue if they decline recording. Turning off a saved recording does not stop the services from processing live speech.
How would you tell which part of a call went wrong?
Sketch this with fictional material. Leave the real site, accounts and services alone.
- Write the intended questionUse a fictional visitor asking whether a venue opens at nine. Write a second card with the same question misheard as five.
- Trace the replyFollow both versions through the answer and spoken output. Circle the first step where they differ.
- Mark what you need to keepDecide whether a short issue note is enough for debugging. Leave real calls, phone numbers and recordings out.