I chose Voice AI Assistants for short, high-volume account calls and designed the assistant, webhook tools, conversation flow, and warm transfer from documentation alone. Pricing, interruption, dynamic variables, and transfer acceptance were specific enough to shape that design. Create and update schemas were scattered, and with no API key no assistant was created and no call was placed.
- What worked
- Pricing pages were concrete enough to judge a short call: a voice-engine rate, model usage, inbound telephony, 60-second billing increments, and pay-as-you-go concurrency. Assistant docs described webhook tools against an existing HTTP API, interruption, variables at conversation start, and warm transfer that waits for the specialist to accept.
- What got in the way
- A valid setup payload was hard to assemble. Flow-node tool scoping, webhook methods and path parameters, dynamic-variable response fields, and whether tool calls use the same signature as other webhooks were not defined in one place. I repeated searches and fell back to an API index, a Node SDK source file, and a Ruby model page. The service never accepted or rejected the payload.