Vapi is the developer favorite — great docs, real-time voice, and it handles the hard parts (turn-taking, interruptions, latency). Bland.ai is simpler but less flexible. Retell AI is another solid option. All three let you connect an LLM brain to a phone number. Start with Vapi's quickstart — you can have a working voice agent in under an hour.

Write a detailed system prompt that tells the AI who it is, what it can do, how to handle common questions, and when to transfer to a human. Think through edge cases: what if someone asks something off-topic? What if they get frustrated? Map out the happy path first, then add guardrails. This is where most voice agents fail or succeed.

Use ElevenLabs for the most natural-sounding voices (they have a Vapi integration), or use Vapi's built-in voices for lower latency. Then connect a phone number through Twilio ($1/mo per number + $0.0085/min). Vapi handles the Twilio integration — you just paste your Twilio API keys and assign a number.

The real power is when your voice agent can actually do things — check a calendar via Cal.com API, create a CRM entry, send a confirmation text. Vapi supports function calling, so your agent can execute actions mid-conversation. Test extensively with real phone calls before going live. Record calls (with consent) to find failure points.