What do you need to build an AI voice agent?

Four layers: a voice platform (Retell AI handles STT, LLM orchestration and TTS), telephony (Twilio numbers with A2P registration for SMS), an automation layer (n8n or similar) for everything the agent triggers, and a CRM (often GoHighLevel) where leads and bookings land.

How long does it take to build a production voice agent?

A working demo takes a day. Production — edge cases, interruption handling, calendar integration, compliance, monitoring — typically takes 1–3 weeks depending on how many systems the agent touches.

Is Retell AI better than Vapi or Bland?

They're all capable. I ship mostly on Retell because its conversation-flow model, component system and debugging tools fit how I work, and clients like the call analytics. The right answer depends on your latency, pricing and integration needs.

How I Build Production AI Voice Agents with Retell AI (My 2026 Stack)

TL;DR: My production voice-agent stack is Retell AI (voice + conversation orchestration), n8n (every action the agent triggers), GoHighLevel (CRM + calendars + SMS), and Twilio (numbers + A2P). A demo takes a day; production takes 1–3 weeks, and almost all of that time goes into edge cases, integrations and testing — not the "AI part."

I've shipped voice agents for US startups like CallSetter AI and Fortell AI — receptionists for hospitals, vet clinics, real-estate teams and car garages, appointment setters handling hundreds of calls a month. This is the stack and process I actually use, including the parts that usually go wrong.

The four layers of every production voice agent

1. Voice platform — Retell AI. Retell handles the hard real-time plumbing: speech-to-text, LLM orchestration, text-to-speech, interruptions, latency. I build agents as conversation flows with components — reusable blocks for things like caller verification or booking — rather than one giant prompt. One giant prompt demos well and collapses in week two.

2. Telephony — Twilio (or GHL's LC Phone). Numbers, SIP, and critically A2P 10DLC registration. If your agent (or its follow-up automations) sends SMS in the US without a registered brand + campaign, carriers silently filter your messages. Register on day one; approval takes days, not hours.

3. Automation — n8n. The agent is only as useful as what it can do. Retell fires webhooks (call started, ended, analyzed; mid-call function calls) and n8n turns them into actions: check calendar availability, create CRM contacts, send confirmations, alert a human when the caller is angry. I keep all business logic here, not inside the voice platform — swapping or A/B-testing agents later becomes trivial.

4. CRM — GoHighLevel. Where leads, pipelines, calendars and campaign follow-ups live. The agent books straight into GHL calendars; missed calls trigger textback automations; every call leaves a contact record with the transcript attached.

The build order that works

Script the happy path as a phone call, on paper. Read it aloud. If a human receptionist wouldn't say it, cut it.
Build the flow in Retell with components per job-to-be-done: greet → intent → qualify → act → confirm → end. Keep each component's prompt short and testable.
Wire one action end-to-end (usually calendar booking) before adding more. A booking that works beats five integrations that almost work.
Then the edge cases — this is most of the work: callers who interrupt mid-sentence, background noise, "can I talk to a human" (always give an escape hatch), voicemail detection, the caller who answers every question with another question.
Instrument everything. Retell's call analytics plus a simple n8n log into a sheet/DB: outcome per call, drop-off point, function-call failures. You cannot improve what you can't see.

Testing: the part everyone skips

At Fortell I built a pipeline that could create and test agents in minutes using Claude Code plus browser automation — because manual testing is where voice projects die. The minimum viable version for any project:

A test-call script — 15–20 scenarios you run against every change (happy path, interruption, wrong info, human handoff, silence).
Transcript review after every deploy — read 10 real calls. You'll find failure modes no test predicted, usually in the first three.
A rollback plan — keep the previous agent version one click away.

What actually breaks in production

Latency stacking: each function call adds delay; three sequential lookups mid-sentence and the caller thinks the line died. Parallelize in n8n, or have the agent narrate ("give me two seconds while I check the calendar…").
Calendar chaos: timezones, double-booking windows, "next Tuesday." Never let the LLM do date math — resolve dates in code and confirm back explicitly.
The unregistered-SMS trap — see A2P above. The system "works" in demo and silently drops confirmations in production.
Over-automation: the best agents know when to stop. Detecting frustration and routing to a human saves more deals than any prompt improvement.

What this costs to run

Usage lands around $0.07–$0.20 per call minute all-in (platform + LLM + telephony), before whatever you charge for the build and the monthly service. I've broken the full economics down in How much does an AI receptionist cost in 2026?

I take on voice-agent and automation projects through Null Studio — if you're a business that wants one built, or a founder who needs the whole pipeline, book a call. And if you freelance in this niche: these jobs move fast on Upwork — I built Upwork Scout to catch them the minute they post.