ElevenLabs Voice Agent Tutorial: 6 Steps & Per-Minute Costs (2026)

Building an ElevenLabs voice agent takes six steps: define its job, write its system prompt and knowledge base, pick a voice, set its channels (phone, widget, WhatsApp), test calls, then deploy. Billing is per minute — 15 minutes free, 75 on Starter ($1 first month, shown until Nov 8 when checked), $0.08 per extra minute. Checked 4 Nov 2026.

How does an ElevenLabs voice agent work?

An ElevenLabs voice agent is a pipeline, not a single model. Your caller speaks; ElevenLabs’ own speech-to-text transcribes it; a large language model (you can bring your own — the platform is LLM-agnostic) writes the reply; ElevenLabs’ own text-to-speech speaks it back. Between those stages sit two of the platform’s own models: a turn-detection model that decides when the caller has finished speaking, and Expressive Mode, which reads the caller’s emotion — the docs cite frustration, urgency, and relief — and adjusts the agent’s tone to match.

You deploy the finished agent across phone, web, WhatsApp, and email. Telephony runs through Twilio, Genesys, Telnyx, or any SIP-compatible PBX; a website widget embeds the agent on your pages; and the Speech Engine API adds a voice layer to an existing chatbot with one prompt, billed at $0.08 per minute. Every conversation bills by the minute from your plan’s included minutes — which is why this tutorial tracks cost at every step.

Build a voice agent in 6 steps

Six steps from an empty project to a deployed agent. Planning steps cost nothing; only test and live calls consume minutes.

1. Step 1 — Define the job and the success metric. Write one sentence: what the agent does and how you will know it worked (“books a test drive; success = appointment in the calendar”). A tight job definition keeps calls short, and short calls are cheap calls. Cost: $0 — planning consumes no minutes.
2. Step 2 — Write the system prompt, knowledge base, and guardrails. The system prompt sets identity and behavior; the knowledge base holds the facts the agent may quote (hours, prices, policies); guardrails set the lines it must not cross (no medical advice, no discounts beyond 10%). Keep all three tight — a chatty agent makes long calls. Cost: $0 — writing consumes no minutes.
3. Step 3 — Pick the voice and write the greeting. Choose from 17k+ voices across 90+ languages, then write a 2-sentence greeting that states who the agent is and asks one question. A greeting that asks one clear question cuts rambling openings. Cost: $0 until you test — voice selection itself is free.
4. Step 4 — Configure the channels. Start with the web widget (fastest to test), add phone via Twilio, Genesys, Telnyx, or SIP when you are ready for real callers, and note WhatsApp and email for later. Set a per-call minute cap now — it is the cheapest guardrail you own. Cost: $0 — configuration consumes no minutes.
5. Step 5 — Test with 5 scripted scenarios. Run the happy path, the confused caller, the angry caller, the edge-case question, and the goodbye. Budget about 2–3 minutes per scenario — roughly 10–15 minutes total, which fits inside the 15 free minutes if you stay disciplined. Fix the prompt between scenarios, not during them. Cost: ~10–15 minutes of your free 15.
6. Step 6 — Deploy, then watch the transcripts. Go live on one channel first, review the first 20 call transcripts, and tighten the system prompt where calls ran long. Long calls are the cost driver — every minute of rambling is a minute billed. Cost: per-minute billing from here on — 15 free minutes, then $0.08/min via Speech Engine or the next plan up.

What voice agents cost per minute

Agent billing is per-minute from your plan’s included minutes. The official ElevenAgents pricing, checked 4 Nov 2026:

Past your included minutes, the Speech Engine API bills $0.08 per additional minute ($0.160 burst pricing). Text messages to the agent are included from Starter up, and the commercial license is included from Starter up — the free plan’s 15 minutes are for testing, not commercial deployment.

Which plan fits your call volume?

Do the minute maths before you pick. A receptionist agent handling 10 calls a day at 3 minutes each burns 30 minutes a day — about 600 minutes a month, which lands between Creator (275) and Pro (1,238); most small pilots fit Starter’s 75 minutes (about 25 three-minute calls). A support agent taking 100 calls a month at 4 minutes each needs 400 minutes — Creator plus extra minutes, or Pro for headroom. And remember concurrency: 4 simultaneous calls on free, 6 on Starter, 10 on Creator, 20 on Pro — a marketing campaign that spikes to 8 parallel calls needs Creator even if the minute total looks small.

5 rules for cheaper agent calls

• Cap every call. Set a per-call minute limit at configuration — the single most effective cost control.
• Write a short system prompt. Chatty agents make long calls; every unnecessary sentence in the prompt risks minutes on every call.
• Ask one question at a time. Greetings and prompts that ask one thing get short answers; open-ended invitations get monologues.
• Test inside the free 15. Five disciplined scenarios at ~2–3 minutes each fit the free tier — fix the prompt between tests, not during.
• Read transcripts weekly. The calls that ran longest show you exactly which prompt lines to tighten.

Common problems and fixes

The agent talks over the caller

Turn detection is mistiming the handoff. Shorten the agent’s sentences — long sentences invite interruption — and check the turn-taking settings in the agent configuration. If callers speak slowly or with pauses, the model may read a pause as a finished turn; tightening the prompt’s speaking style usually fixes more than any setting.

Calls run long and burn minutes

The system prompt is too permissive. Add an explicit instruction to be brief, give the agent a 3-step call structure (greet, resolve, close), and set a per-call minute cap. Review the longest transcripts — the rambling almost always traces to one prompt line you can cut.

The agent invents facts

The knowledge base is thin or the guardrails are missing. Put every quotable fact (hours, prices, policies) in the knowledge base, add a guardrail that forbids answering outside it, and add an escalation rule: unknown questions go to a human, not to improvisation.

You hit the concurrent-call ceiling

Free allows 4 simultaneous calls, Starter 6, Creator 10, Pro 20. If callers get turned away during spikes, that is the ceiling, not a bug — move up a plan or spread the load. For the full cost picture behind plan jumps, see our agents cost guide.

Who should NOT build a voice agent here

Skip ElevenLabs agents if your call volume is tiny and fixed — a 5-minute IVR menu does not need a conversational agent, and per-minute billing will annoy you. Skip it if your compliance team needs data-residency paperwork signed before any trial; the enterprise tiers cover SOC 2, HIPAA, and regional residency, but that conversation happens on paid plans. And skip it if you need more than 20 concurrent calls on day one — Pro caps at 20, and beyond that you are in custom-quote territory. For API-first builds, the API quickstart shows the developer path instead.

Frequently asked questions

How do you build a voice agent on ElevenLabs?

Define the agent’s job, write its system prompt plus a knowledge base of quotable facts and guardrails, pick a voice from 17k+ options, configure a channel (web widget first, then phone via Twilio/SIP), test five scripted scenarios inside the 15 free minutes, then deploy with a per-call minute cap. No engineering is required for the standard path.

How much does an ElevenLabs voice agent cost per minute?

Minutes are included by plan: 15 free, 75 on Starter ($6/month), 275 on Creator ($22/month), 1,238 on Pro ($99/month). Past the included minutes, the Speech Engine API bills $0.08 per additional minute ($0.160 burst). A 3-minute support call therefore costs about $0.24 in extra-minute terms.

Can I try ElevenLabs voice agents for free?

Yes — the free plan includes 15 agent minutes, the no-code builder, the knowledge base, and the web widget. That fits about five 2–3 minute test scenarios. The free minutes carry no commercial license, so they are for testing; commercial deployment starts on Starter, which adds the license, text messages, and 75 minutes.

What affects the cost of running a voice agent?

Call length first — billing is per minute, so a chatty system prompt is a standing cost. Then call volume, then concurrency: 4 simultaneous calls on free, 6 on Starter, 10 on Creator, 20 on Pro, and spikes past your tier force an upgrade. Set a per-call minute cap, keep prompts brief, and review long transcripts weekly.

Which channels can the agent use?

Phone (via Twilio, Genesys, Telnyx, or SIP-compatible PBX), a website widget, WhatsApp, and email — one agent serves all of them. Developers can also reach the agent through WebSocket and REST APIs or add a voice layer to an existing chatbot with the Speech Engine API’s single prompt.

Is ElevenLabs HIPAA compliant for voice agents?

ElevenLabs lists HIPAA compliance, SOC 2 Type II, ISO 27001, regional data residency, and Zero Retention Mode among its enterprise controls. These are paid-tier conversations — if your use case needs signed compliance paperwork before any trial, talk to sales before building, because the free 15 minutes will not cover that process.

Sources checked

How agents work, channels, and platform features: official ElevenLabs voice agents page (elevenlabs.io/voice-agents), checked 4 Nov 2026. Per-minute pricing, included minutes, and concurrent calls: official ElevenAgents pricing (elevenlabs.io/pricing/agents), checked 4 Nov 2026. Speech Engine per-minute rates: official ElevenLabs API pricing (elevenlabs.io/pricing/api), checked 4 Nov 2026.