ElevenLabs Agents: How to Build a Voice AI Agent

ElevenLabs Agents, now branded ElevenAgents, let a business build a voice AI agent that talks, listens and answers from its own documents. Setup happens in the dashboard: pick a template, write a first message and system prompt, choose a voice, add a knowledge base, then test. Every plan includes call minutes, from 15 on Free to 12,375 on Business, with extra minutes at $0.08 each.

Last updated: October 2026. Plans and credit costs checked against ElevenLabs’ official pricing page.

What are ElevenLabs Agents?

ElevenLabs Agents are conversational voice assistants built on the ElevenLabs platform that combine speech recognition, a language model and text to speech in one live loop. The product launched as Conversational AI in November 2024 and is now called ElevenAgents. Typical uses are support lines, booking assistants, website helpers and in-app voice guides.

“a developer platform for launching interactive voice agents”

Source: Wikipedia, ElevenLabs

The difference from plain text to speech is the loop. A narration job turns a finished script into audio once. An agent listens to a caller, decides what to say, and replies in a voice within a fraction of a second, over and over until the conversation ends. That is why latency, concurrency and per-minute pricing matter far more here than in the ElevenLabs text to speech guide.

In one line: an ElevenLabs agent is a voice, a brain and a knowledge base wired into a real-time conversation.

What can an ElevenLabs agent do?

An ElevenLabs agent can hold spoken conversations, answer from uploaded documents, collect structured data from callers and connect to phone lines. The agents pricing page lists text to speech, speech to text, knowledge bases, retrieval-augmented generation and telephony as included features, and there is no limit to the number of agents an account can create.

Here are the 5 core capabilities included with every agent:

• Text to speech output: Replies are spoken in any voice from the account, including library voices and custom voices, so the agent can match a brand’s tone instead of a generic assistant sound.
• Speech to text input: Caller audio is transcribed in real time so the language model receives clean text, which lets the agent handle natural speech instead of keypad menus.
• Knowledge bases: Documents, FAQs and links uploaded to the agent give it company-specific facts, so answers about products, hours or policies come from approved material.
• Retrieval-augmented generation: RAG pulls only the relevant passages from a large knowledge base into each reply, keeping answers grounded without stuffing every document into the system prompt.
• Telephony: Agents can take and place calls, and ElevenLabs lists no telephony fee of its own; external providers such as Twilio or a SIP trunk bill separately.

Those five building blocks cover most support and booking use cases without extra vendors.

How to build your first ElevenLabs agent step by step

Building a first ElevenLabs agent takes six dashboard steps, following the official quickstart: create the agent, set the first message, write the system prompt, choose a voice, add knowledge, and test. No code is needed for a working prototype, and a web widget can put it on a site the same day.

1. Sign in to ElevenLabs, open the Agents dashboard, name the new agent and start from the Blank template.
2. In the Agent tab, fill in First message, the greeting the agent speaks when a conversation starts.
3. Write the System prompt that defines the agent’s role, tasks, tone and what to do when a question is unclear.
4. Open the Voice tab and pick a voice, balancing quality against response time.
5. Upload documents, FAQs and links in the Knowledge Base section.
6. Click Test AI agent, talk to it as a customer would, and refine the prompt from what goes wrong.

The quickstart’s own example is a support assistant for a fictional company, which is a sensible first project because success is easy to judge: did the caller get the right answer or not?

Six steps produce a testable agent; the real work is refining the prompt afterward.

How do you write a good system prompt and first message?

A good system prompt for an ElevenLabs agent is a short operating manual: who the agent is, what it may and may not do, how it should sound, and how to recover from confusion. The first message is the opening line, so it should name the business and invite the caller to state a need in one breath.

Here are the 5 elements a strong agent prompt should contain:

• Role and scope: One sentence naming the company and the agent’s job, such as booking appointments, keeps the model from drifting into topics it should hand off to a human.
• Spoken-style rules: Instructions to keep answers short and conversational matter because long paragraphs that read well on screen sound robotic and slow when spoken aloud on a call.
• Clarifying behavior: The official example asks the agent to request clarification politely when a question is unclear, which prevents confident wrong answers on noisy or partial audio.
• Knowledge boundaries: A line telling the agent to answer only from the knowledge base, and to admit when information is missing, reduces invented prices, policies or dates.
• Escalation path: Clear guidance on when to offer a callback, an email or a human transfer gives frustrated callers an exit instead of a looping conversation.

A greeting such as naming the company and asking how the agent can help mirrors the quickstart’s pattern and works for most businesses. Test it out loud: if the line takes more than a few seconds to say, shorten it.

Short, scoped prompts produce faster, safer agents than long, clever ones.

Which voice and model fit a voice agent?

The right voice and model for an ElevenLabs agent are the ones that keep replies fast enough to feel natural. The quickstart warns that higher-quality voices and models can increase response time, so agents usually favor real-time models such as Flash or Eleven v4 Turbo over the slower, most expressive narration models.

Source: ElevenLabs models docs (elevenlabs.io/docs/models) and API pricing (elevenlabs.io/pricing/api), checked October 2026. Latency excludes application and network latency.

All four real-time options cost the same per character, so the choice comes down to speed versus language coverage and expressiveness.

For a brand voice, a custom voice works too: the ElevenLabs Voice Design walkthrough shows how to create one from a text prompt, and the ElevenLabs model guide compares the full lineup.

Pick Flash for the snappiest English or multilingual calls, Turbo when 90+ languages matter.

Adding a knowledge base and RAG

A knowledge base is the set of documents, FAQs and links an ElevenLabs agent can draw on when answering. Without one, the agent relies on the language model’s general knowledge, which is where wrong prices and invented policies come from. With one, and with RAG enabled, answers are grounded in the company’s own approved material.

The quickstart’s advice is practical: upload the content customers actually ask about, then keep it current. A short, accurate FAQ beats a large folder of outdated PDFs, because retrieval can only surface what is there. When a product, price or opening hour changes, update the knowledge base the same day, or the agent will keep repeating the old answer on every call.

Good knowledge sources include help-center articles, return and shipping policies, service menus with current prices, and booking rules. Marketing copy is less useful, since callers want specifics rather than slogans.

The knowledge base is the agent’s memory, so its accuracy sets the ceiling for every answer.

How to measure agent quality with Analysis and Call history

Agent quality on ElevenLabs is measured with the Analysis and Data collection settings, then reviewed in Call history. Analysis applies evaluation criteria to every conversation and marks each one as success, failure or unknown with a rationale, while Data collection extracts structured fields such as the caller’s question for later review.

The quickstart’s example criterion is named solved_user_inquiry, which captures the single question that matters for support: was the problem resolved? Adding a Data collection item with an identifier like user_question and a string type builds a running list of what callers ask, which doubles as a to-do list for the knowledge base.

A simple weekly routine works well. Filter Call history for failures, read the transcripts, and fix the cause: a missing document, a vague prompt rule, or a voice that speaks too fast. Then rerun the Test AI agent flow before pushing changes live.

Define success once, collect the questions, and let Call history drive every prompt revision.

Deploying an agent: web widget, apps and phone

Deploying an ElevenLabs agent means connecting it to wherever customers already are: a website, a mobile app or a phone number. The dashboard provides an embeddable web widget, official SDKs cover web and mobile apps, and the API reference documents telephony integrations with Twilio, Exotel and WhatsApp for call-based agents.

Here are the 5 deployment channels available:

• Website widget: The dashboard’s embed snippet places the agent on any page after swapping in the agent ID, the fastest route from prototype to a live customer-facing assistant.
• React and web apps: The @elevenlabs/client and @elevenlabs/react packages connect a custom front end to the agent session, suited to products that need their own interface design.
• Native mobile apps: ElevenLabsSwift, the Android Kotlin package, the React Native SDK and the Flutter package elevenlabs_agents bring voice conversations into iOS and Android apps.
• Phone calls: Telephony integrations with Twilio and Exotel route inbound and outbound calls to the agent, with the phone provider billing its own usage directly.
• WhatsApp: The API reference lists WhatsApp among telephony integrations, opening a channel where many customers already message businesses in markets outside North America.

Developers wiring agents into a backend should start with the ElevenLabs API getting-started guide, since the same API key powers both.

Start with the widget to validate demand, then move to phone or app channels once the agent performs well.

How many agent minutes does each ElevenLabs plan include?

Every ElevenLabs plan includes a monthly bundle of agent call minutes and a cap on concurrent calls. Free includes 15 minutes and 4 concurrent calls, Creator includes 275 minutes and 10 concurrent calls, and Business includes 12,375 minutes and 40 concurrent calls. Enterprise minutes are not published and are negotiated directly.

Source: ElevenLabs agents pricing (elevenlabs.io/pricing/agents) and pricing page (elevenlabs.io/pricing), checked October 2026.

Included minutes scale roughly in step with price: Pro costs about 4.5 times Creator and includes about 4.5 times the minutes, plus double the concurrent calls.

Minutes, not credits, are the planning unit for agents, so estimate monthly talk time before choosing a plan.

ElevenLabs agents pricing beyond included minutes

ElevenLabs agents pricing beyond the included bundle is $0.08 per additional call minute. Burst mode, which lets an agent exceed its concurrency limit up to three times, charges the excess calls at double the standard rate, $0.16 per minute. Text messages cost $0.003 each, and ElevenLabs adds no telephony fee of its own.

A worked example shows where the plan break-even sits. A small clinic on Creator ($22) that handles 400 minutes of calls in a month uses its 275 included minutes, then pays for 125 extra minutes at $0.08, or $10, for a total of $32 before LLM costs. At 1,238 minutes a month, Creator would cost $22 plus 963 extra minutes ($77.04), or $99.04, essentially the same as Pro at $99, which includes those minutes and doubles concurrency.

“Configure, deploy and monitor conversational agents.”

Source: ElevenLabs

One cost line stays vague. ElevenLabs lists language model costs as at cost and varying by model, but publishes no per-model numbers, and its FAQ wording conflicts on whether those costs are billed on top or deducted from credits. Check the usage page after a few days of real traffic before forecasting. Full plan math is in the ElevenLabs pricing guide.

Overage at $0.08 a minute is predictable; the LLM line is the one to watch.

Which plan fits your agent volume?

The plan that fits an ElevenLabs agent is the one whose included minutes and concurrent calls match peak usage, not average usage. Concurrency decides how many callers are served at once, minutes decide the monthly bill. A rule of thumb from the break-even math: once monthly talk time passes roughly 1,238 minutes, Pro beats Creator plus overage.

Here are the 5 plan fits for common agent workloads:

• Free for prototypes: Fifteen included minutes and four concurrent calls are enough to build, test and demo an agent, but free use is limited to non-commercial purposes under the Terms.
• Starter for side projects: Seventy-five minutes, six concurrent calls and a commercial license at $6 suit a low-traffic website widget or a small business pilot with occasional callers.
• Creator for small teams: With 275 minutes and ten concurrent calls for $22, Creator covers a modest support line, and overage at $0.08 per minute absorbs busy months.
• Pro for steady call volume: Pro’s 1,238 minutes and twenty concurrent calls at $99 fit businesses with daily call traffic, where Creator overage would cost about the same anyway.
• Scale and Business for operations: Scale’s 3,738 minutes and Business’s 12,375 minutes, with 30 and 40 concurrent calls, serve call centers and multi-location brands running agents all day.

Start one tier lower than the forecast and let overage reveal real demand before upgrading.

What are the limits and open questions?

The main limits of ElevenLabs Agents are concurrency caps, response-time trade-offs and unclear language model billing. Agents also depend heavily on prompt and knowledge-base quality, so a poorly configured agent will sound confident and still be wrong. None of these are blockers, but each needs a plan before a public launch.

Here are the 5 constraints to plan around:

• Concurrency ceilings: Calls above the plan’s concurrent limit need burst mode, which allows up to three times the limit but charges those excess calls at double the minute rate.
• Latency trade-offs: Higher-quality voices and models can slow replies, according to the quickstart, so expressive narration models are usually a poor fit for live phone conversations.
• LLM cost visibility: Language model charges are listed only as at cost and varying by model, so budgets should include a buffer until real usage data is available.
• Third-party phone billing: Twilio, SIP trunks and similar providers invoice separately from ElevenLabs, so the true cost per call includes the carrier’s per-minute charges too.
• Enterprise terms unpublished: Enterprise call minutes and pricing are not listed publicly, so large deployments need a sales conversation to confirm volume, concurrency and compliance terms.

Plan for burst, carrier fees and LLM costs, and the per-minute math stays honest.

Bottom line: should you build a voice agent with ElevenLabs?

Building a voice agent with ElevenLabs makes sense for teams that want natural-sounding voices, real-time models and a no-code builder in one place. Included minutes on every plan, a $0.08 overage rate and no ElevenLabs telephony fee keep costs readable. The weak spot is language model billing, which ElevenLabs does not itemize publicly.

“Agents turn ElevenLabs voices from a production tool into a live customer channel.”

Source: GrowthStackKit, ElevenLabs Review

For how agents fit alongside narration, dubbing and cloning, see the full ElevenLabs review. The fastest way to judge fit is still a prototype: build on Free, run real questions through Test AI agent, and upgrade once the transcripts look right.

Prototype free, measure with Call history, then pay for minutes once the agent earns them.

Frequently asked questions

The 12 most-asked questions about ElevenLabs Agents.

What are ElevenLabs Agents?

ElevenLabs Agents, now branded ElevenAgents, are conversational voice assistants that combine speech to text, a language model and ElevenLabs text to speech in a real-time loop. The product launched as Conversational AI in November 2024. Businesses use agents for support lines, bookings, website help and in-app voice guidance.

How much do ElevenLabs agents cost?

Every plan includes agent minutes: 15 on Free, 75 on Starter, 275 on Creator, 1,238 on Pro, 3,738 on Scale and 12,375 on Business. Additional minutes cost $0.08 each, burst calls cost $0.16 per minute, and text messages cost $0.003 each. Language model costs are listed as at cost.

Can I build an ElevenLabs agent for free?

Yes. The Free plan includes 15 agent call minutes and 4 concurrent calls per month, enough to build and test a prototype. Free use is limited to non-commercial purposes under the ElevenLabs Terms, so a customer-facing agent for a business should run on a paid plan, starting with Starter.

Do I need to code to build an ElevenLabs agent?

No. The dashboard lets anyone create an agent from the Blank template, set a first message and system prompt, choose a voice, upload a knowledge base and test it with the Test AI agent button. Code is only needed for custom front ends, mobile apps or backend integrations.

Does ElevenLabs charge for telephony?

ElevenLabs lists no telephony fee of its own, and telephony is included on every plan. External providers such as Twilio or a SIP trunk bill their own charges directly, so the full cost of a phone call includes the carrier’s rate on top of agent minutes.

What is burst pricing for ElevenLabs agents?

Burst lets an agent exceed its plan’s concurrency limit by up to three times during traffic spikes. Calls above the normal limit are charged at double the standard rate, which works out to $0.16 per minute instead of $0.08. It prevents dropped calls at the cost of a higher rate.

How many agents can I create?

ElevenLabs states there is no limit to the number of agents an account can create. Concurrency and credit limits still apply across all of them, so ten agents on one plan share the same concurrent-call cap and included minutes rather than each getting its own allowance.

Which model is best for ElevenLabs agents?

Real-time models fit agents best. Flash v2.5 lists about 75ms model latency across 32 languages, Eleven v4 Turbo about 100ms with 90+ languages, and Eleven v3 Conversational about 280ms with 70+ languages. The quickstart notes that higher-quality voices and models can increase response time.

Can ElevenLabs agents answer from my documents?

Yes. Agents include knowledge bases and retrieval-augmented generation, so documents, FAQs and links uploaded in the Knowledge Base section ground the agent’s answers. Keeping that content current matters, because the agent can only retrieve what has been uploaded and will repeat outdated information until it is replaced.

How are LLM costs billed for ElevenLabs agents?

ElevenLabs describes language model costs as at cost and varying by model, but does not publish per-model rates. Its FAQ wording is inconsistent about whether these costs are billed on top or deducted from credits, so the safest approach is to check the usage dashboard after real traffic.

Can ElevenLabs agents make phone calls?

Yes. The API reference documents telephony integrations with Twilio and Exotel, plus WhatsApp, for inbound and outbound conversations. Telephony is included on every plan without an ElevenLabs fee, while the phone provider bills its own usage. A web widget and app SDKs cover non-phone channels.

When should I upgrade from Creator to Pro for agents?

Creator includes 275 minutes for $22 and Pro includes 1,238 minutes for $99. At roughly 1,238 minutes a month, Creator plus $0.08 overage costs about $99.04, the same as Pro, which also doubles concurrent calls from 10 to 20. Beyond that volume, Pro is cheaper.