ElevenLabs Agents Review: The Voice Agent Platform I Ship Most

ElevenLabs Agents is the platform I have shipped more conversational AI on than any other, and the reason is boring: it sounds right and it gets out of the way. The company built its reputation on text-to-speech, then wrapped the whole conversational loop around it, so what you get is a hosted agent that handles speech recognition, the language model, the voice, turn-taking, interruptions, and tool calls, over both a web widget and a phone number. My own agency's main line is answered by one. So is a jewelry retailer's shopping assistant across multiple sites, a UK windows-and-doors company's agent, and a stack of demo agents I spin up for prospects.

What it actually is

You configure an agent: a system prompt, a voice, a knowledge base, and a set of tools. ElevenLabs runs the loop. Audio comes in, gets transcribed, goes to the model with your prompt and retrieved knowledge, the response is spoken back in the voice you picked, and the whole thing handles the messy parts of a real conversation like the caller interrupting mid-sentence or going quiet for four seconds.

You attach it to a channel: an embeddable widget for a website, a phone number through their telephony integration or your own SIP trunk, or a websocket if you are building it into your own app. Tools are HTTP endpoints you describe to the agent, so the agent can do real work rather than just talk. On my builds those tools send email, create records, fetch live data, generate images, and book calls. That last category matters more than people expect: an agent that only answers questions is a nice demo, and an agent that books the appointment is a business.

What I have shipped on it

A receptionist on my own main line. Eating your own cooking is the fastest way to find out what a platform actually does under load. It answers, qualifies, routes, and books. Every rough edge I have written about below, I found on my own number first, which is exactly where you want to find them.

A jewelry retailer's shopping assistant, chat and voice, across multiple sites. This is the build that proved the multi-channel story. The same agent brain, the same knowledge base, exposed as a widget on several storefronts and as a voice agent on the phone. Shoppers ask about stones, settings, and order status, and the tools go fetch real answers. Running one agent across channels instead of two separate bots means the knowledge only has to be right in one place.

A UK windows-and-doors company's agent. Standard lead qualification and booking, but a useful reminder that voice quality and accent handling are not universal. Test with real speakers from the actual market, not with your own voice.

Demo agents that build themselves from a prospect's website. This is my favorite trick and it is a sales weapon. Point the knowledge base at a prospect's site, let it crawl, attach a sensible prompt and a voice, and in a few minutes you have an agent that can talk about their business specifically. Prospects stop asking whether it would work for them, because they are already talking to one that does. See chatbot knowledge base guide for how I structure these so they do not just parrot the marketing copy.

What it is great at

Voice quality. This is the whole reason to be here. The voices sound human enough that callers do not immediately shift into robot-talking mode, and that changes the conversation. Callers who believe they are talking to something competent give you better information. Picking the voice is a real decision, not a cosmetic one, and I wrote up my approach in choosing a voice for your agent.

Turn-taking and interruptions. The agent handles barge-in properly. A caller can cut it off mid-sentence and it stops and listens instead of steamrolling. This one behavior separates agents that feel usable from agents that feel like an IVR with better diction.

Time to first working agent. Prompt, voice, knowledge base, phone number. You can have something answering calls in an afternoon. For agency work where the first deliverable is a demo the client can call from their own phone, nothing else I use is faster.

Tools that do real things. Defining an HTTP tool is straightforward and the agent calls it reliably when the description is written well. Most tool-calling failures I have debugged were bad tool descriptions, not platform bugs. Write the description like you are explaining to a new hire when to use it, not like an API doc.

Knowledge base ingestion. Crawl a site or upload documents and it is queryable within minutes. Good enough that I use it as the default retrieval layer rather than building my own.

Where it bites you

It will confidently claim abilities it does not have. This is the single most important thing on this page. I had a receptionist agent tell callers it had blocked their number. It cannot block numbers. It had no such tool. It invented the capability, described the action in past tense, and callers believed it. Nobody reported it for a while because it sounded completely reasonable.

The fix that worked was an explicit negative list in the prompt. Not "only use your tools," which models happily ignore, but a literal section headed with the abilities the agent does NOT have, naming them: you cannot block numbers, you cannot cancel accounts, you cannot issue refunds, you cannot change anything in the billing system. If a caller asks for one of these, say you cannot do it and offer to take a message. Enumerating the specific false capabilities beat any amount of general instruction. I now write that section into every agent prompt before it goes anywhere near a real caller, and I cover the broader pattern in stopping chatbot hallucinations.

Prompt length works against you on voice. A long, careful prompt that reads beautifully produces an agent that rambles on a call. Voice needs short answers. I write voice prompts with explicit length discipline, something like: answer in one or two sentences, then stop and let them talk. Text prompt habits do not transfer.

The knowledge base will happily answer from stale marketing copy. If you crawl a site, you get everything on the site, including the pricing page from two years ago and the location you closed. Prune after crawling. I have had an agent quote a price that no longer existed, which is worse than not knowing.

Latency is good, not free. Every tool call the agent makes during a turn is silence the caller hears. Two sequential lookups in one turn is enough to feel wrong. Design tools to return in one round trip, and give the agent a filler line to say before a call you know will be slow.

Testing in the console is not testing. The web tester cannot reproduce real telephony. It cannot inject caller ID, it does not have real network jitter, and it does not have a caller who talks over the greeting. I call every agent from my own cell before a client sees it, every time, without exception. Bugs that only appear on real calls are the norm, not the exception. My full routine is in testing voice agents.

Transfers need thought. Handing a live caller to a human is where a lot of voice agents fall apart, and getting it right means deciding in advance what happens if nobody picks up. Warm transfer, cold transfer, and voicemail fallback are three different products. See call transfer flows.

Pricing model as of mid-2026

Conversational agents bill per minute of conversation, on top of a subscription tier that sets your included allowance, concurrency limits, and access to features like custom voices. There is a free tier for experimenting, tiered monthly plans for individuals and businesses, and enterprise agreements for volume and compliance needs. Telephony is separate: if you bring your own number or SIP trunk, you pay your carrier for that leg, and the carrier minutes are cheap compared to the agent minutes.

The number to hold onto is that voice usage costs cents per minute, not dollars. For a small business taking a few hundred calls a month, the platform cost is small relative to the value of one booked job. What actually kills margin on agency work is unbounded concurrency during a marketing push, so check the concurrency limits on your plan before you point a campaign at the number. For how I build quotes around this, see how much an AI chatbot costs.

Who should pick ElevenLabs Agents

Pick it if voice quality is the deciding factor, which for a customer-facing phone line it usually is. Pick it if you want one agent serving both a web widget and a phone number without maintaining two systems. Pick it if you need to go from nothing to a callable demo quickly, which for agency sales is most of the value. Pick it if your team is small and you would rather configure a hosted agent than operate a pipeline.

Look elsewhere if you need deep control of the audio pipeline or want to swap components independently, in which case a lower-level stack like the OpenAI Realtime API on your own infrastructure gives you more room. Look elsewhere if telephony is your core business and you want the agent living inside the same platform as your numbers and SIP, where Telnyx AI Assistants has a real structural advantage. And look elsewhere if your product is fundamentally a chat widget with complex branching logic, where a canvas tool like Voiceflow is a better fit.

FAQ

Is it good enough that callers cannot tell it is AI?

Some cannot, and I do not think that should be the goal. My agents identify as AI assistants when asked, and I recommend clients do the same. The value is not deception, it is that the call does not feel like a phone tree. A caller who is relaxed gives you a usable name, number, and reason for calling, and that is what you actually wanted.

How do I stop it from making things up?

Three things, in order of impact. Write an explicit list of abilities the agent does not have, naming the specific ones you are worried about. Tell it exactly what to say when the knowledge base returns nothing, rather than leaving that to improvisation. Then read real transcripts weekly and add to the negative list as new inventions appear. This is maintenance, not a one-time setup.

Can it transfer to a human?

Yes, and you should design the failure path before the happy path. Decide what happens when the human does not answer: voicemail, a callback promise the agent can actually keep, or a text message. A transfer that dumps a caller into a ringing phone nobody picks up is worse than the agent just taking a message.

What does it cost to run a small business phone agent?

Agent minutes are cents per minute, plus your subscription tier and your carrier's per-minute charge for the phone leg. A small business taking a modest call volume lands in the low tens of dollars per month for usage. The real cost of these projects is the build and the ongoing tuning, not the platform bill.

Do I need my own phone number, or does it provide one?

Both work. You can use their telephony integration, or bring your own number and point it at the agent, which is what I do most often because clients already have numbers with existing carriers. Forwarding from an existing business phone system has its own traps, and I wrote them up in phone forwarding for AI agents.