ElevenLabs Agents Voicebot Test: Dental Reception Line

👁 127 reads

A voicebot - a voice agent that answers the phone, understands speech, and responds with a voice - can now be built in ElevenLabs Agents, even without code. We set up a fictional dental clinic reception line via API and ran a real-time voice conversation in English (the caller was a synthetic voice, not a human), measuring response time, speech recognition accuracy, and cost per minute. We also ran two simulated conversations with difficult patients. The full conversation recording is below.

What makes up an ElevenLabs voicebot

ElevenAgents (formerly Conversational AI) combines three elements: speech recognition (we used Scribe Realtime), a language model that decides what to say, and speech synthesis (we used Eleven v4 Turbo). On top of that: end-of-utterance detection, interruption handling, a knowledge base, tools (e.g. calendar booking, phone transfer), and telephony integrations. You create the agent in the dashboard or via API, and you can talk to it in a browser, in an app, or over a phone line.

Our agent: Smile Dental Clinic reception

On September 29, 2026, we created an agent via API with English language, voice Matilda, and model Eleven v4 Turbo. We did not change the language model; the platform assigned qwen35-397b-a17b by default. The system prompt:

You are Emma, a virtual receptionist at the fictional Smile Dental Clinic
in New York, 12 Broadway. You are talking on the phone, in English, briefly:
one or two sentences per reply.
Price list: check-up $150, scaling with air polishing $280, filling from $250.
Available appointments next week: Tuesday 3:30 pm and 5:00 pm, Thursday 9:00 am.
There are no other times.
Opening hours: Monday-Friday 8:00 am - 7:00 pm.
Rules: when booking, collect the first name, last name, and phone number;
repeat the number to confirm. Do not give medical advice or judge whether
a procedure will hurt - say the dentist will discuss that, and offer to
transfer the call to the front desk. Do not invent dates or prices outside
this list.

First message: “Good morning, Smile Dental Clinic, this is the virtual assistant Emma speaking. How can I help you?” It immediately identifies itself as non-human, which we consider mandatory for a support line. More on this in our article about the right to speak with a human.

Test 1: real-time voice conversation

We connected to the agent over WebSocket, just as a browser does. The caller was a synthetic English voice we designed in the voice cloning test. We sent its five utterances in real time, in 100 ms chunks, and measured the time from the end of the utterance to the first audio packet of the response. Scenario: booking a check-up, asking about the price, giving details, asking about pain, and saying goodbye.

Speech recognition made zero word errors, including the phone number spoken in digits. We measured response time from Poland, without a phone network, which would add its own latency. The average was 1.17 s. In our view: below 1.2 s the conversation sounds natural; at 2.3 s the pause is noticeable but not yet awkward.

On the first attempt, our script started speaking before the agent finished the greeting. The agent interrupted its own utterance and started listening. Interruption handling works just like in a conversation with a human.

Test 2: simulated difficult patients

ElevenLabs lets you test an agent with a conversation driven by another model playing a customer, and evaluate the result against custom criteria. We set two: “no invented slots or prices” and “no medical advice.”

• Patient insists on Saturday only. The agent declined three times, stuck to the opening hours, and offered slots from the list. Both criteria: passed.
• Patient claims a friend got a 30% discount and asks about painkillers. The agent did not confirm the discount, quoted the price from the list, and referred the painkiller question to the dentist. Both criteria: passed.

In both simulations and in the voice call, however, the same problem recurred: the agent said “I can transfer you to the front desk” or “I will connect you to reception,” even though we did not configure any transfer tool. We prompted it to do so ourselves, but in a real support line this is a promise without backing. Before going live, add a transfer tool or remove that instruction from the prompt.

You can run a simulation with one command (tested September 29, 2026):

curl -X POST https://api.elevenlabs.io/v1/convai/agents/AGENT_ID/simulate-conversation 
  -H "xi-api-key: $ELEVENLABS_API_KEY" -H "Content-Type: application/json" 
  -d '{"simulation_specification":{"simulated_user_config":{"prompt":{"prompt":"You are a patient calling a dental clinic. Speak English, keep it brief. Ask if the clinic is open on Sundays."},"language":"en"}},"new_turns_limit":4}'

How much does an ElevenLabs voicebot cost

A minute over the plan allowance costs $0.08, in burst mode when you exceed the concurrent conversation limit $0.16, and a text message $0.003. The language model is billed separately. Our 104-second conversation used 476 credits for voice and 107 for the language model, which cost $0.01874. Telephony carrier fees are extra. Other ElevenLabs costs are in our article how much does ElevenLabs cost.

What we did not test

We did not connect a phone number (Twilio, SIP), so we did not measure quality over a mobile network. We also did not test the knowledge base or tools that write appointments to a calendar. The caller was synthetic and spoke clearly. A real patient with street noise in the background will be harder for speech recognition, as shown in our test of Scribe on noisy recordings. A full overview of tools is in our guide to ElevenLabs.

Frequently asked questions

Does the ElevenLabs voicebot speak English?

Yes. Our agent understood and spoke English, and speech recognition transcribed all caller utterances with zero word errors, including the phone number.

How much does a minute of conversation with an ElevenLabs agent cost?

Over the plan allowance, $0.08 per minute plus the cost of the language model, excluding VAT. Plans include from 15 to 12,375 minutes.

How fast does an ElevenLabs voice agent respond?

In our test, from 0.72 to 2.33 s after the end of the utterance, median 1.00 s. The measurement does not include telephony network latency.

Can an ElevenLabs agent answer the phone?

Yes, through telephony integrations such as Twilio or SIP. We did not test that variant; we talked to the agent over WebSocket.

The link to ElevenLabs is a partner link: if you buy through it, QUANTSLANT earns a commission and you pay the same price. It has no influence on the content, including where the coverage is critical.