The AI Tool ShelfMissed calls are missed customers. A small café, salon or clinic can't always have someone free to answer the phone, especially at busy hours. AI voice agents promise to fix that: an assistant that picks up, answers common questions and takes bookings in a natural voice. To see how close that promise is to reality, we built a simple phone assistant for a fictional café with ElevenLabs' agent platform, ElevenAgents, and tested it with a real voice call. Here is exactly how to do it, what went well and what you need to fix before using one with real customers.
What is an AI voice agent?
An AI voice agent is a conversational assistant you can talk to out loud. It listens, understands what the caller wants, and answers with a synthetic voice in real time. Unlike old phone menus ("press 1 for opening hours"), the caller can speak normally, ask follow-up questions and even interrupt.
For a small business, the typical jobs are:
- Answering frequent questions: opening hours, location, prices, menu or services
- Taking reservations or appointment requests
- Collecting a caller's details so a staff member can call back
- Covering calls outside opening hours
What you need
- An ElevenLabs account. You can start on the free plan, which includes a limited amount of agent conversation time for testing.
- Ten minutes and a computer with a microphone.
- A short list of the questions your customers ask most often.
Step 1: Open the agent platform
ElevenLabs is split into platforms. The creative tools (text to speech, dubbing, music) live in ElevenCreative, while conversational agents live in ElevenAgents. Use the platform switcher at the bottom of the left menu to move to ElevenAgents, then create a new agent.
Step 2: Write the instructions (system prompt)
The system prompt is where you tell the agent who it is, what it knows and how it should behave. This is the most important part of the setup. Here is the prompt we used for our test café:
Keep it specific. List your real hours, your real services and what the agent should do when it doesn't know the answer. Vague instructions lead to vague answers.
Step 3: Set the first message, voice and language
The first message is what callers hear when the agent picks up. Keep it short and natural:
Then choose a voice from the library (we used a calm, "smooth and trustworthy" male voice) and the main language. You can add more languages if your customers speak several. We also left interruptible switched on, so callers can cut in while the agent is talking, just like in a real conversation.
Step 4: Test it with a real voice call
Publish the agent, open the preview and start a voice call from your browser. Here is an excerpt from our actual test call, lightly shortened:
Hello! Thanks for calling Atlas Café. How can I help you today?
Hi. What time, uh, do you open?
We open every day at eight in the morning and stay open until ten at night. So you're welcome to drop by any time in between! Is there anything else I can help you with?
I would like to book a table for four people tomorrow at seven.
Of course, I'd be glad to help with that! Could I get your name, please, so I can put the reservation under it?
Yes.
Great, thank you! And what name should I put the reservation under?
Davis.
Perfect, thank you, Davis! So just to confirm, that's a table for four people tomorrow, Tuesday the twentieth of October, at seven in the evening. Does that all sound right?
Okay.
Wonderful! So I have a table for four people under the name Davis, tomorrow, Tuesday the twentieth of October, at seven in the evening. Does everything look right to you?
Thank you.
You're very welcome! Just to be sure, I have your reservation as a table for four under the name Davis, tomorrow, Tuesday the twentieth of October, at seven in the evening. Is that all correct?
What worked well
- It understood natural speech, including hesitation ("what time, uh, do you open?").
- It stuck to the facts we gave it and answered the opening-hours question correctly.
- It turned "tomorrow" into a real date (Tuesday, October 20) and repeated the booking details clearly.
- It recovered from an unclear answer. When we answered "Yes" instead of giving a name, it politely asked again.
What needed fixing
Two issues showed up that any business should handle before going live:
- It kept re-confirming. After we said "Okay" and "Thank you", the agent repeated the booking details three times instead of wrapping up. The fix is to tell it what a confirmation sounds like and when to end the call (see below).
- The booking wasn't saved anywhere. Out of the box, the agent only talks. To record real reservations, it needs to be connected to something that stores them, such as a booking system, a calendar or a simple webhook to a spreadsheet. ElevenAgents includes tools and integrations for this, but it takes extra setup.
Here is the line we would add to the system prompt to fix the looping confirmation:
Step 5: Add knowledge and connect a phone number
Once the basic conversation works, you can make the agent genuinely useful:
- Knowledge base: upload your menu, price list or FAQ so the agent answers from real documents instead of only the prompt.
- Tools and integrations: let the agent save bookings or send details to your systems.
- Phone numbers: connect a phone line so customers can actually call the agent, instead of only testing it in the browser.
Test every change with a few calls of your own before real customers hear it.
Things to keep in mind
- Be transparent. Let callers know they are speaking with an AI assistant. Rules on disclosure and call recording vary by country, so check what applies to your business.
- Always offer a human fallback. The agent should be able to take a message or pass the caller to a person when it can't help.
- Watch the costs. Agent usage is billed by conversation time. The free plan is fine for testing; a busy phone line needs a paid plan, so check current pricing on the official site.
- Review conversations. Read the transcripts regularly to spot questions the agent answers badly, then improve the prompt.
The verdict
In ten minutes and without writing any code, we had a phone assistant that understood natural speech, answered questions correctly and took a booking politely. It is not a finished product out of the box: it needed one extra instruction to stop repeating itself, and it needs an integration before it can save real bookings. But for a small business that misses calls, an AI voice agent is now a realistic option to test, and the free plan makes it easy to try before committing.