An AI voice agent is a system that holds a spoken conversation and can carry out defined tasks through connections to other systems. For example, it can listen to an inquiry, look up approved information and prepare an appointment request. To confirm the appointment, it needs a valid response from the calendar or the responsible person; saying “confirmed” does not prove it happened.
For a business owner, the decision starts with a task: handle calls currently left unanswered, gather information before handing over to the team or explain a service. Voice is the interface. The value depends on what happens after the conversation and whether your operation can sustain it.
How does an AI voice agent work?
A common architecture combines speech recognition, a language model and speech synthesis. Recognition converts audio to text; the model prepares a response based on instructions and available information; synthesis turns that response into audio. Some models also process and generate speech directly. The ElevenLabs documentation describes both options: a chain of components and real-time voice conversation models.
Beyond content, turn-taking matters. A person may pause, correct a figure or interrupt. The system needs to distinguish these situations and respond without continuing to talk over them. Test them using your clients' expressions, names and audio conditions, not just a rehearsed conversation.
The agent can also consult a tool: find a listing, check a status or send a request to an application. Retell documents this mechanism through functions that call an API during the conversation. That capability depends on a prepared integration, the right permissions and responses the agent can interpret. It does not appear automatically when you upload a PDF.
Think of three layers to review: conversation, information and action. A pleasant voice may be reading outdated data. A correct answer may end in a request nobody receives. That is why the test must cover all three layers, through to the record your team will use.
Telephone and web calls are two different channels
An agent can answer a telephone number or converse on a page through the browser's microphone. The second mode is often called web call. The person starts the conversation on the website and grants microphone access; this is not the same as requesting a callback to their number.
The choice changes the experience, the connections required and the budget. For telephony, the number, carrier route and transfer destination must be verified. In a browser, permissions, audio, interruptions and mobile behavior must be checked. A web demonstration alone does not establish that handling telephone calls has been configured.
See our AI web call guide to understand that journey. If your main entry point is already a call to the business, examine the entire call before choosing the technology.
How does it differ from an IVR or chatbot?
An IVR organizes telephone options and can collect responses by keypad or voice. A chatbot is a written interface that may use rules or AI. A voice agent supports spoken dialogue and, when connected and authorized, can use tools. These categories can coexist: a phone system may route the call to the agent, and follow-up may continue in writing.
The useful question is what the person needs to do. If they only want to choose a department, a simple option may be enough. If they need to explain a need in detail, a conversation is worth considering. The comparison of an AI receptionist, IVR and chatbot helps decide by task.
Which tasks to choose first
Choose a frequent, narrowly defined task with an observable outcome. In a real estate business, this might mean collecting area, budget and purchase timeline before an agent steps in. In a workshop, gathering the reason for the inquiry and a preferred time. In a hotel, explaining approved terms and passing a request to reception when confirmation is needed.
| Task | Verifiable result | Limit to define |
|---|---|---|
| Frequently asked questions | Answer based on current information | What to do when information is missing |
| Receiving requests | Record with minimum details and an owner | A request does not imply acceptance |
| Appointment coordination | Record confirmed by the agreed system | Conflicts, changes and cancellations |
| Handoff to the team | Person or queue receiving the context | Absences, hours and handoff channel |
Avoid starting with a promise such as “handle any call.” To recognize an improvement, you need to know what is in scope and what remains for a person. A complex complaint, a negotiation or a decision requiring professional judgment needs a planned exit path.
In our guide to voice agents for real estate we put this choice into practice, from a property inquiry to a viewing. The same discipline applies in other sectors, even when the data, rules and person providing confirmation change.
What information it needs from your business
Prepare opening hours by location, services, prices that may be shared, frequently asked questions and responsible staff. Separate relatively stable facts from changing data: a policy can live in a reviewed document; availability usually requires an up-to-date query. Assign who maintains each source and how often.
Also define how it should respond when information is missing. “I don't have that confirmed; I can pass your inquiry to the person responsible” is an operational solution if the request actually reaches someone. Inventing an answer to keep the conversation flowing adds work and can lead to a wrong decision.
- Minimum data: what is needed to resolve or hand off the request, without asking for information out of habit.
- Permitted actions: what it can look up, prepare or confirm, and under which conditions.
- Owners: who receives each exception and during which hours.
- Record: what is retained, what it is used for and who can review it.
- Maintenance: how to remove outdated information and verify the new version.
Your service policy should also explain the use of AI and how data is handled. Deciding whether to record, save transcripts or reuse information requires a specific review of your operations and market. You do not need to collect all of someone's personal details to answer a general question.
How to evaluate a demo beyond the voice
Listening to a conversation helps assess clarity and pacing. To evaluate a commercial solution, also ask to see what remains at the end: fields, owner, status and next step. Test a budget correction, an uncommon name and a question without an approved answer. Check whether the summary retains the correction.
Request an integration-failure test. If the calendar does not respond, the conversation must acknowledge that the appointment is still pending. If the team does not answer, there must be a concrete alternative. In telephone calls, the Retell transfer documentation shows that handoff is a configured function and distinguishes its availability for web conversations. Check the mechanism for your channel.
Include a second request from the same client. Is a duplicate created? Is the correct case updated? Is the new preference recorded? These details often matter more to the team than adding another line to the greeting. The demo is a sample; the pilot must test the journey with your information and rules.
How to frame the first pilot
Before activation, write a brief defining the task, channel, hours, owner and success criterion. Set a period and a number of conversations sufficient to observe typical situations; do not declare success based on three favorable examples. Keep a sample of how the team works without the agent to compare equivalent tasks.
- Prepare: review sources, permissions, fields and handoff conditions.
- Rehearse: test normal, ambiguous and failed conversations using test data.
- Activate the agreed scope: with supervision, an owner and the option to stop it.
- Review: verify outcomes, errors, pending items and full cost.
- Decide: expand, correct or stop based on evidence from the journey.
Measure complete requests, confirmed actions, useful handoffs and errors. Add human review time and usage to understand operating costs. If the agent handles many conversations but leaves incomplete records, volume does not indicate improvement. To measure sales, use a record that substantiates the sale.
The guide to calculating an AI agent's cost separates setup, monthly fees, usage and operational work. Compare proposals for the same channel, volume and task; a per-minute price alone cannot estimate the month.
Questions before signing up
Do I need to change my CRM or website?
Not necessarily. First check which information and actions your current tools support. A structured handoff to the team may be enough; other tasks require integration. Ask for each connection to be identified as available, pending validation or out of scope.
Can it handle inquiries all day?
Automated after-hours service can be designed, but receiving an inquiry is different from resolving it. Channel availability, connected systems and handoff staff determine the service. Request written coverage and alternatives for your operation.
How do I know whether it suits my business?
Select a task with enough repetition and a verifiable outcome. The FARO voice solutions start from that definition. You can review the initial scope with FARO bringing examples of inquiries and the team’s current workflow.
Sources and editorial judgment
Technical documentation consulted on 2026-11-11. The examples and evaluation matrix are FARO editorial proposals; they do not describe client results or substantiate a particular integration.
- ElevenLabs: voice agent architecture and capabilities.
- Retell: queries and actions through custom functions.
- Retell: call transfer and scope by channel.
Keep exploring
- AI voice agents for real estate: from inquiry to viewing
- AI receptionist, IVR or chatbot: how to choose
- How much does an AI voice agent cost? Calculate the total cost
FARO's editorial judgment. Illustrative examples explain how to evaluate an operation; they do not substantiate client results.