Voice AI Agent

The Voice AI Agent answers the phone. It picks up inbound calls on a number you own, speaks to the caller in real time, works through whatever you have told it to find out, and either books an appointment, captures the details and promises a callback, or transfers the call to a person. It can also place outbound calls triggered by a workflow, for example to follow up a web lead within a minute of the form submission.

Mechanically it is the same agent definition as the text agent, wrapped in speech. Audio from the caller is transcribed, the agent decides what to say, and the reply is spoken back with a synthesised voice. Everything in between, the knowledge, the instructions, the booking permission, the refusals, comes from the agent you authored. What changes is the medium, and the medium is unforgiving: there is no scroll-back, no read receipt, and a two-second pause feels like a dropped call.

The right mental model is a trained receptionist on a headset, not an IVR tree. You are not designing a menu with numbered options. You are describing a job and a set of boundaries, and then listening to recordings until the agent handles the real distribution of callers. Voice amplifies both the value and the risk of AI: when it works, nobody reaches voicemail again; when it misfires, the caller hears it immediately and so does your client.

Voice and delivery

• Voice selection, speaking rate and accent
• Greeting script, including whether the agent discloses that it is an automated assistant
• Filler and acknowledgement behavior so pauses do not read as dead air
• Interruption handling: whether the caller can talk over the agent and have it stop

Call handling

• Which inbound numbers the agent answers, and whether it answers immediately or after a set number of rings
• Operating hours, and whether calls outside them go to the agent, to voicemail, or to a forwarding number
• Maximum call duration and what the agent does as it approaches the limit
• Behavior on silence, background noise and no-speech detection

Conversation objectives

• The information the agent must collect before it offers anything
• The order it works through topics, and what it should do when the caller jumps ahead
• Topics to refuse and claims it must never make

Booking and actions

• Which calendar it may book into and how it reads out available times
• Whether it may reschedule or cancel
• Fields and tags it writes to the contact record during or after the call

Transfer and escalation

• Warm or cold transfer to a person or a ring group, and the numbers involved
• Triggers for transfer: caller request, unresolved question, detected frustration, specific keywords
• Fallback when the transfer target does not answer

After the call

• Recording and transcript retention, and whether a recording notice is played
• Call summary and outcome written back to the contact or opportunity
• Workflows triggered by call outcome, duration or disposition

1. Sort the telephony first. Have a working number in the account, confirm inbound calls reach it, and confirm any required registration or verification for that number is complete before you involve AI at all.
2. Check the legal position on call recording and automated-caller disclosure for the jurisdictions the client operates in, and decide the disclosure wording with the client, in writing.
3. Author or clone the agent in AI Studio, then narrow its knowledge to what a caller would plausibly ask on the phone. Long written answers do not survive being read aloud.
4. Write the greeting. It carries the business name, the disclosure if required, and one open question. Keep it under about two sentences.
5. Pick the voice and set the speaking rate. Then call it yourself and listen, because voices that read well in a preview can sound wrong at speed.
6. Configure interruption handling and silence timeouts. An agent that talks over callers or waits too long is unusable regardless of how good its answers are.
7. Set the call objectives as a short ordered list, and keep it short. Three or four things is realistic on a first call; eight is not.
8. Connect the booking calendar and set how times are offered. Reading out a long list of slots fails on the phone; offer two or three and let the caller pick.
9. Configure transfer: the destination, the trigger conditions, what the agent says before transferring, and what happens if nobody picks up.
10. Set the maximum call duration and the graceful wrap-up behavior.
11. Decide what is written back after the call: summary, outcome, fields and tags, plus the workflow that fires on each outcome.
12. Test by calling as several different kinds of caller: the easy booking, the one with an accent or background noise, the one who interrupts, the one who asks a price question, the one who immediately asks for a human, and the wrong number.
13. Listen to every recording from the first day of live traffic, not a sample. Voice problems cluster in the first minute of the call and are easy to fix once heard.
14. Roll out gradually: overflow or after-hours calls first, then all calls once the client is comfortable. Keep a documented way to switch the agent off instantly.

Voice sits beside the text agent rather than replacing it, and it leans heavily on Calendars for booking and Workflows for everything that happens after the call.

After-hours and overflow answering

Calls ring the office first and reach the agent only when unanswered or out of hours. The agent handles routine bookings and questions, and anything else becomes a task with the transcript attached for the morning. Nobody reaches voicemail, and the client keeps control of the calls they want to take themselves.

One-minute outbound follow-up on web leads

A form submission triggers a workflow that places an outbound call within a minute. The agent confirms interest, asks the qualifying questions and books. If the lead does not answer, the workflow falls back to SMS and the text agent picks up the conversation.

Appointment confirmation and recovery

Before each appointment a workflow calls to confirm. The agent reschedules on the spot where it can, and where the caller wants to cancel outright it moves the opportunity and releases the slot. Reducing no-shows is usually the easiest voice use case to justify to a client.

• Latency is the whole experience. A delay before each reply makes an otherwise good agent feel broken. Judge the agent by calling it, never by reading transcripts alone.
• Transcription is the weak link. Names, street addresses, email addresses and spelled-out details are frequently misheard. Have the agent confirm anything it will write to the record, and prefer sending a text for anything that must be spelled.
• Accents, noise and speakerphone. Real callers phone from cars and worksites. Expect worse recognition than your own test calls suggest.
• Recording and disclosure law varies. Consent requirements for recording, and rules about disclosing an automated caller, differ by jurisdiction and change. This is a compliance decision for the client, documented, not a default to leave on.
• Outbound calling carries more regulatory weight than inbound. Time-of-day restrictions, do-not-call obligations and consent requirements apply to the call itself. Never enable outbound campaigns without checking.
• Transfer failure is invisible. If the transfer destination is busy or unmonitored, the caller is dropped and nothing in the record says so clearly. Always configure and test the no-answer fallback.
• Voice minutes and concurrency are metered. Long calls and simultaneous call volume consume capacity, and a runaway outbound workflow can burn through it fast. Cap durations and rate-limit outbound.
• Reading lists aloud does not work. Availability, service menus and policies have to be summarized for speech, not recited.
• Callers test the agent. Some will try to confuse it, and a few will be abusive. Your refusals and your transfer triggers are what keep that from becoming a story the client hears about.
• It is not a substitute for a receptionist in every context. Emergency, clinical triage and high-emotion calls should route to a person by rule, not by the agent's judgment.