A voice agent is not a cheaper receptionist, and selling it that way is how it ends up doing the wrong job badly. It is coverage: for the third call arriving while the first two are in progress, for 9pm, for the fourth person this week asking about parking. What it buys is that the person at the desk reaches the calls that actually need a person, sooner.
Definition
AI voice agent
Software that answers a phone call, holds a spoken conversation, and completes an action at the end of it: booking an appointment, creating a record, or transferring to a person. Unlike a phone menu it does not make the caller navigate options; unlike voicemail it can finish the thing the caller rang about.
What the person at the desk is genuinely better at
Everything that requires reading a situation. The caller who is upset and has not said why. The one whose question is technically about scheduling and actually about money. The regular whose name matters. The call that does not fit any pattern, which is a larger share of calls than any flowchart suggests.
Judgment, discretion and relationship are not a weaker version of what an agent does. They are a different thing, and the agent's job is to protect the time for them.
What an agent is genuinely better at
- Being available. No queue, no lunch, no handover. The third simultaneous call is answered the same way as the first.
- Repetition without degradation. The two hundredth time it explains where to park, it explains it the same way.
- After hours. A useful share of calls arrive when nobody is there, and voicemail converts a caller into a task nobody wants.
- Writing things down. Transcript, summary, outcome and next step into the CRM, every time, without anyone typing.
Escalation is the whole design
The triggers get written before the build and they belong to the business, not to us. A workable default set:
- 1.Anything clinical, legal or otherwise professional.
- 2.Anything about money: refunds, disputes, pricing outside the published list.
- 3.Any complaint, at the first sign of one.
- 4.Anything the caller has now said twice, which usually means it was not understood the first time.
- 5.Anything the agent has not understood.
The design assumption behind that list is that an unnecessary escalation costs a minute and a missed one costs a customer. Tune toward escalating.
Should it say it is an AI?
Yes. Build agents that identify themselves. The alternative works right up until someone notices, and then it costs more trust than the deception ever saved, and people do notice.
In practice disclosure is rarely the problem. Callers object to being trapped: looped through the same question, unable to reach a person, made to repeat themselves. An agent that says what it is, does the routine thing quickly and hands over cleanly does not trigger that objection.
What "booking" has to mean
It has to mean a slot is held. An agent that takes down a callback request has moved the work rather than done it. Someone still has to ring back, and the caller still does not know when they are being seen.
That means live calendar access, read and write. If the scheduling system cannot expose that, it is worth knowing before the project starts rather than at the end of it.
What it costs, honestly
There are two costs and they are different in kind. The build is a one-off. The running cost is telephony and model usage, billed by the minute on accounts that should be in your name, which means a busy month costs more than a quiet one, and a badly built agent that lets calls run long costs more than a well built one that does not.
Anyone quoting a flat monthly number for unlimited voice minutes is either capping something they have not mentioned or pricing in a margin on usage. Ask which.
When it is the wrong first build
If most of your inquiries arrive as forms and messages rather than calls, a voice agent is an expensive way to improve the smaller channel. Count where inquiries actually come from before buying the thing that handles the one people talk about most.
