operations
The bar isn't a perfect receptionist
Worried an AI will embarrass your business on the phone? Good — that instinct is what makes a voice agent safe. Here's what it should refuse to do.
Most owners who hesitate on this aren’t worried it won’t work. They’re worried it will work badly in public — that it will quote a price it can’t honor, promise a truck that isn’t coming, or argue with someone whose basement is flooding.
That worry is correct, and it’s the right thing to design around. But notice where it points. It’s an argument for a tightly bounded agent, not an argument for voicemail. Voicemail cannot embarrass you, in the same sense that a locked door cannot embarrass a shop.
Judge it against what it replaces
The instinct is to compare a voice agent to your best employee. Do that and it loses — of course it does. Your best employee has been doing this eleven years.
The honest comparison is to the thing actually handling those calls right now, which is nothing. Against nothing, the bar is low and worth stating plainly. A voice agent has to:
- Answer, so the caller doesn’t move on
- Sound like a competent person representing your business
- Get the details a dispatcher would need
- Know the edge of its own knowledge and stop there
- Hand off anything it shouldn’t touch, immediately and clearly
That’s it. Nothing on that list requires brilliance. All of it requires discipline.
A good agent is defined by what it refuses to do
The quality of a voice agent is mostly a function of its constraints. The things it should decline are more important than the things it handles:
It shouldn’t quote a firm price sight-unseen. A range with an on-site confirmation is honest and closes fine. A hard number invites an argument in someone’s driveway on a Tuesday morning.
It shouldn’t promise an arrival time it can’t verify. Offering a window your schedule actually has is a booking. Inventing one is a complaint you’ll receive later.
It shouldn’t give safety or medical advice. “If the smell gets stronger, hang up and call 911” is a boundary. Anything past that is not its job.
It shouldn’t negotiate. Discounts, price matching, warranty disputes — a person owns those, and the caller should be told a person will call.
It shouldn’t handle an emergency alone. Escalation rules exist precisely so urgency exits the automated path immediately.
Each of those is a place where an unconstrained agent could genuinely embarrass you. Each is also a place where a bounded one simply says it will get a person involved — which is not a failure, it’s the correct answer.
“I don’t know” is a feature
There’s a version of this technology that improvises, and it is the one people are actually afraid of. An agent working from a defined set of business facts — your hours, your service area, how your pricing works, what you don’t take — behaves differently. Asked something outside that, it takes a message with the specific question recorded.
That’s a materially better outcome than what voicemail produces. Your callback starts from “she wants to know whether you haul construction debris from a second-floor walkup” instead of “she asked a question, call her back.”
An agent that knows twenty things about your business and admits the twenty-first is outside its knowledge is more useful than one that knows five hundred and guesses at the rest.
Where it genuinely shouldn’t be the answer
Being straight about this matters more than the pitch:
- Active emergencies. Escalate to a human on the first sign. Set those rules before you turn anything on.
- An angry existing customer. Somebody who is calling about a job that went wrong should reach a person, and should be routed there without having to explain themselves twice.
- Anything with legal or safety exposure. If a wrong answer creates liability, it isn’t a candidate for automation.
Write the escalation rules first. Before you tune a greeting or pick a voice, define what counts as urgent in your trade and what happens when it’s detected — collect and flag, or attempt a live transfer. An agent that books beautifully and mishandles one gas-smell call has cost you more than it made.
The failure modes aren’t symmetrical
An imperfect voice agent’s bad day is a slightly stilted call that still ends with a lead in your queue and a name you can follow up.
Voicemail’s bad day is a person who needed help, got a beep, hung up, and hired someone else — and you will never learn their name, so you will never know it happened.
Only one of those failure modes shows up in your revenue.