The short version
An AI answering service is not competing with your receptionist. It is competing with your voicemail, and voicemail loses every time. On a real plumbing line over three months, the agent answered every call inside a second, booked the routine work without help, and had to hand off roughly one call in six. The value was almost entirely concentrated in the hours nobody was covering: after five, weekends, and the lunch hour. The way to decide whether it is worth paying for is not to ask whether it sounds human. It is to count the calls that currently ring out, multiply by your close rate and your average job value, and compare that number to the monthly fee. For most shops taking twenty or more calls a week, it is not close.
Why we tried it at all
We did not go looking for this. A client with four vans kept losing work in the late afternoon and had already tried the two obvious fixes. They paid for an answering service for six months, which took messages nobody returned fast enough. Then they tried rotating phone duty among the office staff, which worked until the first week somebody was off sick. They asked us whether the robot thing was real yet. We said we did not know, and offered to find out on their line rather than guess on a call.
So for ninety days, every call that rang more than four times on their main number went to a voice agent instead of voicemail. Their office kept first crack at every call. The agent only ever got what the humans did not pick up, which turned out to be the most honest way to test it, because it removed the argument about whether it is better than a person. It was never up against a person. It was up against nothing.
What it handled without help
The routine stuff, which is more of the phone than anyone likes to admit. Booking a maintenance visit. A customer asking when the tech is arriving. A repeat customer with a slow drain who wanted the next available slot. A landlord booking access for a tenant. In each case the agent asked for the address, the problem, how long it had been going on, and whether anyone would be home, then read back a time and put it on the calendar.
The transcripts are boring to read, which is the highest compliment available here. Nobody wants a charming receptionist. They want the appointment on the right day with the right address and a note that the side gate is locked.
It also killed the sales calls. Roughly one in nine inbound calls to that number was somebody selling SEO, merchant services or a directory listing. The agent handled all of them politely and none of them reached a human. The office manager mentioned this more often than she mentioned the bookings.
What it fumbled
Two calls in the first month went badly enough that we had to change the script. One was a customer with a strong accent that the agent misheard twice, and rather than transferring, it asked a third time. That is the failure mode that turns a neutral experience into a story someone tells. We changed the rule to two attempts and then a transfer, no exceptions.
The second was a caller who was clearly upset about a previous job. The agent tried to book a new appointment, which is exactly the wrong response to a complaint. We added a rule: any call that mentions a previous visit going wrong stops the script and rings the owner directly. There is no version of a voice agent that should be handling a complaint.
Across three months, about one call in six was handed off for something other than an emergency. That is a fine number, and anybody quoting you a lower one is either counting differently or selling something.
Where all the value actually was
Not in the middle of the day. Between nine and three, the office answered nearly everything, and the agent took maybe two calls. The entire return sat in three places.
After five in the afternoon, which is when homeowners get home and start calling about the thing that has been bothering them since morning. Weekends, which had previously been an on-call mobile that people were reluctant to ring. And the lunch hour, which nobody in any shop wants to own.
That is the honest sales pitch, and it is smaller and more specific than the one you get on a demo. You are not automating your office. You are buying coverage for the hours where the alternative is a beep and a request to leave your name.
Doing the maths on your own line
Ignore the marketing numbers. Yours are in your phone provider's call log and they take an hour to pull.
Count inbound calls that went unanswered, plus anything under fifteen seconds, for one month. Those are your misses. Multiply by the share of answered calls that become booked jobs, which you know roughly. Multiply by your average job value. That is the low estimate of what the misses cost, low because urgent callers skew toward the jobs you actually want and they are the least likely to leave a message.
Then assume the agent recovers somewhere between a third and half of them, because some callers will hang up on any non-human and some jobs are not real. If half of your recovered work still dwarfs the monthly fee, buy one. For the shop we tested on, the fee was somewhere under three percent of the recovered value, which is why the trial did not end after ninety days.
The disclosure question
Everyone asks whether customers can tell. Some can, most do not think about it, and a small number ask directly. We set every agent we build to answer that question honestly and offer a human, because the alternative is a customer who finds out later and feels handled. In ninety days, four people asked. One laughed and carried on booking. Nobody cancelled.
The bigger risk is not disclosure, it is tone. A voice agent that is overly chirpy on a call about a flooded basement reads as insulting. Short, calm and factual survives situations that friendly does not.
What we would not use it for
Quoting. The agent should never say a number, because a number said on a call is a number the customer heard as a promise. Gather, book, hand off.
Outbound calling. We do not build it and we will not, and not because of the technology. Getting a call from a machine you did not ask for is a different experience from having your own call answered by one, and the second is the only one we are comfortable shipping.
Anything where the intake has legal weight. There are trades where the first conversation carries obligations, and a script is not the right tool for that.
What to insist on if you buy one from anybody
Transcripts of every call, available to you, not summarised. Most of the value in the first fortnight is reading what it actually said and rewriting the bits that made you wince.
A hard transfer rule after two failed attempts. Ask any vendor what happens when the agent does not understand. If the answer is vague, the answer is that it keeps going.
Calendar booking, not message taking. A service that only takes a message is a voicemail that costs four hundred dollars a month.
A flat fee, or at least a clear ceiling. Per-minute pricing means your busiest week produces both your best revenue and your worst invoice, and you will find yourself hoping for a quiet month.
Your number stays yours. If setting it up requires porting your main line to a vendor, walk away. Forwarding is reversible in a minute; a port is not.
One number to watch after it goes live
Booked calls is the headline, and it is not the number that tells you whether the thing is healthy. Watch the handoff rate instead, month over month. In the first fortnight ours sat near one in three, because the script was missing questions the office asks without thinking. By month two it was one in six and it stayed there.
If that number climbs after launch, something changed on your side and nobody told the agent: a new service, a seasonal promotion, a second location, a price change people are calling about. The agent does not learn your business by itself, and treating it as a thing you set up once is the most common way these go stale.
Check it on your own site
Pull one month of inbound call data
Timestamp, duration and outcome for every call. If your provider cannot give you that file, replacing the provider is a better first move than buying an agent.
Chart the misses by hour
Unanswered plus anything under fifteen seconds. You are looking for the two or three hours carrying most of the damage. Those hours are the entire business case.
Write your intake script by hand first
The four questions your office asks before a job can be booked, in order, in your words. If you cannot write them down, no agent will ask them properly, and the same exercise improves your human intake anyway.
Run it on overflow only for the first month
Humans keep first crack. The agent gets what rings out. You learn what it can do with no risk to the calls you were already handling well.
Read every transcript for two weeks
All of them, not a sample. Every rule worth having comes out of a call that went slightly wrong, and two weeks is enough to find them all.