Roughly 62 percent of calls to small businesses go unanswered during busy stretches, and about 85 percent of the people who land in voicemail hang up without leaving a message. They do not call back. Industry estimates put the average annual cost of those dropped calls above $26,000 per business, and past $126,000 for service companies with high job values. That is not a marketing problem. It is a staffing problem that now has a software answer, and the software got cheap enough in the last three weeks to be worth a second look.
Early September brought a wave of voice-agent launches priced for operators instead of call centers. ThunderPhone opened at two cents per minute. Amazon’s Nova 2 Sonic, the speech-to-speech model most of these platforms are quietly built on, runs about 1.5 cents per minute in raw model cost. Keiki shipped a customer-facing agent you deploy across phone, chat, and email from one workflow. The pitch is the same everywhere: nobody misses a call again.
The pitch is mostly true. The pricing is mostly not. Here is how to figure out whether a voice agent pays for itself at your shop, and how to test it in fourteen days without signing an annual contract.
Run the number before you look at a single vendor
Most operators evaluate voice agents by watching a demo. Do the arithmetic first, because the arithmetic decides the answer and takes eleven minutes.
Pull your call log. Every business phone system exports one, and if yours does not, your carrier’s portal will. You want four numbers from the last 90 days:
- Total inbound calls.
- Calls that rang out, hit voicemail, or were abandoned before pickup.
- Of those, how many came in outside staffed hours.
- Your average job value or first-year customer value.
Now multiply. A shop taking 500 calls a month with 10 percent going to after-hours voicemail loses 50 buying-intent calls monthly. Not all of them would have booked. Assume a third would have, at a $500 average ticket, and the gap is about $8,300 a month in work that walked. Against that, a voice agent covering nights and weekends costs real money but nowhere near $8,300.
If your number comes back under $1,500 a month, stop reading and go fix something else. Voice agents are worth the setup friction when missed-call revenue clears four figures monthly. Below that, a $40 answering service or a well-written text-back autoresponder wins on effort alone.
The advertised rate is not the rate
Two cents per minute is a real number describing an unreal product. It prices one layer of a stack that has five. Telephony carries the call. Speech-to-text turns audio into words. The language model decides what to say. Text-to-speech says it. The platform handles routing, calendar writes, CRM logging, and the dashboard you actually use.
Buy only the model layer and you pay a penny or two per minute, then spend eight weeks wiring the rest yourself. Buy the finished product and you land between $0.25 and $0.50 per minute on managed platforms with scheduling and CRM built in, and between $0.50 and $1.50 per minute once you count setup fees, number porting, and the minimum-commitment tiers most vendors push at signup. Infrastructure-layer tools that let you assemble your own stack start at $0.05 to $0.15 per minute and assume you have someone who can assemble it.
For a local business, the managed tier is almost always the right buy. Run the cost at your volume, not the vendor’s example. Fifty after-hours calls a month at four minutes each is 200 minutes. At $0.40 all-in, that is $80 a month against $8,300 in exposed revenue. The ratio is so lopsided that haggling over per-minute rates is the wrong fight. The right fight is whether the agent books the job.
Pilot on nights and weekends only
The single biggest mistake operators make is pointing a new voice agent at their main line on day one. Do not. Your staffed hours already work. Test where the alternative is voicemail, because voicemail is a floor so low the agent cannot underperform it.
A fourteen-day pilot looks like this. Day one, pick one managed platform and sign up month to month. Do not port your main number. Buy a new number from the platform and set your existing phone system to forward to it after hours only, which is a five-minute change in RingCentral, Dialpad, Google Voice, or whatever you run. Day one still, write the agent’s brief: your hours, your service area, your three most common call reasons, your pricing ranges, and the exact sentence it says when it does not know something. That last line matters more than the rest combined. “Let me have Trevor call you first thing tomorrow, what number is best” beats a confident wrong answer every time.
Day two, connect the calendar and the CRM. Every managed platform does Google Calendar and the big field CRMs out of the box. If yours needs a developer for this step, you bought the wrong tier. Day three, call your own agent nine times. Play the easy customer, the angry one, the one with an accent, the one who interrupts, the one asking a price question you refuse to answer over the phone. Fix the brief after each call.
Then let it run for two weeks and read the transcripts every morning with coffee. Twenty minutes a day. Transcripts are the whole product. They tell you what your customers actually ask, which is usually not what you assumed, and that alone has repriced more than a few service menus.
Measure booked jobs, not calls handled
Every voice-agent dashboard leads with calls answered and average handle time. Both are vanity. An agent that answers 50 calls and books zero is worse than voicemail, because the caller now believes they were helped.
Track three things on a sheet of paper taped next to the phone. First, appointments created by the agent that showed up. Not appointments created. Appointments that showed. Second, callbacks you had to make because the agent took a message it should have resolved. Third, complaints. One customer telling you the robot was frustrating outweighs ten clean transcripts, because the nine silent ones may have felt the same and just left.
The published benchmarks say 24/7 coverage lifts booked appointments about 27 percent within 90 days, and that after-hours agents recover 15 to 20 percent of calls that used to die in voicemail. Treat those as a ceiling someone else hit, not a forecast. Your own fourteen days of showed-up appointments is the only number that decides renewal. If the agent books four jobs in two weeks at your average ticket, the math closed itself. If it books zero and generates six callbacks, kill it and keep the phone forwarding to your cell.
One signal to watch over the next 30 days: the managed platforms are racing each other down toward the model cost, and Nova 2 Sonic at 1.5 cents a minute sets a floor nobody can undercut for long. Expect at least one major field-service CRM to fold a voice agent into its base subscription at no per-minute charge this quarter. If you are three weeks from renewing your CRM, ask your rep directly whether voice is coming and whether it is included. That question has gotten expensive to skip.
Atlas Unchained builds and runs this stack for local operators who would rather book the job than configure the software. If you want the pilot run for you, with the brief written and the transcripts reviewed, reach out. And subscribe below for one operator-grade AI play every weekday morning.
