AI receptionist or a human answering service: how to pick without guessing

The short answer, before you scroll

Pick based on how many of your calls need a decision made on the spot. If most of your inbound calls are people asking for a time slot, a location, your hours, or leaving their name and problem, an AI receptionist handles that all day and costs you a flat monthly number. If most of your calls are people asking what it costs, whether you can come today, or whether you'll take their insurance, you need a human, and you'll pay by the minute for the privilege.

Everybody argues about whether the robot sounds human. Wrong fight. Your customer doesn't care if the voice is synthetic. They care whether they got an answer. A friendly human who has to say "I'll have someone call you back" loses to a bot that books the appointment, and a bot that says "I'll have someone call you back" loses to everything.

I run an AI receptionist named Maddox on our real business line. Not a demo. The actual phone. I've read the transcripts, and I've moved call types off of it when they weren't working. That's where the rest of this comes from.

Sort your calls into four buckets first

Before you shop anything, spend one week writing down what people actually call you about. Four tally marks on a sticky note by the phone. The buckets:

• Intake. New person, has a problem, wants somebody to know about it. Name, number, address, what's broken, when they're free.
• Scheduling. Book, move, confirm, cancel. The caller already knows what they want.
• Quote and dispatch. "What's this going to run me?" "Can you be here this afternoon?" "Is that covered?" Somebody has to make a judgment call.
• Noise. Vendors, spam, solar guys, the SEO company that calls three times a week.

Now count. If intake plus scheduling is the bulk of your week, you're an AI candidate. If quote and dispatch is the bulk, you're a human-service candidate, or you need a person on your own payroll answering the phone. Most of the trades shops I sit with around Johnstown and Greeley are heavier on intake than they expect, because the customer with a water heater leaking into the garage isn't price shopping. They're panic shopping.

Where a human answering service still wins

Three places, and they're real.

Judgment. A roofer's phone rings at 7pm after a hailstorm. The caller is upset, half the story is missing, and the right answer depends on reading the person. A human picks up on that. Current AI does not, no matter what the sales page says.

Rare, high-stakes calls. Clinics especially. If one call in two hundred is somebody who needs to be told to hang up and call 911, you don't put that on software and go to bed. You route it to a person and you write the script with your own lawyer.

Low volume. If you get eight calls a week, a per-minute service might genuinely cost you less than a flat software subscription. Do the math before you get clever.

The flip side: you are paying by the minute, so the service is structurally motivated to keep talking, and the agent knows nothing about your business beyond the script you gave them. Ask any answering service how they handle a question that's not on the script. The honest ones say "we take a message."

Where an AI receptionist wins

Volume, speed, and consistency. It answers on the first ring at 2am on a Sunday with the same information it gave at 9am Tuesday. It never has a bad day. It doesn't quit in March. And it writes down every single call, word for word, which turns your phone into a data source instead of a mystery.

That last part is underrated. When I started reading Maddox transcripts, I learned things about our own intake I would never have learned otherwise, including which questions callers ask that we had no good answer for. You cannot get that from a human service's message log, because the log is a summary written by somebody who wasn't paying full attention.

Where it breaks, in my experience: anything that requires the answer "it depends." Pricing with variables. Scheduling against a crew calendar that lives in somebody's head. Anything emotional. When I saw those show up in the transcripts, I routed them to a person. Not because the technology failed, but because I'd asked it to do a job that needed a decision-maker.

One more thing nobody mentions: an AI receptionist is only as good as what you feed it. If your hours, service area, and pricing ranges aren't written down anywhere, the bot is going to improvise, and you're going to hate it. Half of setting one up is finally writing down how your business works. That's a feature.

The actual cost math

Human answering services typically bill per minute or per bundled block of minutes, and as a rule of thumb that lands somewhere in the one-to-two-dollar-per-minute range depending on the market and how much training you want. AI receptionists typically bill a flat monthly fee, sometimes with a per-minute overage.

So run the hypothetical. Say you take 120 calls a month and the average call is three minutes. That's 360 minutes. At $1.50 a minute you're at $540 a month, and that's before the after-hours surcharges and the holiday rates most contracts carry. A flat AI subscription at, say, $200 a month handles the same 360 minutes and the 361st for free.

Now flip it. Say you take 25 calls a month at three minutes. That's 75 minutes, $112 at the same rate. The flat subscription just got expensive.

The number that actually matters isn't either of those. It's what a missed call costs you. If your average job is $800 and you miss four calls a week that would have closed at one in three, you're leaving roughly $4,200 a month on the table. Against that, arguing about $200 versus $540 is meshuggah. Both options beat voicemail. Pick the one that fits your call mix and move on. If you want a second set of eyes on which of your numbers is actually bleeding, that's what the free Business Checkup is for.

The hybrid most small shops land on

Here's the setup I see work most often, and it's what I'd build for a plumbing outfit in Loveland or a clinic in Windsor:

1. AI answers every call on the first ring, always.
2. It handles intake and scheduling completely. Name, number, address, problem, slot booked, confirmation text sent.
3. Anything with the words "how much," "today," "emergency," or "insurance" gets warm-transferred to a real person, which during business hours is you or whoever's at the desk.
4. After hours, those same calls go to a short human service plan or a text to your cell, depending on whether you want to sleep.
5. Every transcript lands in your inbox or a shared folder daily.

You end up paying a flat fee for the volume and per-minute for the exceptions, which is the right way around. The exceptions are the calls worth money.

Text is the third lane and almost nobody uses it. SMS open rates run far above email. A customer who won't wait on hold will absolutely text a photo of the leak. Build that door too.

How to test either one in 30 days without signing anything long

How do you eat an elephant? One bite at a time. Don't rip out your phone system. Do this instead:

• Week zero. Forward only your overflow and after-hours calls to the new service. Your main line stays exactly as it is. If the thing is terrible, your business doesn't notice.
• Negotiate month-to-month. Every vendor in this space will do a 30-day term if you ask. If they won't, that tells you what their retention looks like. Ask directly: "What happens if I cancel in week three?" Don't accept a yes-or-no answer, make them explain the mechanics.
• Record your own baseline. Before you start, count missed calls and voicemails for one week. You can't tell if it worked if you don't know where you started.
• Set one number as the test. Appointments booked from inbound calls. Not "call quality," not vibes. Booked jobs.

Thirty days, two services, same forwarding rules, same scorecard. Then cancel the loser. If a vendor's pitch can't survive that, we'll run it through a hype audit and tell you what's real.

What to listen for in the first week of transcripts

This is the part everybody skips, and it's the whole value. Block 45 minutes on Friday and read every transcript from the week. You're looking for five things:

• Where the caller repeated themselves. That's a comprehension failure or a bad script. Fix the script.
• Where the caller got transferred and shouldn't have. Those are answers you should have written down.
• Where the caller hung up. Note the timestamp. If they bail at 40 seconds, your greeting is too long.
• Questions you have no answer for. Every one of these belongs on your website and in your pricing page. If three people a week ask what a service runs, publish the range.
• Calls that weren't customers. Build a block list. You're paying for those minutes on a human service.

You'll fix more in that one Friday session than the vendor will fix in six months. The transcripts are the asset. The answering is just the delivery mechanism.

And yes, make sure your call recording and transcript practice complies with Colorado law and, for clinics, with HIPAA. Colorado is a one-party consent state for recording, but a health practice has a different bar. Ask the vendor for a signed business associate agreement in writing before a single patient call touches their system. If they hesitate, you have your answer.

Questions I get about this

Todd Schreier has owned a bakery, built escape rooms, rehabbed houses and sold businesses. He runs Schreier Group with his wife Amanda out of Johnstown, Colorado. Call him at (970) 591-4263.