“AI receptionist” is a job title doing a lot of marketing work. It sounds warm and competent, and it tells you almost nothing about what you’d be paying for. In the first post of this series I argued that AI takes over the logistics inside sales and leaves the promises to people. This post is that argument at ground level: the shift log of an AI receptionist in a small business, hour by hour, followed by the section vendor websites skip, which is where it breaks and a person has to take over.
Same disclosure as last time: I build these systems for a living, so read everything here with that in mind. And to be precise about what follows: this is not a case study with a happy client attached. It’s a job description for a clinic or a contracting business. If a vendor’s pitch can’t survive being written as one, that tells you something too.
The shift log, hour by hour
6:52am. A call before opening. A homeowner whose water heater died overnight, or a patient whose weekend toothache got worse. The phone gets answered on the second ring, the caller learns whether this is something you handle and when you can see them, and the appointment lands on the calendar before anyone unlocks the office. That caller was working down a list. The list ended at the first business that answered.
10:15am. The twelve questions from the first post, again: price range, hours, insurance, how long does it take. In my market, the Baja California and San Diego corridor, those questions arrive in Spanish, in English, and often both inside a single conversation. A receptionist that handles one language is answering half a phone line.
1:30pm. A patient’s Tuesday fell apart and Thursday’s appointment no longer works. Most no-shows start exactly here: someone’s week changed, and fixing it required a phone call during the exact hours they couldn’t make one. A reminder that says “reply here to reschedule” turns a hole in Thursday’s schedule into a booking for next week instead of an empty chair and a shrug.
9:40pm. The Saturday-night form fill from the first post. Answered in seconds instead of Monday morning, asked the qualification questions (budget, timeline, is this something you actually do), and offered real open slots on a real calendar. The buyer’s research moment and your response finally happen at the same time.
3:00am. Nothing happens. Which is part of the job: being on the phone in case something does, at an hour no business owner should staff and no employee wants.
Read back through the log and check it against the rule from post one. Every entry has a checkable done state: the call got answered, the answer was correct, the appointment exists. Standard outcomes, machine’s side of the line. Nothing on that log persuaded anyone of anything.
The handoff is the design decision that matters
Sooner or later a conversation goes past the standard cases, and the system has to pass the caller to a person. How that handoff works matters more than which AI model sits underneath, and it’s the part I’d inspect first in anyone’s product, including mine.
A good escalation looks like this: the human picks up already knowing the caller’s name, what they asked, what they were told, and where the conversation stopped. The caller never repeats themselves. A bad escalation is one you already know from every phone tree you’ve ever hated: transferred, then asked to start over. That design makes the caller pay the cost of the machine’s limits, and callers remember who made them pay it.
When you evaluate a vendor, ask to watch an escalation, not a demo. The demo shows the happy path. The escalation shows whether anyone designed for the day the happy path ends.
Where it breaks
I promised in the first post to show reasoning instead of asking for trust, so here is the part of the job this software does badly. There are four situations, and they share a shape.
The angry caller: someone calling to complain is not looking for information. They want a person to receive the complaint and own it. A perfectly polite bot makes an angry caller angrier, because politeness without authority reads as deflection. The right behavior is narrow: recognize heat early, apologize once, and get to a human fast, or take a message with a committed callback time and make sure the callback actually fires. Anything cleverer backfires.
Questions it must refuse: a clinic’s line gets asked whether a symptom can wait until Monday. A contractor’s line gets asked whether a wall is load-bearing. Any answer, even a careful one, is practicing medicine or engineering over the phone, and the correct response is a refusal plus a route to the professional. The uncomfortable part, from my side of the fence: a language model’s default instinct is helpfulness. Making it refuse reliably takes deliberate engineering, and a vendor who hasn’t done that work is selling you a liability with good manners.
The caller who needs to be heard more than answered: some calls to a clinic are fear phrased as a scheduling question. A machine will book the slot efficiently and miss the person entirely. I don’t have a clean fix for this one. Escalating on obvious cues helps; some of it is simply the ceiling of the technology, and pretending otherwise would break the rule this series runs on.
The ambiguous request: “Can you match this tile?” “Does my insurance cover this specific procedure?” AI is good at producing a plausible answer and bad at knowing when plausible isn’t right, and these are the moments that weakness was made for. A wrong confident answer costs more than “let me have someone look at that and call you back,” every time.
All four are judgment moments, and judgment moments belong to a person. A well-built system earns its keep by knowing which moments these are and getting out of the way. So when you evaluate one, weigh the refusals more heavily than the answers.
What it costs against what a hire costs
If you have a good receptionist, this software is not their replacement, and a vendor who implies otherwise hasn’t watched a good receptionist work. During business hours, a person at the desk who knows the regulars and reads the room beats the machine at the parts of the job that matter most.
The honest math is about the other hours. A week has 168 of them. A full-time front desk covers about 40. For the remaining 128, which is every evening, weekend, lunch rush, and sick day, the desk is empty and the phone still rings. What you’re buying is coverage for hours you were never going to staff, at software prices rather than salary prices. Framed that way, the only thing it competes with is voicemail. What that coverage should cost, and how to run a pilot that proves it before you commit, is post four’s job; the framing to hold onto now is that the comparison is against your unanswered calls, not against anyone’s paycheck.
Six questions to ask any vendor, including me
- Show me an escalation, not a demo. A real conversation where the system hit its limit and handed off. Did the human get the context? Did the caller repeat themselves?
- What’s on the refusal list? Which questions will it decline on purpose, and what happens next for the caller when it does?
- Walk me through an angry caller. If the vendor’s answer is that the AI de-escalates, keep your wallet closed.
- What does it do when it doesn’t know? Acceptable answers involve saying so and routing to a person. Improvisation is a defect, however fluent.
- Which languages can it hold, and can it switch mid-conversation? If your customers switch, single-language coverage is half coverage.
- Which number on my calendar does this move, and how do we measure it from my current baseline? No number, no deal.
A vendor who gets uncomfortable at the first two questions has told you what the demo was hiding. Ask me the same six. A system built the way this series recommends passes them; a system built for the demo doesn’t.
The takeaway
An AI receptionist is a coverage machine with good manners. It answers at 6:52am and 9:40pm, in both of your market’s languages. It reschedules instead of losing the booking. It qualifies the Saturday-night lead while your competitor’s voicemail collects another message nobody will return. And it breaks on anger, on questions it should refuse, on people who need a person, and on ambiguity, which is why the handoff to a human is the part worth inspecting hardest. The line from the first post holds at the front desk: standard outcome, the machine owns it; promise or judgment call, a person owns it.
Next in the series: speed to lead, the one claim in AI sales you can verify with a stopwatch, and the free audit you can run on your own business this week.
