ai-receptionist
How an AI Receptionist Actually Works (and Where It Fails)
Every vendor demo sounds the same. The AI answers the phone, sounds pleasant, books the appointment, and the room nods. Then the owner signs up, turns it on, and discovers the difference between a demo and a phone system that earns its keep for twelve months. That difference is not the brand of AI. It is whether anyone understood the machinery well enough to set it up and supervise it.
So here is the machinery, in the order a call experiences it, with the failure points marked along the way. By the end you should be able to judge any receptionist product, including ours, on what it does rather than on how the demo sounds.
The path a call takes, second by second
A customer calls on a Tuesday night because their water heater is leaking. Here is what happens before anyone at the business knows a call came in.
The phone rings twice at most, and the AI answers. Most owners hear that and think nothing of it. It is actually one of the strongest reasons to use one, because the business has no choice about answering; the system picks up whether someone is on a ladder, in bed, or on another line.
As the caller speaks, speech-to-text turns the words into text in near real time. Everything downstream depends on this step. If the transcription mangles "sump pump" into something unrecognizable, the best language model on earth cannot recover the meaning. The model reads the transcript, works out what the caller wants and what to do next, and composes the reply. Then the voice speaks it, and the loop repeats, every few seconds, until the call ends in one of four ways: a booking, a message taken, a handoff, or a goodbye.
That loop is the whole product. The parts inside it decide whether it works for your business or becomes a bad experience at two in the morning.
The five parts that decide quality
Strip any receptionist product down and you find the same five components. Each one fails in its own characteristic way, and knowing the failure modes is most of what you need as a buyer.
The first is the ears. Speech recognition on a phone call works with narrowband audio, which is a technical way of saying phone calls sound worse than your laptop. Add road noise, wind, a shop compressor, or a caller who mumbles, and transcription gets harder. Good systems handle this well and still miss sometimes; the mark of a good one is what happens after the miss, which is a clean retry: "Sorry, did you say the pump or the furnace?" The one thing no vendor can fix is a caller shouting over a leaf blower. Some calls will not be transcribable, and your setup should assume that.
The second is the brain, the language model that decides what to say. It is following instructions you or your provider wrote, plus whatever it has been given about your business. This is where tone lives. A system can transcribe perfectly and still sound like a robot reading a script, or it can sound like a competent office manager. The difference is almost always the instructions, not the model.
The third is the script, the written instructions that tell it what to do. What to ask, in what order. What it may promise and what it may not. When to stop and hand a call to a human. When a caller is describing an emergency that needs a real person immediately. This document matters more than any other part of the setup, and it is also the part most owners never see. Ask for it. If a vendor will not show you the script, they are selling you a black box, and you will have no way to fix its behavior later.
The fourth is the hands, the connections that let it act. A calendar it can book into. A contact record it can save the caller's details to. A text or email it can send. Without these, the receptionist can only talk, which means it takes excellent messages and nothing more. The tools are what turn a call into a booked job.
The fifth is interrupt handling, which most buyers never think about until they hear it go wrong. A real caller does not wait politely for the machine to finish. They interrupt, they change direction, they say "no wait, actually Tuesday is bad." The system needs to stop talking when they do, and it needs silence detection so it knows the caller has finished a thought. It also needs to recognize voicemail greetings and fax tones so it does not leave a rambling message to a machine. Listen for all three in any trial call.
What "trained on your business" actually means
Vendors love that phrase, so it is worth decoding. Usually it means some combination of three inputs: documents you upload, your website, and the written instructions. The documents are typically a services list with prices or price ranges, hours, service area, policies, and answers to the questions callers ask every week.
This works well when the material is clean, and badly when it is not. The classic failure is a knowledge base with old prices in it: last year's rates, a draft quote, two documents that disagree. The AI will pick one and speak it with total confidence. If your receptionist ever quotes a price you did not approve, check the knowledge base before you check the model. In my experience the error is nearly always stale or contradictory source material, which is why the setup phase is as much about pruning as about adding.
The other input, your website, is usually helpful and occasionally a trap. A site that says "we install and repair" without saying what you refuse makes the AI answer questions you would rather it deflected. Anything you would not want a new hire to promise a customer belongs out of the material it can read.
What it costs to run, and where the money goes
Because the components are separate products, the bill usually arrives in parts: a platform fee for the service itself, a per-minute charge for the minutes it handles, a per-minute charge for speech to text, and a per-minute charge for the AI voice, which varies by provider since the same call costs different amounts to speak depending on the voice engine behind it. If the receptionist sits on a real phone number, there may also be a line fee. Some vendors bundle these into one rate and some pass them through. Ask for the split, because bundled pricing hides the one lever you control: how long calls last. A script that gets to the point in ninety seconds costs a fraction of one that rambles its way to the same booking. Test calls cost almost nothing. Production calls are where the budget goes, which is one more reason to do the test phase properly.
What a good call sounds like
An example makes the machinery easier to judge. A caller at 9 PM: "Hi, my water heater is leaking all over the basement floor." A well-scripted receptionist does three things in the first two exchanges. It acknowledges the urgency and asks the one question that matters: "Is the water valve shut off, or is water still running?" It offers a real slot: "I can put you down for first thing tomorrow at 8 AM." And it collects what the technician needs: the address, a callback number, and whether anyone will be home.
The same call with a bad script sounds different. It asks five questions in order, including two the caller cannot answer without going downstairs. It quotes a price it was never authorized to quote. It finishes with "Is there anything else I can help you with today?" The caller hangs up and dials the next plumber.
The difference is not intelligence. It is the script's sense of what the call is for. Every receptionist product is only as good as the document behind it, which is why the setup phase below deserves the time.
Where it fails, honestly
A receptionist that never fails does not exist. The honest ones fail in patterns you can predict, and each pattern has a known fix.
First, the messy call. Two people talking, a dog, a contractor van idling. Transcription degrades and the AI may ask the caller to repeat themselves. This is annoying and correct. The alternative, guessing, is worse. If a system you are testing never asks anyone to repeat anything, that is a reason to be suspicious, not impressed.
Second, the question it has no answer for. Permit questions, a warranty clause nobody documented, a dispute about a job from two years ago. A well-scripted system declines gracefully: "That one is above my pay grade, let me take your number and have the owner call you back." That is a success, not a failure. The failure is improvising an answer. Your script should make refusing explicit, because the model will fill gaps if you let it.
Third, wrong confidence. The rarest and most expensive failure mode: the AI states something false, politely, with no hesitation. Good setups reduce this with guardrails in the instructions, a narrow scope of what it may discuss, and a hard rule that anything money-related gets deferred to a human. This is also why call recordings matter. You cannot fix what you never hear.
Fourth, the emotional call. A customer is furious about a four-thousand-dollar bill. A good script hands that call to a human immediately, and it says so out loud rather than pretending to sympathize while stalling. Some products do this well. Some do not, and the difference shows up in reviews, so test it before you trust it with your angriest customers.
Fifth, consent and disclosure. Recording rules vary by state, and there is a growing expectation that an AI identifies itself if a caller asks. Your script should include a plain acknowledgment and your policy should say who reviews recordings. None of this is difficult. It just has to be decided by a human, in advance.
One more thing it will never do: the judgment calls. It cannot diagnose a dying furnace over the phone, price a complex retrofit, or decide to refund an angry customer. Those stay with you, which is fine. The tool's job is to make sure nothing falls through the cracks so that your judgment is spent on the calls that need it.
Setup, realistically, in four phases
The setup is where products succeed or fail, and it has a shape:
Phase one is intake. Someone writes down the services, the service area, the hours, the policies, and the twenty questions callers actually ask. Then the phone number and calendar get connected so the system can answer and book. This phase is mostly writing, and it is worth doing slowly, because everything after it inherits these decisions.
Phase two is test calls, and it is the phase most owners skip. Call it like a customer would, from a noisy truck if you can manage it. Twenty to thirty calls. Ask the questions a real caller would. Try to break it on purpose. Log every stumble: the misheard word, the promise it should not have made, the moment it rambled. Then fix the script and call again. A product that does not allow this phase, or makes it awkward, is telling you something.
Phase three is a soft launch. Keep your phone ringing to a human first, with the AI catching overflow and after-hours calls only. This is the safest possible handoff, and it gives you two or three weeks of real recordings to review before it handles anything alone. Watch what callers ask that you did not anticipate. Fix the script again. The AI is now earning its keep on the calls you could never answer anyway.
Phase four is the full handoff, once the recordings have earned trust. And setup is never really over: a monthly review of recordings, a look at the calls it could not handle, and a script update when services or prices change. That cadence is the difference between a tool that stays sharp and one that drifts into quoting last year's prices.
What it is not
A few honest limits, because every vendor demo skips them. It is not a technician and cannot diagnose anything. It does not know your profit margins unless you tell it, and it should never negotiate. It costs money per minute or per call, so a business with light call volume may not justify it, which is why the honest answer for some owners is a simple after-hours text-back instead. And it will occasionally frustrate a caller who simply wanted a human. The way to blunt that is to make the handoff easy, immediate, and shameless: "Absolutely, I can have the owner call you back in the morning, or you can press one to leave a message right now."
Where it sits among the alternatives
An owner comparing options is really choosing between four ways to handle calls nobody answers. Voicemail, which is cheap and loses most callers, because the ones with emergencies call the next listing. A human answering service, which works, usually runs several hundred dollars a month, and can only relay information rather than book into your calendar. A part-time front-desk hire, which gives the best experience during business hours and does nothing after 5. And the AI receptionist, which answers every call in seconds, books from your real calendar, and works at midnight, but needs a script and supervision to stay sharp.
The right answer depends on volume. A shop that misses a handful of calls a week may do fine with a text-back and a good voicemail greeting. A shop drowning in overflow, or one that lives on after-hours emergency calls, gets the most from the AI option. The honest comparison is not "AI versus human." It is "what happens to the calls nobody answers," and the businesses that win that comparison are the ones that measure it.
What to ask any vendor before you sign
- Can I make my own test calls during the trial, from a noisy environment?
- Show me the script. Who can edit it, and how fast does a change go live?
- What happens when it does not know an answer? Play me the recording of that moment.
- What exactly can it book, and into what calendar? Can it see real availability?
- Who reviews the recordings, and how often? What is the escalation path when it fails mid-call?
- How are callers told they are talking to an AI, and what are my recording obligations?
A vendor who answers those easily will be easy to work with. One who answers in adjectives will be hard to fix later, because the script is the product and you have not been allowed to see it.
The verdict
An AI receptionist is one of the better deals in service-business software right now, and it is not magic. It answers every call, books from real availability, and takes messages a human would have missed. It also mishears things, cannot handle anger, and will occasionally need you to fix the script. The businesses that get value from it are the ones that treat the first month as a setup period with real supervision, not a fire-and-forget install.
If you want help evaluating one, or setting one up with the script done right the first time, that is part of what we do. Book a growth assessment and we will walk your current call flow, or take the growth scorecard first and see how your lead handling scores.
Want a second opinion on your marketing?
Marvin Growth Partners works with service businesses across Michigan and nationwide. Book a Growth Assessment or take the Growth Scorecard to see where your lead flow stands.