What happens when an AI receptionist can't answer a call?

It says so, out loud, and then it follows the rule you wrote for that exact case. A well-built AI receptionist does not guess when a question falls outside what it has been trained on. It tells the caller it does not have that answer, then does one of four things you chose in advance: warm-transfers to a person, sends an SMS or email alert with the transcript attached, takes a detailed message, or books a callback slot. BotsDontSleep runs this 24/7 at $0.60 CAD per answered minute, billed by the second. No caller is dropped into voicemail without a follow-up, and the call you could not answer arrives in your inbox as a transcript rather than a mystery.

What does an AI receptionist do when it does not know the answer?

It admits it, and that is the design rather than a fallback. An agent asked something outside its training — your warranty terms, an invoice number, whether a part is on the shelf — says it does not have that and moves the call to the next step. It never invents an answer. That one behaviour separates a system you can put in front of a paying customer from one that will eventually quote a price you do not charge.

What happens after the admission is not left to the software to decide. Before launch you write rules for four situations: the question it cannot answer, the caller who asks for a person, the call that sounds urgent, and the call that arrives outside your hours. Each gets a destination, and the agent follows it every time without needing to be asked twice.

The distinction people underestimate is that "take a message" and "transfer to Sam's mobile" are both defensible answers to the same call at two in the afternoon. Picking the wrong one either costs you a job or rings Sam while he is under a car. The rules are yours for that reason.

What are the four ways a call can actually go wrong?

They are worth separating, because they have different fixes and only one of them is about the AI being clever enough.

• A knowledge gap. The caller asks something the agent was never given. Fixed by adding the answer, which takes minutes.
• A recognition error. It heard the street name or the phone number wrong. Fixed by confirmation habits, not by more training.
• A preference. The caller understood fine and wants a human anyway. Fixed by transferring quickly instead of arguing.
• An outage. Something in the chain is down. Fixed by where your calls fail to, which is a carrier setting.

Most of the anxiety business owners bring to this question is about the first one. In practice the second and third are far more common, and the fourth is the one worth checking before you sign anything.

Who decides when the call goes to a person?

You do, in writing, during setup. The escalation rules are configuration rather than a judgment call the model makes on the fly, and they are the part of the build we spend the most time on.

1. List the calls you never want handled by software. Write down the enquiries where you want a person on the line regardless of how well the agent could cope. Emergencies, complaints, anything involving money already paid. For a medical or veterinary line this list is usually longer than the owner first expects.
2. Give every item on that list a destination. A mobile number for a warm transfer, an SMS to a group, an email to the office, or a callback slot booked into a real calendar. A rule with no destination is a rule that fails quietly.
3. Set the hours each destination is live. Sam's mobile between seven and six on weekdays; after that, an alert and a first-thing callback. This is where most of the value of an after-hours answering service actually sits, because the same call needs two different answers depending on the clock.
4. Decide what happens when the transfer is not picked up. The caller is still on the line and the person you routed to is not answering. The agent should come back, apologize plainly, take the details, and log it as a missed handoff rather than looping the ring forever.
5. Read the first week of transcripts and move the line. Week one is when you find out you set the urgency threshold too low and are being paged about scheduling questions, or too high and a real emergency sat in a message queue. Both are ten-minute corrections at that stage.

What happens if it mishears a name, a street, or a number?

It confirms before it commits. Anything that ends up written into a booking — surname, callback number, address, part number — is repeated back to the caller, digit by digit for numbers. A caller who says "no, it's Beaudoin with a D" corrects it in the moment, which is the only point at which the correction is free.

Speech recognition is where the real-world failure rate lives, and it is worse than the demos suggest for the reasons you would guess: a moving car, a shop floor with a compressor running, a caller on speakerphone. Proper nouns take the worst of it. We would rather name that than let you find it on a Tuesday, and it is the same reason a good human receptionist spells names back.

The safety net underneath is the transcript. Every call produces one, attached to whatever booking or message came out of it, so a booking that went to the wrong Beaudoin is visible the same day rather than at the appointment. Calls are also scored, which means bad handling shows up as a pattern in the review rather than as one customer who never called back. If you want the fuller version of where these systems still show their seams, that is the subject of whether an AI receptionist actually sounds human.

What happens if the phone line, the internet, or the AI goes down?

Your calls fail to wherever you pointed them, and you choose that during setup rather than discovering it during an incident.

Because your existing number is forwarded rather than replaced, the forwarding rule at your carrier is the last line of defence and it is entirely under your control. If the agent cannot be reached, the call continues down the chain you configured: your own mobile, a second number, or your original voicemail box. Unforwarding takes about the same ten minutes that forwarding took, and no part of that requires us.

We do not publish an uptime percentage, because a figure like that from a vendor is unverifiable and you should treat everybody's with suspicion. What is checkable is the design: your number stays yours, the forwarding is yours to change, and there is a defined destination for the case where nothing on our side answers. Ask any provider what happens to a live call at that moment. A vague answer is the answer.

How does that compare to voicemail or a live answering service?

The interesting column is not what happens on a good call. It is what happens on the call nobody could resolve.

The row that decides it for most people is the third one. A failed call you can read is a fixable problem. A failed call you never hear about is just a customer who went elsewhere.

What does a handed-off call actually cost?

The arithmetic is short because there is one rate. At $0.60 CAD a minute, a 90-second call that ends in a transfer costs 90 cents. A 40-second call where the agent says it cannot help and takes a message costs 40 cents. A month of 300 answered minutes is $180 whether those minutes ended in bookings or in handoffs.

There is no separate charge for a transfer, no per-message fee, and no minimum, so a call that "failed" is priced the same as one that booked a job. That is deliberate: charging more for escalation would give us a reason to escalate. The full breakdown across different call mixes sits on our answering service cost page.

When is a person simply better than an AI receptionist?

Four cases, stated plainly, because you will find them anyway.

When the call is bad news. A complaint about work already paid for, a billing dispute, a cancellation with anger behind it. The agent can take these and route them accurately, but a caller who is upset wants a person who can decide something, and being handled competently by software is not the same as being heard.

When judgment has to happen before anything is scheduled. Triage in medical clinics is the clearest case: deciding whether a symptom is an emergency is a clinical act, not an intake question. Our agents book, screen, and escalate on the rules a practice gives them, and the practice keeps the deciding. That split is why the medical answering service pages for Ottawa and elsewhere are written around escalation speed rather than around call handling.

When you genuinely want a human voice on every single call. Some businesses do, and for them a staffed service is the better purchase. Our comparison with Abby Connect says so directly rather than pretending the trade-off does not exist.

When your call volume is tiny and you already answer it. Fewer than about twenty calls a month, all of them reaching you, means the phone in your pocket is already doing this job. Buying software to fix a problem you do not have is a cost with no return attached.

How do you test the failure path before you buy?

Call it and try to break it, which takes about ninety seconds. Our demo agent is Maryann, on +1 437-494-9110.

Ask her for your last invoice number. She cannot possibly know it, and what she does next is the whole evaluation: a system that says it does not have that and offers a message is one you can trust on a live line, while a system that produces a plausible-sounding number is one that will eventually do that to a customer. Then ask for a human and see how quickly she stops trying to handle you. There is a longer list of things worth trying in what to ask an AI receptionist.

Setup is $0, there is no contract and no monthly minimum, and most businesses are live in about one week. The failure rules are written during that week, with you in the room, and they stay editable afterwards.