What happens if an AI can't answer a call?

Something else picks the call up, and which something depends on which failure just happened. "Cannot answer" covers five separate events on a business line — a question outside what the agent knows, a transfer nobody picks up, a caller who will not deal with software, a call that drops mid-sentence, and a network that is down — and each one has a different fallback behind it. A well-built AI receptionist never leaves a caller with nothing: it says plainly what it cannot do, follows the rule you wrote for that case, and puts the call in front of a person as a transcript rather than a mystery. BotsDontSleep runs this 24/7 at $0.60 CAD per answered minute, billed by the second. One of the five cannot be fixed by anything you buy, and that one is worth understanding before you sign up for anything.

What does "cannot answer" actually mean on a phone line?

Five different failures, and they behave nothing like each other:

• The question is outside what the agent knows. Your warranty terms, an invoice number, whether a part is on the shelf.
• The transfer is not picked up. The agent did its job, routed the call correctly, and the phone at the other end rang out.
• The caller will not deal with software. They ask for a person in the first sentence and mean it.
• The call drops. A dead spot, a tunnel, a battery at two percent, a hand on the wrong button.
• The network is down. Your carrier, not your provider. Nothing on your line is reachable by anyone.

Only the first of those is about the AI at all. The other four are about what you put behind it, and they are the ones that decide whether a caller comes back. Most writing on this question, ours included, answers the first and stops.

What happens when the question is outside what it knows?

It says so, out loud, and then does one of four things you chose in advance: warm-transfers to a person, sends an SMS or email with the transcript attached, takes a detailed message, or books a callback slot. It never invents an answer, which is the behaviour that separates a system you can put in front of a paying customer from one that will eventually quote a price you do not charge.

We have written that case out in full, including how to choose between the four destinations, in what an AI receptionist does when it cannot answer the question. Everything below is about the failures that happen while the agent is working exactly as designed.

What happens when the transfer rings out?

This is the most common real failure on a live line, and it is almost never the software's fault. The agent applied your rule, dialled the number you gave it, and the person you named did not pick up. On a roofing line in storm season the estimator is on a roof; at 4 p.m. on a Friday the office is already empty. The rule was correct and the destination was not there.

What happens next is a decision you make before launch, not one to leave to the software. The behaviour worth setting: two attempts of roughly twenty seconds, then the agent comes back onto the line, tells the caller plainly that it could not reach anyone, takes the details itself, and logs the call as a missed handoff rather than an ordinary message. A missed handoff is an event you can go and look at on Monday. A message in a pile of forty others is invisible.

Two settings make this survivable. Give each destination hours — a mobile that is live between seven and six, an alert and a first-thing callback after that — and give the important call types a second name and number to try when the first does not answer. A rota written as roles is a diagram. A rota written as names, numbers, and hours is a plan.

What happens if your carrier or the platform goes down?

The honest half first: if the network carrying the call is down, nothing answers it. Not an AI, not a staffed answering service, not the phone on your desk. Any vendor promising otherwise is selling you something it does not control.

Canada has a recent illustration of the scale. On 8 July 2022 a Rogers network failure took wireless and wireline service off the air for more than 12 million customers from 4:58 a.m. Eastern until services were gradually restored by 7:00 a.m. the following morning, and 9-1-1 calling stopped working during it (CRTC-commissioned assessment prepared by Xona Partners, published 2024). Every answering arrangement in the country that sat behind an affected line failed at the same moment, whoever was providing it.

Platform-level failure is a smaller and more manageable case. Your calls arrive because your number forwards to us, and forwarding rules live at your carrier, not with us. Set the no-answer fallback on the day you forward the line — a mobile, a second number, or plain voicemail — and an unreachable platform degrades to whatever you chose rather than to silence. Your published number never changes hands, and unforwarding takes about the same ten minutes that forwarding took. For a no-heat line in Edmonton in January, deciding that fallback in advance is worth more than any feature on the list.

What happens if the caller refuses to talk to an AI?

The agent tells callers it is an AI by default, so most are refusing knowingly rather than discovering it halfway through. When someone asks for a person, the useful measurement is not whether it eventually complies — everything complies eventually — but how many turns it spends trying to help first. One is fine. Three is a caller who is annoyed before a human has said hello.

Set that threshold deliberately and set it low for the call types where a request for a person is itself information: complaints, anything involving money already paid, anything a caller is embarrassed about. Outside the hours a person exists, the honest version is a stated callback window rather than a transfer that is going to ring out. Thirty minutes, actually kept, buys more patience than "right away" ever does.

There is a version of this question where the answer is that you should not buy an AI receptionist. If a human voice on every call from the first second is the product you want, that is a staffed answering service, and our Abby Connect comparison sets out where a staffed desk is the better purchase.

What happens if the call drops mid-sentence?

The caller's number is captured when the call connects, not when it ends, so a call that dies at the forty-second mark still arrives as a transcript with a number attached and everything said up to that point. That is the difference between a dropped call and a missed one: you can ring back and open with what they already told you.

The rule to set here is what the agent may do with a half-finished booking. A confirmed appointment is written when the caller has agreed to a specific time and heard it read back — not before. A call that drops during the offer of two slots should leave a message and a callback task, never a provisional booking nobody knows about. Phantom appointments are worse than no appointment, because your calendar tells you a lie you then plan around.

How does each way of answering your phone fail?

The last row is not a typo. It is the row most comparison tables leave out.

Does a call the agent could not finish still cost money?

Yes, and this is worth stating plainly. Billing is by the answered second at $0.60 CAD a minute, so a ninety-second call that ends in a message and a callback costs about ninety cents. A twenty-second wrong number is about twenty cents. There is no separate charge for a transfer, a failed transfer, or a call that drops, and there is no setup fee or monthly minimum underneath any of it — but the failed calls are on the bill alongside the successful ones. Set that beside the value of one recovered job and the arithmetic usually settles itself; the full breakdown is on our answering service cost page.

What none of this fixes

Four things, stated plainly, because the fallback chain is not magic.

A carrier outage is not survivable by anything you can buy. If your network is down, your calls are not arriving anywhere, and the only real mitigation is a number on a second carrier that you publish somewhere your customers actually look.

A fallback chain is only as good as its destinations. The most common cause of a call going nowhere is a mobile number in the rota belonging to somebody who left in March. Nothing in the system knows that. Re-read your escalation list twice a year.

If nearly every one of your calls needs a person within a few seconds — a crisis line, a trading desk, a business where the first thirty seconds carry legal weight — you are buying the wrong category, and a staffed service is the honest answer.

And if you take a handful of calls a week and you answer nearly all of them yourself, a five-branch fallback chain is more machinery than your problem deserves. The phone in your pocket is already doing the job. Come back to this when the volume changes.

How do you test the failure path before a customer finds it?

Testing the happy path proves nothing; every provider sounds good on a booking that goes to plan. Run these six calls at your own line instead, and read what lands afterwards.

1. Ask it something it cannot possibly know. Your last invoice number is the classic. It should say it does not have that and offer a message or a transfer. A plausible-sounding number ends the evaluation.
2. Ask for a person, and count the turns. Say you want to speak to someone. Count how many times it tries to help you first. That number is what an irritated customer will experience.
3. Let the transfer ring out on purpose. Put the on-call phone face down and do not answer. You are testing whether the agent comes back to the caller or leaves them listening to a ring tone.
4. Call while your own line is engaged. Then call twice more at the same time from two other phones. Concurrency failures are invisible until the day they are not.
5. Hang up in the middle of a booking. Cut the call at the point where you have given a name and a time but confirmed nothing. Then look at your calendar and check that nothing was written into it.
6. Do all of it again at eleven at night. The overnight rules differ from the daytime ones, and overnight is when your destinations are least likely to answer.

Then check what actually arrived: the inbox, the SMS, the calendar, and the phone belonging to whoever was supposed to wake up. The gap between what you expected and what landed is your real configuration, and week one is when correcting it costs ten minutes rather than an awkward conversation six months later.

Our demo agent Maryann answers at 437-494-9110 at any hour. She is trained on a business that is not yours, which is what makes her worth breaking. Run the first three calls at her before you talk to anybody about pricing.