Which calls can an AI receptionist take messages for?

Almost all of them, and the honest way to decide is by what a wrong field costs rather than by a published accuracy figure. An AI receptionist can take the message on any call where the useful content reduces to a few short, confirmable fields — who called, the number to ring back, what they want, when, and how urgent it is. That covers bookings, reschedules, quote requests, status checks, supplier calls and most overflow. BotsDontSleep answers on the first ring, 24/7, at $0.60 CAD per answered minute, billed by the second. The calls to keep off an automated message are the ones where the message itself is the deliverable: a witness statement, a clinical history, a first notice of loss. There the caller's exact words carry weight months later, and a structured summary is not a record.

What makes a call safe to hand to an automated message?

Three questions, asked of the call type rather than of the software.

Does the message reduce to fields? A plumbing enquiry does: name, number, address, what is leaking, whether the water is off. A caller describing a dispute with a contractor over eight months does not reduce to fields, and squeezing it into them is where the value gets lost.

Is the message an instruction or a record? An instruction tells someone to do something — ring this number back, move this appointment, send a van. It is judged by whether the action happened. A record is read later by someone who was not on the call, possibly by an adjuster or a court, and it is judged against what the caller actually said. Automated intake is good at instructions.

Is a wrong field recoverable? If the surname comes through misspelled on a booking, the appointment still happens and someone fixes the spelling at the desk. If the callback number is wrong on a new enquiry, the lead is gone and you will never know it existed. Both are the same transcription error; only one of them is expensive.

Most business calls pass all three comfortably. The point of asking is to find the small number that do not, before they are routed the same way as everything else.

Which call types can an AI receptionist take messages for?

The top six rows are the overwhelming majority of inbound volume at a small business, and they are all instruction calls with short fields. A dental front desk is mostly rows two and three. The calls that get missed during a procedure are almost all a name and a Thursday, which is why dental call answering in Toronto is a routing decision rather than a hard technical one. The last row is the one that needs a different answer, and it is covered further down.

Row seven is worth its own note, because screening is the only row where accuracy barely matters. A robocall that gets hung up on costs nothing whether the system understood it or not.

How do you price a wrong field?

With arithmetic, using your own numbers rather than an industry average.

Start with the cost of the call. A 90-second message call at $0.60 CAD per answered minute, billed by the second, is 90 cents. Per-second billing is what makes that figure usable — on a service that rounds every call up to the next minute, the same call is billed as two. The rate and the billing rules are the two halves of that number and should always be read together.

Now price the protection. Reading a ten-digit callback number back to the caller and asking how a surname is spelled takes about eight seconds, which at this rate is eight cents. So the only question is whether the thing being protected is worth more than eight cents. For a new enquiry, the answer is not close. For a supplier confirming a delivery window, the readback is eight cents spent on nothing, and you can switch it off for that call type.

That asymmetry is the whole decision. Confirmation is cheap enough to apply to every call type where a wrong field loses money, and cheap enough that getting the judgment wrong in the careful direction costs almost nothing. Put your own average first-job value in place of a borrowed statistic and the sums are not difficult: whatever that number is, eight cents is a rounding error against it.

One more figure belongs here. A business taking 300 answered minutes a month pays about $180 — 300 multiplied by $0.60 — with no setup fee, no monthly minimum and no contract. There is no plan to size and no overage band to fall into, which removes the usual reason for routing fewer call types to a service than you should.

Which fields still need confirming out loud?

Four kinds, consistently, and they are the same four that catch human operators.

Anything alphanumeric comes first: a plate, a unit number, a policy number, an order reference. Surnames are second, particularly those not spelled the way they sound. Street names are third, because callers say them quickly and there is nothing in the sentence to correct them from. Fourth is a phone number delivered in one unbroken run, which is how most people say their own.

Everything else in an ordinary message — "the sink is leaking and there is water on the floor" — is the easy part. Modern speech recognition handles conversational speech about an ordinary problem well; it is the short, high-value, context-free strings that need the confirmation step. The mechanics of that, and how to test it on a live line, are in our longer piece on how accurate AI phone receptionists are at taking messages. Deciding which fields a given call type must capture before any of this matters is covered in what information an answering service should collect.

Does the call type change when callers switch languages?

The field list does not. The confirmation step matters more.

Calls are handled across 90+ languages, including a caller switching mid-sentence, which is ordinary in a lot of Canadian markets. On a bilingual property-management line in Montreal, a maintenance request needs the same five fields whether the tenant describes the problem in French, English, or both in the same breath. What changes is that street names and surnames are harder audio in a second language, so the readback is doing more work and should not be trimmed to shorten the call.

The case that does change category is the record call in a second language. If the words are the deliverable and the caller is not speaking their first language, that needs a qualified human interpreter and a person taking the statement, not an automated intake in 90 languages.

Which calls should never go to an automated message?

Four cases, stated plainly, because they are real and they do not shrink with better software.

When the message is the deliverable. A witness account, a first-notice-of-loss statement, a clinical history where the exact phrasing is read back months later by someone deciding something. These need a trained person taking a structured statement, and often a recording plus a human transcription pass. The split runs through the middle of a single line rather than around it: on a medical line in Winnipeg, bookings, reschedules and callback requests are safe to automate while symptom intake stays with a person. Those are different call types, and nothing requires them to be routed the same way. Where a message has consequences under rules that apply to your profession, that is a question for your own compliance advisor rather than for a phone system.

When the caller cannot name the problem. Someone distressed, contradicting themselves, describing something they do not have words for. Capturing and escalating is not the same as helping, and a person is better at the work of getting to the actual question.

When a meaningful share of your callers will not deal with an automated voice. Then accuracy is beside the point: the message is perfect and the caller hung up before giving it. This is worth testing on your own line rather than assuming either way, in both directions.

When you cannot write the routing rule down yourself. If you cannot say in a sentence which calls are urgent at 2am and who they should reach, no system will infer it. That is a twenty-minute document, and it is the actual prerequisite.

If two or three of these describe most of your volume, a staffed service is the better purchase and we would say so. Our Abby Connect comparison sets out where a human receptionist on every call earns the difference in price. Emergency routing specifically has its own guide: how urgent calls are captured and transferred.

How do you sort your own call types in an afternoon?

Five steps, with a call log and a sheet of paper.

1. List last month's inbound calls by reason, not by caller. Your carrier's admin portal has the log. You will usually finish with six to nine reasons, and two of them will be most of the volume.
2. For each reason, write the fields the message must carry. Three to six fields each. If you cannot list them, that reason is really two reasons and needs splitting.
3. Mark what a wrong field costs. Lost lead, empty chair, wasted van, annoyed customer, nothing. Write the actual consequence rather than high, medium or low.
4. Draw the line. Fields plus a recoverable error means automate, with a readback where a wrong field costs money. Words-as-record, or an error you would never find out about, means a person takes that call.
5. Test your two biggest reasons on a live line before changing anything. Call an automated receptionist as your own customer and check what the message says. Ours is Maryann, on +1 437-494-9110, at any hour. Ask her something off-script and listen to what she does at the edge of what she knows.

The output is a one-page routing table, which is also exactly what any provider needs from you at setup. Doing it first means you are buying against your own list instead of a sales script.

What is the smallest sensible way to start?

Point one call type at it — usually after-hours and overflow, because those are the calls currently going to voicemail — and leave your front desk exactly as it is. You are testing calls you are already losing, not replacing calls you are already winning. Add the next call type when the first one's messages have been right for a fortnight, which is a judgment you can make from the transcripts rather than from a dashboard.