What information can an AI receptionist collect?

More than a message pad, and less than a form you would fill in yourself. An AI receptionist can collect any spoken field you write down for it — name, callback number, reason for the call, urgency, address, appointment type, file or policy number, vehicle, pet, unit — on top of a set of things it captures on every call without asking: the number that dialled, the time, the language spoken, a recording, a transcript, and a written summary. BotsDontSleep does this on every call, 24/7, at $0.60 CAD per answered minute billed by the second. The limit is not what it can ask. It is how many questions a caller will sit through before hanging up.

What does an AI receptionist collect without asking a single question?

Six things, before the caller has said anything at all.

The number that dialled, the date and time to the second, how long the call ran, which of your lines it arrived on, the language the caller opened in, and — once the call ends — a recording, a full transcript, and a written summary of what was agreed. None of that costs the caller a question. It is a by-product of answering the phone with software instead of a notepad.

That set is more useful than it sounds, because it is the part that is complete. A front desk having a bad Tuesday writes down four fields out of six. This writes down six out of six on every call, including the ones at 2am and the ones where the caller hung up after eight seconds. Over a month it becomes something you can count: how many calls arrived, when they cluster, how many were robocalls, how many came from outside your service area, how many were the same person calling for the third time. Most of what after-hours answering gives you while the office is dark is this kind of data before it is anything else.

The caller ID is the one field to be careful with. It is the number that dialled, which is not always the number the caller wants you to ring back — a receptionist calling from a clinic's main line, someone on a spouse's phone, a blocked number with nothing behind it at all. That is why the callback number is asked for out loud even though the system already has one.

What can an AI receptionist collect by asking?

Anything you can write down as a question. The ceiling is your patience, not the software's.

The standard set runs to four fields: who is calling, the number to reach them on, what they want, and how urgent it is. After that it is entirely yours. Address and cross-street. Appointment type and preferred window. New or existing customer. File, policy, invoice or patient number. Make, model and plate. Unit number and building. A pet's name, species and age. How they heard about you. Whether there is an active contract. Whether they need an interpreter. Whether they can be texted rather than phoned back.

Each field is a question asked in your own words, in the order you set, with the ones you marked required repeated if the caller skips past them. It also asks conditionally, which is where the real gain sits: a caller who says "leak" gets the shut-off question and the ceiling question, and a caller who says "quote" does not. Those branches are written during onboarding alongside your hours, services and transfer rules, in the same session covered in how an AI answering service learns your business.

How many questions can it ask before the caller hangs up?

Fewer than you would like. Six to eight spoken fields is the practical ceiling on a first call, and the last two are already costing you people.

Here is the arithmetic, using a two-minute average call as a planning figure rather than a measured one. Each extra question adds roughly eight to twelve seconds once you count the caller thinking, answering, and the agent confirming. Four extra questions is about forty seconds — at $0.60 CAD a minute that is forty cents a call, or $12 across thirty calls. The money is not the problem. The hang-ups are. Somebody who phoned to ask whether you are open on Saturday will not answer eleven questions to find out.

The fix is to split the two jobs. Ask the minimum that lets you act — name, number, what they want, how urgent — and send the rest as a link by text before the call ends, so the caller fills in the long fields on their own phone, in their own time, with their insurance card in front of them. New-patient detail comes back more complete that way than by reading a form aloud down the phone, which is how new-patient intake in Ottawa is usually set up.

Can it verify what the caller says, or does it only write it down?

Some of it, and the gap between the two is where most bad messages come from.

Three things can be checked while the caller is still on the line. Format, first: a Canadian postal code has a shape, a phone number has ten digits, a date has to exist. Readback, second: the agent repeats the number digit by digit and the surname letter by letter and asks the caller to confirm, which catches the errors a transcript alone never will. Cross-reference, third: a file, policy or patient number can be matched against the record in your system, so the message arrives attached to the right customer rather than to a name that resembles them.

Everything else is unverified by definition. When a caller says the leak started yesterday, or that they are an existing customer, or that their appointment was at three, all the agent can do is record what was said. That is equally true of a person at a desk, and the honest version is to mark the field as reported rather than confirmed. How far the readback step actually moves the numbers is the subject of how accurate AI receptionists are at taking messages.

Can it collect fields specific to my trade?

Yes, and the trade-specific fields are usually the ones that decide whether a message is worth anything at all.

Emergency plumbing intake in Hamilton needs the address, whether the water is off, where the shut-off valve is, and whether a ceiling is involved — four answers that decide whether a van leaves tonight. A veterinary line needs the species, the age, when the symptoms started, and whether the animal is eating, which is the basis of urgent triage for veterinary clinics. A law firm needs the matter type and the other party's name before anything else, because a conflict question asked late is a conflict question not asked. An auto shop needs the plate or the VIN and the noise the customer is trying to describe.

None of these are add-ons and none of them cost extra. They go into the script during setup and get changed afterwards, when you find the field you forgot. That is usually week two.

How do the options compare on what they capture?

The staffed rate above is the general Canadian market range rather than one provider's published card. Check any provider's own pricing page for current figures before setting it against the number here.

What should an AI receptionist never collect over the phone?

Card numbers, and anything else you would not want sitting in a transcript.

We do not take payment card details on calls. A card number spoken aloud lands in a recording and a transcript, and there is no version of that we are comfortable operating on a client's behalf. Callers who want to pay get a payment link by text instead, which keeps the number out of the call entirely.

The same reasoning runs wider. Social insurance numbers, full dates of birth where a partial would do, passwords, and clinical detail beyond what is needed to judge urgency all belong in a system built to hold them rather than on a phone line. On a medical line, "chest pain, started an hour ago" is an escalation trigger and belongs in the message. A medical history is not, and that conversation belongs with a clinician. What is held, who can open it, and how long it is kept is set out in how call data on an AI phone system is stored.

Callers are told the call is recorded, and the agent discloses that it is an AI by default. What your particular line has to announce is a question for your own advisor rather than for us.

What can an AI receptionist not collect?

Four things, and the fourth is the one that costs you.

It cannot collect what the caller will not say. Somebody who wants to speak to you rather than to a machine gives a name and a number and nothing else, and no script fixes that. It cannot collect anything from a caller who hangs up either — you get the number and the time and nothing more, which is still more than voicemail leaves you, and still not a lead.

It cannot reliably collect long strings spoken quickly over a poor line. Alphanumeric references, unusual surnames, and street names it has never encountered are where errors concentrate. The readback catches most of them and not all of them. A vocabulary list of your own products, streets and staff names is the largest single thing you can do about that, and it costs an afternoon.

And it cannot collect what nobody asked for. This is the failure that goes unnoticed for months, because nothing looks broken: the messages arrive, the fields are filled, and a field you never put on the list is simply absent from every call since launch. A person at a desk improvises — they hear something odd in a caller's voice and ask about it. A script does not, unless you wrote the question. An AI receptionist working from a thin intake list will be wrong consistently and at volume, which is worse than being wrong occasionally. Whoever writes that list is doing the real work here, and it is not the software.

How do you write your own intake list?

Five steps, and an hour with last week's calls beats any template.

1. Pull twenty recent calls and write down what you actually needed from each one. Not what a form asks for — what you had to know before you could do anything useful.
2. Mark each field required or optional. Required means the agent asks again when a caller skips it. Most lists have too many required fields; three or four is usually right.
3. Write the conditional branches. Which answer opens which follow-up question. This is where a generic script becomes yours.
4. Decide what gets read back. Callback numbers always. Anything alphanumeric. Appointment times.
5. Move the long fields to a text link. Insurance details, addresses with unit numbers, anything the caller can read off a card at their own pace.

Then test it from a phone the system does not recognize, once in hours and once late at night, and look at what actually lands in your inbox. Ours answers on +1 437-494-9110 if you would rather run that test on a line that is not yours first. Asking four questions or twelve costs the same $0.60 CAD a minute either way, and the answering service cost breakdown has the rest of the arithmetic.

Deciding what to ask is a separate job from deciding what to do with the answer. If you have not settled the first one yet, what an answering service should collect on every call is the field list to start from.