What you are actually buying
An AI receptionist is not one product. It is a stack, and each layer carries a price. There is a phone number and the carrier minutes behind it. There is the speech and language model that holds the conversation. There is the logic that decides what happens once the caller has said what they want.
Then there is the part that gets under-quoted. Someone has to write the script, test it against awkward calls, and change it when your hours or services change. That work is labour. Labour does not get cheaper because the model got cheaper.
When a quote looks unusually low, find out which layer it leaves out. A small monthly fee can mean the tuning is yours to do.
Call volume sets the meter
Conversation is usually metered by the minute, either billed directly or bundled as included minutes. So two numbers matter. How many calls you take, and how long a typical call runs.
Call length varies more than people expect. A booking that confirms a slot is short. A quote request, where the caller describes the problem at length, runs much longer. Take your monthly call count, multiply by a realistic average, then add a margin, because callers ramble.
Ask whether partial minutes round up. Ask whether abandoned calls and voicemail drops are billed. A system that covers after-hours calls only will use a fraction of the minutes of one answering the main line all day.
Workflows and integrations drive the build
One workflow is cheap. Answer, capture a name and number, send a text summary to the owner. That is a small build, and it can be running in days.
Cost climbs with each path you add. Booking into a live calendar. Checking whether an address falls inside your service area. Quoting a price band. Routing a 2am emergency differently. Writing the result to a CRM. Each path needs its own logic, its own failure handling and its own testing.
Integration difficulty is the wild card. A mainstream calendar with a documented API is routine work. Field service software with a closed system, or an API licence sold per seat, is not. Ask whether the vendor has connected to your specific software before. Ask what happens when that connection breaks mid-shift.
Setup fees against monthly fees
Two shapes are common. A larger one-time setup fee with a lower monthly, or little upfront with a higher monthly that amortises the build. Neither is dishonest. They allocate the same work differently.
The low-setup shape usually carries a term commitment, because the vendor is financing your build. Read the term and the early-exit clause before you compare on monthly price alone.
Usage charges sit on top of both. Ask for the overage rate per minute. Ask the price of outbound calls and SMS, and whether transcription or a call summary costs extra. Ask what a change request costs after go-live. That last one is where a cheap system quietly becomes an expensive one.
Questions to ask before signing
Who owns the phone number. If the vendor buys it, get written confirmation that it can be ported out to a carrier you choose. A number you cannot take with you is a stronger lock-in than any contract clause.
What happens to recordings and transcripts. Ask where the data is stored and how long it is kept. Ask whether it trains anyone's models, and how a caller's deletion request would be handled. Under Canadian privacy law the accountability sits with the business that collected the information, and a vendor contract does not move it.
What the exit looks like. Ask about the notice period. Ask what export you receive of call records and captured leads. Ask whether the prompts and workflows you paid to build belong to you. Put these questions before the demo, not after it.
When the cheaper option is better
Sometimes the honest answer is that you do not need this. If your line rings a handful of times a week, this is the wrong purchase. A human answering service, or a disciplined callback routine, costs less and handles nuance better.
Be sceptical of the pays-for-itself arithmetic. It assumes a missed call would have closed at the same rate as an answered one. Someone ringing four contractors at 9pm is not the same lead as a referral who asked for you by name. The arithmetic often still works. It works on a smaller number than the pitch implies.
There is also a lower rung worth trying first. Voicemail transcription sent to your phone. A booking link published where people can find it. A rule that a missed call gets a text back within ten minutes. That combination costs very little, and it recovers some of the same value. If it fails because nobody keeps to the rule, the constraint is capacity, and automation becomes a reasonable answer.
Price an AI receptionist over twelve months at your own call volume, and ask who owns the number before you ask what it costs.
Tell us what you want to improve.
A short call to work out what the job is and put a number on it.