How do I train an AI phone system for my business?

Training an AI phone system is a writing exercise rather than a machine-learning one. You supply five things — your service list, your hours, your prices, the questions your front desk asks every day, and the rules for who gets put through to a person — and the provider turns them into call flows you review before a customer ever hears them. With BotsDontSleep.ai that costs about two to three hours of your time, the tuning is included at $0 setup, most businesses are live in about one week, and the agent then answers 24/7 at $0.60 CAD per answered minute. It does not teach itself from your callers, which means every word it says is a word somebody chose on purpose.

What does training an AI phone system actually mean?

It means writing your business down, once, in enough detail that a stranger could answer your phone from it.

That is duller than the phrase suggests, and it is accurate. There is no period where the system sits on your line absorbing your habits. Somebody types your service list into a configuration, decides what the agent says when a caller asks a price you do not publish, and names the person who gets woken at midnight. The quality of the result tracks the quality of that material almost exactly.

Owners underestimate this part because it looks like admin. A shop that hands over a price list and a set of hours gets an agent that quotes prices and states hours. A shop that also hands over the four questions its service advisor asks before booking gets an agent that produces bookings the shop can service. The gap between those outcomes is about ninety minutes of somebody's afternoon.

What the system does with that material once it has it — the tuning, the connecting, the review — is covered in how an AI answering service learns your business. This page is the other half: what you have to do.

What do you have to write down before it can answer a call?

Five documents, and you almost certainly have four of them already.

• Your service list, with what each job involves. Not the marketing names — the operational ones, with how long each takes and what it excludes. A caller asking for a "check-up" and one asking for an "assessment" may need different slot lengths, different staff, and different questions.
• Your hours, including the awkward ones. Regular hours, seasonal hours, statutory holidays, the week you close in July, and the difference between closed and closed-but-reachable-in-an-emergency. Seasonal trades get this wrong most often, which is why HVAC companies spend their first tuning session on the calendar rather than the script.
• Your prices, and the jobs nobody may quote by phone. The second list matters more than the first. An agent that quotes a number on a job that has to be seen creates an argument you will have in person, later, with a customer who has already decided what it costs.
• Your intake questions. What you need to know before a booking is worth having: address, vehicle, unit number, symptom, insurer, whether they have been in before. For a cleaning company quoting an estimate, square footage and frequency do most of the work.
• Your escalation rules. Who gets interrupted, for what, at which hour, and what happens when that person does not pick up.

Four of those come off your website, your booking system, and your intake forms. The fifth has usually never been written anywhere.

Which of your rules only exist in someone's head?

The valuable ones, generally, and they are the reason a discovery call exists at all.

Every business runs on a layer of undocumented judgment. Whether you take work outside your service area if the job is big enough. The customer whose calls always go straight to the owner, no matter what. Whether a Saturday callback counts as after-hours. What you say to somebody asking for a price you deliberately do not publish.

None of that is on your website, and an agent cannot follow a rule it was never given. When these surface three weeks after launch instead of during tuning, they surface as a call that went wrong — a bad way to discover that everyone in the office had a slightly different answer to the same question.

The fix takes twenty minutes. Sit with whoever answers the phone most and ask what they get asked that the website does not cover. Write the answers as sentences, not categories. "We will book a quote visit but we do not price roof work over the phone" is usable. "Handle pricing enquiries appropriately" is not.

How do you train an AI phone system, step by step?

Six steps. Almost all of your share of the work sits in the first three.

1. Gather the five documents. Service list, hours, prices, intake questions, escalation rules. Send what you have rather than what you wish you had — a scruffy price list beats a tidy one that arrives in three weeks.
2. Write down the rules that live in someone's head. Twenty minutes with whoever answers the phone. This is the highest-value thing you will do, and it is the step people skip because it does not look like setup.
3. Decide what the agent may do without asking you. Which appointments it can book outright, which questions it may answer, which prices it may state, and which calls must reach a person immediately. This decision shapes the result more than the wording of anything.
4. Hand it over and let the tuning run. Call flows get built, wording and tone get set, and your calendar or business-management system gets connected. This part is ours, included in the $0 setup fee. How the setup runs covers the sequence end to end.
5. Run your worst calls at it before launch. Not the easy booking — the warranty argument, the caller who will not say what they want, the person asking for a price you do not publish. Read the transcripts and change any phrasing you would not say yourself.
6. Read the first week and correct it. Every call arrives as a transcript. Week one is when the greeting turns out to be slightly wrong and a job type is booked into too short a slot, and both are ten-minute corrections at that stage.

Steps one through three are yours and take about two hours. Step five takes an hour. The rest is ours.

How do you test it before a real customer hears it?

By trying to break it, in the specific ways your callers break your front desk.

Scripted test calls prove the agent can do what it was built to do, which is the least interesting question about it. The useful tests are the awkward ones. Interrupt it mid-sentence and see whether it stops or finishes the thought. Change your mind halfway through a booking — ask for Tuesday, switch to Thursday — and see whether it follows or starts again. Ask it something it cannot possibly know, such as your last invoice number, and listen for whether it says so or produces a plausible figure. That last one is the whole test. An agent that invents an answer under pressure will eventually do it in front of a customer.

Then give it the call your own staff dread. For a dental clinic in Toronto that is often the patient in pain asking to be squeezed in today; for an auto shop in Hamilton it is the caller describing a noise and expecting a number. Where the agent hands off tells you more than what it says.

You can run all of this on our demo agent before buying anything. Maryann answers at 437-494-9110 at any hour, trained on a business that is not yours — which is exactly why the test is fair.

What has to change after the first week of real calls?

Something, always. A launch that needs no corrections means nobody read the transcripts.

The corrections are small and they repeat across businesses. A greeting three words too long. An intake question nobody can answer on the spot, so every call stalls in the same place. A job type with a slot length fifteen minutes short of reality. An urgency threshold set too low, so somebody is paged about scheduling questions at eleven at night, or too high, so a real emergency sat in a message queue until morning.

None of those are failures of the technology, and none are visible before real callers arrive. They are the second half of training, and the reason to read week one call by call rather than checking a dashboard summary. After that, maintenance is light: new prices, new services, seasonal hours, and new transfer rules are updates rather than rebuilds, and they take effect on the next call.

Who does the training work — you or the provider?

It depends on what you bought, and the difference is measured in your hours rather than in dollars.

The honest reading of that table is that a build-it-yourself platform can be cheaper per minute and moves every row into your column. If you are technical and enjoy this, that is a real option and we would not argue you out of it. If your day is already full, the row that decides the purchase is the one about Friday night.

When is training an AI phone system not worth the effort?

Four situations, and we would rather name them here than discover them together in week three.

When nobody will own the material. Training has an owner or it rots. Prices change, services get added, someone new joins the on-call rota, and an agent working from last spring's price list is worse than voicemail because it is confidently wrong. If there is no one who will spend twenty minutes a month keeping it current, this is not the right purchase.

When every answer depends on judgment. If your quoting genuinely cannot be reduced to rules — bespoke work where the number depends on things only you can assess — then all anyone can train is intake. That can still pay for itself. It is a much smaller win than the pitch usually implies, and you should buy it knowing that.

When the volume does not justify the afternoon. Under about twenty calls a month that you answer nearly all of yourself, two hours of documentation buys you very little. The phone in your pocket is already doing the job.

When you want a person on every call. No amount of tuning turns this into a human being. If a live voice on every single call is the requirement, a staffed answering service is the better fit, and that is a straight recommendation rather than a hedge.

What does the training cost, and what does keeping it trained cost?

Nothing, and then nothing.

Discovery, call-flow mapping, tuning, calendar and system integration, your review calls, and the corrections you ask for in week one are all included at $0 setup. There is no onboarding fee, no configuration charge, and no per-language pass — all 90+ languages come with the same tuning rather than a separate build. Ongoing changes are included too: a new price, a new service, or a new transfer rule is a message rather than a project.

The only invoice is $0.60 CAD per answered minute, billed by the second, with no contract and no monthly minimum. Three hundred answered minutes a month is about $180, and a two-minute booking call is about $1.20. Set that against one job you would otherwise have lost to voicemail, and the two hours of writing stops being the expensive part of the decision.