Can AI answer your phone? What an intake agent does, and what it must not do

Yes. An AI agent can answer a business phone, hold a real conversation, collect the details, check a calendar and book the job, at 2am, on a Sunday, while you are on a roof in Barrie or under a sink in Windsor.

The useful question is not whether it works. It is which calls it should take and which it must hand over, because the second list is what separates a system that wins work from one that quietly loses customers who were trying to give you money.

First, do the arithmetic on what you are actually losing

Everybody selling this quotes a statistic about missed calls. Ignore them and use your own numbers, because the answer swings by an order of magnitude.

Three figures decide it:

1. Missed calls per week. Take it from your phone bill or your mobile log, not from memory. Owners underestimate this badly, because a call missed while you are working does not feel like an event.
2. The share that would have become work. Be pessimistic. Some of them called the next business on the list and some were not customers at all.
3. Your average job value.

Then: missed calls a week, times 50, times the conversion share, times the job value.

Four missed calls a week, one in five converting, at an average job of $600, is about $24,000 a year. Against an agent build of $4,000 to $15,000 that pays back inside a year and keeps paying.

Four missed calls a week, one in twenty converting, at $150 a job, is $1,500 a year. That business should record a better voicemail greeting and spend the money somewhere else. This is the honest half of the pitch and almost nobody makes it.

What an intake agent genuinely does well

The work an AI agent is good at is the work that has a shape.

• Answering on the first ring, at any hour. This is the whole advantage. It costs the same at 2am as at 2pm, and it does not get worse on the fortieth call of the day.
• Collecting the same six fields every time. Name, callback number, address, what is wrong, when it started, and whether it is urgent. A human on a bad day forgets the callback number. The agent does not.
• Answering the forty questions that arrive every week. Hours, service area, whether you take that brand, what a call-out costs, how far north of Sudbury you will travel.
• Booking against a real calendar and writing the result into your system, so the job exists in one place rather than on a sticky note.
• Summarising and routing. A written summary in your inbox two minutes after the call, with the recording attached and the urgency flagged.

That last one is worth more than it sounds. Most trades businesses do not have a missed-call problem so much as a missed-detail problem, and a written record of every call ends a whole category of argument.

What it must not do

Three categories, and they belong in the scope document rather than in a conversation after the build.

Anything that needs judgement about money

Pricing an unusual job, agreeing a discount, or deciding whether to fit somebody in ahead of a booked customer. An agent that quotes a number it should not have quoted has created a commercial problem, and the caller will reasonably hold you to it.

Anything with a safety edge

Gas, water, heat in January, electrical burning smells, anybody describing an emergency. These calls need a defined escalation to a human phone that rings, and the escalation has to be tested rather than assumed. If your business handles that class of call, this is the single most important paragraph on this page.

The call where the caller is upset

An agent can detect it. It cannot fix it. The correct design is a fast handover with the context already captured, so the person who picks up is not asking somebody angry to start again.

The Canadian rules that shape the build

This is where an intake agent differs from the same product sold in the United States, and it is the part most demos skip.

The agent is recording, so consent rules apply

An AI intake agent transcribes by design. That makes the guidance from the Office of the Privacy Commissioner of Canada on recording customer telephone calls directly relevant, and it sets four expectations: record only for a purpose a reasonable person would consider appropriate, tell the customer that the call is being recorded and why, offer an alternative if they object, and use and retain the recording only for the stated purpose.

In practice that is one sentence at the start of the call and one decision about retention. Say what is being recorded and why, and set a retention period rather than keeping every call forever. Neither is difficult. Both are frequently missed.

Say that it is an AI, even though you do not have to

There is no general rule in Ontario or federally that requires you to disclose a non-human voice on a call your customer placed to you. Disclose it anyway. It costs one clause in the greeting, and it removes the outcome where a caller works it out halfway through and feels handled.

The rules that do govern automated voice, the CRTC’s Unsolicited Telecommunications Rules, apply to calls your business makes rather than to calls it receives. The CRTC opened a review of those rules under Compliance and Enforcement Notice of Consultation CRTC 2026-132 in 2026, which considers among other things whether callers must be told when a non-human voice is used. Nothing in that proceeding changes what applies today, and the current position is worth checking with the CRTC before you rely on it. Nothing here is legal advice.

PIPEDA reaches everything the agent hears

A caller giving their address, their phone number and a description of a problem inside their home is handing over personal information, and the Personal Information Protection and Electronic Documents Act governs what your organisation then does with it. That decides where the transcripts live, how long they are kept, and which vendors are eligible in the first place. The longer version is in what PIPEDA means for the AI tools your business uses, and the Privacy Commissioner’s May 2026 finding against OpenAI is the reason the vendor question is now a real one rather than a formality.

Outbound is a different legal animal entirely

Using the same agent to call people back who did not ask to be called moves you from inbound service into telemarketing, where the National Do Not Call List and the Unsolicited Telecommunications Rules apply. If a supplier offers you outbound follow-up as an easy add-on, treat it as a separate decision with separate advice.

What it costs, and what runs on top

• Build: $4,000 to $15,000 per agent, from AI agents and intake. What moves the number is how many systems the agent has to write into, whether it books against a live calendar, and how many escalation paths it needs.
• Running costs, monthly. Telephony minutes, speech processing, model usage and any seat licences. On a small trades deployment this is usually in the low hundreds rather than the low thousands, and it scales with call volume rather than with headcount.
• The thing to hold back. Do not build the agent and the customer-facing chat and the internal assistant at once. Build the phone, run it for six weeks, read the transcripts, and let what you find decide the second one.

When I would tell you not to buy one

• Under about two calls a day. The exposure does not justify the build. Fix the voicemail greeting and add a callback promise with a real time in it.
• When the calls are all different. A business where every call is a bespoke technical conversation does not have an intake problem. It has a scheduling problem, and that is a different build.
• When nobody will read the summaries. An agent that captures 40 leads a week into a system nobody opens has replaced a missed call with an ignored one.
• When the actual bottleneck is quoting. Plenty of businesses answer the phone fine and then take four days to send a price. That is covered in the four automations that pay in a trades business, and it is usually the better first build.

If you are unsure which of those you are, that is the question the $999 assessment exists to answer: three named jobs, hours and dollars against each, and the ones I would leave alone. Statistics Canada put Canadian business AI use at 19.2% in the second quarter of 2026, and the single largest reason non-adopters gave was that AI was not relevant to what they do. A phone that rings out after five is the most common proof that it is.