The hard part of booking service calls automatically is not the conversation. It is everything that has to be true before the booking is real — a genuine slot, a captured address, a stated consent — and everything that happens after, where Canadian anti-spam law starts to apply.
Key takeaways
The question behind "AI dispatcher for plumbers" is usually simpler than it sounds: the phone rings after hours, nobody answers, and the caller phones the next plumber. An intake agent that answers, asks the right four questions and puts a real slot on the board solves that. It is worth being precise about what it is doing, because the failure modes are specific.
It is producing a record, not a conversation. In Dynamics 365 Field Service's work order documentation that record carries a service account, an address, a work order type and incident types that determine the tasks, products and services required. Whatever front end you use — voice, web chat, a form — the test is whether the output is complete enough to dispatch without a callback.
Four fields decide that in service trades: what is broken, where exactly, how urgent, and who is paying. An agent that captures the first three and misses the fourth has produced a lead, not a booking. Write the required fields down before you evaluate any product, then ask the vendor to fail a call that is missing one.
An agent that says "we can come Tuesday morning" from a static price list is guessing. A booking is real only when the agent is checking the same availability the dispatcher sees. The mechanics are documented: Microsoft's schedule assistant ranks resources by availability, skills and location and estimates travel time before a booking is made, and warns that booking outside its recommended slots means capacity, work hours and time windows are not verified or enforced.
The practical consequence for an unattended intake agent is that you should give it a narrow, safe booking window — a defined set of slots reserved for intake — rather than the run of the whole board. Anything it books inside that window is real. Anything outside it goes to a human in the morning.
This is the step most "AI receptionist" write-ups skip. The confirmation text, the reminder the day before, the "we have a cancellation this afternoon" message — those are messages sent to an electronic address, and Canada's Anti-Spam Legislation governs them. The CRTC's spam and malware pages, which set out how it enforces CASL, group the obligations under consent, identification, unsubscribing and record keeping, and its compliance tips specifically cover consent, contact lists, an internal do-not-call list, corporate compliance programs and accurate records.
The nuance that catches service businesses is the difference between the two kinds of consent, and whether a message is transactional or commercial. Our sister firm's note on implied versus express consent under CASL is the better place to work that out than a vendor's onboarding checklist. The design rule is simple enough: have the agent state, in words, what it is going to send and how to stop it, and log that it did.
Build the consent step into the script, not the settings page
Have the agent say what it will send and to which number, then capture the answer as a field on the record — not as an inference from the fact that the call happened.
Keep an accurate record of when and how consent was given. CRTC's compliance tips list accurate records as one of the practices it looks for.
Give every automated message an unsubscribe path and identify the business sending it.
Keep an internal do-not-contact list that the agent reads before it dials or texts.
Most voice agents transcribe, and a transcript of a customer describing their flooded basement at their home address is personal information. The Office of the Privacy Commissioner's overview of PIPEDA sets out the framework that applies to commercial activity across most of the country, and its workplace privacy guidance covers the employee side if your own technicians are recorded too. Decide before launch what you record, why, who can hear it, and how long you keep it — and say so at the start of the call.
PIPEDA also attaches a specific consequence if those recordings are ever exposed: the same law that governs collecting them requires an organization to report a breach of security safeguards to the Privacy Commissioner and notify the affected individuals whenever the breach poses a real risk of significant harm, and to keep a record of every breach — even ones below that threshold — for two years. A voice agent that records a caller describing a flooded basement at their home address is exactly the kind of file that obligation is written for; build the retention period and the access controls in from the start rather than discovering the two-year record-keeping duty after an incident.
A four-truck plumbing company in Hamilton takes about 40 calls a week, a third of them outside office hours. Today those go to voicemail and roughly half never call back. The build that works looks like this.
The agent answers after two rings, identifies the business, states that the call is recorded and why, and asks four questions: what is happening, the service address, whether water is currently running, and whether this is a rental or an owner-occupied home. If water is currently running it does not book — it pages the on-call technician and tells the caller someone will phone back within fifteen minutes.
For everything else it offers two slots from a reserved intake block, confirms the address back, asks permission to text a confirmation and a reminder, and creates the work order with the transcript attached. In the morning the service manager sees six structured requests instead of six voicemails, and the only judgement call left is whether the Tuesday 08:00 job needs the senior technician.
Note what the agent never does: quote a price, promise a fix, or decide that a job is an emergency worth an after-hours premium. Those are commercial decisions with contractual consequences, and they belong to a person.
The realistic comparison is not agent versus person; it is agent versus voicemail at 22:00. Test it by measuring the completion rate of after-hours calls before and after, not by asking people how they feel about AI. The same PIPEDA framework applies here too: a caller asking whether the call is recorded, or asking that a recording be deleted, is exercising a right the Act already gives them, and the agent’s script should have a clear answer rather than escalating a routine question to a person.
Design the handoff first. Any call the agent cannot map to its four fields should transfer or take a message, and you should be able to see every call it declined to book. If a vendor cannot show you that list, they cannot show you their failure rate.
It moves the coordinator from answering to reviewing. The value shows up in the calls that would otherwise have been missed entirely, and in the mornings that start with a sorted list rather than a full voicemail box.
A 30-minute call is enough to tell you whether AI pays for itself here.
