It can answer, identify, capture and route — and it must never say whether something is covered. Placing and advising on insurance is licensed work, and that single line decides the whole design.
Key takeaways
An agency owner asking about an AI receptionist is rarely asking about reception. They are asking whether the calls that arrive after five o’clock, during a storm, or while both licensed staff are on renewals, can be handled without adding a salary. The honest answer is that about half of that call volume can be, and the other half cannot be, and the split is not about difficulty. It is about licensing.
In Ontario, the Financial Services Regulatory Authority of Ontario is the regulator for the property and casualty, and general insurance sector, and its stated role is to make sure the agents and businesses that sell that insurance meet Ontario’s qualification and ongoing licensing requirements. In British Columbia, the Insurance Council of BC sets the licence qualifications and requirements for individuals, partnerships and corporations applying to hold an insurance licence. Every province has an equivalent body. None of them contemplates a licence being held by software.
Everything on one side of the line is information handling: who called, about which policy, about what event, and what happens next. Everything on the other side is advice: whether a loss is covered, what a deductible will be, whether a limit is adequate, whether to report a claim at all. A voice agent that strays into the second category is not merely wrong occasionally — it is producing the exact record a claimant’s counsel wants.
That is not theoretical. A broker who fails to obtain coverage that reasonably meets a client’s needs, or fails to clearly explain what is not covered, can be liable in negligence separately from any claim against the insurer. The standard is what a reasonably competent broker would have done given what the client said about their risks, and the evidence is the broker file — applications, emails and recorded conversations about coverage options. An automated voice that answers a coverage question is writing into that file.
Identify and log. Match the caller against the book, confirm the policy in question, and record the reason for the call in structured fields rather than a voicemail nobody transcribes.
First notice of loss. Work a fixed script — date, time, location, what happened, other parties, injuries, whether emergency services attended, whether mitigation is under way — read the answers back, and file them. This is the single highest-value use, because the facts of a loss decay fast and the call almost always arrives outside business hours.
Certificate and proof-of-insurance requests. Capture the holder, the required wording and the deadline, and route to the person who issues it. The request is clerical; the document is not.
Triage by urgency. A water loss at two in the morning and a billing question at two in the morning need different escalation. A rules-based triage on a handful of call types is more reliable than a general-purpose assistant improvising.
Write the note. A structured summary into the broker management system, attached to the right file, with the recording reference. Most agencies do not lose calls; they lose the detail of calls.
Answering a call somebody made to you is not telemarketing. Placing calls is, and this is where agencies get themselves into trouble with a tool that is perfectly safe on the inbound side.
Under the CRTC’s Unsolicited Telecommunications Rules, an automatic dialing-announcing device (ADAD) is defined as any automatic equipment capable of storing or producing telephone numbers and used to convey a pre-recorded or synthesised voice message. A synthesised-voice outbound campaign is therefore an ADAD, and a telemarketer must not initiate a telemarketing call by ADAD unless the consumer has given express consent naming the specific number. Critically, the rules state that the ADAD rules apply whether or not the call is exempt from the National Do Not Call List. So the existing-business-relationship exemption that lets you phone your own renewal list does not rescue a synthesised-voice renewal campaign — the consent requirement is separate.
Where an ADAD is used without any attempt to solicit, the same rules set the operating conditions: calls restricted to 9:00 a.m. to 9:30 p.m. on weekdays and 10:00 a.m. to 6:00 p.m. at weekends in the recipient’s local time; a clear opening message identifying the person on whose behalf the call is made, with an email or postal address and a local or toll-free number; display of the originating number; no sequential dialling; and equipment that disconnects within ten seconds of the recipient hanging up. Those addresses and numbers must stay valid for at least sixty days after the call.
The telemarketing rules themselves add registration with the National DNCL operator and an internal do-not-call list on which a name stays for three years and fourteen days from the request — and the CRTC publishes guidance for telemarketers, including for the insurance sector. If the follow-up is by email or text rather than voice, Canada’s anti-spam legislation sets three requirements for a commercial electronic message: obtain consent, provide identification information, and provide an unsubscribe mechanism — with the onus on the sender to prove consent was obtained. The CRTC enforces that legislation.
Call content at an insurance agency is dense personal information — addresses, vehicles, health details in an accident report, sometimes financial information. Under the Personal Information Protection and Electronic Documents Act the organisation remains accountable for that information whether a human or a machine collected it, and what counts as personal information is broader than most owners assume.
The Office of the Privacy Commissioner’s principles for responsible, trustworthy and privacy-protective generative AI are the practical checklist here: be open about what is collected, how, when and why; make sure outputs that could significantly affect someone are identified as machine-generated; limit collection and use to the identified purpose and avoid function creep; and where the system forms part of a decision, keep records good enough to answer an access request. If your agency does not already have one, whether your business needs a privacy policy is the first thing to settle, not the last.
A six-person brokerage in southwestern Ontario runs one CSR on the phones and diverts to voicemail after 5:30. At 6:40 on a Tuesday a sprinkler line lets go at a commercial client’s unit.
The voice agent answers on the first ring, recognises the number against the book, confirms the named insured, and runs the loss script: when, where, what, has the water been shut off, is a restoration contractor on site, is anyone hurt, is the building accessible. It reads the answers back, files a structured first notice on the commercial file, texts the on-call producer, and sends the client a written summary of what was recorded.
The client then asks whether the tenant’s stock is covered. The agent says it cannot answer coverage questions, that a licensed member of the team will call at 8:15, and that the details are already on file. The next morning the producer opens a complete notice of loss with a timestamp, instead of a 40-second voicemail saying “call me, there’s water everywhere”.
Nothing in that sequence required a licence. Everything the client actually wanted to know still did.
Inbound only, after hours only, one call type. Write the refusal script first — the exact wording used when a caller asks a coverage question, requests a change to a policy, or asks for a quote — and only then write the greeting. Review the first two weeks of transcripts line by line with a licensed staff member, and count how many times the tool was asked something it should refuse. That number, not call volume, tells you whether the script is right. The same review discipline is what makes a sign-off workflow work for written output.
Can it bind coverage or issue a certificate?
No. Selling and advising on insurance is licensed activity — in Ontario FSRA regulates the agents and businesses that sell property and casualty insurance and in British Columbia the Insurance Council of BC issues the licences. A tool can capture the request and route it. A licensed person does the act.
Do we have to tell callers they are speaking to a machine?
Treat it as required. The privacy commissioner’s openness principle asks that people be told what is collected and why, and that outputs with significant impact be identified as machine-generated (OPC guidance). Beyond compliance, a caller who thinks they are speaking to a licensed broker will interpret silence as confirmation, which is precisely the failure mode you are trying to avoid.
Can we use it to call our renewal list?
Not without express consent for a synthesised-voice call to that specific number. The ADAD rules apply even where the National DNCL exemptions would otherwise cover a call to your own client, so the existing relationship does not answer the question.
What about Quebec and the other provinces?
The federal telemarketing and anti-spam rules apply nationally, but insurance licensing and conduct are provincial. Name your own regulator before you design the script; the boundary this article describes is the same everywhere, but the body enforcing it is not.
A 30-minute call is enough to tell you whether AI pays for itself here.
