A few months back, a clinic owner in Trois-Rivières told me his story. He had signed with an AI voice agent vendor that advertised $0.07 a minute. Three months later, his invoice had climbed to $0.31 a minute — premium voice fees, LLM tokens, telephony, "advanced orchestration." He was locked into a 12-month contract.
That kind of surprise is everyday life for Quebec SMBs shopping for an AI voice agent in 2026. The market is exploding — recent projections put it at $2.4 billion US today, heading to $47.5 billion US by 2034. The best tools resolve 90 to 95 % of calls without a human. But the gap between the solution that transforms your business and the one that turns into a nightmare comes down to one thing: the questions you ask before signing.
Here are the ten questions a serious vendor can answer without flinching. If yours dodges three or more, close the file.
1. What is the total cost per call, all-in, with three worked scenarios?
The $0.05/minute trap is everywhere. Retell published a 2026 analysis of VAPI showing that the advertised rate becomes $0.25 to $0.33 per minute in real production, once you add premium voices, LLM tokens (GPT, Claude, Gemini), telephony, and orchestration fees. Some vendors stack five or six cost lines on the invoice.
What you want to hear: one all-in per-minute or per-call rate, with three scenarios — baseline volume, seasonal peak, heavy use — projected over 12 months. A real partner hands you a costed document, not a glossy brochure.
Red flag: "It depends on your setup, we'll figure it out after signing." That's exactly how you get squeezed. We broke down the real cost structure of an AI voice agent in Quebec in 2026 with the four pricing layers no vendor mentions upfront.
2. Does your agent understand Quebec French — and handle FR/EN code-switching?
This is THE question American blogs systematically miss. An agent trained on international French doesn't catch "icitte", "tantôt", "on va-tu", or the fluid mid-sentence jump to English that Montreal callers make all the time. Word Error Rate (WER) can double if the speech recognition model wasn't trained on authentic Quebec French.
What you want to hear: a live demo, with a real regional accent, including one or two FR → EN → FR pivots. The vendor should be able to share WER metrics measured on real Quebec calls, not generic benchmarks.
Red flag: "Our model speaks 50 languages." Sure, but does it speak Quebec French? Ask for an anonymized transcription of a real customer call handled in Quebec.
3. What is your measured latency — not marketed latency?
In a human conversation, we leave 200 to 300 milliseconds between turns. Beyond 500 ms, the illusion breaks and the caller hangs up. AI voice agent latency gets measured on three levels: API response time (target: < 200 ms), perceived wait including silence detection (< 400 ms), and end-to-end conversation delay (< 600 ms).
What you want to hear: specific numbers, measured in production, not spec-sheet numbers. The best vendors monitor latency in real time and can export their P50/P95 medians from the last 30 days.
Red flag: "Our agent is ultra fast." Too vague. Recent gpt-realtime models from OpenAI show an 18.6 percentage point improvement in instruction-following — but that only matters if they're properly integrated into the vendor's stack.
4. Where is the call data stored, and how do you comply with Law 25?
Quebec's Law 25 imposes strict obligations: explicit consent, limited retention, right to erasure, breach notification. A vendor storing call recordings on US servers without a Canadian data residency option exposes you to fines of up to $25 million or 4 % of worldwide revenue.
What you want to hear: Canadian hosting (or at minimum a Canadian residency option), configurable retention policy, documented process for handling access and deletion requests, SOC 2 or equivalent certifications. Bonus: a named privacy officer you can actually contact.
Red flag: "Our servers are GDPR compliant." European GDPR and Quebec's Law 25 look alike but aren't identical. The vendor must demonstrate specific understanding of the Quebec regime.
5. What native integrations do you have with the tools I already use?
An AI voice agent that doesn't connect to your calendar, CRM, or industry software is just a fancy answering machine. The real savings come from end-to-end automation: the agent takes the booking, logs it in Dentitek or Cliniciel, sends the SMS confirmation, updates the record in Zoho, and notes everything in the customer file.
What you want to hear: a precise list of native integrations (Google Calendar, Microsoft 365, HubSpot, Zoho, Salesforce, plus the Quebec-specific industry software relevant to your sector). Also ask whether integrations are API-native or routed through tools like Make.com or Zapier (which add latency and failure points).
Red flag: "We can integrate with anything." Too vague. Ask: "Show me, on your dashboard, a live integration with [your tool]." If it takes three weeks to configure, it's not a native integration.
6. Can I edit scripts myself, or do I go through your team?
Two schools here. First: you handle everything yourself in a DIY interface — sounds flexible, but in practice 80 % of SMBs lack the time or expertise to test their own conversation flows and end up with a poorly calibrated agent. Second: a team configures and maintains the agent for you — you gain time and quality, but depend on vendor speed.
What you want to hear: a hybrid model. Your team can adjust simple things (business hours, greeting, qualification questions) in a dashboard, and the vendor handles heavy configuration, integrations, and ongoing optimization.
Red flag: "Submit a ticket to our team, we respond within 5 business days." For an SMB, that's unacceptable. Aim for 24 to 48 hours on common changes.
7. How do you handle escalation to a human — and with what context?
No AI voice agent, even the best one, handles 100 % of calls. The real question is what happens to the 5 to 10 % it can't resolve. If the agent hangs up or puts the caller on hold with no transcript, you lose the customer. If the call transfers to a human with a real-time contextual summary, the conversation stays smooth.
What you want to hear: smart transfer to the right extension, with a real-time transcript, an intent summary, and ideally a recommended action. Some vendors even transfer with customer details already captured.
Red flag: "We can transfer to your main line." That's not escalation, that's a call forward. This comparison between chatbot and AI voice agent for Quebec SMBs spells out why handoff handling often separates a successful project from a failed one.
8. What is your trial policy: duration, commitment, cancellation fees?
A vendor who demands a 12-month commitment before letting you test is telling you something. The best platforms accept a two-week to one-month pilot on a capped call volume, with no long-term commitment. You should be able to measure concrete results (resolution rate, satisfaction, calls captured) before signing an annual contract.
What you want to hear: low-cost pilot, month-to-month billing with no long lock-in, clear cancellation fees (ideally none after 30 days), portability of your data and configured flows if you leave.
Red flag: "Our standard contract is 36 months." For a technology moving this fast — ElevenLabs shipped five major updates in April 2026 alone, including multi-agent workflows and new LLMs like Gemini 3.1 Pro and Qwen 35 — locking you in for three years makes no sense.
9. What uptime guarantees (SLA) and what French-language support?
Your AI voice agent is often the first point of contact between your business and your customers. A one-hour outage mid-day costs you 30 to 60 missed calls depending on volume. A serious SLA guarantees 99.9 % uptime (max 8 hours of downtime per year) with automatic credits if the threshold is missed.
What you want to hear: written SLA, public real-time status dashboard, documented incident notification process, and human support reachable in French during your business hours. A direct Slack channel or an emergency phone number for active accounts is a concrete plus.
Red flag: email-only support with "response within 72 business hours." When your agent crashes at 9 a.m. on a Tuesday, you can't wait until Thursday afternoon.
10. Who handles initial configuration and ongoing maintenance?
This is probably the most important question — and the most forgotten one. Many vendors sell you the platform, then leave you alone with the technical documentation. Configuring an AI voice agent that actually works in production requires expertise in prompt engineering, flow design, API integrations, and monitoring. It's not a one-afternoon job.
What you want to hear: a turnkey model where a team configures, tests, launches, and maintains your agent. You provide the business context, they deliver an operational agent. Then they optimize continuously based on real call data.
Red flag: "We'll send you your account credentials, good luck!" For an SMB with no in-house AI team, that's a recipe for failure. At Agent IA Vocal, for instance, all configuration, CRM and calendar integrations, and continuous optimization are handled by the TECHMA team — you never open a technical interface yourself.
The checklist to print before your next discovery call
Quick recap. Before signing, your vendor should be able to answer these ten questions in writing, with numbers and examples. If two or three answers are fuzzy, dig deeper. If five or more are evasive, move on.
The Quebec AI voice agent market is no longer a tiny insider world. In 2026, we're seeing specialized vendors who understand local realities — Quebec accent, Law 25, French-language industry software, French-speaking support. We're also seeing US platforms translate their marketing brochure and hope it's enough. The difference lives in the details: the ten questions above are your filter.
If you're still weighing technology approaches — voice agent, text chatbot, traditional IVR menu — the arrival of multi-agent systems completely reshapes the 2026 picture and is worth reading before you decide.
One last tip: always ask to speak with two existing customers of the vendor — ideally in your industry, ideally in Quebec. Ten minutes on the phone with someone who has used the solution for six months will teach you more than two hours of sales presentation. And if the vendor refuses to put you in touch with customers? You just answered the eleventh question, the one nobody told you to ask.