Friday morning, 9:47 a.m. A dentist in Laval is watching a demo from an AI voice agent vendor. The agent books appointments perfectly. Confirms the time. Even suggests an alternative. The dentist signs on Tuesday.
Wednesday, 11:03 a.m. A real client calls: "Uh yeah, hi, I had an appointment Thursday but my wife already took Friday off, so can we move it, but like not before 2 because the kids get back from school, and by the way do you still take the Zellers card?"
The agent loops back to the start. Asks for the name again. The client hangs up. The receptionist, who has fielded 4 other calls like this that morning, calls the dentist. The contract is cancelled by Friday.
Cost of the mistake: $1,200 in setup fees, three weeks to rebuild the routine, and a dent in trust toward AI that won't heal quickly.
The real question isn't "does the agent give an impressive demo." It's "does it survive a real client."
Why 8 out of 10 demos are rigged (without anyone lying)
It's not that vendors lie. They just show the happy path: the client who says their name clearly, gives a precise date, confirms, hangs up. Thirty seconds. Clean.
The problem? Real Quebec clients don't talk like that. They change their minds. Interrupt. Switch between French and English mid-sentence. Ask off-topic questions. Leave 4-second silences that trick the agent into speaking over them.
A recent QA framework for voice agents shows that the difference between an agent that holds up and one that breaks plays out across four layers: infrastructure, execution, user reaction, business outcome. The demo covers the first layer. The other three, you discover in production. Too late.
So here's the simple idea: run the tests during the demo, not after you sign.
Here are the 5 scenarios every Quebec SMB should play live before committing. Not ask to see — play yourself. In real time. With the vendor listening.
Scenario 1: The mid-course change of mind
You're booking an appointment. Halfway through, you change the service. Real example: "I'd like to book a cleaning for Thursday — actually no, make it a full exam for the same day, and actually move it to Friday."
What you're testing: conversation recovery. The agent must adapt, not restart.
Red flag: the agent asks for your name or number again, or restarts the script from the top. On Reddit, it's complaint #1 against voice agents in 2026 — users describe agents that "loop back to the beginning" the moment a caller modifies any detail. It's the silent pain point no one tests and that sinks 60% of deployments.
A good agent keeps context across the last 5-6 turns and updates only the variable that changed. Ask the vendor how their agent handles entity tracking. If the answer is vague, be suspicious.
Scenario 2: The client who interrupts
The agent starts listing available slots. You cut in mid-second-option: "Yes, Thursday at 2, perfect."
What you're testing: barge-in. The agent must stop within 200 ms and catch your response.
Red flag: the agent keeps talking over you, or worse, forces you to wait until it finishes its sentence. ElevenLabs, in its April 7, 2026 update, specifically strengthened multi-agent support with per-agent scoped analysis — proof that even serious vendors recognize turn-taking remains a blind spot.
In a hair salon where the client is holding coffee and responding fast, an agent that talks too long is a client who hangs up. If the demo forces you to wait for the agent to finish every sentence before answering, it's over.
Scenario 3: The completely off-script question
Your theoretical script: book an appointment. Your test question: "By the way, are you open December 24th?" or "Do you still take the gift card from my old vet?"
What you're testing: graceful handling of the unknown. Two acceptable behaviors:
- The agent transfers to a human with clear context ("I'll put you through to reception, mention your gift card question to them").
- The agent admits it doesn't know and offers an alternative ("I don't have that info, but I can have someone who does call you back within the hour").
Critical red flag: the agent invents an answer. That's called a hallucination, and in an SMB, it costs dearly. A voice agent that makes up a price or invents a policy can generate a legal bill much steeper than the AI contract itself. We've written about how to avoid these hallucinations in a voice agent — required reading before any signature.
Scenario 4: The Quebec-style French-English mix
You start in French, then mid-sentence you slip in an English expression. "Oui je voudrais un appointment pour ma fille, preferably après l'école."
What you're testing: linguistic robustness. A third of Quebec SMBs serve bilingual clients, and calls are rarely monolingual in practice. On the Island of Montreal, 40% of commercial conversations contain at least one code-switch per call.
Red flag: the agent replies in dry English, or worse, confirms the wrong time because it misread "pm" vs "after." Recent STT (speech-to-text) models — including the new ones mentioned in ElevenLabs' 2026 changelog — handle code-switching better, but you need to verify, not assume.
Ask the vendor which STT model the agent runs on and whether it's been tested specifically on Montreal-style franglais. If the answer is "we use Whisper by default" with no nuance, be careful.
Scenario 5: The 4-second silence
You deliberately drop silence mid-response. As if thinking, or checking a paper calendar. 3 seconds. 4 seconds.
What you're testing: agent patience. A good agent waits about 800 ms to 1.5 seconds, then gently prompts ("Take your time," "I'm still here"). A bad agent talks over your thinking, or worse, hangs up thinking the call ended.
Red flag: the agent fires a new question while you're still thinking. Callers feel it as "the agent is rushing me, not listening." Result: the client hangs up and calls back for a human — which erases all the value of the AI.
The latency bar not to cross, which we've documented in detail, is 800 ms for Time-to-First-Word (TTFW). But watch out: an agent that's too fast to speak into silence is also a defect. Patience matters as much as speed.
The 4 numbers to demand in writing
While you're running these 5 tests, ask the vendor to put — in writing, in the contract — the following 4 numbers:
TTFW < 400 ms. Time before the agent's first word. Above 800 ms, your client perceives a "machine that's lagging."
Turn latency P95 < 800 ms. Latency in 95% of cases. A vendor who only gives an average is suspicious — averages hide the worst calls.
Intent accuracy > 95%. Percentage of times the agent correctly understands the request. Below 90%, you lose more clients than you gain.
Containment rate > 70%. Percentage of calls resolved without human escalation. Below 50%, your ROI evaporates.
These numbers come from industry benchmarks, but for a Quebec SMB handling 200-400 calls per week, these are the thresholds not to drop below. OpenAI has actually updated its Realtime API with latency improvements that now allow well-configured agents to drop under 500 ms on full-turn latency — proof these numbers are achievable, not utopian.
The final trap: demand the audio, not just the transcript
Last piece of advice, and it's the one 90% of buyers forget: ask to receive the audio recording of your 5 tests, not just the written transcript.
Why? Because the transcript tells you what was said. The audio tells you how it was said. The hesitations, the robotic tone, the awkward pauses, the agent's "ums" — all of that disappears on paper. And it's exactly what your clients will feel.
If the vendor refuses or says "we don't keep audio for privacy reasons," that's a red flag. Under Quebec's Law 25, you have the right to receive your own call recordings — it's good governance practice.
What to take away
A 3-minute demo can hide 3 months of failed production. The 5 scenarios above take 12 minutes to play, and they tell you in real time whether the vendor is serious or selling dreams.
If you want to go further, we've compiled the full list of 10 questions to ask the vendor before signing. The 5 scenarios in this article are the live test that complements those questions — one validates the pitch, the other validates reality.
And above all: run these tests in the presence of the staff who'll use the tool. A receptionist who watched the agent bug out during Scenario 1 will have much more patience for improving it later — and much less surprise when a real client breaks the flow.
The difference between an AI project that succeeds in a Quebec SMB and one that fails isn't budget. It's rigor at the demo stage.
One last thing: book a 20-minute slot for the test, not 3 minutes. Give yourself room to play all five scenarios twice, take notes, and ask follow-up questions between each one. Most vendors will happily give you the time if you frame it as "I want to understand the edge cases before I commit." If they push back or try to rush you, that answer itself is information — it tells you how they'll behave once you're locked into a contract. A serious partner welcomes tough tests. A shaky one dodges them.