Wednesday, 9:47 a.m. An operations director at a manufacturing SMB in Drummondville dials the number of an industrial equipment supplier on Quebec's South Shore. He has three questions, not one. Forty-three seconds later, he hangs up — frustrated. The contract he was about to renew — $14,000 over twelve months — just quietly shifted to a competitor. The AI voice agent on the other end didn't scream, didn't lie, didn't crash. It did exactly what it was asked to do: follow the script. The problem? The customer hadn't read it.
This case is not hypothetical. It mirrors a pattern we keep seeing among Quebec SMBs that deployed an AI voice agent in 2025 without planning for the moment a real human — with questions asked out of order, hesitations, and unpredictable requests — actually calls. Below is the full post-mortem, transcript included, and the three flaws to fix before the same thing happens to you.
The setup: an SMB that was doing everything "right"
The company (let's call it Beaumont Metallurgy — name changed, details anonymized) manufactures custom components for the food-processing sector. Eighteen employees, roughly $4.2M in annual revenue, about 140 active customers. In February 2026, the general manager decided to deploy an AI voice agent to filter inbound calls while her assistant, on maternity leave, went unreplaced. The vendor, Ontario-based, delivered the agent in six days. Price tag: $287 per month, 3,000 minutes included. On paper, the math added up.
The agent had been trained on five intents: appointment booking, quote request, order follow-up, product information, escalation to management. Clear intents, tight script, a decent French synthetic voice. The first three weeks were quiet. Nobody noticed that 94% of inbound calls matched those five intents — because callers with other needs hung up without leaving a trace.
The transcript: 43 seconds that flipped the contract
On March 11, 2026, Patrick L. — operations director at a pork-processing plant in the Saint-Hyacinthe region — called in. Here's what the recording revealed:
Agent: "Hello, welcome to Beaumont Metallurgy. I'm the virtual assistant. How may I help you today?"
Patrick: "Yes, hi, this is Patrick from Salaison St-Hyacinthe. I'm calling about my purchase order 4412, but at the same time I need a quote on a part we've never ordered, and I also need to talk to Julie about the delivery delay from last week, because it cost us."
(2.3 seconds of silence. Agent processing.)
Agent: "Of course. You'd like an order follow-up. Can you confirm the purchase order number?"
Patrick: "I just gave it to you, 4412. But my real problem is the delay."
Agent: "Purchase order 4412 is currently in production. Scheduled delivery is March 14. Is there anything else?"
Patrick: "Yes, I want to speak to Julie."
Agent: "I can help you with: booking an appointment, requesting a quote, an order follow-up, or product information. What would you like?"
Patrick: "Transfer me to Julie, I just said so."
Agent: "I can help you with: booking an appointment, requesting a quote…"
Click.
Patrick then called two competitors. One of them answered on the twelfth ring — with a human. The $14,000 contract moved that same week. Beaumont Metallurgy never found out. The agent hadn't triggered any alert, because no "human transfer" intent had fired long enough to be logged.
Flaw #1: the agent handled a compound request as a single item to parse
The caller stated three intents in one sentence. That's the conversational norm in human interaction — not the exception. The agent, however, was trained on single-intent-per-turn logic. As soon as Patrick mixed an order check, a quote request and a transfer demand, the agent picked the first intent detected (follow-up) and discarded the other two. Worse: it never confirmed that filtering to the caller. Patrick assumed the agent had heard him in full.
This is exactly what voice-agent failure studies call "script collapse" — the moment a rigid flow breaks as soon as the user steps off the rails. And 62% of B2B conversations in Quebec involve at least two simultaneous intents in the first turn. If your agent can't detect them, it starts 6 calls out of 10 with a structural handicap.
Flaw #2: no human escape hatch
Patrick explicitly asked to speak to Julie. Twice. The agent had an "escalation to management" intent — but it was only triggered by specific phrases ("I want to speak to a manager", "put me through to someone"). "Transfer me to Julie" didn't match. The intent-detection model wasn't thinking in proper nouns — it was keyword-matching.
To understand why this kind of escalation must be tested before deployment, we laid out the 5 scenarios every Quebec SMB must play out before signing with a voice-agent vendor. "The caller asks for a person by first name" is one of them — and for good reason.
Flaw #3: the recording existed, but nobody listened to it for three weeks
The vendor did provide transcripts. In a dashboard accessed through a portal, in English, with a CSV export limited to the five programmed intents. An "intent-failure" category didn't exist in the dashboard. Patrick's call was logged as "order follow-up — resolved". Because the agent had, technically, given the delivery date for PO 4412. Mission accomplished, according to the vendor's metric.
This measurement bias is the silent source of most post-deployment disappointment. As long as you're not tracking the undetected intents, you'll never know how many contracts you're losing. It's the same bias that feeds broader reliability issues — we covered the ways a voice agent can invent pricing and cost you dearly in a separate post.
The three fixes to demand in 2026
Fix 1: a model trained for multi-intent requests
The new generation of real-time conversational models — OpenAI's gpt-realtime, released this year, and ElevenLabs' Conversational AI 2.0 — natively handle multi-intent detection within a single turn. It's no longer a nice-to-have. Your agent should be able to say: "I hear you — you need the status of PO 4412, a quote for a new part, and to speak with Julie. Which do we handle first?" That's a full paradigm shift from the single-intent logic that dominated 2023 to 2025.
Fix 2: a human escalation ramp that understands proper nouns
A decent agent in 2026 should recognize an escalation request even when it doesn't use the scripted phrase. "Transfer me to Julie", "I want to talk to a real person", "Put your boss on the line" — all three must trigger the same path. And if the named person is unavailable, the agent should offer a concrete callback in a real time window, not a vague "we'll call you back soon".
Fix 3: an "unmatched intent" metric on the dashboard
Require from your vendor a daily indicator: calls where the primary intent didn't match the catalogue. It's the metric that best predicts silent revenue leakage. At Beaumont Metallurgy, that number would have shown 14% in the first month — a red flag for anyone paying attention. Nobody was paying attention.
What this changes for your SMB
An AI voice agent in 2026 is no longer a glorified IVR. It's a conversational system that needs to be configured, tested and monitored by a team that understands both the technology and your Quebec business reality. That's why at TECHMA we never deliver an agent as "plug-and-play, configure it yourself." Our teams handle the CRM integrations, the off-script scenario tests, and the failure-metric setup — because a poorly monitored agent costs far more than the monthly fee it advertises. Before signing with a vendor, ask them the 10 questions that separate a real partner from a reseller.
Patrick L. never called Beaumont Metallurgy back. The contract is gone. But those 43 seconds of recording are worth their weight in gold for any other SMB that wants to avoid repeating the story. The question isn't whether your AI voice agent will meet a Patrick L. one day — it's what's going to happen when it does.
Three questions to ask yourself before next Monday
If you already run an AI voice agent, or you're about to sign with a vendor, run this quick self-check tonight. It takes less than fifteen minutes, costs nothing, and will tell you more about your actual risk than any sales pitch.
First, open the last 30 days of call logs and count how many conversations lasted under 60 seconds without producing a lead, a booking, or a case number. In healthy deployments, that figure sits under 8%. In unhealthy ones, it's 20% or more — the silent leakage zone. If the dashboard doesn't let you filter like that, you already have your answer about the monitoring problem.
Second, try this test call from your cell phone: "Hi, I need to follow up on an invoice, change my credit card on file, and talk to [name of a real person in your company]." Listen to what happens. Does the agent acknowledge all three requests? Does it try to route you to the right human when you ask by name? Does it offer a specific callback window if that person is unavailable? Any "no" here points to a concrete fix to demand from your vendor.
Third, ask your vendor to show you the list of "unmatched intents" from the last 30 days. If they can't produce it in under ten minutes, you're running blind. Good vendors have this ready — it's how they improve the agent. The ones who can't are either hiding it, or they're not measuring it at all. Neither answer is acceptable for a tool that sits between your customers and your revenue.