I ran a small test last week. I called ten Quebec SMBs that proudly advertise "we have an AI voice agent." I asked each of them the same thing: "Hi, can you dial my supplier's number for me, navigate their phone menu down to billing, and put me on the line with an agent?"
Eight out of ten did the same thing: a polite pause, then a version of "I'll take your message and someone will call you back." That is not an AI voice agent. That is voicemail with better grammar.
Worse, these SMBs are paying $200 to $600 per month for it. Meanwhile, their better-equipped competitors — starting April 7, 2026 — are running agents that press DTMF keys, branch their configuration like Git, and politely refuse to go off the rails. Three capabilities that are not a technical detail. They are the new floor. And if your vendor is not talking about them, you are buying 2024 at 2026 prices.
The 5-second test that exposes glorified voicemail
Before you sign anything, run these three scenarios during the vendor demo. You will see the difference between a chatbot that talks and an agent that acts.
Scenario 1 — The internal transfer. "I want to speak with someone in billing." A real agent transfers the call with context already passed. Glorified voicemail says "I'll take your number and they'll call you back."
Scenario 2 — The outbound call on your behalf. "Can you call my supplier at 514-555-0199 and check if part number 4827 is in stock?" A real agent dials, navigates the menu, waits, asks, notes the answer. Glorified voicemail says "I cannot make outbound calls."
Scenario 3 — The phone menu. "When you reach Bell for my account, press 2 for French support, then 3 for billing." A real agent sends DTMF tones. Glorified voicemail literally says the word "two" out loud into the phone.
Three scenarios, three seconds each to judge. If your current agent fails two out of three, you don't have an agent. You have a voicemail system with a less robotic accent.
Why April 2026 changed everything (and why no one in Quebec is talking about it)
On April 7, 2026, ElevenLabs shipped a major update to its Conversational AI API: DTMF support, Git-style branching, scoped analysis by test folder, built-in safety guardrails. Around the same time, OpenAI locked in gpt-realtime at 250 ms latency — MultiChallenge score of 30.5% versus 20.6% for the December 2024 model. I broke this down in detail in my analysis of the April 2026 shift for Quebec SMBs.
These three releases did not land together by accident. They mark the end of one era — the "AI voice agent that answers" — and the start of another: the agent that executes real tasks. Gartner now projects that 40% of enterprise applications will feature task-specific AI agents by the end of 2026.
The problem is that most vendors in Quebec keep selling the old generation while pocketing the premium of the new one. Here are the three capabilities they don't love explaining.
Capability #1 — DTMF: the agent that actually dials digits
DTMF stands for Dual-Tone Multi-Frequency. It is the sound your phone makes when you press a key. The full lexicon of terms like this one is in our AI voice agent glossary for Quebec SMBs.
Before April 2026, most AI voice agents did not know how to send real DTMF signals. If you asked them to press "2", they literally said the word "two" into the phone. Result: they crashed the moment they had to navigate a real IVR — Desjardins, Bell, your wholesale supplier. ElevenLabs added native support for digits 0-9, *, #, and 0.5 s / 1 s pause characters, which finally lets the agent traverse any interactive voice system.
What does that mean for a Quebec SMB in practice?
A dental office verifying insurance with SSQ. An HVAC tech stuck on hold with the compressor supplier for 14 minutes. An insurance broker calling Intact for a quote. In all three cases, a human spends 10-15 minutes per call just navigating the menu. If your agent does it for you — while you are serving a client standing in front of you — you recover between 7 and 10 hours per week. At $35/hour fully loaded, that is $1,120/month you stop burning on hold music.
Capability #2 — Branching: experiment without breaking prod on Friday night
Developers have had this for 15 years with Git: you create a branch, test a change, compare it to the live version, merge if it works better. Before April 2026, an AI voice agent was cowboy territory: you tweak the prompt, push to prod, pray it doesn't break Friday reservations.
ElevenLabs' new branching feature changes the model. You now have a "prod" version answering real clients, and a "test" version where you try a new tone, a new bilingual handoff policy, a new transfer logic. You send 10% of traffic to the branch for a week. The metrics speak. You merge — or you kill it.
For a Quebec SMB this means two concrete things. First, you can finally fix FR/EN code-switching failures without taking the service down (we went deep on this in our piece on bilingual voice agents that crash in 3 seconds). Second, you can prove a change worked with numbers, not a gut feeling. It is the difference between "I think clients are happy" and "first-call resolution went from 61% to 74% on branch v2."
If your current vendor says "we'll push an update tonight, should be fine," you are still in 2024.
Capability #3 — Guardrails: the agent that politely refuses to drift
The recurring voice-AI scandal is the agent that invents a price, promises an appointment that doesn't exist, or — worst case in Quebec — leaks a client's personal info to another caller in direct violation of Bill 25. Until recently, the only defense was to write a very long system prompt and hope the model held its line under pressure.
The new safety guardrails change the approach. You declare hard rules, outside the prompt, evaluated on every turn. Examples for a Quebec SMB:
- "Never quote a price above $500 without human confirmation."
- "Never confirm an appointment with less than 2 hours' notice without reading a no-show policy warning."
- "Never read out an address, date of birth, or health card number aloud — even if the client asks."
Paired with scoped analysis — one dashboard per agent or per branch, instead of one giant bucket — these guardrails turn Bill 25 compliance into something you can demonstrate to an auditor, not just promise verbally. It's the difference between telling a lawyer "we think we're compliant" and showing them a log sorted by rule.
The honest counter-argument — and why it collapses in 12 months
I hear this a lot: "Masdouk, my SMB doesn't need all that. I just want someone to answer the phone."
I hear you. But listen to what the data says. According to GoodCall's breakdown of the agentic AI framework, businesses deploying an agent that executes tasks (booking, transfer, outbound call, IVR navigation) see a 25-35% conversion uplift. Those deploying an agent that only answers questions plateau at 8-12%.
On 100 inbound calls per week, the "orchestrated" SMB converts 25-35 more clients. If your average ticket is $180, that's $4,500 to $6,300 per week. Meanwhile the "passive" SMB pays the same monthly fee for 8-12 conversions. In 12 months, the gap will be so brutal that clients will default to the SMB that treats a call as an action, not a memo.
The choice isn't "spend more or spend less." The choice is "pay the same price for a 2024 agent or a 2026 agent." If your vendor keeps billing you for the passive version, they pocket the difference.
The SMB grid: what this actually changes on your invoice
Here is what we see on the ground, weighted average across 14 recent Quebec deployments:
The right column costs $100-200 more per month. It recovers 14-22 hours of human labor — roughly $700 to $1,100 at loaded cost. Payback hits in under two weeks. We walk through exactly how to measure that payback in our 14-day deflection test for an AI voice agent.
What TECHMA installs — and why we refuse passive setups
Our position is blunt: we no longer install glorified voicemail. If a client comes to us wanting an agent that "just answers questions," we first walk them through what they are leaving on the table. Then we install, configure, integrate with their CRM, wire up the SIP telephony, write the guardrails, run the A/B for two weeks. The client touches nothing. Our team handles the full setup, monitors the first two weeks, and delivers a turnkey dashboard.
If you want a serious comparison, read our detailed AI voice agent vs human receptionist math. You'll see the financial gap widens even further when the agent is orchestrated instead of passive.
Verdict
If your vendor isn't talking about DTMF, branching, and guardrails, you are buying a rotary phone in 2026. That is their right. Changing vendors is yours. The competitors who already made the jump to active orchestration are quietly capturing the calls your passive agent is only transcribing.
The real test, from today forward, is the five-second one: "can you dial a number, navigate a menu, refuse a price I did not authorize?" Three yeses, you have an agent. Three nos, you have expensive voicemail.