"For French, press 1. For English, press 2. For sales, press 3. For after-sales service, press 4. To speak with a representative, please stay on the line — your call is important to us."
If you've ever called your telecom provider, your dental clinic, your car dealership, or any reasonably structured business in Quebec, this script is familiar. And you know what happens next. The customer hits 4, lands in a submenu, hits 2, listens to two minutes of elevator music, then hangs up. A 2026 NICE study just put a number on this: 67% of consumers abandon the call during IVR navigation, and systems with more than 10 nested levels show abandonment rates of 30% to 50%.
For a Quebec SMB handling 800 calls a month, that's up to 400 people who wanted to buy, book an appointment, or get support — and who hung up before any human, or even a voicemail, ever heard them.
The new architecture that's been spreading across the industry since spring 2026 is called the multi-agent handoff pattern. Instead of a rigid touch-tone tree, you get an AI triage agent that listens to what the customer is saying, classifies the intent with an LLM, then hands the call off to a specialist AI agent — or to a human, with the full context passed along. According to the Metrigy CX Optimization 2025-26 study, 37.6% of companies plan to fully replace their IVR with an AI triage system, and among the best (the "Research Success Group"), that figure jumps to 62.5%.
This article breaks down how the architecture actually works, why it's a game-changer for a Quebec SMB, and what you should demand from your voice AI provider before signing.
Why your current IVR is an abandonment trap
A classic IVR (Interactive Voice Response) works like a multiple-choice form, but spoken. The customer is forced to translate their need — "I want to cancel my Thursday appointment" — into a sequence of keypresses the system understands. The problem: nobody thinks in keypresses. People think in sentences.
The friction stacks up in three stages:
1. The mental translation. The customer has to guess which category their problem falls into. "Cancel an appointment" — is that "customer service" or "account management"? When the menu is poorly named — and it almost always is — the customer picks at random, lands in the wrong place, and backs out.
2. The depth of submenus. Each extra level is a drop-off point. Systems with more than 10 nested layers reach 50% abandonment. That's not a bug, it's ordinary cognitive fatigue.
3. Zero context transferred. When the customer finally reaches a human, they have to re-explain everything. The file number, the context, the history. The IVR only passed along a label: "this call is for service A." For dental clinics, as we recently documented, that lost context costs up to $100,000 a year.
Multiply this by 800 calls a month and you get the macro number: based on recent calculations published in our breakdown of the true cost of missed calls in Quebec, an average SMB loses $126,000 per year just from calls that nobody handles end-to-end.
The multi-agent architecture: three tiers, one conversation
The handoff pattern replaces the menu with a three-tier architecture. Think of it like a well-run hotel front desk.
Tier 1 — The triage agent (the concierge)
This is the first voice AI agent that picks up. Its job is hyper-specialized: listen to what the customer says, understand the intent in natural language, and route. It doesn't sell, doesn't troubleshoot, doesn't book appointments. It sorts.
In practice, the triage agent runs on a fast LLM (GPT-Realtime-2 or Gemini 3.1, for example) with a short prompt and classification tools. The customer says "I have a billing issue on my May statement," the agent replies "got it, transferring you to our billing specialist," and the handoff happens in under half a second.
Tier 2 — The specialist agents
Each specialist is trained on a narrow domain: appointment booking, technical support, billing, lead qualification, order taking, and so on. Narrow means better. An agent that only books appointments at a dental clinic knows every treatment code, Dr. Tremblay's Tuesday availability, the cancellation policy, and the right tone to use when an anxious patient is calling about a root canal.
ElevenLabs actually published in May 2026 a new visited_agents field in its conversation response, which lets you trace exactly which specialist agents participated in a call — useful for diagnostics and internal billing.
Tier 3 — Human handoff (the rescue)
If the conversation goes sideways — anger, out-of-scope request, legal, urgent — any specialist agent can escalate to a human with the full context passed along. Not just "this call is yours," but "this call is from Jacques Lavoie, customer since 2022, double-billing issue on the May 12 transaction of $89.99, already tried solution X which didn't work, customer tone: frustrated but constructive."
According to LiveKit's analysis, this kind of well-implemented handoff cuts call abandonment by up to 40% compared to a classic IVR transfer. The difference is simple: when the human picks up, they don't need to ask the basic questions, and the customer doesn't feel like they're starting over.
The before/after numbers for a Quebec SMB
Contact centers that switched from classic IVR to a multi-agent architecture in 2025-26 report (sources: Metrigy + industry feedback):
- Call abandonment: from 35% down to 5–10%, about 75% lower
- Average handle time: -20% to -35%
- CSAT (customer satisfaction): +15 to +25 points in the first 90 days
- Cost per call: around $0.40 vs $7–12 for a fully human call
For a Quebec SMB that used to sit at 50% abandonment (typical of small clinics, garages, salons, and restaurants sharing a single number), going down to 8% means recovering 336 of the 800 monthly calls. If 30% of those recovered calls convert into appointments or sales with an average ticket of $150, that's $15,120 in extra monthly revenue. Against a typical SMB voice AI investment of $500 to $1,200 a month, the payback is usually reached in 2 to 3 weeks for high-volume businesses.
What it changes for a bilingual SMB (and Law 25)
The multi-agent architecture also solves two Quebec-specific problems that classic IVRs handled poorly.
Bilingualism. In a traditional IVR, you have to maintain two parallel menu trees — one French, one English — and any change has to be made in both. With a triage agent, the LLM switches between languages automatically, and each downstream specialist can be either bilingual or dedicated to one language. Even better, as we showed for Montreal hotels and restaurants deploying in 8 languages, the multi-agent architecture lets you add specialists per language without redoing the routing logic from scratch.
Law 25. Quebec's regulation on personal information requires you to be able to document data flow. A classic IVR just tells you "call routed to service A at 2:32 PM." A properly instrumented multi-agent architecture tells you: "triage classified as billing, transferred to billing-specialist, who accessed file #4827 (name, email, truncated card number), no plaintext transcript stored, 30-day retention." That's the difference between a black box and a log book you can show an auditor.
Platforms like Telnyx with their multi-agent handoff announced in May 2026 and ElevenLabs with their Conversational AI 2.0 (which now includes a custom turn-taking model and a RAG knowledge base) are starting to expose these traces natively.
The pitfalls to avoid when choosing this architecture
Not every vendor selling you "multi-agent" delivers the same thing. Three questions to ask before signing:
1. Is context actually passed at handoff, or just the label?
Ask to see a full transcript of a transfer. If you only see "→ transfer to booking-specialist," that's a disguised IVR. If you see the conversation history, the identified intent, the customer profile, and the emotional state passed to the next agent, that's a real handoff.
2. How long does the transfer take?
A good handoff happens in under 500 ms. Past one second, the customer perceives a cut and reflexively either hangs up or repeats themselves — exactly like in a bad IVR. Modern platforms expose this metric in their logs.
3. What's the human escalation plan?
If every specialist agent has its own "human button" calling the same cell phone, you've got a bottleneck. Good design: each domain has a dedicated escalation, with routing rules by urgency type and time of day. And the human should receive a written summary before taking the call — not after.
How TECHMA deploys this for SMB clients
Multi-agent architecture is part of TECHMA's standard deployment for 2026, because it pays off even at low volumes. Our team configures the triage, the specialists, the handoff rules, and the human escalations, plugs it all into your CRM or your appointment software, and runs two weeks of simulation tests before going live. You don't have to understand LLM details or prompt engineering — that's exactly the value of working with a specialized team rather than cobbling together a self-service solution.
The typical transition unfolds in four phases: audit of your current call flow (1 week), design of the triage and specialists (1 week), parallel testing alongside your existing IVR (2 weeks), and full cutover. During the testing phase, you see before/after numbers on the same call types, which makes the decision to retire the IVR much easier.
The takeaway
The classic IVR was designed in 1989, at a time when voice recognition didn't work and the natural option was asking the customer to punch keys. Thirty-seven years later, we no longer have any excuse for making 67% of our callers hang up in a menu tree. Multi-agent architecture — AI triage, focused specialists, instrumented human handoff — pushes call abandonment into single digits, doubles customer satisfaction, and pays for itself in 2 to 6 months for most Quebec SMBs.
If your main number still says "for French, press 1," you're leaving calls — and therefore revenue — on the table every day your menu is live. The right moment to make the jump is before your competitors do. As of May 2026, plenty of them already have.