The phone rings. An English-speaking customer on the line. And your receptionist answers... in French. Three seconds. That's all the time a caller needs to decide whether to stay on the line or hang up.
According to 2026 data on the Montreal bilingual market, 38% of callers hang up when greeted in the wrong language, and 52% will choose a competitor who speaks their language. For a Quebec SMB serving customers in Montreal, Quebec City, Trois-Rivières, or Gatineau — or simply a mix of local and tourist clientele — that's a daily bleed.
And yet most SMBs we meet still try to cover this with a single French-speaking receptionist who "gets by in English" — or worse, with a 2010s-era IVR menu prompting "Press 9 for English."
The good news: on May 7, 2026, OpenAI released GPT-Realtime-Translate, a voice model supporting 70+ input languages and 13 output languages in real time, preserving the speaker's original tone (official OpenAI announcement). Combined with the GPT-5-class reasoning of GPT-Realtime-2 — which we covered in detail — it's now possible, and affordable, to deploy a truly bilingual AI voice agent for a Quebec SMB.
Below are the exact 6 steps we follow at TECHMA to configure a bilingual French-English AI voice agent for our SMB clients in Quebec. No DIY here: this guide exists to show you what we do for you, not to ask you to do it yourself.
Why a bilingual AI voice agent isn't a "nice to have" in Quebec
A full-time bilingual receptionist in Montreal costs roughly $45,000 to $55,000 per year, before benefits. Outside Montreal — in Repentigny, Saint-Jérôme, or Trois-Rivières — the challenge isn't the salary, it's finding someone who speaks both languages as fluently as your customers demand.
A properly configured bilingual AI voice agent will:
- Pick up in 1 second, in the caller's language, 24/7.
- Switch between French and English mid-sentence if needed.
- Never carry an accent that frustrates either an Anglophone or a Francophone.
- Document every call for compliance (useful for Law 25).
- Cost a fraction of a human receptionist.
For a side-by-side comparison between a human receptionist and an AI voice agent in Quebec, see our 7-criteria analysis. Now, the 6 deployment steps.
Step 1: Choose the right base model (and why it matters more in 2026)
In 2026, you have essentially three families of serious multilingual voice models for SMBs.
1. GPT-Realtime-2 (OpenAI) — Conversational model with GPT-5-class reasoning, 128K context window, capable of holding a long conversation without losing the thread. Ideal when the agent has to qualify, ask follow-up questions, or handle complex flows (medical scheduling, custom quotes).
2. GPT-Realtime-Translate (OpenAI, May 7, 2026) — Dedicated to live translation between 70+ input languages and 13 output languages. This is the one that changes the game for bilingual SMBs. Priced at $0.034 per minute. Whereas GPT-Realtime-2 "speaks" multiple languages via prompt, Realtime-Translate translates in real time while preserving the speaker's tone.
3. ElevenLabs Agents — Voice-agent-native platform. The May 2026 release added MCP, guardrails, and versioning — we walked through them in this article. ElevenLabs also offers a dedicated bilingual answering service page.
Our pick for 90% of Quebec SMBs: GPT-Realtime-2 as the conversational brain, with voices synthesized through ElevenLabs for Quebec-French naturalness. We bring Realtime-Translate into the stack only when a 3rd or 4th language is needed (Spanish or German tourists in hospitality, for example).
Step 2: Define the language-detection strategy
Three approaches, depending on the context.
A. Auto-detection on the first word. The agent opens with a short bilingual greeting: "Bonjour, hello, how may I help you today?". The model detects the caller's first reply and continues in that language. This is the default approach in Microsoft Dynamics and the one we use for most clinics and restaurants.
B. Phone-number routing. Two distinct numbers (one FR, one EN) route to the same agent with a different language instruction. More rigid but ideal for bilingual brands like Montreal real-estate agencies running differentiated campaigns.
C. Explicit ask. "Would you prefer to continue in French or English? / Préférez-vous continuer en français ou en anglais?". Slower (adds 4–6 seconds) but useful for formal contexts — legal services, notaries.
For most SMBs, we go with auto-detection. And we specifically train it with Quebec-accented samples — otherwise the model occasionally mistakes a Québécois-pronounced "yes" for French.
Step 3: Configure voice and tone for each language
This is where most amateur bilingual configurations fall apart. A voice that sounds professional in French can sound robotic in English — or vice versa.
What we do at TECHMA:
- Two distinct voices in ElevenLabs, calibrated for Quebec French (not Parisian French) and North American English. The rolled "r" of a Parisian voice instantly breaks the illusion for a Montrealer.
- Localized vocabulary: "stationnement" vs. "parking lot", "dépanneur" vs. "convenience store", "cellulaire" vs. "cell phone". We write two full scripts, never a literal translation.
- Tu vs vous: in Quebec French, "vous" is the norm for a professional agent. In English, we adapt to the sector (formal for notaries, more casual for hair salons).
- Code-switching: in Montreal, customers often alternate between the two languages in a single sentence ("Je voudrais booker un rendez-vous pour next Tuesday"). The agent must follow without "resetting."
Step 4: Build prompts and tools in both languages
An AI voice agent is more than a voice — it's also a brain that checks calendars, queries availability, and sends SMS confirmations. All those tools (function calls) need to be dual-language.
Concretely:
- The system prompt is written in English (LLMs reason better in English), but with explicit instructions: "You will respond in the customer's language. If they speak French, use Quebec French conventions (vous, dépanneur, stationnement)."
- Tool descriptions are also in English, but parameters accept both languages. A
book_appointmenttool acceptsservice_namein French OR English and maps internally. - Error messages the agent may speak are pre-translated, never generated on the fly. This prevents translation hallucinations on specific service names ("déneigement de toiture" doesn't reliably translate to "roof snow removal" in every context).
Step 5: Law 25 compliance (and why bilingualism makes it harder)
Quebec's Law 25 requires that personal information collected over the phone be stored and processed with explicit consent and clear documentation. For a bilingual AI agent, this means:
- Bilingual privacy notice delivered at the start of each call: "Cet appel sera traité par un agent IA et enregistré pour la qualité du service / This call will be handled by an AI agent and recorded for quality assurance."
- Data hosting in Canada or in the US with an adequacy agreement. Most SMBs skip this until an audit catches them.
- Retention policy: how long do you keep transcripts? We default to 90 days, except specific cases (medical: 7 years).
- Right of access: if a customer asks for their data in a specific language, you need to deliver it in that language. Our stack automatically tags the language of each interaction.
It's dry, it's boring, but this is what separates a deployment that lasts from an agent that gets shut down after a complaint.
Step 6: Test, measure, adjust
We never deploy to production without this step — and this is where most amateur configurations derail.
Metrics we track for a bilingual agent on our clients' lines:
- Language-detection accuracy on the first word (target: > 96%).
- Mid-conversation switch rate (how often the agent has to change languages mid-call, and how many switches succeed).
- Hang-up rate in the first 5 seconds (target: < 8%).
- CSAT by language: customer satisfaction measured separately in FR and EN. If the gap exceeds 10%, there's a voice or prompt issue in the weaker language.
- Cost per resolved call: to compare against your current receptionist.
For the first two weeks, we listen to every conversation (with Law 25 consent, of course) and tune the prompts continuously. From week 3, we sample 10% of calls.
How much does it actually cost?
For a ballpark, here's what we typically charge for a bilingual SMB deployment in Quebec in 2026:
- Initial setup (one-time): $3,500 to $6,500 depending on complexity (CRM integration, calendar, SMS).
- Monthly operational cost: $250 to $800 depending on call volume (includes GPT-Realtime-2 + ElevenLabs + hosting + TECHMA monitoring).
For an SMB currently paying $50,000 per year for a bilingual receptionist (and still missing 20% of calls after 5 p.m.), ROI typically lands in 4–7 months.
Recap
A bilingual AI voice agent in Quebec in 2026 is six things we assemble for you:
- The right base model (GPT-Realtime-2 + ElevenLabs for 90% of cases).
- A language-detection strategy adapted to your customer base.
- Two distinct voices calibrated for Quebec French and North American English.
- Dual-language prompts and tools.
- Law 25 compliance from second one.
- Testing, testing, and more testing.
All of that — we do for you at TECHMA. You give us your numbers, your services, your business hours; we handle configuration, testing, and ongoing tuning. If you run a bilingual SMB in Quebec and you're still losing calls because no one picks up in the right language, talk to us: we can deploy a bilingual AI voice agent in under 14 days.