Friday, 3:47 p.m. A customer calls your garage in Boucherville. He hit a curb, his tire is making a weird noise, and he wants to know if it's safe to drive home to Saint-Hyacinthe. Your AI voice agent reassures him, takes his details, books him in for Monday. Good answer — except the customer hung up 30 seconds earlier thinking, "If only I could have shown him the tire."
That frustration? In April 2026, ElevenLabs just killed it. And most Quebec SMBs don't even know it exists.
The opinion that ruffles feathers: voice-only is now a handicap
Let's say it plainly, because nobody else does. An AI voice agent that only handles voice in 2026 is the equivalent of a website that doesn't display images in 2010. It works. But it's costing you contracts you'll never even know you missed.
On April 7, 2026, ElevenLabs added sendMultimodalMessage to the useConversationControls hook and exported the MultimodalMessageInput type. Translation: your customer can now text a photo during a phone call, and the AI voice agent sees it, analyzes it, and responds by voice. Telnyx did the same thing on the telecom side with MMS support during active calls. It happened with no media fanfare in Quebec. It's redefining what a service call means.
Why your current providers aren't telling you about this
Three reasons, and none of them work in your favor.
First reason. The majority of voice agent integrators in Quebec built their setups 12 to 18 months ago around a voice-only stack (STT → LLM → TTS). Activating multimodal requires reworking the Twilio webhook, swapping the LLM for one with vision (Gemini 3.1 Pro, GPT-4o, Claude Sonnet), and adding the image analysis pipeline. It's not a toggle. It's three honest days of integration work.
Second reason. As long as nobody offers it, nobody pays for it. The typical Quebec SMB doesn't know their AI voice agent could receive images. So they don't ask. So the integrator doesn't bill it. The status quo is comfortable — for the provider.
Third reason, and the most honest: not everyone needs it right now. An SMB taking restaurant reservations in Drummondville probably doesn't need multimodal this year. But an insurer, an auto body shop, a plumber, a vet clinic? They need it now.
The 5 Quebec industries where multimodal changes the game in 2026
1. Insurance and claims. Customer calls at 10 p.m. — their car was just hit in an IGA parking lot. With a multimodal agent: photo of the damage, photo of the other vehicle's plate, photo of the joint statement. The agent validates the elements, opens the file, dispatches the tow truck — without any human stepping in before Monday morning. Quebec insurers who adopt this in 2026 will gain market share. Those who wait will lose it.
2. Auto mechanics and body shops. "My engine sounds weird" is useless on the phone. A photo of the dashboard with warning lights on, or a short video of the engine bay, turns a vague diagnostic into a real estimate. Quebec dealerships that combine an AI voice agent with multimodal recover the estimates they were losing on missed calls and qualify the work before the customer even reaches the counter.
3. Veterinary clinics. A dog who ate something dubious, a wound that's worrying, weird behavior filmed on the owner's iPhone. Veterinary clinics already triaging via AI voice agent can now layer visual triage on top. That's the difference between "bring him in now" and "keep him calm, we've seen worse, take this Tuesday slot."
4. Plumbing and contractors. 60% of plumbing emergency calls in Quebec resolve one way or another based on a single thing: can the contractor evaluate urgency without a site visit. A photo of a pipe spraying water versus a faucet that's dripping changes the emergency rate and the priority on the call list. For a solo contractor or a small crew, that's three or four site visits avoided per week.
5. Retail and post-sale support. Customer calls because the product isn't working. Photo of the error screen, photo of the serial number, photo of the damaged box from Canada Post. The agent opens a return ticket, generates the return label, sends everything by email. No human. At 7 p.m. on a Saturday.
The numbers nobody quotes enough
A few hard data points to size the stakes.
The AI virtual receptionist market hit USD $4.64 billion in 2026 and is growing at 9.8% annually. But that number hides a fracture: voice-only market share is plateauing while multimodal is capturing the growth. Providers who don't pivot will see their customer cohort age without renewal.
On the operational side: 27% of SMB calls go unanswered, costing the average business about USD $126,000 per year. Add to that the fact that 40 to 60% of emergency calls (plumbing, mechanics, roadside) get qualified differently when the customer can send a photo — meaning routed differently, meaning priced correctly. That's gross margin slipping past you without you even seeing it.
And the cost of adding multimodal to an existing agent? Between $0 and $0.02 per processed image depending on the LLM. At 100 images per day, that's CAD $60 per month. ROI is trivial the moment a single avoided plumbing site visit per week is on the table.
What doesn't work (yet) — the honest section
Not to sell you the dream: three concrete limits in May 2026.
The network. Multimodal during an active call requires the customer to send the image via MMS or via a web link clicked during the call. In Quebec, on Bell, Rogers and Vidéotron, MMS works fine — but 8 to 12% of customers have business SIP-only phones where MMS doesn't go through. You need a fallback (short SMS link to an upload page).
Added latency. Image analysis by a vision-capable LLM adds 1.2 to 2.5 seconds depending on the model. On an agent already running at 300-500ms of voice latency, that delay feels long if the agent stays silent. Solution: the agent says "perfect, I'm looking at the photo, give me 2 seconds" — a human behavior, not a robot one.
Quebec Law 25. Receiving and processing a photo means processing personal information (potentially biometric if a face is visible). Explicit consent, retention, right to deletion — all of it must be settled before deployment. Not blocking, but it must be documented. Our TECHMA team handles this for you during the deployment.
What you need to decide before the end of the quarter
Here's what we'd tell an SMB owner in Quebec in May 2026, in 30 seconds:
If your business generates revenue on calls where "seeing" is faster than "describing" (mechanics, plumbing, insurance, vet, post-sale), you have 6 to 9 months before competitors who switched to multimodal eat your emergency calls. Not because they're better. Because they're more practical — and at 10 p.m. on a Saturday, the customer picks practical.
If your business is mostly standard appointment booking (restaurant, salon, short call intake), your voice-only AI voice agent is still doing the job. You can wait until 2027 to go multimodal — but by then, it'll be table stakes and nobody will be charging extra for the novelty.
The architectural angle: multimodal is not an add-on
One more thing, because we're already seeing providers "adding" multimodal by duct-taping it on. Multimodal changes how the AI voice agent thinks, not just what it hears.
Concretely, that means the LLM must be natively multimodal (Gemini 3.1 Pro, GPT-4o, Claude Sonnet) — not a text model with an OCR layer slapped on top. It means the multi-agent architecture must include a specialized agent for visual triage. And it means the system prompts, tools, and guardrails — all of it must be rethought to assume the image is part of the call's context.
If someone offers you multimodal for an extra CAD $99/month, ask two questions: "Which vision-capable LLM?" and "Does the image context carry across the entire conversation, or only the turn where the image was sent?" If the answer is fuzzy, that's surface-level multimodal. It's useful, but it's not what we're talking about here.
The verdict
In May 2026, the AI voice agent in Quebec is no longer just a phone that picks up. It's a communication channel where voice, text, and image blend in the same conversation. SMBs who adopt this now take a lead measured in market share, not tech gadgets.
And the ones waiting until all their competitors already have it? They'll pay the catch-up bill three times over — in lost customers, in migration cost, and in awkward explanations to customers who wonder why the competitor across the street is just more practical.
Multimodal is not the next step. It's the current step. The only remaining question is: is your SMB in, or is it watching from the window?
External sources: ElevenLabs — Introducing Multimodal Conversational AI · Telnyx — AI Agents now support inbound MMS during live voice calls · ElevenLabs Changelog — April 7, 2026.