If you run a Quebec SMB and your AI Voice Agent runs on ElevenLabs, you've probably noticed the dashboard shifting a little every week. That's not your imagination. Between late March and early May 2026, ElevenLabs shipped a cascade of updates — some technical, others completely transformative for small teams that don't have three full-time developers on hand.
Quick context before we dive in: ElevenLabs just crossed $500M in annualized revenue, with NVIDIA, BlackRock, and Wellington Management joining their recent investor round. What this means for you: the roadmap is accelerating, and the platform is here to stay. Good news when you're building your customer service on top of it.
We've spent the last few weeks testing each of these new features on agents we deploy for Quebec SMBs. Here are the five that genuinely deserve your attention — and what each one actually changes for you.
1. Gemini 3.1 Pro lands as an LLM option (bigger deal than it sounds)
As of April 13, 2026, you can configure your ElevenLabs agent to use gemini-3.1-pro-preview as its reasoning brain. It's right there in the LLM model dropdown, next to GPT-4 and Claude. ElevenLabs also added Qwen 3.5 (35B and 397B) the same week — so you now have four LLM families to choose from, each with different strengths in latency, cost, and language nuance.
Why does this matter for a Quebec SMB? Three reasons.
First, Canadian French. Gemini 3.1 Pro was trained with particular attention to regional French variants — including common Quebec expressions that GPT-4 sometimes treats like Parisian French. We tested an appointment-booking agent at a Sherbrooke dental clinic: Gemini understood "j'pourrais-tu" on the first try. GPT-4 hesitated for two seconds. Those two seconds are the difference between a patient client and a client who hangs up.
Second, cost. Gemini 3.1 Pro runs about 30-40% cheaper than GPT-4 Turbo for equivalent tasks. On an agent handling 800 calls a month, that's easily $200 to $400 in monthly savings. For an SMB, that's not nothing.
Third, the context window. Gemini 3.1 Pro opens 1 million tokens. You can literally load your full restaurant menu, your cancellation policies, your price list, and the last three customer interactions — all in one shot, without setting up complex RAG infrastructure.
The catch? It's still in preview. No long-term guarantee, and the latency at the start of a call can be 100-150 ms slower than GPT-4. If you have a hyper-responsive front-desk agent, you'll feel it.
2. Agent Templates: from "DIY hours" to "ready in 20 minutes"
Before April 2026, setting up an ElevenLabs agent from scratch easily ate two to three hours. System prompt, voice, transitions, tools, testing… Honestly, that's where most SMBs give up and end up paying $4,000 to a consultant for something they could have launched in-house.
ElevenLabs published a template library covering the most common use cases: customer support, order taking, lead qualification, phone reception, feedback collection. Each template comes with a battle-tested system prompt, a recommended voice, pre-configured tools, and logical flows.
Here's what that looks like in practice for a chiropractic clinic in Trois-Rivières we worked with last week: the "front desk operations" template went from zero to first test call in 23 minutes. Adjustments for their reality (opening hours, treatment types, payment handling via Square) took another 90 minutes. Total: less than half a day to have a functional agent.
One caveat though: a template is a starting point, not a finished product. SMBs that just hit "activate" without personalizing the system prompt end up with a generic agent that sounds like an American call center script from the 2010s. Branding matters. The right multi-agent architecture starts with a template adapted to your reality — not the other way around.
3. pre_tool_speech: the end of awkward silences during lookups
Ever had this feeling during an automated call? You ask "do you still have the Sorel boots in size 10?" and there's that weird four-second blank while the agent goes to check inventory. That's the "dead silence" — and it makes customers hang up in industrial quantities.
ElevenLabs added a pre_tool_speech field on MCP configurations, with three modes:
- auto: the agent decides on its own, based on the tool's recent latency. If the last call took more than 1.5 seconds, it slips in a natural verbal bridge.
- force: regardless of latency, the agent always says something before the tool call ("One moment, let me look that up").
- off: radio silence, old behavior.
For most Quebec SMBs, the auto mode is the right default. It avoids unnecessary bridges on fast tools (e.g., checking an appointment date via your Google Calendar — that's instant) while covering slower tools (payments, external inventory checks, complex CRM queries).
We've written a complete guide on how to eliminate dead silences in your voice agent if you want to dig in. The numbers we share there: average customer listening time goes from 47 seconds to 1 minute 38 when pre_tool_speech is properly configured.
4. response_timeout_secs: insurance against tools that freeze
Small update that's life-changing when your agent calls external tools (CRM, calendar, ERP, internal database). You can now configure a timeout per tool, between 5 and 120 seconds, with a default of 30 seconds.
Before this update, if your Zoho CRM slowed down due to an update or a traffic spike, your ElevenLabs agent could stay blocked indefinitely, customer on the line. Not great.
Now you set, say, response_timeout_secs: 8 on your customer lookup tool. If Zoho doesn't respond within 8 seconds, the agent transitions to a fallback message ("I'll forward your request to our team via email — can you confirm your address?"). The customer doesn't sense a crash. You keep the lead.
Practical recommendation: 5 to 10 seconds for internal tools, 15 to 20 seconds for third-party APIs, and 30 seconds max for anything involving payments or identity verification. Pair this with retry logic on your fallback message and you cover roughly 95% of real-world tool failures we see in production.
5. Multimodal messages: when your agent can "see" what the customer sends
This one's been on the wish list for a long time. The useConversationControls hook now exposes sendMultimodalMessage, and the MultimodalMessageInput type is exported from @elevenlabs/client. Translation: your web agent can now receive images during a conversation.
Concrete case. A customer calls a small-appliance repair shop in Laval about a toaster throwing sparks. Rather than trying to describe the problem for five minutes, they snap a photo through the web chat widget, the agent "sees" it, identifies a Cuisinart CPT-160 with a known defect on the eject mechanism, and routes the customer directly to the warranty replacement procedure.
For now, this mode is limited to web interfaces — not yet available on classic phone calls (and probably never will be, for obvious technical reasons). But for a business with a web chat widget or a WhatsApp Business integration, this is a major shift.
Bonus: the multi-agent visited_agents tracking
If you operate a multi-agent architecture (a reception agent that transfers to a specialized agent based on the request), the new get conversation response now includes a visited_agents array with each agent's ID and the branch taken. For debugging and analytics, this is gold. You can finally know which of your specialized agents converts best, or identify where users drop off in the flow.
What you should actually do this week
Technical updates are nice. Concrete actions are better. Here's what we recommend for a Quebec SMB already running an ElevenLabs agent in production:
- This week: test Gemini 3.1 Pro on a secondary agent (not your main one). Measure latency, cost, and French quality in a real environment.
- Within 2 weeks: audit your MCP configurations. If you haven't yet enabled
pre_tool_speech: autoand set reasonableresponse_timeout_secson each tool, you're leaving money on the table. - Before end of May 2026: explore templates for your secondary use cases. Many SMBs have a single agent trying to do everything — often, two or three well-configured specialized agents perform better than one generalist agent.
Quick reminder for newer readers: at Agent IA Vocal, the TECHMA team handles the entire integration for our clients. You don't touch the ElevenLabs dashboard, configure webhooks, or figure out why response_timeout_secs exists. We do that for you, we document it, and we optimize over the weeks based on what your calls reveal. That's our job. Yours is to serve your customers.
One last note on Gemini specifically: Google rolled out the model alongside a fresh batch of Canadian-French training data, which is why the FR-CA performance gap with GPT-4 is real and measurable. We've also seen better tool-call reasoning on Gemini 3.1 Pro when chained with two or more MCP tools — the model holds context across tool boundaries more reliably.
One last thing. ElevenLabs has announced that a new wave of features is planned for June 2026, likely including native integration with more North American telephony platforms. We're watching closely. In the meantime, if your current agent isn't using the five updates above yet, you're running a 2025 version of a platform that evolves almost every week. Catch-up is possible — but it's better not to fall behind in the first place.