On May 25, 2026, ElevenLabs shipped text_behavior_overrides. Here is why Quebec professional firms need to rethink their multichannel strategy. In Trois-Rivieres, a notary, Mrs. Lavoie, recently told us a scene that was almost comical until you thought about the client on the other end. Her AI agent was replying to an SMS as if it were on the phone: long paragraphs, formal greeting, ceremonial tone, then a URL recited like it was being read out loud, character by character. The client just wanted to know whether to bring an original ID for the Tuesday 2:30 PM signing. He got a mini-letter instead.
Let us be honest: this is exactly the kind of detail several vendors swept under the rug for two years with a tempting promise — one single agent for voice, SMS, and chat. Easier to sell. Easier to deploy on paper. But in a professional firm, simple does not always mean good. A notary does not speak the way they text. An accountant does not write SMS the way they answer a 5 PM call during filing season. A lawyer, even less so.
The May 25, 2026 ElevenLabs update settles the debate fairly clearly. When a platform introduces distinct behaviors by channel, that is not a cosmetic detail. It is a technical admission: speaking the same way everywhere is often bad design. For an AI voice agent multichannel Quebec setup, the real question is no longer "can we unify?" but "what should we separate to stay useful, clear, and prudent?"
What text_behavior_overrides actually changes (and why May 25, 2026 is a pivot date)
The May 25, 2026 update, visible in the ElevenLabs changelog, adds something very specific: text_behavior_overrides. Put differently, you can now define a different text behavior based on the conversation initiation source — for example voice, sms, chat, or widget.
Technically, this is a mapping per ConversationInitiationSource to BehaviorOverride objects. The parameters that matter most for a firm are concrete:
- verbosity: how detailed the response should be;
- output_format: the expected output shape;
- interaction_budget: the conversational latitude the agent gets.
It can sound trivial from afar. Between us, it is not. For a long time, many teams tried to dodge the problem with a single prompt like: "be warm on the phone, brief on SMS, structured on chat." It works… until it does not. Because a model does not always interpret those instructions with the consistency a professional firm demands.
With overrides, you no longer politely ask the system to be different per channel. You configure it to be. Major distinction. On voice, you can allow more flexibility, more rephrasing, a higher interaction budget. On SMS, you tighten the length, force a more direct format, cut the fat. On chat, you can target readable structure with short steps or two-block answers.
May 25, 2026 becomes a pivot date for a simple reason: it marks the moment multichannel architecture stops being a matter of taste and becomes a matter of governance. If ElevenLabs makes this granularity official, it is because they have seen what everyone else sees on the ground. For firms that already audit their conversations with conversation tags, the value is immediate: you can finally compare behaviors by channel without bolting on fuzzy rules.
The ElevenLabs overrides documentation goes in that direction. It does not just say "you can customize." It shows, in practice, that a text message and a voice conversation should not be treated as clones. For a professional services AI voice agent, that is almost an official recognition of a principle that careful integrators were already applying by hand.
The 'one agent for everything' trap that pro firms discover too late
The problem with "one agent for all channels" is that it looks smart until the day it produces an awkward reply in front of a real client. Not in a lab. Not in a demo. In real life, a Wednesday morning, when reception is buried and two people are in hearings or with clients.
First concrete mistake: the agent that treats the phone like a text channel. We have seen it more than once. A client calls, asks where to upload a document, and the agent answers: "Here is your link: h-t-t-p-s colon slash slash…" That loses everyone. A human instinctively understands that on the phone, you rephrase: "I can send it to you by SMS or email right now." It is not just a comfort issue. It is basic ergonomics.
Second mistake: SMS that becomes an essay. The client writes: "Can I drop by tomorrow?" They receive four paragraphs with a formal greeting, context, terms, confidentiality reminder, and closing salutation. The truth is that a good firm SMS usually runs between 140 and 320 characters, not 980. Beyond that, you stop helping the person; you transfer the reading burden to them. In a unified SMS voice chat setup, that is where the myth cracks.
Third mistake: chat that imitates speech too much. Replies with "uh," "so," "I'll look into that," small hesitations that can be charming on voice but read as vague in writing. On a web widget, a client asking about fees, processing times, or document lists wants something readable. Not a transcript. When you work on voice naturalness, especially with GPT-Realtime-2 preambles, you are trying to make voice more human. But what sounds good on the phone can read poorly in chat.
In a notary, accounting, law, or brokerage firm, those gaps are not just annoying. They can become risks. A vague SMS phrasing about a limitation period, an over-expansive reply about required documents, a too-casual tone on something touching professional secrecy — all of that creates confusion, and sometimes worse. The danger is not always dramatic. It is often mundane. One extra word. A miscalibrated tone. A legal nuance delivered in the wrong format.
The consistency problem is also underestimated. When a firm has seven people — two assistants, one partner, one accounting tech, three professionals with different risk thresholds — a too-uniform agent ends up displeasing everyone. Reception finds it too long. The professional finds it too light. The client finds it odd. After 6 to 8 weeks, the story becomes "the AI does not work," when the real issue was the conversational architecture from day one.
A notary accountant lawyer AI setup does not need a robot that speaks the same everywhere. It needs a system that understands the channel changes the shape, the rhythm, and sometimes even the caution of the reply. That is where the ElevenLabs update ends a comfortable marketing illusion that rarely held up on the ground.
Voice, SMS, chat: three ergonomies, three brains or just one?
The right question is not "one agent or three?" The right question is: how much real divergence exists across your channels? For a Quebec firm, you can make the call with three simple criteria.
1. Call and message volume
If your firm gets 8 to 15 calls a day and a handful of confirmation texts, a single agent with tuned behaviors might be enough. If you are closer to 47 calls, 22 SMS, and parallel web chat requests, the operational pressure changes everything. The higher the volume, the more channel gaps cost you. A slightly too long SMS reply multiplied across 180 conversations per week becomes a structural irritant.
2. Legal sensitivity and compliance
Law 25, professional secrecy, personal data, consent, traceability. In a pro firm, those are not footnotes. If your exchanges regularly touch identity documents, financial details, file matters, or deadlines with legal weight, separating by channel often becomes preferable. Why? Because you want to limit ambiguity. A voice agent can be warm and triage-oriented. An SMS agent can be dry, useful, almost telegraphic. A chat agent can be more structured and better fenced.
3. Available technical bandwidth
Be realistic. The more you segment, the more you govern. At Agent IA Vocal, deployments and integrations are handled by the TECHMA team for the client. That is precisely what enables either a single-agent setup with overrides or a stricter multi-agent architecture. If you have neither time nor appetite to manage fine variations, a well-configured single agent can be an honest compromise. If your firm demands high precision, separation is usually still the better call.
The contrarian point is this: no, "one brain" is not automatically smarter. Sometimes it is just easier for the sales demo. In reality, voice, SMS, and chat are three different ergonomies. Not three skins on the same conversation. Voice tolerates transitions, controlled hesitations, rephrasing. SMS rewards brevity. Chat demands readability. Wanting to unify at all costs often means asking the system to make mediocre compromises everywhere instead of being excellent somewhere.
What this means concretely for a 4 to 15-person firm in Quebec
For a typical 4 to 15-person Quebec firm, our take is fairly clear: the dedicated per-channel approach still wins in most cases. A voice agent with a higher interaction_budget, able to guide, clarify, and follow up cleanly. An SMS agent with verbosity set to "low," short replies, action-oriented. A chat agent with a clean format, free of oral tics. That is usually the configuration that produces the least friction and the fewest surprises.
Why? Because at that size, the firm already lives real context variations. The assistant wants to confirm an appointment. The notary wants to filter emergencies. The accountant wants to avoid over-detailed answers on an unconfirmed tax situation. The lawyer wants to remove any ambiguity about what counts as advice. The broker wants to answer fast without over-promising. One agent can do all that, yes, but it has to juggle a lot of nuance.
That said, the ElevenLabs May 2026 update genuinely shifts things for smaller setups. For a 1 to 3-person firm with few active channels and fairly repetitive requests, a single-agent setup with text_behavior_overrides finally becomes a serious option. Not perfect. Serious. That is an important distinction. You can share the same knowledge base, the same guardrails, the same business logic, while modulating the output by channel without rebuilding from scratch.
Financially, stop treating this as a large-enterprise luxury. The Forrester study on Voice AI ROI, relayed by NextLevel, reports a 3-year ROI between 331% and 391%, with a breakeven around 3.2 months. Obviously, a Quebec City or Trois-Rivieres firm is not a multinational. But those numbers reinforce something simple: when routing, qualification, and baseline replies are well configured, the impact shows up fast.
The keyword is "well configured." Not "turned on." Not "wired up quickly." At Agent IA Vocal, the TECHMA team handles the integrations based on the firm's actual reality, precisely because the details matter. A poorly tuned AI agent verbosity on SMS can cost appointments. A wrong output format on voice can stretch calls by an average of 42 seconds. Multiply that across 5 days, then 11 working months, and the cost of bad design quickly outruns the cost of a better-thought architecture.
The conversation every pro firm should have this week
If you already have an agent in place, or if you are evaluating a project, there is a discussion to have right now with your integrator — whether that is TECHMA or another partner. Not in three months. This week.
- What is my current setup? One agent with shared logic, or several per channel?
- Do we have real text_behavior_overrides? Or just a generic "be brief on SMS" line in the prompt?
- How many channels are actually being served? Voice only? Voice and SMS? Web chat too?
- What is the Law 25 SLA for automated SMS output? Who validates, who logs, who reviews?
- What monitoring is in place? And how often do you review channel gaps?
On that last point, a simple weekly audit ritual can make all the difference. Twelve minutes, well invested, often beat an improvised overhaul after a client complaint. Especially if you combine it with conversation tags to spot per-channel drifts.
The bottom line is simple. The market sold the idea that one agent for everything was a sign of maturity. On May 25, 2026, ElevenLabs showed instead that maturity is accepting the difference between channels. For Quebec pro firms, the decision is no longer theoretical. If your agent speaks to SMS like the phone, or writes in chat as if improvising out loud, you are not missing a marketing layer. You are missing a serious conversational architecture.
That is exactly where an AI voice agent multichannel Quebec setup stops being a nice gadget and becomes a real firm asset.