There's a moment in every conversational project when someone says: "what if we made this voice?" Good question.

The default answer, voice everywhere, enabled on every page, for every user, is the wrong one.

We know how that ends: an agent talking while the user reads a spec sheet. Interrupting a comparison to ask an open-ended question. Requiring people to speak out loud in an open-plan office. More irritation than fluidity.

Voice isn't a universal mode. It's a contextual choice. And that choice deserves a policy, not a toggle.

Voice isn't a universal mode. It's a contextual choice.

Voice takes up mental bandwidth

That's its power and its limit, and everything else follows from it.

In discovery, a user arriving with no specific intent, hesitating between categories, voice is ideal. It steers without pushing, reframes fuzzy needs, builds a relationship. Exactly what a good salesperson would do.

In analysis, it becomes an obstacle. A user comparing two spec sheets wants to read, pause, go back. The brain can't process spoken language and dense text at the same time. An agent talking through that moment isn't advising, it's disrupting.

Add the social context: open office, public transit, café. Speaking out loud there is impossible or uncomfortable for most people. That's not an edge case, it's a huge share of real-world usage.

"Voice everywhere" ignores all of it. It turns voice's greatest asset, presence, into a deterrent.

An activation policy, not a setting

The answer isn't to walk back from voice. It's to calibrate it by context, like an experience policy.

At Scenaro, four modes cover the spectrum. Full duplex, agent and user speak and interrupt each other naturally, for discovery and advisory. Discreet mode, the agent listens to voice but replies in writing, for public settings and content meant to be read. Push-to-talk, the user triggers when they choose, for support and noisy environments. Text→audio, the user types, the agent speaks, for busy hands.

These aren't toggles. They're design commitments, activated by context: full duplex on the homepage, discreet mode on a product page, push-to-talk in support. Urbansider, a travel concierge in active production, ships this "discreet mode" to its travelers, because nobody talks out loud to their phone on the Paris métro.

And context shifts mid-session. A user arriving in discovery benefits from full duplex; ten minutes later, deep in spec sheets, the same user needs the agent to go quiet or switch to text.

On Scenaro, these transitions are defined at the scenario level, not just the page level.

That's a lesson learned in production. On Club Français du Vin, our first experience, the voice conversation ran continuously, and some users wanted to browse in silence. Urbansider, the next generation, draws the conclusions: the agent can be paused at any moment, the map can be explored without narration, and every screen offers a button to ask the agent to narrate, when the user decides. Voice stays permanently available. It's no longer permanently imposed.

Voice stays permanently available. It's no longer permanently imposed.

Technical tempo follows usage tempo

The voice mode also determines the stack. A discovery journey in full duplex demands a realtime architecture, where interruptions feel natural. A support journey in discreet mode tolerates an STT→LLM→TTS pipeline, when the agent replies in writing, the user reads at their own pace and latency matters less.

Each brand picks its stack per scenario, across dozens of providers, with no lock-in, and the estimated cost of each configuration displayed in the editor. The voice policy is a trade-off between context, budget, and fluidity, made scenario by scenario, from the Cockpit.

Voice activity detection and background noise filtering complete the picture: an agent that triggers on ambient noise, or cuts the user off at the first hesitation, ruins the best activation policy.

Voice is a powerful tool. Its effectiveness depends on knowing when to pick it up, and when to put it down. Knowing when not to speak is a design skill, every bit as much as knowing what to say.