The announcements all sound the same: "our platform is now multimodal." Dig in. You'll find a voice/text toggle on an existing chat interface. The agent answers in voice when you speak to it. That's it. The mechanics of the experience haven't moved an inch.

Box checked. Deck updated. And the experience is still monomodal, with sound.

Real multimodal is a behavioral design problem. Nobody will sell you that as an API, because it can't be enabled. It has to be designed.

The three-question test

Want to know if an experience is actually multimodal? Three questions are enough.

When the agent says "here are the three wines I'd recommend," do the product cards appear on screen while it's speaking, or after?

When the user opens a product card or adds to cart while the agent is talking, does the agent know?

When the user says "compare them," who decides to show a comparison view, and when?

A voice/text toggle fails all three. That's not a nuance: it's the entire difference between a conversational experience and a chatbot with speakers.

That's not a nuance: it's the entire difference between a conversational experience and a chatbot with speakers.

Every micro-decision is a design choice

Synchronizing what the agent says with what the screen shows isn't an integration problem. It's a problem of behavior, timing, and intent.

Does the interface wait for the end of the sentence before rendering, or trigger as soon as the intent is detected? What does the agent do if the user taps mid-sentence? What happens when two signals arrive at once?

None of these decisions are documented in any language model's API reference. Product teams who've tried to assemble them from scratch know the reality: weeks of R&D on states, transitions, and edge cases, for a single journey.

The voice ↔ screen choreography

At Scenaro, voice and screen share the same real-time channel. That's not an architectural detail, it's the foundation.

When the agent talks about a product, it can display the card on screen, at the right moment, with the right intent. And the channel runs both ways: when the user scrolls, opens a card, adds to cart, the agent knows and adjusts what it says. Two performers listening to each other and adapting in real time, the definition of choreography.

Visual state flows back to the agent as checkpoints: a session can be resumed where it left off, with an agent that knows what the user has already seen.

Urbansider, a Paris travel concierge, pushes this model in production: interactive map, trip timeline, immersive video, multimodal itinerary. The agent doesn't describe an itinerary, it displays it, evolves it during the conversation, and knows what the traveler is looking at.

The experience is even structured in phases, like a real appointment with a concierge: a brief to capture the intent, a search working in the background, a staged reveal on map and timeline. And on every screen, buttons let the user interact on their own terms, trigger the narration, explore a neighborhood, adjust the brief, without voice ever being forced on them. Club Français du Vin measured the same principle in e-commerce: a marked improvement in engagement.

The agent doesn't describe an itinerary, it displays it, evolves it during the conversation, and knows what the traveler is looking at.

A craft, scenario by scenario

This choreography doesn't get generated. It gets designed, scenario by scenario: the order of events, the priority of information, when the agent speaks and when it lets the screen take over.

Even the voice stack is a per-scenario choice. On Scenaro, each brand picks its architecture, realtime for fluid discovery with natural interruptions, or an STT→LLM→TTS pipeline when flexibility matters most, across dozens of providers, with no lock-in. An agent that takes more than a second to respond can't orchestrate a screen in real time: the architecture choice is an experience choice.

Nobody can make these choices for you. They constitute your experience. And your experience is your brand.