DEV Community: Rémi Henriot

Voice, Data, GDPR: What Brands Need to Clarify Before Going Live

Rémi Henriot — Sat, 22 Aug 2026 12:00:51 +0000

Voice, Data, GDPR: What Brands Need to Clarify Before Going Live

Voice is intimate data, and a generation of assistants treats it like a cookie. Brands that don't ask the data question before the POC will face it after.

Voice is intimate data, and a generation of assistants treats it like a cookie. Brands that don't ask the data question before the POC will face it after.

A voice is not a click. It's a timbre, a hesitation, a purchase intent spoken out loud, sometimes a health detail slipped into a sentence. It's what a customer entrusts to you when they speak naturally.

An entire generation of assistants, designed in the US, for the US market, treats this material like one more cookie: collected by default, sent who knows where, governed by terms nobody reads.

The scenario repeats itself: a team builds a voice POC in three weeks, the demo impresses, the project accelerates. Then the DPO asks their questions. Where are conversations hosted? Who has access? Can they be deleted? Which sub-processor, in which jurisdiction?

Silence. The vendor answers "we're GDPR-compliant", no detail, no contract, no architecture to show. The project dies there. Not because of the DPO: because of a tool choice made without asking the data question.

The order is the problem. Compliance can't be retrofitted, it's chosen upfront, contractually, in the very architecture of the platform.

Treating voice as intimate data isn't a legal constraint. It's a brand position. The customer who speaks to your conversational experience gives you a trust no form ever earns. That trust is protected, or lost all at once.

Brands deploying voice in Europe need three simple answers before signing anything: where the data lives, who can erase it, and what a visitor has to give up in order to speak.

Scenaro's infrastructure is hosted in Europe, in Paris. Conversations don't have to cross the Atlantic to exist.

Conversations are deletable. Not a manual process triggered by a support ticket, a capability of the platform.

And visitor identity stays anonymous: no mandatory login to speak. The returning customer is recognized without ever having handed over their identity. You can remember someone without knowing who they are.

Transparency can go all the way to the user themselves. At Urbansider, in production, a history panel gives travelers direct access to what the experience holds about them, conversations, memories, saved plans, with the ability to delete each item. Trust isn't declared in terms and conditions: it's shown on screen.

We don't sell turnkey compliance, no serious vendor can; every brand remains responsible for its own processing. What we offer is an architecture that makes the conversation with your DPO possible, early, with concrete answers.

Voice is coming to every customer relationship. Brands that treat this data with the respect it deserves will turn it into an advantage. The others will discover the topic during an audit.

Want to discuss Scenaro's data architecture? [Contact us]

The full article lives on Scenaro: Voice, Data, GDPR: What Brands Need to Clarify Before Going Live

If you prefer the series by email: https://scenaro.substack.com/subscribe

From Conversations to Personas: The Hidden Intelligence in Your Conversational Data

Rémi Henriot — Thu, 20 Aug 2026 12:02:12 +0000

From Conversations to Personas: The Hidden Intelligence in Your Conversational Data

Your web analytics measure clicks. Your customers, meanwhile, voice objections, criteria, and doubts, out loud. What your users say to your assistant is worth more than any consumer panel.

Your web analytics measure clicks. Your customers, meanwhile, voice objections, criteria, and doubts, out loud. What your users say to your assistant is worth more than any consumer panel.

You know the user left the product page after 45 seconds. You don't know they were hesitating between two bottles and couldn't find an answer to their food-pairing question.

The first fact costs you a dashboard. The second is worth a catalog decision. And your current tools will never see it.

Bounce rates, time on page, conversion funnels: twenty years of web analytics have perfected the measurement of behavior. And plateaued there.

A click doesn't say why. A session recording shows hesitant scrolls, not the question that went unanswered. A consumer panel? It captures what people think they should answer to your question. Not what they say spontaneously when nobody's asking.

A conversation does. The real objection to a price. The decision criterion no form ever captures. The moment of doubt right before add-to-cart. It's all there, in plain words. Brands that talk with their customers, and soon they all will, collect this material with every exchange.

A conversational experience in production generates thousands of conversations. Qualitative data at scale: the holy grail of marketing research, free, continuous.

But nobody reads ten thousand transcripts. And patterns don't emerge from linear reading. That's the paradox: AI produces data that only AI can process at scale.

On Scenaro, every conversation is stored and summarized automatically. Daily and weekly summaries are produced per scenario: recurring topics, questions that went unanswered, the signal of a gap in your collection, drop-off moments, requests your offering doesn't address yet.

And the analysis isn't generic: the scenario's persona, its character, its objectives, structures the reading. A scenario built to advise isn't analyzed like a scenario built to sell. What you receive isn't a keyword cloud: it's a synthesis shaped by what that scenario was meant to accomplish.

All of it aggregates trends, never individuals. The summaries speak about your customers in the plural, never about one customer in particular.

Your analytics will always tell you what happened. Your conversations tell you why. The question is no longer collecting this material, it arrives on its own. It's reading it. That's what Scenaro does for you, every week.

Want to see what a Scenaro weekly summary looks like for your sector? [Request an example]

The full article lives on Scenaro: From Conversations to Personas: The Hidden Intelligence in Your Conversational Data

If you prefer the series by email: https://scenaro.substack.com/subscribe

Memory Between Sessions: What Your AI Assistant Should Actually Remember

Rémi Henriot — Sat, 15 Aug 2026 12:12:40 +0000

Memory Between Sessions: What Your AI Assistant Should Actually Remember

Your customer spent twenty minutes opening up. Eight days later, your assistant asks "How can I help you?". That's not a technical limit. It's a choice, and it's costing you your loyal customers.

Your customer spent twenty minutes opening up. Eight days later, your assistant asks "How can I help you?". That's not a technical limit. It's a choice, and it's costing you your loyal customers.

The user gave everything: their tastes, Burgundy preferred, €30 budget, family dinners on Sundays. The assistant answered well, advised well. Then it threw it all away.

The customer comes back. "Hello! How can I help you today?"

That's not a bug. It's the market norm: amnesiac, stateless bots that reset to zero every session. The problem isn't GPT-4. It's zero context, zero scenario, zero memory.

Look at the demos from conversational platforms: they all show the first conversation. Never the second.

Because there is no second. An amnesiac bot treats your loyal customer exactly like a stranger, a salesperson who never recognizes anyone. Nobody comes back to talk to someone who forgot them.

Meanwhile, the rest of the digital world remembers. Amazon knows your history. Netflix knows your genres. Spotify knows your weeks. Your brand's assistant forgets. That asymmetry is untenable, and your users feel it: "I would have had to explain everything again" is the number one reason they don't return.

At Scenaro, memory is architected in three levels.

First level: visitor identity persists. The returning customer is recognized, no mandatory login, no friction. The relationship has a thread.

Second level: memories configurable per scenario. You decide, scenario by scenario, what deserves to be remembered, tastes, budget, occasions, and those memories are injected into the next conversation. After each exchange, a structured summary is generated: not a raw log, a usable synthesis.

Third level: session resumption with the state of the interface. The assistant doesn't just remember what was said, it remembers what was seen. "You were looking at the Bordeaux selection": the conversation picks up exactly where it stopped, screen included.

That third level is live in production at Urbansider: a traveler who closes the tab mid-planning finds, on return, a "Resume" button, and the conversation restarts on the exact screen where they left it, brief, map, and timeline included.

Good memory isn't memory that keeps everything. It's memory that keeps what serves the relationship, and stays governable: every element is deletable, and the brand decides what gets remembered.

Governance extends to the user too. At Urbansider, a history panel gives them access to everything the experience holds about them: past conversations, memories, saved plans, all viewable and deletable. That transparency isn't a regulatory burden to endure. It's what makes memory acceptable: people trust an advisor who remembers, not a black box that records.

An assistant that remembers is the difference between a tool people try once and a relationship they maintain. Without a memory architecture designed from day one, you're building loyalty on sand.

Want to understand how to configure memory on Scenaro? [Let's talk]

The full article lives on Scenaro: Memory Between Sessions: What Your AI Assistant Should Actually Remember

If you prefer the series by email: https://scenaro.substack.com/subscribe

Connecting Your Product Catalog to an AI Assistant: It's More Than an API Call

Rémi Henriot — Thu, 13 Aug 2026 12:00:43 +0000

Connecting Your Product Catalog to an AI Assistant: It's More Than an API Call

Plugging Shopify into an AI assistant takes an afternoon. What you get is a search engine in disguise, not an advisor. The difference is architecture, not the API.

Plugging Shopify into an AI assistant takes an afternoon. What you get is a search engine in disguise, not an advisor. The difference is architecture, not the API.

The pitch is everywhere: "connect your catalog in a few hours." A webhook, two endpoints, a demo that handles "show me your running shoes." Sold.

Except the demo lies by omission. What you just wired up is keyword search with a voice. A duct-taped RAG pipeline that matches strings and recites product sheets.

The assistant can answer. It can't advise. It lists. It doesn't recommend. And that difference shows up straight in your conversion rates.

A product API gives you access to data. Not to meaning. Ten thousand SKUs loaded raw into a language model are ten thousand rows of metadata.

Ask that system for "a light wine for a summer dinner" and it greps your descriptions for "light" and "summer." It doesn't understand you mean a low-tannin red, served cool, under €20.

The user asks a customer's question. They get a database's answer. They wanted advice; they got a list. That's where the conversational promise dies, politely, factually, one exchange at a time.

At Scenaro, a connected catalog is called a collection. Shopify, Magento: synchronization is native, the catalog stays current. But the connection is only the starting point.

Every field in a collection can be indexed three ways: full-text for exact terms, semantic for meaning, not words, phonetic for what the ear actually hears. And it's configurable collection by collection, to fit your business.

The result: the agent doesn't match strings. It understands "a light wine for a summer dinner" and goes looking for what that means in your cellar. That's what runs in production at Club Français du Vin, our first client.

And when the request outgrows a product query, advanced search profiles take over. At Urbansider, the Paris travel concierge, the conversational experience doesn't return a list of addresses, it orchestrates the generation of an actual itinerary.

Anyone can connect a catalog. Making an agent understand it, field by field, intent by intent, is architecture. That's what Scenaro builds.

Want to see how catalog indexing works on Scenaro? [Request a demo]

The full article lives on Scenaro: Connecting Your Product Catalog to an AI Assistant: It's More Than an API Call

If you prefer the series by email: https://scenaro.substack.com/subscribe

Multi-Model: The Strategy to Stop Depending on OpenAI

Rémi Henriot — Sat, 08 Aug 2026 12:00:46 +0000

Multi-Model: The Strategy to Stop Depending on OpenAI

70% of AI projects in production run on a single provider. That's not a strategy, it's inertia. And inertia has a cost that gets paid in silence.

70% of AI projects in production run on a single provider. That's not a strategy, it's inertia. And inertia has a cost that gets paid in silence.

GPT-4o pricing swung 50% in eighteen months. Availability incidents hit production services. And every quarter, a competitor ships a model that beats yours on your exact use case.

Staying single-vendor in 2026 is a choice. The problem: almost nobody actually chose it. It settled in.

Lock-In by Default

OpenAI dominated the first cycle of generative AI. Documentation, quality, ecosystem: the gravity was natural. Most projects were built on it, and the inertia persists.

Yet the landscape has changed. Some models now beat GPT-4o on reasoning, others on multilingual or European-language work. On the voice side, STT and TTS alternatives offer radically different cost-latency profiles.

Teams know this. And they don't move. Because in a monolithic system, switching providers takes weeks: abstracting API calls, harmonizing formats, regression-testing every critical behavior, redeploying. The switching cost is so disproportionate that the decision is no longer free.

The Risk That Compounds in Silence

Single-vendor dependency makes no noise. It compounds.

A unilateral price increase: you absorb it. An availability incident: you have no plan B. A usage policy change: you adapt. A better model at a competitor: you watch it go by.

A Provider Registry, Not a Marriage

At Scenaro, multi-provider isn't an option, it's the base architecture. Scenarios point to a registry of decoupled providers: the conversational LLM, the summarization model, and the search models are independent of each other.

Concretely: you swap the STT without touching the LLM. You test a new model on a staging scenario, compare in real conditions, publish if it wins. Switching providers becomes what it always should have been, a setting, not an engineering project.

And the API keys stay yours. Scenaro composes the voice stack; it doesn't resell the provider's voice. You pay OpenAI, Anthropic, Deepgram, or ElevenLabs directly, at actual cost, with no hidden margin on models. The day you want to leave, your keys leave with you.

That's how the production experiences run, Club Français du Vin, Urbansider, Lupi: each scenario assembles the providers that fit its context, and that assembly gets revised without a rewrite.

Scenaro sits above the voice stack, the experience layer, not inside it. That's precisely what makes the stack swappable.

Lock-in never materializes with a bang. It erodes, invoice after invoice, incident after incident. The question isn't whether you should diversify your providers. It's how long you can afford not to.

Explore the 40+ supported providers →

The full article lives on Scenaro: Multi-Model: The Strategy to Stop Depending on OpenAI

If you prefer the series by email: https://scenaro.substack.com/subscribe

Conversational Versioning: Deploy Without Breaking Things

Rémi Henriot — Thu, 06 Aug 2026 12:00:27 +0000

Conversational Versioning: Deploy Without Breaking Things

Tech teams have git. Marketing teams have nothing to version their assistant with. That asymmetry kills more AI projects than any model choice.

Tech teams have git. Marketing teams have nothing to version their assistant with. That asymmetry kills more AI projects than any model choice.

The scene replays every week across dozens of teams: a prompt change applied straight to production, a validated behavior gone, and no way back.

Version control has existed for code for thirty years. Nobody would accept shipping software without history or rollback. Yet that's exactly what we ask of marketing teams operating a conversational experience.

Deployment Without Governance

Engineers know who changed which line, when, and why. They roll back in two minutes. They ship with a safety net.

The marketing team operating an AI assistant edits a prompt in production, no snapshot, no history, no rollback. On most platforms, configurations aren't versioned: they're overwritten. Change tracking, where it exists at all, lives in a Notion file.

It's a structural asymmetry. And it always produces the same outcome: a black-box configuration nobody fully understands, full of behaviors nobody can explain.

Fear Fossilizes the Assistant

A system prompt isn't a sentence. It's weeks of accumulated tuning, instructions, constraints, persona. Changing one line can disable a validated behavior or break a journey that was converting.

Without versions, every change is a bet. So teams stop betting: "don't touch it." The assistant fossilizes out of caution, while offers, seasons, and priorities move on without it.

Git for Conversations

At Scenaro, the Cockpit brings to conversational experiences the discipline git brought to code, without requiring the skills.

Every publication generates an immutable version of the full configuration: prompts, services, scenarios, features. You publish to staging first, validate, then promote to production. You compare two versions. You roll back in one move if a behavior goes sideways.

Promotion is non-destructive: the publication identifier doesn't change, so scripts already embedded on the site stay valid. Marketing publishes autonomously; engineering has nothing to redeploy. Tags even let you prepare a "Christmas" or "sale" version in parallel, without duplicating the scenario.

The effect is mechanical: when risk is controlled, iteration goes up. Teams move from "don't touch it" to confident weekly iteration. That's exactly what separates living experiences, Club Français du Vin, Urbansider, Lupi, from assistants frozen on day one.

And the stakes go beyond tooling. Every brand will soon talk directly with its customers; that dialogue will become an asset as critical as a website. A critical asset without change governance doesn't exist, not for long. Conversational versioning isn't a feature. It's the condition for marketing to run its assistant like a product, without fear, and without tickets.

See how Scenaro versioning works →

The full article lives on Scenaro: Conversational Versioning: Deploy Without Breaking Things

If you prefer the series by email: https://scenaro.substack.com/subscribe

Why Most AI Projects Stop at v1

Rémi Henriot — Sat, 01 Aug 2026 12:00:28 +0000

Why Most AI Projects Stop at v1

Building got easy. Running didn't. Most AI assistants die within 90 days, killed by a question nobody asked before launch: "who's maintaining this?"

Building got easy. Running didn't. Most AI assistants die within 90 days, killed by a question nobody asked before launch: "who's maintaining this?"

The demos impress. Budgets get approved. POCs have been multiplying across marketing teams for eighteen months.

Then the v1 becomes a tombstone. Still live, still billed, less relevant every week. Nobody dares touch it.

The technology didn't fail. It's the after-launch that nobody planned.

The Disposable POC

Industry reports converge: 60 to 70% of enterprise AI initiatives never make it past the pilot phase. Not for lack of ambition. For lack of infrastructure to operate a conversational system over time.

The disposable POC always follows the same script. An agency delivers in six weeks. The demo convinces. It goes live. Then user feedback arrives, the offer changes, a better model ships, and there's no owner, no tooling, no budget to iterate.

The budget covered the build. Nobody budgeted the operation. A conversational experience isn't a deliverable you sign off on: it's a living product you run.

Three Frictions That Turn v1 Into a Tombstone

First: the prompt belongs to engineering. Changing the assistant's behavior requires a ticket, a developer, a deployment. So even when user feedback is unambiguous, nothing moves.

Second: nobody sees anything. Without conversation metrics, there's no way to know whether users are asking unanswered questions, or whether a conversion path is broken somewhere. You're flying blind, which means you're not flying at all.

Third: no governance. Who changed what, when? Without history, every modification is a leap of faith. The assistant becomes a black box nobody dares touch. Six months post-launch, it answers exactly like day one, same mistakes included.

This isn't an LLM problem. The problem isn't GPT-4. It's zero context, zero scenario, zero memory, and zero loop to fix any of it.

Run It, Don't Re-Deliver It

At Scenaro, the premise is the opposite: a conversational experience is a living product, and the infrastructure must make iteration mundane.

In the Cockpit, every publication generates an immutable version. You publish to staging, then production. You compare two versions. You roll back when something breaks. It's "git for conversations", but usable by a marketing team, no tickets, no deployments. That's the concrete answer to "who's maintaining this?": the team that knows the customers, autonomously.

On visibility, conversations are analyzed automatically: unanswered questions, recurring topics, and drop-off moments surface every week. Iteration stops being intuition and becomes a loop fed by reality.

That's the difference between Club Français du Vin, Urbansider, or Lupi, production experiences that keep evolving, and the POCs that fade out in silence.

And there's a reason to build this muscle now rather than in two years. Every brand will soon talk directly with its customers, just as mobile apps forced everyone to "think mobile," the next design reflex will be thinking conversation-first. Teams learning to run conversational products today will be three years ahead of those still re-delivering POCs.

Building...

The full article lives on Scenaro: Why Most AI Projects Stop at v1

If you prefer the series by email: https://scenaro.substack.com/subscribe

Voice Prompt Engineering: A Discipline of Its Own

Rémi Henriot — Thu, 30 Jul 2026 12:00:50 +0000

Voice Prompt Engineering: A Discipline of Its Own

Copy-pasting your ChatGPT prompts into a voice assistant is the first mistake, and nearly everyone makes it. Voice has its own rules, and text never taught them.

Copy-pasting your ChatGPT prompts into a voice assistant is the first mistake, and nearly everyone makes it. Voice has its own rules, and text never taught them.

Teams learned to prompt with ChatGPT. A few iterations, a tone adjustment, and the text assistant behaves. So they transpose: same prompts, microphone attached.

The result: verbose agents that recite lists out loud and leave awkward silences while they work. Users hang up.

Prompting for voice is not prompting for text. It's a different discipline. And it can be learned.

The Copy-Paste That Fails

The problem is structural. There are thousands of resources on text prompting, ChatGPT, Claude, Gemini. Voice gets treated as just another output channel, handled with a "respond conversationally."

That's not enough. A prompt written for writing produces writing, and writing read out loud is exactly what makes users leave.

The discovery always happens in the same place: in production, when people hang up.

What Speech Forbids

Out loud, no markdown. An agent that generates a bulleted list produces absurd synthesized speech: "Dash. First benefit. Dash. Second benefit." That's not a conversation, it's a document reading.

Out loud, no scrolling back. Comprehension is linear and irreversible, the user can't reread. A 40-word sentence that works in text becomes unintelligible in synthesis. Short sentences. A rhythm.

Out loud, silence is a bug. When the agent runs a tool, product search, stock check, three seconds of dead air is enough to make the conversation feel broken. A human would say "let me check." The agent should too.

And out loud, the first sentence decides everything. An opening that sounds like a call center script ("Hello, I'm your virtual assistant, how can I help you today?") produces immediate abandonment. The opener must be a line of conversation, not a service announcement.

How Scenaro Structures the Voice Prompt

At Scenaro, a scenario's prompt isn't a monolithic block. It's structured around two things text prompting systematically conflates.

Character. Each scenario carries a persona: who this agent is, how it speaks, what it never does. A wine advisor at Club Français du Vin doesn't share a register with a travel concierge at Urbansider. Character governs sentence-by-sentence behavior, tone, rhythm, forbidden formats.

Goals. What the agent is trying to accomplish in the session: qualify, recommend, guide toward a purchase. Goals are declared separately from character. Mixing them into one block produces schizophrenic agents, that know what they want but not how to carry themselves, or the reverse.

And one voice-specific rule: Scenaro prompts encourage the agent to keep talking while tools execute. While a search runs, the agent narrates what it's doing. Silence is never left to chance.

Every prompt change is versioned and testable in real voice before publication. Because a voice prompt isn't judged by reading it, it's judged by listening to it.

Prompting for voice is a skill to build, not a reflex to transfer. Teams that treat the two disciplines as one pay for it in production. Teams that listen to their prompts before publishing...

The full article lives on Scenaro: Voice Prompt Engineering: A Discipline of Its Own

If you prefer the series by email: https://scenaro.substack.com/subscribe

Realtime vs Separate Pipelines: Choosing the Right Voice Architecture

Rémi Henriot — Sat, 25 Jul 2026 12:00:26 +0000

Realtime vs Separate Pipelines: Choosing the Right Voice Architecture

Everyone wants realtime. Not everyone needs it. Latency isn't a religion, it's a setting, tuned per use case.

Everyone wants realtime. Not everyone needs it. Latency isn't a religion, it's a setting, tuned per use case.

Since GPT-4o Realtime, "real time" has become the magic word in boardrooms. The demos are stunning: 200ms latency, natural interruptions, a conversation that finally feels like a conversation.

So teams demand realtime. Everywhere. For everything.

That's an architecture mistake dressed up as ambition. You're choosing the thrill of the demo, not the fit of the use case. And the invoice follows.

What the Hype Doesn't Show

Realtime processes audio natively, with no intermediate transcription. It's real, it's impressive, and it comes with three trade-offs the demos gloss over.

Cost. Native speech-to-speech is structurally more expensive than a separate pipeline. At volume, support, lead qualification, the gap isn't marginal, it's budgetary.

Voice. In a pure realtime stack, synthesis is baked into the model. Less TTS customization, no cloned voice. Your brand's voice becomes the provider's voice.

Fit. A user exploring a catalog by voice deserves 200ms. A prospect recording an async message will never hear the difference. Paying for realtime there is burning budget on a thrill nobody feels.

Two Architectures, One Setting, Not a Doctrine

At Scenaro, each scenario chooses its architecture. It's not a platform decision, it's a per-context setting.

The realtime model: native audio, roughly 200ms, natural interruptions. Ideal when conversational fluency is the differentiator, a premium advisor in full duplex, voice-driven product exploration. Trade-off: less control over the voice.

The STT→LLM→TTS pipeline: 800ms to 1.5s, but full control over every layer. You pick your transcription engine, your model, your voice, cloned if needed. Ideal for volume, cost control, and sonic identity.

And in between, the hybrid variant: realtime inference, separate voice synthesis. Realtime speed on reasoning, freedom on voice.

The Right Question

The right question isn't "realtime or not?" It's: what latency, what voice, and what cost fit this specific scenario?

A wine advisor in full duplex on a homepage justifies realtime. A qualification scenario in push-to-talk doesn't need 200ms, a separate pipeline does the job at a fraction of the cost, with the brand's exact voice.

That's why at Scenaro the choice happens scenario by scenario, in the Cockpit. The same conversational experience can run a realtime scenario and a pipeline scenario side by side. You compare, measure, adjust. Scenario versioning even lets you test one architecture against the other on the same journey before rolling it out.

The most expensive mistake never changes: deploying realtime everywhere because the demo impressed, then discovering on the invoice that most conversations never needed it. Voice architecture is decided per use case. Never per trend.

The full article lives on Scenaro: Realtime vs Separate Pipelines: Choosing the Right Voice Architecture

If you prefer the series by email: https://scenaro.substack.com/subscribe

STT, LLM, TTS: The Full Stack Behind a "Simple" Voice Assistant

Rémi Henriot — Thu, 23 Jul 2026 12:00:10 +0000

STT, LLM, TTS: The Full Stack Behind a "Simple" Voice Assistant

"Plug in OpenAI and you're done" is the most expensive lie in voice AI. A voice that answers is three layers, fifteen-odd providers, and a trade-off at every link.

"Plug in OpenAI and you're done" is the most expensive lie in voice AI. A voice that answers is three layers, fifteen-odd providers, and a trade-off at every link.

A visitor asks a question out loud. They get a natural answer in under a second. From the outside, it looks like a product. From the inside, it's a chain: one layer transcribes speech (STT), one generates the answer (LLM), one speaks it back (TTS).

Three layers. Three decisions. And one unforgiving rule: a slow or wrong link degrades the whole chain.

The problem isn't that this complexity exists. It's that most teams only see it after they've made the wrong calls.

The Naive Stack

The reflex is always the same: "We're using OpenAI, right?"

One API key, one contract, one invoice. It's comfortable. And that's exactly the trap: you're not choosing a stack, you're choosing procurement convenience. Every layer inherits a default provider, not the best one for the job.

Why Each Layer Deserves Its Own Decision

Transcription first. An STT model that shines on standard English can stumble on a regional accent or domain vocabulary. And a faulty transcription in produces an absurd answer out, the user doesn't hear "the STT failed," they hear "this assistant is stupid."

Voice next. A voice is a brand asset, just like a logo. "Serviceable" isn't enough for a premium brand: some engines deliver richer prosody, others let you clone a real ambassador's voice. That's a deliberate decision, not a default setting.

The model last. Complex reasoning, long instructions, multilingual contexts: no LLM wins everywhere. The right model depends on the task, and the right LLM tells you nothing about the right STT.

How Scenaro Composes the Chain

Scenaro is built on LiveKit, a battle-tested real-time engine, with fifteen-odd provider plugins: Deepgram, ElevenLabs, Cartesia, Mistral, OpenAI, Google, and more. Each layer is chosen independently, on its own latency-cost-quality trade-off.

The decisive part: the chain is configured per scenario, in the Cockpit. Swapping the STT, the LLM, or the voice is a setting, not a rewrite of the conversational experience. Business logic, scenarios, memory: nothing moves.

This is what runs in production at Club Français du Vin, Urbansider, and Lupi. Same chain, different settings, because a travel concierge and a wine advisor don't share latency constraints or a voice.

And that's the real test of a voice architecture: not the day-one demo, but how easily you change your mind at month three. Teams that freeze their stack at launch discover in production that the default provider wasn't the right one, especially on STT. That recalibration should be routine, not a rebuild.

A voice assistant isn't a product. It's an assembly of decisions. The right architecture gives each layer the best provider for your context, and lets you change your mind without rebuilding anything.

The full article lives on Scenaro: STT, LLM, TTS: The Full Stack Behind a "Simple" Voice Assistant

If you prefer the series by email: https://scenaro.substack.com/subscribe

Perceived Latency: The Invisible KPI Killing Voice Experiences

Rémi Henriot — Sun, 19 Jul 2026 12:00:08 +0000

Perceived Latency: The Invisible KPI Killing Voice Experiences

Your teams track conversion, NPS, session time. Nobody tracks the assistant's silence. Yet that's what drives users away.

Your teams track conversion, NPS, session time. Nobody tracks the assistant's silence. Yet that's what drives users away.

There's a precise moment when a voice experience tips from "fluid" to "broken." It's not the wrong answer. It's not the crash. It's the silence: the user asks a question out loud, and nothing happens.

The threshold is known: roughly 800 milliseconds of unmanaged silence, and the illusion collapses. In a real conversation, a pause that long is filled by a glance, a breath, a "hmm", signals that say "I heard you, I'm thinking." Remove them, and the same pause turns anxious. The user wonders: did it hear me? Is it thinking? Did it crash?

A collapsed illusion of conversation doesn't rebuild. The user won't leave a negative review. They'll just stop using it.

The KPI nobody looks at

The voice hype produced dashboards full of familiar metrics: conversion, NPS, session length, drop-off. Useful, and blind to the problem.

The time between the end of the user's utterance and the start of the agent's response shows up nowhere. Google Analytics doesn't see it. Session recording tools capture it poorly. Yet it's the first quality signal users perceive, before they've heard a word of the answer.

Cognitive science is unambiguous: beyond 700-800ms with no processing signal, the brain reads the pause as an error, not as thinking. The experience doesn't degrade gradually. It breaks.

Technical latency isn't perceived latency

Confusing the two is the most widespread mistake in voice deployments.

Technical latency is actual processing time: transcribing the voice, generating the answer, synthesizing speech. On a classic pipeline, expect 800ms to 1.5 seconds. Not catastrophic in absolute terms.

Perceived latency is what the user feels. And it can be engineered independently of the tech. A human doesn't go mute while looking something up: they say "let me check that…" and keep working. An audio or visual processing signal massively reduces perceived latency, at identical technical latency. With nothing, 800ms is an eternity.

At Scenaro, this principle is built into the agent itself. It exposes a state, listening, thinking, speaking, synchronized with the interface. And it's designed to talk while it works rather than leave dead air. Unmanaged silence isn't an incident: it's a design flaw.

The extreme case is the long wait. At Urbansider, building an itinerary takes real time, tens of seconds of search. Rather than hiding that wait behind a spinner, the experience transforms it: while the search runs in the background, the agent offers the traveler an interlude, a piece of Parisian culture, a neighborhood guide, a French word to learn. Pedagogical content, a relaxed tone, a touch of the unexpected. The wait stops being a hole in the conversation: it becomes a moment of the experience.

Latency is a choice, scenario by scenario

The other mistake is treating latency as a single number to optimize. It's a trade-off, and it's made per scenario.

Two architecture families exist. Realtime: 200-400ms, full duplex, natural interruptions, the architecture of fluid conversation, at the price of less flexibility on voice...

The full article lives on Scenaro: Perceived Latency: The Invisible KPI Killing Voice Experiences

If you prefer the series by email: https://scenaro.substack.com/subscribe

When to Use Voice — and When to Stay Quiet

Rémi Henriot — Fri, 17 Jul 2026 12:00:09 +0000

When to Use Voice, and When to Stay Quiet

"Voice everywhere" is a design mistake. An intrusive voice assistant is worse than a search bar.

"Voice everywhere" is a design mistake. An intrusive voice assistant is worse than a search bar.

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 takes up headspace

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.

Technical tempo follows usage tempo

The voice mode...

The full article lives on Scenaro: When to Use Voice, and When to Stay Quiet

If you prefer the series by email: https://scenaro.substack.com/subscribe