Lucid Intelligence / AI Experience platform

Summary

Lucid Intelligence is the AI Experience (AIX) platform Lucid publicly announced as the foundation for multi-modal natural interactions in its vehicles. As Lead UX Researcher and Lead UX Strategist on this work, I was one of five people who shaped the platform’s definitive product strategy. I led the foundational research, authored several of the strategic framings that informed the platform’s positioning, wrote the business case for an in-house in-car AI assistant, and now define the production prompt governance and conversational quality system for Lucid Assistant.

The reframing

“AI in a car” had a default mental model when this work started. It looked like a voice assistant: a microphone icon, a wake word, a query, an answer. Useful for setting a destination or playing music. Limited for almost everything else.

The interesting strategic question wasn’t whether Lucid should ship a voice assistant. The interesting question was what AI should be in a vehicle context once you stopped treating it as a chat surface.

A car already has a multimodal interaction system: screens, voice, physical controls, vehicle state, location, time of day, who’s driving, who’s in the passenger seat, what’s happening outside. The opportunity wasn’t to add a chatbot on top of that system. It was to make AI the connective layer underneath it, something that could reason across modalities and context, not just respond to prompts.

Articulating that reframing, and grounding it in evidence, became the through-line of the work.

The research

I led a foundational behavioral research program on what an in-car AI assistant should actually do. It was anchored by a quantitative study of 441 drivers across three regions (North America, the Middle East, and Europe), supplemented by qualitative interviews. The study set out to understand the conditions under which drivers would, and wouldn’t, trust an in-car assistant with progressively more responsibility, and which categories of capability would unlock adoption versus which would feel like novelty.

Two ideas came out of that research that ended up shaping the platform’s framing:

The “missing 10%”: large language models handle the conversational surface of most driver interactions well. But there’s a specific automotive slice where they break: offline conditions, vehicle-state awareness, latency budgets that don’t tolerate cloud round-trips, recovery behavior when something goes wrong. That slice is small in surface area but disproportionately determines whether drivers trust the system. The platform’s job is to own that slice with the same rigor a foundation model gets for everything else.

The Must-Not-Break Contract: a framework treating accuracy, latency, offline reliability, and recovery UX as four non-negotiables for in-car AI. Offline reliability in particular is treated as designed, not generated: an explicit, surfaced list of what always works without connectivity, a UX promise rather than a fallback.

The research itself was the first part of the contribution. The framings that came out of it were the second.

The business case

Alongside the research, I authored a ~40-page business case for an in-house in-car AI experience. The deck articulated:

A four-level capability model for automotive voice AI, ranging from Level 0 fixed commands, to Level 1 natural-language car control, to Level 2 multi-turn conversational and contextual interaction, to Level 3 proactive multimodal companion behavior. Competitors were mapped against it. The model became a reference framework in leadership conversations.

A multi-scenario business case spanning Lucid’s own vehicles and potential platform applications beyond them.

A model-agnostic platform layer that sits above foundation models, so that base LLM choice could evolve while experience logic, safety constraints, personality, and automotive correctness remain owned at the platform layer.

The architectural and regulatory case for a hybrid edge/cloud approach across North America, Europe, and China.

This work informed internal leadership conversations on platform and model strategy.

What’s publicly disclosed

At its 2026 Investor Day, Lucid publicly introduced the Lucid Intelligence AI Experience (AIX) platform as “the foundation for multi-modal natural interactions.” Four pillars were disclosed:

• Agentic: built on experience pillars and cultural intelligence
• Personalized: a framework using memories that emphasizes privacy and control
• Seamless: cloud and edge support across touchpoints
• LLM Agnostic: flexible to adapt to advances in a fast-moving space

Lucid also disclosed that Lucid Assistant is the consumer-facing assistant built on the platform, with personalities and roles that preserve the brand while offering user choice, and that the platform anchors UX 4.0, a “voice-first capable, multi-modal UI model.” Lucid Intelligence is on the public software roadmap for Lucid Gravity in 2026.

Full public reference: Lucid Investor Day, March 2026, Derek Jenkins (SVP Design & Brand), slides ~57–60.

Production prompt and conversational quality governance

The work that began upstream as research and strategy now continues downstream as production governance. I lead the production prompt and conversational quality system for Lucid Assistant: prompt architecture, authoring principles, review criteria, versioning, and the evaluation scorecard covering usefulness, clarity, task completion, brand fit, and fallback handling.

Most of what shapes the day-to-day quality of an AI product happens at this layer. The model is one input. Brand voice, personality, tool-call behavior, refusal handling, fallback when something is misunderstood, graceful failure when the system can’t do what the user wants, these are design and product decisions that need someone reviewing them with the same care that visual UI gets.

AI products are won upstream of the model

AI products in consumer environments aren’t won by model capability. They’re won by orchestration, restraint, and trust. Conversational quality is a product discipline, not a model property, and most of what determines an AI product happens upstream of any feature, in the vocabulary and framings a team agrees are worth using.