AI-Native Automotive CX | Mario Tsavalos

01Problem

The client's brand ambition had outrun its experience

The client had built one of Europe's most distinctive automotive brand identities — rebellious, design-driven, emotionally charged. But the customer experience at every real touchpoint was generic: the dealership felt like any other, the digital retail journey was conversion-optimised but soul-less, and there was no consistent thread connecting the brand promise to what customers actually felt.

Meanwhile, the business was moving fast into AI — but without a framework for what good AI-native automotive experiences looked and felt like. There was no design language for conversational UX, no maturity model for LLM readiness, no criteria for what "brand-appropriate" AI behaviour meant.

Brand–CX gap

Strong brand identity, generic physical and digital experience. No consistent thread from awareness to purchase to renewal.

No AI design framework

The business was investing in LLMs without a shared understanding of what brand-appropriate AI behaviour, tone, and guardrails look like.

Fragmented touchpoints

Digital retail, dealership, WhatsApp, renewal journeys — each owned by a different team, with no unified experience logic connecting them.

Low digital conversion

The online retail journey was underperforming. High drop-off at configuration and financing stages with no intelligent assistance layer.

02My Role

What I was accountable for

Over 2.5 years I grew from designer to workstream lead — holding full strategic and delivery responsibility across multiple parallel workstreams. On the conversational-commerce workstream I owned the customer and strategy side, directing a junior colleague who owned systems and integration.

Area	What I owned
AI Experience Design	Designed the AI Assistant's brand personality, conversational flows, prompt architecture, guardrails, and fallback behaviours. Built the 60-question, 5-category maturity framework used across 5 assessment cycles to drive the LLM to go-live readiness.
Innovation & Origination	Self-initiated the Mixed Reality dealership experience (CMO-recognised, five-figure build investment). Originated the AI Commerce concept with a technology partner — estimated mid-to-high six-figure upsell potential for the consultancy.
Workstream Leadership	Owned target vision, customer journey, use cases, and prioritisation on the WhatsApp workstream. Set the structure and directed a junior colleague on systems/integration. Led Refinement and PI Planning independently on-site.
Stakeholder Ownership	Regular contributor to bi-weekly client C-level leadership meetings. Presented to the client CMO and to a senior automotive leadership roundtable.
People Development	Directed and upskilled a junior colleague on the WhatsApp workstream; coached an intern on the XR/VR workstream; onboarded the incoming workstream lead.

03Why AI — Not a Classic Solution

The case for AI-native design

Classical solutions — better FAQs, improved configurators, chatbots with decision trees — had already been tried. They failed not because of technical execution but because they couldn't handle the combinatorial complexity of automotive configuration, the emotional register of high-value purchase decisions, or the real-time nature of WhatsApp commerce.

Why classical approaches failed — and what AI made possible

Classical: Decision-tree chatbot

Broke on configuration complexity. Any car with 12+ option combinations exceeded what a rule-based tree could handle without hitting dead ends. Users dropped off.

AI solution: LLM-based assistant

Handles open-ended queries, maintains context across a conversation, and adapts tone to match the brand — expert but approachable, never corporate.

Classical: Static product pages

High-value purchase decisions require reassurance, comparison, and trust-building that static content cannot provide dynamically in context.

AI solution: One-click AI Commerce

LLM-powered purchase journeys that surface the right product, explain trade-offs, and guide the user from intent to transaction — inside a WhatsApp conversation.

04Process

Research → Flows → Prototyping → Testing → Iteration

Phase 01Research

Understanding the brand–experience gap

Mapped existing CX touchpoints across digital retail, dealership, WhatsApp, and post-purchase renewal. Conducted contextual research with the client sales staff and customers to identify where the brand promise collapsed in real interactions. Key finding: customers could feel the brand in ads — but not in the actual purchase or ownership experience.

Phase 02Define & Frame

Target vision + use case prioritisation

Defined the strategic vision for each workstream — what "brand-specific" means in an AI interaction, in a dealership, in a renewal flow. Mapped and prioritised use cases against technical feasibility (Salesforce, Sprinklr, IT/NTT constraints), balancing ideal UX against what was realistically implementable in the first cycle.

use-case-prioritisation.fig

Phase 03Flows & Wireframes

Mapping conversation flows + interaction logic

Designed the conversational architecture before any visual design — mapping user intents, AI response logic, fallback paths, escalation triggers, and guardrail conditions. Wireframed the WhatsApp commerce flow end-to-end including AI handoff moments.

conversation-flow-v3.fig

Phase 04Prototyping

Interactive prototypes — from Figma to live LLM testing

Built interactive Figma prototypes for the conversational UI and dealership flows, then moved to live prompt testing with the engineering team. Rapid iteration between UX design and LLM output — adjusting prompt structure, system instructions, and conversation logic based on observed behaviour. Presented prototypes directly to C-level — ensuring design decisions were anchored in real use, not slides.

Phase 05Testing & Iteration

60-question AI maturity framework — 5 assessment cycles

Built a structured evaluation framework: 60 questions across 5 categories (brand tone, factual accuracy, scope adherence, fallback behaviour, edge-case handling). Each of 5 test cycles produced a maturity score that drove the LLM from prototype to go-live readiness. The client repeatedly named this testing framework as a decisive building block for the AI's development.

05AI Interaction Design

Prompts, conversational UX, agent behaviour, guardrails

Designing for AI is different from designing screens. The material is language, context, and probability — not pixels. The core challenge was defining what "brand-specific AI behaviour" means: expert but never cold, confident but never over-promising, in-scope but never feeling limited.

// Conversation design example — AI Assistant, branded messaging channel

SYSTEM

You are the Automotive AI Advisor. You are knowledgeable, direct, and brand-appropriate — never corporate, never generic. You help customers understand the vehicle range, configure options, and take the next step. You do not discuss competitors, pricing outside official ranges, or pending legal topics. When uncertain, offer to connect with a human advisor.

USER

"I'm interested in the electric model but I've heard the range is underwhelming in winter. Is that true?"

AI

Fair question — and one we take seriously. Like all EVs, the electric model's real-world range does decrease in cold temperatures — typically 15–25% below the WLTP figure in sustained sub-zero conditions. What helps: the vehicle's heat pump (standard on Performance versions) recovers a meaningful portion of that. Most of our customers find it handles typical German winters well with a bit of pre-conditioning. Want me to show you what range looks like for your specific commute pattern?

USER

"What's the cheapest way to finance it?"

AI

Financing is handled by our advisors who can find the best combination for your situation — leasing, PCP, or direct purchase each have real trade-offs depending on your mileage and tax situation. I'll connect you with someone who can run the numbers properly. Can I ask — are you buying privately or through a company?

Guardrail design

Guardrails were designed as a layered system — not a blocklist. Three types of response handled different situations:

✅

In-scope: Handle directly

Product specs, configuration, range, charging, accessories, test drives, dealership locations, feature comparisons within the brand's range.

⚠️

Boundary: Handle + escalate

Financing details, trade-in values, legal topics (EAA compliance, warranty claims) — AI handles the framing but flags to a human advisor.

🚫

Out-of-scope: Redirect

Competitor pricing, political topics, anything not in the knowledge base. AI acknowledges the question, explains its scope, and offers an alternative path.

06Key Decisions

Options considered — and why we chose this

Option A

Option B

→ Chosen: Option C

Decision: AI Assistant Tone & Personality

Generic assistant personality — neutral, helpful, brand-agnostic. Fast to build, easy to test.

Fully scripted persona with a name and backstory. High consistency but low adaptability.

Brand-voice-first prompt design — no persona name, but strict tone and vocabulary guidelines embedded in the system prompt. Feels true to the brand without feeling theatrical.

Option A

Option B

→ Chosen: Option C

Decision: Dealership Innovation Format

Digital screen wall — interactive product specs and video content. Low friction, proven format.

AR on mobile — customers use their phones to explore vehicles. High drop-off risk.

Mixed Reality headset experience — customers view real accessories on physical vehicles via MR overlay. CMO-recognised, five-figure investment secured, dealership rollout planned.

Option A

Option B

→ Chosen: Option C

Decision: WhatsApp Commerce Architecture

Decision-tree bot — scripted flows, low AI dependency, fast to ship.

Full LLM autonomy — AI handles all interactions with no defined scope boundary.

Constrained LLM with defined escalation paths — AI handles discovery and guidance, humans handle financing and closing. Balances quality with what Salesforce + Sprinklr could realistically support.

07Collaboration

How the cross-functional team worked

This was never a purely design project. Delivery required tight coordination across product ownership, engineering constraints, data science, and business stakeholders — often simultaneously.

Product & Strategy

Translated business requirements into design scope. Led use case prioritisation with Project Owner (Verena Venne) and Delivery Lead (Benedict Baur).

Engineering & IT

Constant alignment with NTT and Salesforce/Sprinklr integration teams. Designs balanced ideal UX with system realities — particularly on the WhatsApp workstream.

Data / AI

Worked directly with the LLM engineering team on prompt architecture, output evaluation, and the maturity assessment framework. Design drove AI quality criteria.

C-Level Client

Regular contributor to bi-weekly client leadership meetings. Presented the XR concept directly to the client's CMO — which triggered the build investment.

08Impact

What changed — and what it made possible

5-fig.

Build investment secured for Mixed Reality XR — dealership rollout planned

€8M

Annual merch revenue target tied to automotive roadshow concept

6-fig.

Estimated upsell potential — AI Commerce technology partnership

60Q

AI maturity framework built — drove LLM from prototype to go-live across 5 cycles

The Mixed Reality concept didn't just land in a deck. It earned CMO recognition, a build investment, a dealership rollout plan, and external visibility — referenced by the CMO in a business podcast and picked up for a META video. Colleagues positioned it internally as a rare case that genuinely combines creativity and business impact.

09Learnings

What worked, what we adapted

↑

Brand voice is the hardest thing to get right in AI. Technical quality — accuracy, coherence, recall — is table-stakes. The real design work is making an LLM feel like the brand. That requires iterating on system prompts with the same rigour as visual design: multiple drafts, live testing, consistent evaluation criteria.

↑

C-level buy-in comes from making ideas tangible fast. The Mixed Reality concept succeeded because I moved from idea to prototype to room presentation quickly — before the concept could be questioned out of existence. Speed of materialisation is a design skill.

→

Feasibility constraints are design constraints. The WhatsApp architecture needed three iterations before landing on something that balanced ideal UX with Salesforce/Sprinklr realities. Designing without understanding system constraints produces beautiful work that never ships.

→

Guardrails need as much design thought as the happy path. Early versions of the AI assistant felt frustrating in edge cases — the out-of-scope responses were abrupt and broke the brand voice. Designing the "no" with the same care as the "yes" is what made the experience feel coherent.

↑

The maturity framework outlasted the project team. By building a structured, repeatable evaluation model — not just ad-hoc feedback rounds — the LLM testing became something the client could own and run independently. Reusable frameworks compound in value over time.