AI
guided
recommendations
The most confusing step in configuring a car, redesigned with an optional AI route: tell us what matters and we find the version that fits. Shipped live to three markets, measured properly, and read honestly.
Overview
An optional AI route through the hardest step of the configurator: users state what matters, the AI parses real derivative data and recommends at most two. Designed, localised and shipped live to three markets as a properly measured experiment, then read honestly.
Client
Carwow
Role
Product Designer + Design Engineer
Timeline
2026
Platform
Web · UK, DE, ES
Team
Design lead + cross-functional squad
The screen where car buying goes to die
Trim and engine selection is where configurator journeys quietly fall apart. Names like "Sport Line" and "M340i xDrive" carry no inherent meaning, the data is dense, and a fifth or more of users simply abandon rather than choose. The instinctive fixes all add information: better descriptions, feature breakdowns, popularity badges, jargon-free labels. We had tested versions of these, and the pattern was consistent. More information tends to increase abandonment, because it hands the user more homework. The real problem was never a shortage of facts on the page. It was that the onus sat entirely on the user to map complicated data onto their own life, repeatedly, across every trim and engine.
So we inverted it: the user tells us what matters, and the system does the mapping. I designed an optional AI route alongside the manual one, and the sharpest design position was what it is not. It is not a chatbot. One structured input, tags for the common needs plus free text for everything else, then the AI parses real derivative data and returns at most two recommendations with a route back to the full list. Knowledgeable users keep their manual path untouched. I designed every state including the unglamorous ones, the processing moment, the no-match case, and a failure state that apologises like a person. Then I localised the lot across UK, German and Spanish markets, down to per-market decisions about the AI's persona and the legal disclaimer.
Prompt design + experimentation
2026
Ship it, measure it, tell the truth
This was live-fire AI, not a demo. Working with engineering on the technical investigation and proof of concept, the whole experiment was cost-engineered to run on roughly $300 of inference across three countries and several weeks, a constraint that shaped the prompt design as much as any style guide did. We defined the metrics before launch: chargeable enquiries as primary, configurations and signups as guardrails, and a full interaction layer underneath, from route choice to which of the two recommendations users trusted more. Then we shipped it 50/50 to live traffic in three markets, close to nine hundred thousand users across the arms.
The evidence came back and I read it straight. Enquiries did not move in any market. The UK was neutral throughout, but the guardrails caught significant upstream harm in Germany and Spain, where fewer users completed configuration at all. A weaker process ships the win anyway or buries the result; ours held the line, and the territory split became the most valuable finding, reshaping how we think about AI assistance across markets rather than assuming one behaviour fits all. I would rather have this study in my portfolio than a fake win. Designing the feature is the visible half of the job. Defining honest guardrails, and letting the evidence beat your attachment to your own work, is the half that makes it product design.
In numbers
3
Markets A/B tested
890k
Users in the experiment
$300
Total inference cost