Cutting vehicle valuation costs by 74,000%
Carbay • 2025
Home screen showing current claim status with progress bar and next-step preview, plus quick access to suggestions and FAQs.
Context
Carbay hired me as founding designer to build a bridge between private sellers wanting a fast sale and dealers who couldn't get consistent stock: a consumer site for valuations and listing, and a dealer site for live bidding three days a week. Before building it out, the job was to prove Carbay could compete, and that the economics would hold up against real users.
Problem
The market already had strong players doing versions of this, so the first risk was demand, not build. I ran a 3-screen acid test on PPC traffic, taking users all the way to a valuation screen before honestly telling them we didn't want their vehicle — a clean read on intent with no real cost.
The second risk was cost. Valuation from suppliers quoted £2.25, £1.05, and £3.00 per check, steep for a pre-revenue startup, and nothing stopped a user requesting the same valuation repeatedly (without further development) . Left unchecked, valuations alone could become the biggest line item before Carbay had proven a business.
Append claim timeline showing completed, in-progress, and upcoming stages with dates, from parts ordered through to bike delivery.
"Repairs Are Underway" status page in plain English, with estimated completion date and a link through to the full progress timeline.
Notification inbox showing the full claim history as individual plain-English updates, each timestamped, from bike collection through to repairs in progress.
Solution
I built a Python script that used an AI model to pull live listings from the wider market and return a structured JSON valuation. Tested against the same vehicles suppliers had quoted, the outputs were consistently close, sometimes exact, at roughly £0.00004 per call versus up to £3.00.
We launched with it. The AI number was always framed as a valuation in principle, not final, the real price was still settled at auction. That framing meant it only had to be fair and directionally right, and users treated it that way.
This wasn't sellable as a standalone valuation API. A product like that must stand on its own accuracy — nothing else catches a bad quote before it costs money. Carbay could tolerate an imperfect AI quote because the auction was the real check; the AI just had to be close enough for the auction to correct it. The to "do things that don't scale" move: right for Carbay's stage as it leaned on an existing backstop instead of replacing the need for one. Past a certain volume, that dependency breaks, bad quotes compound faster than the auction would be able to absorb, and you need quotes trustworthy on their own, with no human or auction behind them.
Impact
The acid test converted at 16%, ~8x the ~2% baseline, proving demand before a single supplier contract was signed. The AI valuation then let Carbay launch and test against real cars in the market without carrying supplier costs it couldn't yet justify. The approach has since been extended into other categories, including motorcycles.
74,000%
cheaper valuations than the most expensive supplier quoted
8x increase
the ~2% market baseline on conversion
AI solution
validated against real suppliers, later extended into other vehicle valuations