Driving fast and slow: what building a vehicle market index assisted by AI taught us

Last week we published the first edition of the Motorway National Car Index: a quarterly read on the UK used car market, built entirely on our own platform data. You can have a look: here.

The idea is simple. Rightmove has a house price index. Linkedin publishes regular workforce data reports. Both get cited in their respective industries quarter after quarter because they’re consistent, named, and reliably published. We’ve helped over 1m people sell their car online through daily live dealer auctions, which means we see something almost nobody else does: what dealers are actually willing to pay for a car this week and what trends are shaping the C2B automotive market.

There’s also an emerging new reason to look at this type of thought leadership in a category and that’s how content is now being used by LLMs themselves. The answer to “what’s happening to used car prices?” increasingly isn’t a search results page as it once was — it’s an AI-generated answer, and those answers cite sources. A named, recurring, methodologically transparent index is the kind of thing that gets cited, rather than one-off press releases and it also ties into how our brand will show up on new channels like ChatGPT.

What I want to write about isn’t really the findings from our index though (although they are very interesting!). It’s how quickly we were able to build it in the age of AI, and why that speed turned out to be the least interesting part of the project.

What Q3 told us

Used EV values are rising. Volumes on Motorway have more than doubled year on year, up 115%, and the average price paid for an EV is up 6% over the same period. After eighteen months of headlines about collapsing EV residuals, that’s a turn worth noting as we look towards the next decade and net zero targets start to become more real as we approach 2030.

Chinese brands are arriving in the used market. We had three Chinese brands in our daily auctions last year — this has doubled to six in Q3 2026. Still small compared to the traditional brands, but they continue to make inroads into the UK market and we expect this trend to continue into 2027 and beyond.

The build: hours, not weeks

Internally, the index isn’t a static spreadsheet as we might have had in years gone by. It’s a live artefact built with Claude, sitting on top of our data warehouse, that anyone involved can open, filter and export from internally.

From a data and analytics perspective, this felt like a game-changer with AI. If we or our stakeholders wanted to see what an EV-only view looked like — same breakdowns, makes, models, regions, bids per car, but filtered to electric — that was a same-day turnaround, not a new ticket in a backlog for an analyst to pick up for a few days when they could. If we wanted a full export across a year rather than top-15 lists, in a single prompt it was done. The comms team could pull their own CSVs rather than queueing behind an analyst, giving them autonomy with the data like never before.

Historically, a publication like this would have meant a dashboard build, a requirements doc, and several rounds of “can you also add…” and would have taken multiple weeks end-to-end. Instead we were able to turn it around in hours and get the data needed for whatever new idea we wanted to look into, whenever it was needed.

The bit that actually took the time

This all sounds great but here’s the thing I’d want anyone reading this to take away: AI helps you generate the numbers dramatically faster, but we found validating them didn’t get faster at all. Going fast surfaced problems earlier to be addressed, but it also meant we could produce a plausible-looking, externally-shareable table in twenty minutes — and plausible-looking is exactly the failure mode you need to worry about more and more with AI becoming ubiquitous.

Blanks that mean two different things. Empty cells in a top-15 table meant “dropped out of the top 15 models”, not “sold none in the period”. Misinterpreting these could give you a very different idea of how the marketplace was moving — and you need that human judgement to get it right.

Totals that didn’t reconcile. Our fuel table and our make/region/age/colour tables carried different totals, because a few hundred records had a null fuel type. Each file was internally consistent, so nothing was wrong — but “why are these two numbers different?” is the first question you have to ask as the human-in-the-loop, and get a good answer for.

Quarter definitions. Calendar quarter or financial quarter? Which cut-off date does the methodology note state and how much could this vary the numbers by? Seems tiny, but could alter a story on car brands that are quite rare.

None of those are AI problems. They’re the same, tedious data governance problems we’ve always had. What’s changed is that the gap between “I have a question” and “I have a number” has collapsed, so the definitional work is now the whole job rather than something that happens quietly during the build — and we need to all get better at knowing what is expected of us as the humans-in-the-loop to get the most out of our new AI ways of working.

How we handled it

Pair programming — nothing external is approved with only a single pair of (human) eyes. Every figure was cross-checked by an analyst who hadn’t built the query. We validated against our semantic layer; where a stat didn’t exist there, someone checked it manually in the warehouse. Fast generation raises the bar on reproduction and review, it doesn’t lower it.

Raw and adjusted, side by side. Our raw averages reflect whatever mix of cars came through the platform that quarter. We built a mix-adjusted view too, normalising for age and mileage, and looked at both. Where they disagreed, the disagreement was the story — where different averages could be hiding stats we might otherwise have missed.

Benchmark against someone who already does this. We checked our EV findings against other industry standards, like CAP pricing, before publishing. If we’d diverged wildly, we’d want to know why before a journalist asked.

And one that’s more a new habit we are building: the Analytics team reviewed not just the numbers, but how they were being characterised in the copy. A correct figure can still support a wrong sentence, especially when AI is involved. Aligning on the story is now part of an Analytics team’s work, not a separate Comms step downstream after a handoff. Building on our already strong relationships with the PR and Comms team, we had to adapt to new ways or working with AI.

Final takeaway: Why data governance becomes more important than ever in the age of AI

To summarise: AI tooling made our new vehicle index quick to build, and AI powered marketing channels are increasingly a large part of why it’s worth building. The only part that hasn’t sped up — and mustn’t — is where a human decides which number is right, and what the right story to take from the data really is. Without data and AI governance, you just speed up the time to market, but critically not the time to real value for your business.

Read the Medium article here https://medium.com/building-motorway/driving-fast-and-slow-what-building-a-vehicle-market-index-assisted-by-ai-taught-us-a2362be1d01e