Most organisations with a website of any real size have the same quiet problem. Over years of projects, redesigns and quick fixes, the same kind of component gets rebuilt again and again, slightly differently each time. Nobody set out to create six versions of an accordion. It just happens.
We ran into this recently with a large UK higher education institution. They'd already started on the foundations of a design system, but their site had grown organically for years, and nobody had a clear, honest picture of what was actually on it. That was the real starting point: not "here's a design system, help us tidy it up," but "we know roughly where we're heading, but we don't yet know what we've actually got."
You can't design what you can't see
It's tempting to jump straight to the fun part: define the components, write the standards, roll out a pattern library. But you can't sensibly decide what the "one true accordion" should look like if you don't know how many versions already exist, where they live, or why they drifted apart in the first place.
So before any design system work could really begin, the team needed a clear, honest picture of what was actually on the site. Not a guess. Not a sample. A confident answer.
Prototyping at the speed of conversation
This is where AI came into the picture. We saw a need for a tool that could create that viewpoint and give us confident answers. It could take weeks to build ourselves, so we experimented with AI to explore how it could help solve our problem.
Rather than writing a spec and commissioning a build, the audit tool took shape through conversation: try something, run it against the site, see what comes back, adjust, try again. That loop was fast enough to test ideas the same day they came up.
A couple of moments from that process stuck with us. Early on, the tool flagged navigation and footer elements on every single page, which buried the patterns that actually mattered under noise that didn't. Once we could see that happening, stripping it out was obvious. But it's not something you'd think to write into a spec in advance; you only spot it once you're looking at real results.
Something similar happened with how the tool handled repetition. The first pass flagged every instance of every pattern it found, giving hundreds of near-identical alerts that a reviewer had to wade through. So the audit tool was adjusted to recognise a pattern it had already seen, and only raise something when a genuinely new variation turned up.
None of that came from the AI deciding what was useful. It came from people looking at real output and making a call. AI helped things move faster. It didn't do the judging.
What it turned into
The tool that came out of this groups similar components together, highlights unique patterns, and lets someone quickly mark each one as worth keeping, merging, retiring, or looking at more closely. It enables teams to move from “here’s what we have” to “here’s what we’re going to do about it.”
It wasn't built only for designers, either. Developers use it too, comparing how a component has actually been implemented across the codebase and spotting where the markup or styling has quietly diverged over time, which is its own kind of drift worth catching.
Everything runs locally, so nothing is sent off to a third party. That mattered here, given how sensitive data can be in education. It's a sensible default for most organisations, whatever sector they're in.
What it gave the team was evidence. A real, shared picture of what existed, built on those foundations, so they could move towards a finished design system with confidence, not guesswork.
Not a shortcut around expertise
We didn't use AI here because it's the trend of the moment. We used it because it did a specific job well: getting a working tool built fast, so the team could get on with the real work sooner.
We don't hand judgement over to a machine. AI takes on the heavy, repetitive lifting, so people stay free to apply their expertise and decide what's actually worth keeping.
If you're facing something similar, a site that's grown faster than anyone's kept track of, we'd be glad to talk it through. Get in touch.
