Two designers independently found that AI tools misread our component library. The bloat humans route around is the same that breaks machines, and it tipped the call from patch to rebuild.
Until this week, the case for rebuilding our component library kept losing, and it deserved to. The argument was familiar: the library is bloated, it predates modern auto layout, and ten years of accretion left components carrying options nobody remembers choosing. Familiar, and survivable. Humans had learned every workaround, so patch-and-live kept beating rebuild on the roadmap, the way it does in most orgs.
Then two of our designers, working separately, filed the same finding: AI design tools misread the library. Generate from our components and the output comes back wrong, structure invented, variants misapplied. The bloat a human routes around on autopilot defeats a reader that has no autopilot. Ten years of workaround knowledge turns out to be undocumented, and the machines just told us.
That table was built 5 years ago. Imagine what’s been built in the last three to six months.
This is why it’s a leadership problem and not a cleanup ticket. Your design system’s readership doubled. It used to be designers and engineers, who compensate; now it’s every AI-assisted designer, every PM with a prototyping tool, every codegen pipeline, and they read the system literally. Bad source compounds with every generation built from it. One of our designers is experimenting with a context file that briefs the tools on the system’s intent, and it helps, but it’s a translation layer taped over a source problem.
So the consensus tipped from patch to rebuild, starting where age and traffic intersect: the table, which carries roughly half of our busiest surfaces. The general form of the lesson travels beyond us. Audit your system as if a machine were the reader, because one now is. Where a human compensates and a machine fails, you’ve found the real debt, and its interest rate just went up.
The teachable part
Audit your design system as if a machine were the reader: run your AI tools against your components and log what they misread. Where humans compensate and machines can’t, that’s the real debt. Price the rebuild against every future generation built from bad source.