<aside> <img src="/icons/bookmark_blue.svg" alt="/icons/bookmark_blue.svg" width="40px" />
AI is everywhere. It’s great at making new things. But can it make existing ones better?
</aside>
<aside>
</aside>
At Enlyte, our icon library had grown organically for over a decade. While we technically had 300+ icons, the set had drifted over time — inconsistencies in weight, spacing, sizing, and metaphor created friction for both designers and product teams.
This case study covers how I used AI as an analysis and production accelerator to audit and modernize iconography, reducing the projected effort for a full refresh by ~70%.
<aside>
</aside>
The icon set didn’t appear overnight. It grew slowly over the years, shaped by different designers, evolving styles, shifting guidelines, and burning deadlines. As a result, the icon system became partially misaligned: mismatched weights, inconsistent sizes, conflicting metaphors, and occasionally, they even actively fought each other when used side by side.
Iconography misalignment carried a cost:
✕ Higher maintenance effort with every update ✕ UI inconsistencies shipped into product experiences ✕ Slower design velocity (more double-checking and workarounds) ✕ Declining trust in the icon system over time
As AI tools became more accessible and more capable, I revisited a long-forgotten backlog item that had been sitting untouched for 5+ years: Iconography Refresh. It had always been deprioritized — the estimate was enormous. Updating the entire icon set manually just wasn’t realistic. So I asked a practical question: could AI change the economics of this problem?
My initial hypothesis was simple:
→ Feed the entire icon set to AI → Add instructions to make the icons aligned and consistent → **Update the design and coded assets → Save months of manual work for the design system team!
In reality, the first attempt didn’t end with a breakthrough. It ended with my laptop overheating at the 2nd step (oops!) – the icon set was too big, the prompt was too general, and the idea – as appealing as it was – was too ambitious. That was the moment I realized this wasn’t going to be a one-prompt miracle.
<aside>
</aside>
To run the audit, I didn’t rely on a single tool. I used ChatGPT, Claude, and Copilot in basic chat mode (agent workflows and research features weren’t available yet). I fed the models PNG exports to compare icons visually, raw SVG code to inspect structure, and file naming conventions to analyze how the library was organized.
AI accelerated the work humans struggle with at scale:
✓ clustering similar icons ✓ detecting repeated patterns ✓ flagging inconsistencies ✓ surfacing redundancies