We took a major process in the app to redesign. The kind of flow that touches multiple user roles and has years of accumulated complexity. We started the way you’re supposed to: collected and analysed the current data, drafted the existing process with the biggest pain points marked, built the PRD, and verified with engineering which automation routes could address those pain points.
Then we asked AI to draw the improved flows based on the documentation and generate mockups powered by the design system.
What came back
The flows were logical and the mockups were quite consistent. AI treated every automation route in the documentation as feasible and built the flows around them, but without questioning whether engineering could actually deliver any of it.
That’s a third thing AI didn’t do, and we almost missed it: it assumed the documentation was ground truth. We still haven’t verified those routes with the engineering team.
What AI didn’t do
Beyond the feasibility assumption, two more things were missing.
It brought no outside inspiration. No patterns borrowed from competitors, no solutions adapted from other industries, nothing we hadn’t already described in the brief. AI optimised what we gave it. It didn’t look sideways.
Second: it didn’t question the flow itself. It followed the sequence we described and improved each step, but never suggested rearranging them. Never said “you could collapse these three steps into one” or “this entire flow could work differently.” It treated our process shape as a given and made each piece better within that shape.
In hindsight, that makes sense. We gave it a brief, and it executed the brief. That’s what it does.
Going back to Figma
We took things back to brainstorming sessions and Figma, step by step. Started drawing and analysing the flow the old-school way, with the AI-generated diagrams and documentation open as reference.
The first thing we did was play with rearrangement and cutting down the steps. Move this step before that one. Things you do naturally when you’re dragging boxes around a canvas but that AI never proposed because it was following our described sequence.
Then came the bigger move: eliminating noise. We went through every screen and cut unnecessary text, or removed components that didn’t earn their place. For each remaining step we ran a “what if” analysis: what if we removed this entirely? What if this happened automatically? What if the user never saw this screen at all?
That’s where the real improvements came from. Not from optimising the existing flow, but from questioning whether each piece of it needed to exist.
What we took away
When you rely on AI for the design, you get boxed in. You see a well-structured, feasible output and it’s easy to accept it because it already looks better than what you had. You get biased toward the solution in front of you.
But AI’s output is also a strong starting point to critique. You have a “somehow good” flow already built, so you can only do better from there. You’re not starting from a blank canvas, you’re starting from a draft that handles the basics, and your job is to challenge it.
Did it speed things up? Massively. The combination of AI-generated flows and manual rework in Figma was faster than either approach alone would have been.
Would I hand off this kind of process redesign entirely to AI? It depends. If I haven’t got the resources to give it proper attention, yes: because even the AI-only version improves things. But it won’t reach what the human-plus-AI version achieves, because AI won’t question your brief. It’ll execute it faithfully, and sometimes faithful execution is exactly what you don’t need.
