Designing backward: reviewing what an AI team already built
What it looks like to be the designer on a project where the interface existed before I did
About this project
This is a vendor onboarding platform, launching first into health related industries, with plans to expand into others afterward. It's still in development, and most of the MVP is already built.
Here's the part that made this project different from every other one in this collection:
It was built primarily by an AI team, using agentic coding, before I ever touched it.
My process here ran in reverse. Instead of designing screens and handing them to developers, I came in after most of the functionality and flows already existed, and my job was to review it, enhance it, and decide whether it was actually ready to sign off on.
Role: Product Designer (UX/UI)
Work: UX audit of existing flows, onboarding redesign, systems thinking, workflow design
The completion indicator was actively misleading, showing 100% when it wasn’t
AI assisted document mapping already reduced busywork, but there were no clear signals of onboarding progress
15 steps onboarding form
The AI mapping already removed a lot of the busywork. But removing steps isn't the same as making the process make sense.
How the system presented the process, how it signaled where you were and what was left, was genuinely confusing.
There were no clear signals of progress, and the completion indicators were actively giving people the wrong idea, which I'll get into later. The original form had 15 separate steps to get through, mapped or not.
Job to be done
“As a vendor onboarding for the first time, I want to know clearly where I am and what’s left, so I don’t lose trust in a process I’ve already put real information into.”
My hypothesis
If I collapsed the perceived structure without cutting real requirements, and gave an honest completeness signal, people would get through with more trust.
The process
The questions that actually shaped this
How might we make a long process feel short, without cutting requirements?
How might we give people an honest, accurate sense of real completeness?
How might we let AI handle the busywork of data entry, without also handing it the job of explaining the process to a person?
What “working” would actually look like
The process feels short and clear, even though requirements didn’t shrink
People can always tell how complete their profile really is
Vendors complete onboarding without dropping off
The team keeps shipping at the pace AI development allows
Compressing 15 steps into 3
The original form walked people through 15 separate steps or sections. I restructured the entire flow into three main steps: upload, fill in the remaining fields, and review and submit.
Upload is where the AI mapping does its work first. Fill in the remaining fields covers whatever the system couldn't confidently map on its own. Review and submit is the final check before it's sent.
I did consider collapsing the middle step further, into one continuous form. I didn't. Merging everything into a single long scroll would have made it harder, not easier, for people to track their own progress, so I kept it split into 4 to 5 sections within that step, small enough to feel manageable, structured enough that progress still feels visible.
Progressive disclosure, applied everywhere
Across the entire onboarding experience, I only show what's relevant to the step someone is actually on, instead of surfacing every field and requirement up front.
This is a big part of why 15 steps of underlying complexity can now feel like 3.
Fixing a dashboard that was quietly lying to people
Before I got involved, the dashboard would claim a vendor's profile was 100% complete the moment every required field was filled in.
Technically, that was true. In practice, it was a false signal: a vendor could see 100%, assume they were fully done, and never realize they'd skipped optional fields that would have made their profile genuinely stronger.
I fixed the indicator so it reflects real completeness, not just the required field threshold, so "complete" actually means complete.
Fixing a dashboard that was quietly lying to people
Part of my process is I audit the entire flow for the onboarding. My auditing process for this is quite different to how I used to do it. Before, I will manually record everything: from user flows, manual screenshot and consolidate findings.
Now, part of this is I used AI to automate the auditing process. I gave it a step by step process of what to follow, what to include in testing, and strictly save a screenshot of every states. I run it overnight then next day, I was able to compile and validate this myself.
“It’s less about being the first to imagine the screen, and more about catching the gap between something that technically works and something that earns trust.”
This project taught me something the others didn't: what a designer's job actually looks like once AI can build the thing, and even handle a chunk of the actual data entry, before a designer ever sees it.
It's less about being the first person to imagine the screen, and more about catching the gap between something that technically works and something that actually earns someone's trust, like a completion indicator that was accurate on paper and misleading in practice.
I don't think that makes the work smaller. I think it's a different kind of judgment call, and one I expect to make a lot more often from here.