AI-Assisted Survey
Response Analysis Tool
An internal AI-assisted research tool for healthcare research teams — turning open-ended responses into structured insights, faster, and without losing researcher judgment.
Challenge
Manual analysis couldn't scale.
Third-party AI couldn't be trusted.
Researchers were hand-categorizing thousands of open-ended responses across hundreds of surveys — a process that was slow, inconsistent, and nearly impossible to scale as study volumes grew.
External AI tools looked promising on paper but failed two tests that matter most in healthcare: they couldn't be customized to each team's taxonomies, and they raised serious data privacy concerns. No response could leave the building.
Users
Designing for researchers, not consumers.
Before opening Figma, I spent two weeks with the Research & Reporting team — sitting through analysis sessions, reading old export files, and asking what "done" actually felt like. Four patterns emerged, and they shaped every decision that followed.
Researchers don't want AI to replace their judgment.
They want a system that handles the heavy lifting
while keeping them in control at every step. — Core insight from discovery research
Workflow
A workflow, before any pixels.
I mapped the end-to-end analysis pipeline first — from raw input to final report — and made one design rule non-negotiable: AI does the first pass, humans always close the loop. Every interface decision downstream was judged against it.
Testing
Three findings that reshaped the design.
After the first prototype I ran usability testing with 5 internal researchers — the people who live with open-ended data every day. I wasn't testing the visuals. I was testing how researchers behave when AI does part of their job.
Final Design
From AI output to structured insights — in one workflow.
The shipped product moves researchers through five unmissable steps. Each solves a real friction point from testing. Together, they collapse a 7-step manual workflow into 3.
AI first-pass categorization
The researcher selects a question, kicks off analysis, and the system categorizes all open-ended responses — running entirely on internal infrastructure, never sending data to third-party APIs.
Single Editing — response by response
Category tags appear inline beneath each response. Researchers can add, remove, or reassign tags while reading the full context — the most common mode for careful, judgment-heavy review.
Bulk Editing — responses at scale
Once patterns emerge, researchers switch to Bulk Editing Mode — selecting multiple responses at once and applying categories in a single action. Faster decisions, same level of control.
A safety net before destructive actions
Re-running AI can overwrite manual edits — the single biggest source of anxiety in testing. A clear warning modal, an explicit loading state, and a success confirmation turn an irreversible action into a controlled one.
Validate & export
Before anything leaves the system, the Results Preview surfaces category distribution, categorized vs. uncategorized totals, and key metrics. Users review the breakdown, then return to the main dashboard to verify responses before export — answering the universal pre-export question: did I miss anything?
Impact
The tool shipped.
The results spoke for themselves.
The AI-Assisted Survey Response Analysis Tool became part of the Research & Reporting teams' standard workflow at launch. Same rigor, half the time, zero third-party data exposure.
What I learned
The hardest part of AI design isn't the interface.
It's building trust. Clarity, reversibility, and evidence matter more than intelligence — the best AI UX is one where the user never feels like they've lost the pen. Design for uncertainty. Add the safety net earlier than you think you need to. And remember: users don't need AI to be perfect. They need to understand what it's doing, feel in control of the outcome, and have a clear path to correct it when it's wrong. Get those three right, and efficiency follows.
