Redesigning a 35-Minute Onboarding Application
The numbers up front
The problem
Tear it down and start over is rarely a feasible plan of action. In this case, a health insurance company, Medbest, and its onboarding application was a viable candidate: 62 pages. 75 questions. 35 minutes average completion time. Conversion was trending downward, on pace to drop 9% year over year, and 10% of applicants who did make it through still needed a human to intervene before the application could be processed.
Finding where the 35 minutes actually went
With the blessing of stakeholders I started my discovery. Before touching a single screen, I needed to know where time and people were actually being lost, not where I assumed they were.
Funnel and behavioral data. I worked to pull drop-off points and time-on-task by section from Google Analytics. This is where the qualitative signal got quantitative confirmation: one section, the Medical Questionnaire, accounted for 38% of total time spent in the application, 13.3 of the 35.1 minutes, against 21% for the next-largest section. Not 38% of the questions. Thirty-eight percent of the time.
That number changed the scope of the whole project.
Interviews. Eight recent applicants were selected to for lightweight sessions built around empathy mapping. I wasn't testing screens yet, I was listening for where people got frustrated, confused, or stalled out entirely.
The transcripts were fed into a LLM for synthesis, which I followed behind to spot-check for data accuracy. Even though it was a smaller interview set the time savings was worth bypassing the manual labor.
Journey mapping. Using a LLM and existing flow diagrams, I had the as-is flow mapped end-to-end and layered emotion and empathy data from the interview findings directly onto it. The map made the pattern visible in a way a spreadsheet couldn't: friction wasn't evenly distributed across the 62 pages. It clustered.
The decision not to rebuild everything
A full re-architecture of all 62 pages was still on the table. I made the case to target the Medical Questionnaire specifically, because the data said that's where the leverage was. It would have taken longer than a quarter, needed more than three developers, and touched systems, payment, eligibility, account creation, that weren't actually causing the drop-off.
Everything else in the flow got interaction-cost cleanup, not a redesign: workflows made continuous instead of fragmented, redundant content condensed, dead weight removed. The Medical Questionnaire got the real investment.
The wrinkle would be that Legal and Clinical teams' priorities would create strict constraints to navigate. Both teams were adamant about preserving the existing question set and disclosures to prevent undue risk. Because any claim filed during the policy's first 36 months had to map directly back to health history provided at onboarding, the lengthy question set was held as non-negotiable. As we drilled down there were grandfathered or outdated questions that were able to be discarded.
Redesigning the Medical Questionnaire
The as-is questionnaire asked 75-plus yes/no questions about individual medical conditions, one at a time, in clinical language. Users didn't recognize the terms, couldn't tell how many questions remained, and even had to step away to track down provider contact details. It was well documented, in interviews and in the data, that this section is where fatigue set in hardest.
Using Figma Make, I generated layout variations and evaluated them against usability heuristics to establish the optimal visual hierarchy. I then restructured the required disclosures into intuitive body-system categories (such as cardiac, neurological, and orthopedic) using plain language. To streamline the flow, I designed a progressive disclosure pattern that revealed detailed follow-up inputs only when an applicant selected a relevant condition.
To confirm people could actually find their condition inside these new groupings, I ran a card sort before committing to the taxonomy. Grouping conditions the way I assumed they'd be grouped isn't the same as grouping them the way applicants actually look for them.
Then I tested the interactive prototype in three rounds, five users each, rapid feedback between rounds. I stopped at round three: 90% task success, and every session rated the flow "easy" or "very easy." More than one person used almost the same words unprompted, that the design "just leads you from one thing into the next." That's the reaction I was designing for, a sign the sequence itself felt inevitable, not a compliment on the visuals.
"It just leads you from one thing into the next."
Clinical and Legal stakeholders were satisfied with the level of risk tied to the structural changes of the questionnaire, even letting go of outdated or grandfathered rules.
The system underneath the screens
The primary focus was on the Medical Questionnaire, but there was more work completed in Phase 1. Additional workflows were redesigned for continuity or streamlined, content condensed or reorganized, and the useless was jettisoned with zeal.
To accomplish this, three guiding principles carried across the whole application, not just the questionnaire:
Interaction cost, removed wherever it wasn't earning its keep.
Redundant fields, disconnected steps, content that existed because nobody had questioned it in years, gone or consolidated.
Progressive disclosure.
Complexity introduced gradually instead of dumped on the applicant at page one. Cognitive overload is a conversion killer before it's anything else, and the old flow front-loaded almost all of it.
Explicit progress feedback.
A progress bar wasn't decoration, it gave applicants a way to gauge remaining effort, exactly the information the old flow withheld and the thing people said they wanted most in the early interviews.
Results
- Average completion time dropped from 35 minutes to 16, a 46% reduction, well under the business's 20-minute target.
- Manual intervention fell from 10% to 6%, hitting the target the business had set.
- The $12M projected Year 1 revenue lift was modeled with the Business Analytics team from the conversion gains.
It shipped as a single release; usability metrics came from pre-launch testing, business metrics from post-launch measurement.
Lessons learned
This project ran without an A/B test, one release, measured against the numbers it needed to hit. It also wasn't the last phase; later work picked up smaller friction points that didn't carry the same weight as the questionnaire redesign. If I ran it again, I'd push for a phased rollout on the questionnaire changes specifically, since that's the section carrying the most disclosure risk.
Where I'd put AI, and where I wouldn't. I'd hesitate to put it near the Medical Questionnaire. The whole design problem was proving to Legal and Clinical that nothing gets lost or reinterpreted between what someone types and what gets recorded, and a model inferring medical risk from free text breaks that promise.
Where it does fit: pricing quote workflow. If I had to do it again, I'd transform this interaction into a conversational, highly-personalized experience to better map plan features directly to a consumer's unique lifestyle, financial tolerance, and anticipated health needs. Given the high-risk and complexity of the environment, an escape hatch to speak with an individual would also be a must with respect to guardrails.