Designing for User Autonomy
Designing for User Autonomy
Designing for User Autonomy
Designing for User Autonomy
Designing for User Autonomy
CONTEXT
What FlairX does
FlairX is an Interview-as-a-Service platform. Companies use it to conduct technical interviews at scale, either with AI interviewers or expert human interviewers from FlairX's network.
The platform handles the entire interview workflow: creating custom questionnaires, scheduling candidates, conducting interviews, and providing detailed feedback to hiring teams.
Team
2 Designers, 3 Engineers, PM, CEO
Timeline
3 weeks
Domain
B2B SaaS, HRTech
My Responsibilities
HCI Research, Iterative Design Cycle
THE PROBLEM
Scaling limitations
FlairX had demand. Companies wanted in. But they couldn't scale past 3-4 active clients a day.
The bottleneck: questionnaire creation. Each custom questionnaire took their 3-person ops team 1.5-2 hours to create.
They were using four separate tools: Google Docs and spreadsheets for job descriptions, ChatGPT for generating questions, Sheets for organizing and taking approval from the company and the founder, then FlairX's platform for final input done manually.
Impact
Business impact: Growth was capped. Revenue constrained by ops capacity. Every new client hire delays expansion.
User experience: Ops team felt like human copy-paste machines rather than domain experts. Repetitive, tedious, error-prone.
"We can't take on more than 3-4 clients at once without hiring more people. We're spending more time creating questionnaires than conducting interviews."
-Founder/CEO, initial conversation
THE DISCOVERY
What I learned by participatory shadowing
I had 2 weeks to deliver. Spent the first two days watching the CEO and ops team create questionnaires.
Critical insight
Users weren't avoiding speed; they were avoiding losing control. Every "fast" solution (templates, full automation) took away their ability to shape the output. So they rejected it.
The challenge was making the process faster while keeping users in control.
COMPETITIVE LANDSCAPE
What others were building
Platforms analyzed
HireVue
Karat
BrightHire
BarRaiser
Intervuew
Metaview
Glider
What they had
AI note-taking
Interview recording
Automated scoring
Static templates
What was missing
Question generation
Smart refinement
Human-in-the-loop
Adaptive workflow
Our opportunity
Build AI-assisted flow that balances automation with editorial control; questions aligned to job descriptions, easy to personalize, with users keeping strategic control.
MAJOR DESIGN DECISIONS
Decisions that shaped the solution
Following core decisions that determined whether users would adopt this or reject it like they'd rejected templates.
AI assistant panel
Iter 1
Iter 2
Iter 3
Iter 4
The forcing function
Watching the CEO use the tool for an hour revealed 4 repeated prompts, constraint drove curation instead of infinite options.
What was traded off
Natural language flexibility dropped. But the 4 curated actions cover ~95% of real use cases observed in testing.
Net outcome
Workflow time dropped to 18 min vs open chat. No learning curve. Buttons show what's possible at a glance.
Job Description Parsing + Skills Step
The problem: Manually extracting skills from JDs, typing them into ChatGPT. Time-consuming, error-prone. Questions often misaligned.
What we built: Upload JD → auto-parse skills → review/edit → set weightage/time → generate questions.
This happens before any questions generate. Users shape direction first.
Why this decision
Business impact: Questions aligned to job requirements. Less rework, fewer complaints.
User experience: Control at the right moment. Users loved reviewing skills before committing. Felt strategic, not reactive.
Tradeoff: Added a step. But users found it helpful, not burdensome; saved time downstream.
Why it's best: Turned out to be most valuable. Strategic control + automated extraction.
Structured Flow (Not Canvas)
Initial idea: Notion-style canvas. Drag-and-drop, build anything, full flexibility.
PM's pushback: "Months of development. We have two weeks."
The compromise: Structured approach. Pre-defined sections, guided flow, clear progression.
Why this decision
Business impact: Shipped in 2 weeks vs months. Got to market fast, started learning immediately.
User experience: Lost "build anything" feeling. Gained clarity, users knew exactly where they were. No decision paralysis.
Tradeoff: Gave up flexibility for speed and simplicity. Less magical, more learnable.
Why it's best: Shipped fast, users understood it immediately.
Question Bank Library
The business need: Preserve competitive edge through vetted, high-quality questions.
What we built: Searchable question bank with tagging, filtering by skill/difficulty/type. Browse, favorite, and add to questionnaires.
My honest take: Skeptical this gets used much. Usage will decline. People won't update it. They'll lean on AI generation because it is faster and personalized.
Why this decision
Business impact: Competitive differentiator on paper. Shows expert-curated questions, not just AI content.
User experience: Fallback for users who don't trust AI fully yet, or need some specific question that they have seem before.
Tradeoff: Maintenance effort, potentially low usage.
Why it's included: Business priority, not product necessity. Leadership wanted it. Sometimes you build for strategy despite doubts.
THE SOLUTION
How it works
IMPACT
What changed
For the business:
Same 3-person ops team now handles 12+ clients (was capped at 3-4). No new hires. Growth unconstrained.
For users:
Two weeks after launch, ops team stopped using ChatGPT for questionnaires. Stopped opening Google Docs for this workflow. Just used the assistant.
Time: 2 hours → 20 minutes average.
Adoption: The entire ops team. No training. No mandates. They chose it.
What made the difference: The JD parsing step. Users loved reviewing skills before generating questions. Control at the right moment.
What I'm skeptical about: The question bank. Usage will probably decline as users grow to trust AI generation more.
MY TAKEAWAYS
Reflections
Watch what users do, not what they say
CEO said he wanted "better templates." Watching her revealed she avoided templates and repeated the same ChatGPT prompts. The solution came from observation, not asking.
Constraints force better solutions
Engineering said "no open-ended chat." I resisted, then realized curating most-used actions beat infinite flexibility. Limits reveal what matters.
Control drives adoption more than speed
Users rejected fast solutions that removed control. We shipped the fastest version that kept them in control. That's why they adopted it.
Should've pushed harder for inline editing
We put AI actions in a sidebar. It works, but inline editing would've been more intuitive and saved space. I didn't fight for it. Next time, I'd explore deeper before compromising.