top of page
Screenshot 2026-05-19 at 6.46.37 PM.png

ClearVoice

AI writing assistant for non-native English professionals
Designing for transparency, trust, and voice preservation in AI-powered writing tools.

Role

End-to-end UX research, IA, interaction design, prototyping, usability testing

Timeline

3 weeks
Self-initiated

Tools

Figma, FigJam, Maze, Zoom, Google Forms

platform

Web app
Desktop-primary

Problem Statement

Over a billion professionals write in their second language every day and most AI tools make it worse.

Existing writing assistants correct errors. But they don't explain why, signal how confident they are, or let users stay in control. For people who are already competent but uncertain, this creates anxiety rather than alleviating it.

3 Key design phases
1.2B

non-native English speakers in professional settings

67%

say they rewrite emails multiple times due to tone uncertainty

0

of 5 major writing tools communicate AI confidence levels to users

Design Challenge
Picture2.png

"How do we design an AI writing assistant that gives non-native English professionals genuine confidence not by removing their voice, but by making AI suggestions transparent, explainable, and truly in their control?

Discover & Research

Qualitative research 

Interview guide - full question set

Method: Semi-structured interviews · 8 participants · 45 min via Zoom · Recorded with consent
Participants: Non-native English speakers in US professional roles - software engineers, data analysts, product managers, HR

    Warm-up · 10 min

  • Can you tell me a bit about your current role and what your team does?

  • How long have you been working primarily in English? How would you describe your relationship with written English?

  • Walk me through a typical workday, what kinds of writing do you do most often?

    Probe: emails, Slack, reports, documentation, presentations

  • Who do you typically write to?

    Probe: managers, peers, clients, skip-levels, external partners

  • Roughly how much time per day would you say you spend on professional writing tasks?

     

      Core experience - 20 min

  • Tell me about a recent email or document you wrote where you felt uncertain about something. What was the situation?

    Critical listen for what "uncertain" means to them

  • What specifically were you uncertain about the grammar, the tone, the structure, the word choice, or something else?

  • What did you do when you felt uncertain? Walk me through exactly what happened.

    Probe: Who did they ask? What tools? How many rewrites?

  • Have you ever sent something and then regretted how it came across? Tell me what happened.

  • Is there a type of writing situation you find consistently harder?

    Probe: giving feedback, pushing back on a deadline, writing to a senior leader, cold outreach, apologizing

  • How do you feel right before you hit Send on an important email? Describe the feeling.

  • What does "getting it right" look like for you? How do you know when a piece of writing is good enough to send?

  • Tell me about a time when something you wrote was misunderstood, or was interpreted very differently than you intended.

    Tool behavior - 10 min

  • What tools, if any, do you currently use to help with your professional writing?

  • Can you show me how you typically use it?

  • Ask to screenshare observe actual behavior, not described behavior

  • What do you like about it? What doesn't work, or what has frustrated you?

  • Have you ever disagreed with a suggestion a writing tool gave you? What did you do?

  • If your ideal writing support tool could do anything — with no technical limits — what would it do?

    Closing - 5 min

  • Is there anything about your experience with professional writing in English that you think I haven't asked about?

  • Is there anyone else in a similar situation you'd recommend I speak with?

  • Do you have any questions for me before we wrap up?

Affinity Diagram
Affinity_diagram.png
Competitive Analysis
Competitive_analysis.png

Define

User Personas
Untitled (1).png
Untitled (2).png
How Might We Question

HMW show users why a suggestion is being made, not just what it changes?

HMW let users set context (audience, relationship, goal) before the AI weighs in?

HMW help users learn over time, gradually reducing their reliance on the tool?

HMW communicate AI confidence levels so users can calibrate their trust appropriately?

HMW ensure the user always feels like the author never a passive recipient?

HMW make declining a suggestion feel as valid and dignified as accepting one?

Ideate

Exploring the solution space
Below are the design principles

Transparent by default

Never make a suggestion without explaining it. The AI must show its reasoning, not just its conclusion.

Confidence-graded

Always signal how sure the AI is. Users should be able to trust high-confidence flags and deliberate on uncertain ones.

Voice-preserving

Show what would change and what would stay. The user's phrasing, style, and intent must survive the editing process.

Context-aware

Let the user tell the AI who they're writing to and why. Suggestions without context are guesses.

Three design directions explored

Direction A

Sidebar suggestion panel

All suggestions appear in a persistent right-hand panel. Text in the editor is underlined when a suggestion exists. User expands each card to see explanation and confidence.

✓ Balances visibility with writing flow
✓ Shows all suggestions at a glance

Direction B

Inline popover on hover

Hovering over underlined text triggers a floating card with suggestion, confidence, and explanation. Minimal UI footprint similar to Track Changes.

✗ Users miss suggestions they don't hover over
✗ Loses the "all at once" scan

Direction C

Writing coach chat mode

A chat interface where the user sends their draft to an AI coach, who responds with holistic feedback rather than inline corrections.

✗ Breaks word-processor mental model
✗ Better for long-form documents, not emails

Decision:
Direction A (sidebar panel) - it balances visibility of all suggestions with not interrupting the writing flow.
Direction B risks users missing suggestions.
Direction C is better suited for long-form writing tools.

Primary Flow
Screenshot 2026-05-18 at 11.29.46 PM.png

Design

Wireframe progression

Screenshot 2026-05-19 at 3.39.20 PM.png
Screenshot 2026-05-19 at 3.39.46 PM.png
Screenshot 2026-05-19 at 3.40.04 PM.png

Final hi-fidelity prototype 

Screenshot 2026-05-19 at 6.46.37 PM.png
Accessibility - WCAG 2.1 AA compliance

Color + icon pairing

All confidence badges use both color and a distinct icon (●, ◐, ○) — color-blind accessible

Screen reader labels

All interactive elements have descriptive aria-labels: "Accept suggestion: change 'I think' to 'I believe'"

Keyboard navigation

All suggestion cards are fully keyboard-navigable (Tab to focus, Enter to expand, Escape to collapse)

Touch target size

All buttons meet the 44×44px minimum touch target. Accept / Decline / Tell me more tested on mobile

Color contrast

All text meets WCAG AA minimum 4.5:1 ratio. Body text on card backgrounds tested at 6.2:1

Focus indicators

Visible focus ring on all interactive elements - 3px offset, high-contrast color, not hidden

Test

Usability testing - 6 participants, moderated over Zoom

Each session was 45 minutes. Participants matched the primary persona: non-native English professional, writes daily in English, US-based.
 

Below are the 5 usability findings:

1. Confidence scores were the breakout feature unprompted comments from 5 of 6 participants

Nobody was told to pay attention to the confidence badges. Yet 5 of 6 participants commented on them without prompting. Average time spent reviewing a high-confidence card: 8 seconds. Average time on a low-confidence card: 34 seconds - users deliberated more, exactly as intended. "Oh wow, it tells you how sure it is? I wish Grammarly did this."

2. The context setter caused one point of confusion the word "relationship" was ambiguous

2 of 6 participants weren't sure whether "relationship" meant their professional hierarchy (manager vs. peer) or personal familiarity (do they know this person well?). It was intended to mean the former but the label didn't say so. Fix: renamed to "How well do you know this person?" with options: Well · Professionally · Just met.

3. "Tell me more" was underused but loved when discovered

Only 3 of 6 participants found "Tell me more" on their own during Task 3. The other 3 discovered it only when prompted. The original design used a small gray text link — easy to miss. Fix: changed to an icon-button with visible label, moved above the action row, given more visual weight.

4. Users wanted a "preview all changes" view before committing

Only 3 of 6 participants found "Tell me more" on their own during Task 3. The other 3 discovered it only when prompted. The original design used a small gray text link easy to miss. Fix: changed to an icon-button with visible label, moved above the action row, given more visual weight.

5. Self-reported writing confidence improved by 50% after one session

Pre-task confidence rating (1 to 5 scale): average 2.8/5. Post-task rating after using ClearVoice: average 4.2/5. All 6 participants rated higher confidence after the session. 5 of 6 said they would use ClearVoice over their current tool. NPS: +67.

Iterations made after usability testing

1. Context setter: "Relationship" → "How well do you know them?"

Screenshot 2026-05-19 at 5.06.18 PM.png

2."Tell me more": Gray text link → Prominent icon-button

Screenshot 2026-05-19 at 5.06.28 PM.png

3. Added "Preview final draft" before copy

Screenshot 2026-05-19 at 5.06.44 PM.png

Results

Outcomes
+50%

increase in self-reported confidence after one session

94%

task completion rate across all 4 usability tasks

64%

Net Promoter Score 5 of 6 would use over Grammarly

0/5

competitors offer AI confidence levels - a genuine market gap

What I learned

Designing for AI transparency is different from traditional software

​The hardest design problem wasn't layout or interaction, it was deciding what the AI should say about itself. "High confidence" sounds reassuring. "Low confidence" sounds scary. Language and framing matter as much as visual design. I'd spend more time on UX writing for AI states in future projects.

Control is not the same as complexity

Users didn't want more buttons or more options. They wanted to feel like the AI was working for them, not replacing them. Every design decision kept returning to this. Restraint fewer features, each done with more care was the right call.

bottom of page