Salesforce Externship

73% wanted AI help. 0% trusted it.

We built the bridge.

AI was already inside Slack. Students were already using it. But nobody told them what it was doing with their conversations.

This is how we designed transparency into every interaction.

Real Users Tested

Weeks

Features Shipped

Final Designs

Read the Story ->

The Problem

Three moments where AI breaks

Students trust.

Three moments where AI breaks

Students trust.

We didn't guess at problems. We ran Black Mirror brainstorming sessions to surface the worst realistic scenarios then designed for each one.

Bias Moderation

Two students discussed colonial history for class. AI flagged them both as threats.

AI interprets academic discussion as hostile content. Students are silenced. The chilling effect on free expression is immediate and severe especially for international students whose visa status could be at risk.

Misinterpretation

A new teammate used AI to catch up. The summary was wrong.

Sarcasm, banter, and cultural nuance go right over AI's head. A "classic procrastinator move" joke becomes a documented conflict. Trust collapses at the worst possible moment.

Opportunity

What if AI made privacy education feel like part of the conversation?

Instead of burying controls in documentation nobody reads, we designed privacy awareness directly into the moments students need it

inline, contextual, and human.

Instead of burying controls in documentation nobody reads, we designed privacy awareness directly into the moments students need it

inline, contextual, and human.

The Evidence

We tested it ourselves.
This is what AI thought a joke looked like.

We ran our own team's Slack banter through ChatGPT with and without context. The results became the core proof point for our Engagement Style feature.

Same message. Same conversation.

Completely different AI understanding.

Context isn't optional, it's the entire product. `This real experiment, running our own team's Slack banter through ChatGPT became the core proof point for our Engagement Style feature.`

Design Process

What We tried.

What We killed.

What Survived

Every concept below went through Wizard of Oz testing with 40 real students before we decided its fate.

Cut

Personal Info Tagging

Real-time alerts every time a user typed personally identifiable information, student names, grades, contact details.

"Annoying. It fires on everything. I'd disable it immediately." From User Testing

Envolve

Seriousness Slider

Real-time alerts every time a user typed personally identifiable information, student names, grades, contact details.

"What does the emoji mean exactly?" — Ambiguity killed it. Became the Engagement Style dropdown.

Shipped

Contextual Training Overlay

Real-time alerts every time a user typed personally identifiable information, student names, grades, contact details.

"Finally. I know what it's doing and why." — Became core to all 3 final features.

Method

We couldn't access Slack AI.

So we became it.

🧙‍♂️

Wizard of Oz Testing

A human moderator simulated AI responses in real-time while 40 students believed they were interacting with the actual system. Their emotional reactions were completely authentic.

01 Context is Everything

Users expected AI to understand the situation, not just the words. An alert without context felt accusatory.

02 Tone of AI Voice Matters

Legalistic language triggered distrust. Conversational, peer-like explanations were accepted and welcomed.

Envolve

Seriousness Slider

Real-time alerts every time a user typed personally identifiable information, student names, grades, contact details.

"What does the emoji mean exactly?" — Ambiguity killed it. Became the Engagement Style dropdown.

03 Control = Trust

Real-time alerts every time a user typed personally identifiable information, student names, grades, contact details.

What was tested?

Privacy Alert response, Bias Moderation report path, and Engagement Style setup during channel creation.

Findings from the Testing

Lack of Context Awareness

Users expected clear, contextual alerts within conversations.

Engagement style is crucial

Feedback emphasized the need for natural,

conversational AI tone.

Feedback emphasized the need for natural,

conversational AI tone.

Prompted Exploration of Design Systems

Insights led us to dive deeper into design system consistency and visual hierarchy.

From Rough to Real

Lof-Fi Sketches

Mid-Fi Prototypes

Three Features.

Each feature maps directly to an insight from testing. Nothing assumed.

Everything earned.

Feature 01

In-Line Privacy Alerts

"Users weren't afraid of AI. They were afraid of not knowing what it knew."

Custom Alert Control

Testing revealed alerts were triggering on everything, regardless of context. The fix: let users define what matters to them. Alerts become intentional, not noise.

"The alerts are getting triggered for whatever topic, without understanding the context. I'd just turn it off."

Feature 02

Bias Moderation Transparency

"Users were scared to type anything after being wrongly flagged."

Only Visible To You

"The alerts are getting triggered for whatever topic without understanding the context. I'd just turn it off."

Modal = Dialog

Matches Slack's native modal pattern exactly, same surface.

Feels native, not bolted on.

Matches Slack's native modal pattern exactly, same surface.

Feels native, not bolted on.

Button Hierarchy

Cancel = ghost. Report = solid green. Slack never uses two solid buttons side by side. Neither do we.

Feature 3:

Engagement Style

"The same joke reads as aggressive in a serious channel, and friendly in a relaxed one."

One More Thing

Custom Alerts

Testing broke our assumption. Alerts were firing on everything, frustrating users. So I rebuilt the feature from scratch,

making alerts user-defined, not AI-assumed.

Design System

We didn't design with Slack.
We designed inside it.

Every component we shipped had to feel like it could have come from Slack's own product team. Same tokens. Same interaction grammar.

Same component anatomy.

Aubergine

#4A154B

Green

#2BAC76

Red

#E01E5A

Yellow

#ECB22E

Blue

#36C5F0

Dark Canvas

#1D1C21

Design Decisions

Within System

All Color Usage

Zero custom colors across three features. Every token is Slack's existing semantic palette, used exactly as intended. A design system audit would find nothing unexpected.

Within System

Engagement Style step

Added Step 3 to Slack's 2-step channel creation modal using their exact dialog component. Same container, button pattern, step counter. One new concept, zero structural deviation.

Within System

No custom iconography

Consciously chose not to introduce new icons. Every icon is from Slack's existing set. In a real sprint, new icons require a design system team review, we understood this.

Typography & Spacing

Type Hierarchy

Dialog Header . 700 . 8px

Why was it flagged?

System Label · 700 · 11px · All-caps

⚠ PRIVACY ALERT

Body Text · 400 · 14px / 22px line-height

This message contained language that could be interpreted as derogatory.

Interactive Link · 400 · 13px

AI Explained, Learn how AI makes decisions

Spacing 4px Base Unit

24 px

Dialog horizontal padding

16 px

Between alert label and body

12 px

Alert icon to text gap

8 px

Between radio options

6 px

Border radius all overlays

4 px

Between radio options

Human-Centered Lessons from Building
AI for Students

Turning Ethical Frameworks into Actionable Design Decisions

I took this challenge as an opportunity for exploring the concepts of "ethics by design" and how important to follow the guidelines, especially working on the projects like data privacy & AI

Let's talk design!

For the clutter free future & to make the design feel personal

For the clutter free future & to make the design feel personal

Feel free to leave me an email with your name and contact information.