MailChimp acquired Creative Assistant from Sawa, an AI-powered graphic design tool that automatically generated marketing assets. The technology was promising, as it could create social media graphics, email headers, and ad creative in seconds. But users weren't getting good results.

Users jumped straight into generating designs without providing sufficient brand context

Website scraping pulled brand assets but dropped users into the designer immediately

Incomplete brand kits (missing colors, fonts, or personality attributes) produced generic, unusable designs

Users without websites had no way to use the tool

No validation that brand assets were sufficient before generation

The core issue: Garbage in, garbage out. The AI needed robust brand information to create on-brand designs, but users weren't providing it.

Design an onboarding experience that:

Guides users through building a complete brand kit before generating designs

Validates brand completeness and prompts users to fill gaps

Provides multiple entry points for different user needs (website, sample brand, manual)

Builds trust in the AI by showing users exactly what information it's using

Ships in 3 months with a single designer-developer (me)

This was my first time using React, and I was working in a complex front-end architecture with tricky state management across parent-child components.

I redesigned the onboarding flow to ensure users provided complete brand information before generating designs, with three pathways based on user needs:

For users with existing websites:

Enter website URL

System scrapes logos, colors, fonts, and images

Review scraped brand kit with clear indicators (complete or incomplete)

Edit and refine before proceeding

For users without websites or who want to explore first:

Pre-built sample brand

Fully populated with logos, colors, fonts, personality

Users can customize to match their actual brand

Removes barrier to entry for experimentation

For users who want complete control:

Upload brand assets directly

Build brand kit from scratch

Full editing capabilities

All paths converge on brand kit review:

Logos: Visual display with application rules

Colors: Palette with hex values, prompt to add variety if limited

Fonts: Primary and secondary typefaces

Brand Personality: Sliders for attributes (minimalist, bold, playful, etc.)

Button Styles: Preview of how UI elements will appear

Users can't proceed to design generation until their brand kit is sufficiently complete.

Stack: React, MailChimp Design System

Timeline: 3 months (including React learning curve)

Scope: Designed, coded, and shipped the entire onboarding experience

First time using React (steep learning curve)

Complex component architecture with parent-child interactions

Tricky state management across multi-step flow

Integrating with backend systems for website scraping and AI generation

Worked with one engineer for backend integration

PM briefs from Sawa founder

Design manager guidance

Otherwise operated independently

Shipped onboarding to millions of MailChimp users in 3 months

Improved brand kit completion rates (no metrics available, but qualitative feedback was positive)

Reduced user confusion about why designs weren't on-brand

Enabled users without websites to try Creative Assistant

After I left, the Creative Assistant team shifted to working on GPT-powered text generation and integrated Creative Assistant features into core MailChimp product. In December 2024, standalone Creative Assistant was sunset because all features had been rolled into the main platform.

This is actually a success story: The work was valuable enough to become part of MailChimp's core offering rather than remaining a separate tool.

Users need to see what information the AI is using

Validation and editing capabilities build confidence

Transparent processes reduce 'black box' anxiety

Recovery paths for failures are critical

 

Power users want control (manual upload)

New users want guidance (sample brand)

Most users want convenience (website scraping)

All paths should converge on the same quality bar

 

The challenge isn't generation—it's refinement

UI affordances for editing AI output are critical

Users need to understand when AI will fail

Designing guardrails and constraints is product design, not limitation

 

Building in production code revealed constraints Figma couldn't

Working with engineers as peers (not handoffs) improved quality

Learning React during shipping taught me to design within technical reality

This project was my introduction to AI product design—before ChatGPT, before widespread AI hype, when 'AI graphic design' felt like science fiction. I'm proud that the work was successful enough to be integrated into MailChimp's core product rather than abandoned as a failed experiment.

The onboarding flow I designed addressed a fundamental truth about AI tools: the quality of output depends entirely on the quality of input. By forcing users to build complete brand kits before generating designs, we ensured the AI had what it needed to produce useful results.

Three years later, this lesson still holds: AI product design isn't about the model—it's about the interface between human intent and machine capability.