A Systemic Approach to Human-AI Collaboration
As AI Lab Lead, I conceptualized a generative framework to break a "Maintenance Loop" that consumed 60% of designer capacity. Using cards and invitations products as an experimental pilot, I implemented AI agents and a centralized Prompt Library to transition the team to strategic Art Direction with AI, eliminating the friction of legacy templates to reclaim room for innovation.
Validated via A/B testing, this experiment delivered a 34% reduction in production time and a 45% lift in design freshness. The resulting conceptual blueprint serves as a scalable model for balancing creative integrity with operational speed across the company's product catalog.
Categories
AI Research, Experimental AI Design
Tools
Figma, UserTesting, FullStory
HEAR THE STORY
01
The Catalyst
Why is satisfaction dropping for our cards and invitations?
The Cards and Invitations category was a top seller, but it wasn't evolving. We were wasting the graphic design team talent on manual data entry instead of innovation. This created a 'visual debt' that was actively driving our customers toward competitors.
The templates page
The User Friction Clusters
Research revealed a disconnect between user aspirations and our current reality. While customers were seeking modern aesthetics, our catalog was stuck in 2020 trends.
The Maintenance Loop
Because our system was so rigid, graphic designers spent the majority of their time manually refreshing old seasonal collections (e.g., updating dates from 2024 to 2025) rather than creating new work.
60% Technical chores
(manual updates)
40% Creative
02
Design Mandate
Core objectives of the AI Lab
Based on the customer friction and maintenance loop, the experiment was built around four strategic pillars:
Eliminate technical friction
Modernize Offer Perception
Reclaim Creative Capacity
Establish a Scalable Blueprint
03
Designing for Intent
From executors to art directors
To solve the "Legacy Trap" and high internal production costs, I steered the AI Lab toward LLM (Large Language Models) experimentation to augment rather than replace our talent. Through direct collaboration with the graphic design team, we created a Centralized Prompt Library, a foundational tool that turned the designers' aesthetic expertise into a shared, generative resource, empowering them to scale their creative vision without manual repetition.
Create a wedding invitation with an 'Ethereal Garden Romance' aesthetic. The visual style should feature soft watercolor illustrations of dusty rose florals...
Create a graduation announcement with a 'Modern Deco Academic' aesthetic. The graphic style relies on clean geometric patterns, intersecting architectural lines...
Create a holiday card with a 'Rustic Nordic Winter' aesthetic. The graphic style should feature symmetrical, folk-art inspired patterns of pine trees, snowflakes...
Systemizing design DNA
By codifying the spirit of a seasonal collection into the library, the system gained a contextual understanding of design that empowered graphic designers to generate vast template variations.
From rigid boxes to fluid design
Instead of graphic designers manually updating templates, the system intelligently re-architected card layouts. Whether handling holiday greetings or varied photo orientations, the AI-driven system maintained visual harmony automatically,.
04
Quality Lens
Guarding professional standards with a scoring framework
The most critical output of the Lab initiative was the Quality Control Scoring Key. We created this framework using human judgement to ensure every AI-generated card met professional print standards by defining success across four essential pillars.
VISUAL INTEGRITY
2 / 3
3 / 3
2 / 3
1 / 3
OVERALL SCORE
1.75
While the design is conceptually strong and fits perfectly within the collection, it has critical "blockers" that prevent it from being shipped or printed as-is.
Partially automating checks with AI judges
Our long term-vision was a balanced approach to ensure that we could scale production without losing the human intuition. For example, if a design contained a compromised layout, the set of AI judges would evaluate it against the scoring key, flagging it so that a graphic designer would need to verify, tune and approve the final outcome.
Accessi-Billy
Checking WCAG compliance
ACCESSIBILITY SCORE
1 / 3
Fails text legibility
ACCESSIBILITY SCORE
3 / 3
Fully AA/AAA compliant
ACCESSIBILITY SCORE
2 / 3
Text contrast needs improvement
05
AI Agents
Designing a fleet of specialized roles for partial automation
To move from a conceptual framework to an operational reality, I conceptualized the logic for an initial AI Agent Infrastructure that could act as a "digital workforce" that would handle the heavy technical lifting that previously burdened the graphic design process.
06
Vibe Coded Validation
Using "Vibe Coding" as a new prototyping tool.
To validate the effectiveness of the AI Lab experiment. I used Vibe Coding, a prompt-based functional prototyping method, to recreate the "Template Selection Page" featuring 50 AI-generated postcard templates. Then I ran an A/B test to compare customer perception of these new designs against our legacy inventory.
I chose this approach not only for its speed but as a strategic learning opportunity to begin experimenting and familiarizing with this new class of design tools.
A/B testing the AI generated designs with customers
In collaboration with the graphic design team, we established thresholds for success, focusing on aesthetic quality, user sentiment, and flexibility. The results from the test with customers proved that our generative approach modernized the product's appeal.
"I don't like the colors, and the photo looks AI generated"
"It feels really fun, and it’s a great way to put my puppy’s photo"
"I feel like any of our pictures would look great in it"
"The colors are pretty and I love the cute little drawings"
"It looks like something I just printed out at home"
07
Outcomes & Learnings
Higher "freshness" score
The new experimental experience outperformed the legacy flow, boosting customer satisfaction.
Production Time
Optimized the internal delivery window, allowing for greater collection variety without increasing operational costs.
Learning Curve Gap
While workflows were faster, we identified a significant disparity in AI literacy that requires ongoing, mentorship and training sessions.
Aesthetic Artifacts
AI generated compositions occasionally produce uncanny or weird layouts that require manual oversight to ensure the human touch isn't lost.
08
Looking Ahead
True AI Integration isn't a choice between speed and soul; it’s about building systems where the machine handles the friction so the designer can provide the resonance. This experiment proves we can scale our vision without sacrificing the high quality standards of our craft.
