How do design teams transform their practice with AI?
Key Insights for Design Leaders
- Challenge: Traditional UX design assumes consistency; AI requires embracing productive variability
- Approach: Systematic transformation of Carbon Design System for AI-native experiences (Summer 2023 – May 2024)
- Framework: 7 research-backed principles with 6 implementation strategies operationalised for generative AI design
- Results: 40% HR productivity improvement, new “AI Experience Designer” roles, design transformation at scale
- Go Timeline: 6 months intensive from concept to THINK 2024 debut, “widely considered a triumph of process”
The Crisis: When Design Consistency Meets AI Chaos
By summer 2023, IBM faced a paradox that would challenge their design philosophy. Their Carbon Design System; one of the industry’s most sophisticated design frameworks, was built on a fundamental assumption: interfaces should be consistent and predictable. But the launch of watsonx, IBM’s generative AI platform, shattered that assumption.
Generative AI introduced what IBM researchers termed “generative variability”, the reality that every AI interaction could produce different results. This violated core UX principles that had guided IBM’s design teams for decades. How do you maintain consistency when AI makes every experience unique?
This challenge sparked a pivotal moment for IBM designers, leading to a profound AI design practice transformation.
The challenge was immediate and existential. IBM executives had mandated AI integration across their entire product portfolio through Carbon. The design team had to choose: modify their system incrementally and risk fragmented experiences, or fundamentally reimagine design for an AI-native world.
They chose transformation.
The Response: Reimagining Design Systems for AI
Building the Foundation: Carbon for AI Initiative
IBM’s response was characteristically systematic. In summer 2023, they launched the Carbon for AI initiative, not as an add-on to their existing design system, but as a fundamental rethinking of design for AI-powered experiences.
The structure was ambitious: 8 sponsor-user product teams paired with dedicated UX designers and researchers, all empowered to move at startup speed within enterprise constraints. This wasn’t design by committee; it was design by collaboration with clear accountability.
The centerpiece of their effort became a comprehensive design toolkit that went far beyond traditional component libraries. As Jeannie Servaas, Carbon for AI Implementation Lead noted, “We had to design a language for the unpredictable,” creating new design tokens for colours and attributes specifically for AI interactions.
The AI label component became their foundation, not just a visual element, but a complete rethinking of how designers communicate AI involvement to users. This single component spawned an entire design language for uncertainty.
The Research Foundation: Seven Principles for Generative AI
Parallel to the toolkit development, IBM’s research team, led by Justin Weisz, was conducting a year-long initiative to understand how designers could create effective experiences with generative AI. Their work, later published in ArXiv publication: January 2023 as “Toward General Design Principles for Generative AI Applications,” established seven fundamental principles that emerged from real product team experiences.
These research principles provided the theoretical foundation that would later be operationalised into six practical implementation principles for Carbon for AI.
The Seven Research Principles (IBM Research AI)
Principle 1: Design for Generative Variability
Traditional interfaces assume consistency. AI’s variability is a strength, not a bug. Designers must communicate and embrace this productive uncertainty.
Principle 2: Design for Multiple Outputs & Imperfection
AI produces multiple, probabilistic outputs—some better than others. Designs must present options and communicate that perfection isn’t guaranteed.
Principle 3: Design for Exploration & Human Control
Enable users to explore vast possibility spaces whilst maintaining control. Balance guidance with freedom through iterative refinement.
Principle 4: Design for Co-Creation
Position AI as creative partner, not tool or replacement. Facilitate human-AI collaboration through back-and-forth refinement.
Principle 5: Design for Mental Models & Explanations
Help users develop accurate mental models of AI capabilities and limitations. Provide progressive explanations that build appropriate trust.
Principle 6: Design Against Potential Harms
Proactively prevent harms from hazardous outputs, misuse, or human displacement through content filters, usage guardrails, and human oversight.
Principle 7: Design Responsibly
Ensure AI addresses real user needs whilst minimising potential harms through human-centred, ethical approach sorted collaborative workflows between human judgment and AI capability.
The Six Implementation Principles: From Research to Practice
These research principles were then operationalised into six practical implementation principles specifically for Carbon for AI:
Implementation Principle 1: Intent-Based Outcome Specification
Traditional interfaces assume users know how to achieve their goals through specific interactions. AI requires users to specify what they want, not how to get it. This shift demanded entirely new onboarding approaches and interface patterns.
Implementation Principle 2: Co-Creative Capabilities
Instead of tools that execute commands, AI becomes a creative partner. This required new interaction models that supported collaborative workflows between human judgement and AI capability.
Implementation Principle 3: Managing Generative Variability
The hardest principle for veteran designers to accept. “Generative variability counters traditional UX guidelines that user interfaces should operate in consistent and predictable ways,” the research team found. Success required helping designers embrace productive uncertainty rather than eliminate it.
Implementation Principle 4: Imperfection and Uncertainty Communication
AI outputs are probabilistic, not guaranteed. This demanded new patterns for communicating confidence levels, uncertainty visualisation, and verification prompts for critical information.
Implementation Principle 5: Multiple Possibilities and Exploration
AI’s ability to generate numerous options quickly required new interface paradigms for comparison, filtering, and hybrid approaches that combined multiple AI generations.
Implementation Principle 6: Transparency and Control
Users needed to understand and influence AI behaviour through explainability features, parameter controls, and clear data usage indicators.
Each implementation principle came with four specific strategies, tested with real IBM product teams working on watsonx.ai and internal conversational AI tools; creating 24 distinct implementation approaches in total.
The Implementation: From Theory to Practice
Testing Ground: watsonx Prompt Lab
The principles weren’t theoretical constructs, they were tested under fire with IBM’s watsonx Prompt Lab, an environment for prompt engineering with large language models. The Prompt Lab team became the first to implement the new design approach, creating interfaces that helped users understand the relationship between prompts and outputs.
The team discovered that traditional user testing methods broke down when designing for AI. They couldn’t test for consistency when the whole point was productive variability. This led to new research methodologies that measured user satisfaction with uncertainty rather than its elimination.
“Users needed to calibrate their trust in real-time,” the team found. This insight drove the development of progress indicators that showed not just that AI was working, but what it was considering, making the AI’s “thinking process” visible to users.
Tool Transformation: New Designer Capabilities
IBM recognised that designers needed completely new tools for AI development. Traditional design tools assumed deterministic outcomes; AI required adaptive tooling.
They developed:
- IDE extensions that integrated AI capabilities into design tools
- Low-code visual interfaces for non-technical designers
- SDK integration with design tokens and components for AI features
- Template libraries with pre-built patterns for common AI interactions
But the hidden challenge was role transformation. Successful AI design required understanding foundation models, not just user interfaces. IBM created new role definitions: “AI Experience Designers” who bridged technical and human-centered design.
Daily workflows transformed fundamentally. Instead of wireframing static screens, designers prototyped conversation flows. Instead of style guides, they created “AI behaviour guides” that defined system responses to edge cases.
The Struggles: What Nearly Broke the Transformation
The Trust Calibration Problem
Early implementations suffered from inappropriate trust calibration; either over-promising AI capabilities or making users too skeptical. Justin Weisz explained it as“Users needed to understand both AI capabilities and limitations, not develop blind faith or complete skepticism. Designers had to learn about cognitive biases, trust formation, and how to communicate uncertainty without creating anxiety.”
This wasn’t an abstract challenge. Real user testing revealed that users didn’t trust results that came back too quickly from AI systems. The team had to add progress indicators not because the system needed them, but because humans did.
The Consistency Paradox
Teams struggled with maintaining brand consistency while embracing generative variability. Some early interfaces felt chaotic and unpredictable. The solution required inventing new concepts like “consistency anchors”; elements that stayed the same while AI-generated content varied around them.
As the team learned, “brand consistency doesn’t mean visual sameness. It means consistent values and behaviour, even when the outputs are different each time.”
The Governance Overhead
Implementing responsible AI principles added significant complexity. Every AI feature needed ethics review, bias testing, and explainability documentation. Design reviews that previously took an hour now required half a day.
This wasn’t bureaucratic bloat, it was the reality of designing systems that could impact users in unpredictable ways. The governance framework became as important as the design framework.
The Result: Measurable Transformation
Business Impact
The Carbon for AI initiative delivered measurable results:
- 40% improvement in HR productivity through AI-assisted processes
- 60% productivity gain in application modernisation
- 90%+ automation of contact center cases with conversational AI
But the more significant impact was cultural.
- Time to prototype dropped from weeks to days using AI templates.
- Design iteration cycles accelerated 3x with generative variations.
- Most importantly, designers moved from “controlling every pixel” to “orchestrating human-AI collaboration.”
Process Evolution
The transformation required new research methods. Traditional user research focused on understanding existing mental models; AI design required understanding how users think about AI capabilities and limitations.
New roles emerged organically. The most successful teams developed “AI Experience Designers” who could bridge technical AI capabilities with human-centered design principles.
“Designers have to become part user researcher, part prompt engineer, part system” Charlie Hill IBM Fellow for Strategic Design, watsonx Orchestrate.
Design System Maturation
The work debuted at IBM THINK 2024 and was widely considered a triumph of process. The collaboration between Carbon’s design system team and product subject matter experts created a new model for high-priority projects with short timelines.
More than delivering components, IBM had created a new paradigm: adaptive design systems that could handle AI variability while maintaining brand coherence.
Transferable Lessons for Design Leaders
1. System-Level Transformation Required
IBM’s success came from reimagining their entire design system for AI, not adding AI features to existing patterns. Half-measures would have created fragmented experiences.
2. Research-Driven Principles Beat Ad-Hoc Solutions
The six implementation principles emerged from systematic observation and testing with real product teams. They provided a foundation that could adapt as AI capabilities evolved.
3. Role Evolution, Not Replacement
The most successful teams developed hybrid roles that bridged design and AI understanding. This required new hiring practices and training programs.
4. Governance as Design Foundation
Responsible AI principles weren’t constraints, they were design requirements that shaped better user experiences. Time saved in other parts of design process were invested in ensuring the AI designs were trustworthy.
5. Community Learning Accelerates Adoption
IBM’s Design for AI Guild became a knowledge-sharing hub where teams could learn from each other’s implementations, failures, and breakthroughs.
Implementation Framework for Design Teams
Assessment Phase (2-4 weeks)
- Audit current design system for AI readiness
- Identify team skills and AI literacy gaps
- Map existing design processes that assume deterministic outcomes
- Establish baseline metrics for design velocity and user satisfaction
Foundation Phase
- Develop core AI design principles adapted for your context
- Create initial AI component patterns and behaviors
- Train core team in prompt engineering and AI collaboration
- Establish responsible AI governance framework
Pilot Phase
- Select 2-3 products for AI integration testing
- Implement principles with real user validation
- Develop new research methods for AI experience validation
- Document learnings and refine approach
Scale Phase (Ongoing)
- Roll out AI-native design system updates
- Expand training programs and role definitions
- Build community of practice for knowledge sharing
- Continuously evolve based on AI capability advances
The Broader Implications
IBM’s transformation suggests fundamental changes ahead for design practice:
Design Systems Become Adaptive: Static component libraries evolve into dynamic systems that handle AI variability while maintaining brand coherence.
Designers Become Orchestrators: The role shifts from pixel-perfect creation to orchestrating AI capabilities for user needs.
Research Methods Evolve: Understanding user mental models about AI becomes as important as understanding task flows.
Governance Becomes Central: Every design decision carries ethical implications that must be considered from the start.
Critical Questions for Your Context
Before implementing similar transformation:
- Is your design system ready for uncertainty? Traditional systems optimise for consistency. AI requires designed flexibility.
- Do your designers understand AI capabilities and limitations? Technical literacy becomes essential for effective AI experience design.
- How will you handle responsible AI requirements? Every AI interaction needs appropriate safeguards and transparency.
- Can you maintain brand coherence with generative content? Success requires balancing creativity with consistency.
- What’s your learning velocity? AI capabilities evolve rapidly. Can your design practice keep pace?
From Control to Orchestration
IBM’s Carbon for AI transformation represents more than a design system update, it’s a fundamental reimagining of design practice for an era where uncertainty is a design material.
The productivity gains are impressive, but the real achievement is cultural: transforming thousands of designers from pixel controllers into AI experience orchestrators. They didn’t just add AI to their design system, they rebuilt their design practices for AI.
For design leaders, IBM’s experience offers both validation and caution. The productivity and creative benefits are real, but achieving them requires genuine transformation at every level – tools, processes, principles, and mindset.
The companies that succeed won’t be those that add AI features to existing practices, but those that reimagine their practices for an AI-native world. IBM’s teams have written that playbook through systematic research, principled implementation, and honest iteration.
The future of design isn’t about human versus AI – it’s about designing the collaboration between human creativity and AI capability.
Sources and References
- Justin D. Weisz et al., “Toward General Design Principles for Generative AI Applications” (ArXiv 2301.05578, January 2023)
- IBM Carbon Design System
- IBM Design for AI
- IBM watsonx Platform Documentation
- IBM Design Language
- IBM Global AI Adoption Index (2024)
- IBM THINK 2024 Conference Presentations
This case study is based on publicly available research, IBM’s official publications, and verified transformation metrics. Implementation details reflect IBM’s documented approach to enterprise AI design transformation.
