Gen-AI design process: Rethinking the traditional approach for next-Gen UX
When I first began working on a Gen-AI project nearly a year ago, I frequently reflected on the process for developing it. Initially, I leaned on the Double Diamond model but soon realized it was not ideal and definitely required some level of modification.
The Double Diamond model, popularized by the British Design Council in 2005, originates from a markedly different era in our history — when design thinking and digital paradigms were just beginning. With the rapid evolution of technology, particularly AI, resulting in a major shift in interaction paradigms in recent years, there’s a pressing need to rethink the design process.
Part 1: The need for Reevaluation
If you have worked on the GenAI products/ features before or are just starting, you might find adopting a reverse engineering approach beneficial. This approach involves examining what current technology enables and figuring out how it can be utilized to achieve user outcomes.
Now, if you’ve adopted the Double Diamond as a concrete baseline, you might encounter disappointment, as I did when I found myself fiddling in various directions. While a process is meant to guide progress, an unrealistic one may lead to a certain level of frustration when other stakeholders are not onboarded in that direction. Let’s briefly examine the Double Diamond and the challenges it poses, especially in AI-related problem-solving.
- Excessive Focus on Desirability: The left side of the first diamond is all about understanding human needs (desirability), often at the expense of assessing the feasibility and viability of AI solutions. This could result in time spent exploring human-centric problems without evaluating their practical AI solutions.
- Designed for Deterministic Systems: A deterministic system performs set tasks predictably, while a probabilistic system dynamically responds to inputs with uncertain outcomes. The Double Diamond primarily caters to deterministic systems, struggling to accommodate the probabilistic, iterative nature of AI development. This leads to a mismatch between the linear progression of the Double Diamond and the iterative nature of AI development (More read: A new age of UX).
- Neglect of real-world constraints: In the industry, it’s not uncommon for stakeholders to already come up with predefine outcomes, such as designing an AI chatbot. This approach contradicts the Double Diamond model, which prioritizes user needs first. However, working within these constraints is sometimes necessary. (Talk by Andy Budd — Design’s Mid-Life Crisis)
Nonetheless, we must understand that design is inherently a messy activity and no process can be perfect. The advantage of the Double Diamond is that it offers a structured approach to problem-solving, aiding teams through the complexities of modern product development. Therefore, we’ll use the Double Diamond as a foundation for our process, adapting it for the Gen-AI process to make it more actionable.
Part 2: Proposed Gen-AI Design Process
Let’s briefly examine each of the above steps outlined:
- Define: This initial stage is dedicated to identifying the product’s core domain, prioritizing AI-driven opportunities, determining data sources, and articulating the user problems and objectives intended to be addressed*
- Develop: The subsequent phase entails the collection of diverse data sets, the creation of design prompts, the prototyping of concepts, and the deployment of preliminary AI solutions — which is instrumental in uncovering real data and interaction patterns.
- Refine: The expanded phase of the diamond further broadens the use case through the evaluation of AI performance, the execution of user research activities, and the iterative refinement of designs.
*It’s important to acknowledge, as mentioned previously, that in some cases stakeholders might already have a high-level outcome in mind at this stage. Nonetheless, it’s crucial not to overlook the definition of the core user problem you’re addressing and what success criteria could look like.
Key considerations:
A. Emphasis on the AI-driven opportunities: It is important to clarify that the omission of the Discovery phase does not imply a neglect of Discovery activities. On the contrary, User Research and Discovery are more critical than ever for delivering real value. However, Discovery should not be constrained to a linear progression; rather, it should be an iterative process integrated throughout all stages. (Relevant book read: Continuous Discovery Habits.)
B. Integration of the Refinement phase: This phase is specifically introduced to highlight the probabilistic nature of AI products, emphasizing the need for continuous evaluation and improvement. Once confident in the data output, efforts can shift to scaling the solution and extending it to more use cases within the identified core domain.
C. Acknowledgment of the critical role of data: Design decisions will be deeply impacted by the data’s nature and the model’s responses. It’s crucial to be actively involved in every phase to craft the optimal experiences.
Part 3: Deep-dive for the new Gen-AI Process
Let’s explore each of the process in detail to get started better:
Phase 1: Define
The key steps in this phase could include:
- Identification of the product’s core use case to be addressed with GenAI, focusing on the challenges to be solved.
- Exploration of AI-related opportunities, followed by prioritization based on their potential impact and the effort required for implementation.
- Determination and validation of data sources to ensure their reliability and the quality of the response.
- Formulation of the problem statement, leveraging insights gained through these activities and conducting further discovery work as necessary.
“Defining UX in Gen-AI is an art of balance. While working backwards from what technology can enable, spend considerable time understanding user needs, ensuring technology serves humanity, not the other way around.”
Key Callout:
1. Familiarize with AI and LLMs Basics: Herein, you should familiarize yourself with AI fundamentals and the workings of Large Language Models (LLMs), including their inherent limitations such as the potential for generating incorrect information, known as “hallucinations,” and the impact of biased training data (AI for UX: Getting Started). Recognizing these limitations is essential for evaluating feasibility and prioritizing ideas effectively. Notable use cases of Generative AI include content creation, summarization, discovery, and automation.
2. Innovate with core user needs: An AI solution will not be useful unless it helps users achieve their desired outcomes. Maintain a focus on the fundamental needs of your users, as these remain unchanged even as technology progresses. Utilizing frameworks such as “Jobs to Be Done” could help in monitoring user needs throughout the process (Jobs-to-be-Done Framework). As the saying goes:“People don’t want to buy a quarter-inch drill. They want a quarter-inch hole.” — Theodore Levitt
3. Embrace Data throughout the process: Data plays a crucial role in Generative AI projects, significantly affecting both the direction and quality of outcomes. The principle “Garbage in, Garbage out” highlights the critical need for high-quality data, which should be accurate, complete, consistent, fresh, and unique. Ensure to make informed design decisions by assessing the quality and relevance of the data available for your project. (Google: People + AI Guidebook)
Phase 2: Develop
The key steps in this phase could include:
- Balancing data and user needs by exploring potential datasets, establishing workflows, and creating style guides.
- Defining prompts and response types to match the desired tone and voice of the results.
- For AI features related to creation and discovery, where complexity arises from unpredictable user queries, consider developing prompt suggestions to facilitate easier user articulation.
- Launching initial AI solutions to learn about the inputs and outputs in real time, validate the response resolve open issues and identify unexpected outcomes, while maintaining focus on the ultimate solution. As the saying goes: “Always design a thing by considering it in its next larger context — a chair in a room, a room in a house, a house in an environment, an environment in a city plan” — Eero Saarinen.
“While designing for Gen-AI, remember perfection is a moving target. Launch your MVP early to understand real-world data, ensure the output functions correctly, see how users react, and refine your path forward.”
Key Callout:
- Design prompts together: Collaborate with your team to guide the generative AI’s responses, addressing potential challenges along the way. Involve the copy team throughout the process to enrich the user experience, ensuring it feels personalized rather than overly mechanical. (Google: Generative AI prompt samples)
- Plan for co-learning. Introduce a feedback loop for continuous improvement, similar to Bard & ChatGPT, where user suggestions after each interaction help evolve the model. Proactively inform users of potential errors and provide clear alternatives for when they occur. (Google: People + AI Guidebook)
- Onboard in stages: Since Gen-AI is relatively new, it’s important to educate users about its capabilities, limitations, potential changes, improvement methods, and the logic behind its results. No matter how resilient the system is, errors can occur. Assisting users in recovering from failures can help build trust, even if the responses are flawed.
Phase 3: Refine
The key steps in this phase could include:
- Assisting with human evaluations for reviewing and moderating generated content, guiding it towards ethical and socially responsible directions.
- Analyzing user interactions and conducting extensive user research to identify areas for improvement.
- Discovery of new opportunities and broaden the application to more use cases within the same domain.
“Once confident in the data quality through initial testing and tuning, experiment with refining the solution and exploring additional use cases within the same domain, based on user research and data analytics."
Key Callout:
1. Conduct both Quantitative and Qualitative Research: Similar to other projects, you should perform a mix of qualitative and quantitative research methodologies, such as A/B testing to understand the potential impact on the main KPIs and usability testing, which provides qualitative insights to comprehend user reasoning. (3 Essential research tips for product designers).
2. Monitor Data to Enhance Engagement: If you notice a lot of users bouncing off after the initial interaction, data could be one of the major issues. Keep an eye on false positives & negatives generated, watching out for for unintended consequences: beyond generating false results, issues like duplication and translation errors may arise. (Google: People + AI Guidebook)
3. Expand to New Use Cases Responsibly: With increased confidence, consider scaling to other mediums to address different user problems using the gathered data. Take into account ethical considerations discovered through human evaluation to ensure the design process avoids bias, respects privacy, and prevents harm.
Summary
Through this article, I wanted to offer a glimpse into how I have approached and am approaching the development of Gen AI features. Again, there is no one-size-fits-all solution for everyone but I hope this article helps you in some way during your Gen AI journey. Meanwhile, enjoy navigating this unexplored space.
This article wouldn’t have been possible without the brilliant individuals I’ve had the privilege of working alongside. I extend my heartfelt thanks to everyone who contributed to shaping these lessons over the years. If you have something to add, please feel free to leave a comment on the article or connect with me on LinkedIn.
References and Helpful links:
- AI: First New UI Paradigm in 60 Years
- AI for UX: Getting Started
- Visual storytelling of how LLMs work
- A new age of UX: Evolving your design approach for AI products
- Google: People + AI Guidebook
- Google: Generative AI Fundamentals
- The executive’s guide to generative AI
- Design Council: The Double Diamond
- Improving The Double Diamond Design Process
- Jobs-to-be-Done: A Framework for Customer Needs
