How we approach UX when building AI-powered products

Generative AI is one of those topics that dominates a lot of today’s conversations. We at Whitesmith are actively engaging in this emerging field by nurturing the skills to explore, research, and implement solutions built around Generative AI.

We recognize the evolving nature of Generative and embrace the opportunities it presents. We focus on addressing its limitations, constraints, and challenges. By applying innovative UX approaches focuses on human-centred design principles, among others, we enhance the practical potential applications of Generative AI.

Challenges

There are several challenges and limitations with existing AI models and workflows, but central to these all are the more user-facing and critical challenges, among which some of are:

• Privacy
  End users need to share more personal or sensitive information with AI systems in order to get meaningful results. This is not very user-friendly, especially since most people find it hard to understand how it all works under the hood.
• Reliability & Trust
  AI systems can be unpredictable and inconsistent, especially when dealing with generative AI (even though this is by design) which can erode confidence. Their very nature makes it harder to trust their output, given that current models tend to hallucinate and provide incorrect and inconsistent information.
• Transparency & Traceability
  The inner workings and machinations of widely available generative AI models exist in a complex domain, making them difficult for most users to comprehend.

Opportunities

As we all try to understand where Generative AI fits into our personal and professional lives and how to get the most out of it, certain opportunities present themselves. These opportunities include, but are not limited to, new patterns, possibilities, and accelerating change. We combine new design principles and processes to facilitate and unlock these opportunities.

Design Principles

We have several principles we work with depending on the scope and complexity of the problem we’re trying to solve, but central to all are the following:

• Design for Trust and Transparency
  Deploy tools, workflows, and patterns that explain AI decisions and operations. Implement procedures and systems for continuous feedback loops.
• Design for Safety
  Communicate clearly the limitations of a system and its level of reliability. Explore and engage ethical considerations. Pick the right systems for the right tasks, with the end-users’ safety as a priority.
• Design for Value and Impact
  AI interactions should directly benefit users. Implement solutions that enhance the user experience.

Design Process

Generative AI mimics intelligence, which means its output can sometimes be unpredictable. We previously wrote about exploring if Generative AI can be a good fit for your product. With the hypothesis that Generative AI is a good fit for your product, our UX process is expanded to include the following

1. Research
   Besides understanding users and their needs we also look to understand AI models, their limitations, their strengths and weaknesses along with peripheral tools that can augment and expand the capabilities of our final solution. This allows us to determine a good fit for our intended solution. Eg. Some models have a larger context and hence are more suitable for dealing with large documents, some models are smaller and more focused and suitable for certain specialised tasks, etc.
2. Design
   The lowest-hanging fruit when implementing interfaces for Generative AI is via chatbots. Admittedly, a conversational interface is most suitable; however, we would like to explore more creative potentials of Generative AI beyond just a text-based chat interface. We design voice-only, voice and text, generative user interface, and a hybrid of these options depending on the scope of a problem we’re trying to solve. Our recent research and shared experiences show that people are getting exhausted with the chatbot-only interfaces that have sprung up lately. We’ve found that going beyond chat-only interfaces opens up more possibilities for dynamic and personalised experiences for end users. We believe that Generative AI can provide a user experience with a greater focus around a user’s goal and final outcome. Essentially, we explore design where Generative AI goes beyond just an assistant you chat with, to an assistant that completes tasks on behalf of an end-user.
3. Prototyping
   We create early prototypes using low-code AI tools with the goal of gathering feedback early. This allows us to validate our ideas and make iterative improvements (e.g. prompt engineering, RAG methods, few-shot prompting, style transfer, agents and workflow tools, etc.), ensuring we’re building solutions that meet users’ expectations before committing significant resources.

Conclusion

By using an iterative approach and an underlying mindset of design thinking, we can create new ideas rapidly while reducing the time it takes to validate and improve those ideas into a solution that has found market-fit. The rapid advancement in AI, especially with the almost frequent introduction of new or improved models, means we have to maintain agility and flexibility when we design Generative AI-powered products.