You're Probably Designing AI Products Wrong
The design process for building AI-first products, especially those using LLMs, is significantly different from traditional product design. In the past, the design process was linear and looked something like this:
- The design team would research user needs and current design issues
- They would create low-fidelity and high-fidelity mockups and wireframes
- These designs would be handed off to developers to build
However, when building AI-first products using LLMs, the process should look something more like this:
There are now three key components: design, development, and research. The research component often involves testing numerous iterations of prompts, parameters, and models to create the AI system needed for the product.
Importantly, the design process is no longer linear, but cyclical. It starts with design, perhaps in a tool like Figma or Sketch, but equally important in this early phase is scoping out the research aspect. Sometimes, after building the front-end and back-end around the experience, the prompts or language model calls may not perform as expected when applied to real-world use cases.
To illustrate this challenge, consider a hypothetical flight planning app that uses an LLM to search and summarize flight options. Can the LLM accurately parse and understand data from various flight aggregator sites? How will the app handle cases where the LLM misinterprets the data or provides incorrect summaries? What prompts and parameters will yield the most useful and relevant results for users? Answering these questions requires close collaboration between design and research teams to test and refine the AI capabilities in the context of the desired user experience.
Therefore, in the early phases, research and design need to happen in tandem, with close communication between these teams. Ideally, the same person or team would handle both research and design, but since these require different skill sets, they are often separate units that need to work closely together.
When a new design innovation is proposed, it must be validated by research to determine if it is feasible with the current AI capabilities. The design, development, and research cycle is tightly intertwined, resembling a jumbled ball of yarn more than a clean circle or straight line.
For those building an AI-first product and trying to innovate on the user interface and overall experience, it is imperative to have strong collaboration and frequent communication between the research, development, and design teams. This can save significant time by catching issues early, before investing heavily in prototypes or prompts that may not meet the business requirements in the real world.
Potential Challenges and Solutions
Shifting to a cyclical, research-heavy design process poses some challenges:
- Skill gaps: Designers may lack experience with AI systems, while researchers may be unfamiliar with UX principles. Cross-training and hiring for interdisciplinary skills can help bridge these gaps.
- Longer timelines: The iterative nature of the process may lengthen product development timelines. Setting clear priorities and managing stakeholder expectations is crucial.
- Balancing innovation and feasibility: Bold design ideas may push the boundaries of current AI capabilities. Teams need to find a balance between innovation and what's technically feasible.
Broader Implications
The shift towards AI-first design thinking extends beyond individual products:
- Organizational structures: Companies may need to rethink team structures and incentives to foster closer collaboration between design, development, and research.
- Design education: Design curricula will need to evolve to include AI concepts and hands-on experience with LLMs and other AI technologies.
- Ethical considerations: As AI becomes more integral to product design, teams must grapple with ethical questions around data use, bias, and transparency.
Embracing a cyclical, collaborative approach to design is not just a matter of process, but a fundamental shift in how we think about building products in an AI-driven world.