Generative AI has been a buzzword for quite some time now, but building real-world solutions with this technology has been a journey of discovery and adaptation. I’ve seen a lot of information about how to use generative AI tools to aid the design process, but not a lot on how to design for generative AI experiences. Having wrapped up a recent project that designed and built a number of generative AI proof of concepts that focused on increasing internal productivity, I’d like to share some key learnings and guidance for designers creating similar experiences.
Start with the right problem but make it the right problem that can be solved by generative AI
One of the first lessons learned is the importance of starting with a clearly defined problem. For our project, we collaborated closely with the client to identify specific business and customer issues that could be addressed with generative AI. Understanding what problems generative AI, specifically large language models (LLMs), solve best is a perspective that needs to be brought in from the very start. LLMs are very good at specific things, like summarisation, text classification and answering questions via conversational interfaces, however the more constrained an LLM output needs to be, the more challenging designing and building the ideal solution becomes.
We approached each of our identified problem statements as its own proof of concept, this allowed us to focus our experimentation with the LLM and apply learnings to the next concept. For instance, one of our focuses was on simplifying the search and retrieval process for a 450-page PDF document used by staff. This document contained critical information that staff needed to access quickly in order to provide advice and guidance to customers. The existing process was cumbersome, requiring staff to manually search through the PDF, which could lead to delays and errors. Our challenge was to streamline this process using generative AI, making it faster and more intuitive, while balancing the accuracy of the response provided by the LLM.
The three-lens approach
In any project, balancing desirability, feasibility, and viability is essential to ensure value is delivered. When designing generative AI solutions, this balance becomes even more important because there are new considerations for each lens. We constantly asked ourselves:
- Desirability: Does the solution meet user needs? For instance, can staff easily find the information they need using our generative AI-powered search tool?
- Feasibility: Can the technology support the solution? With generative AI, it was about ensuring that the model could handle complex and nuanced queries effectively.
- Viability: Is the solution sustainable and safe for the business? This was particularly important given the regulated nature of our client’s industry.
Our approach ensured that we weren’t just building something technologically impressive but also something that users would find desirable and that the business could safely implement at scale. During the feasibility analysis, we worked closely with engineers to ensure the generative AI technology could handle domain-specific terminology and adhere to the strict regulatory requirements of our client.
Empathy and iteration
Empathy is at the core of design thinking, and it’s no different when working with generative AI. We spent a significant amount of time understanding the day-to-day challenges faced by the staff. One poignant moment was when a staff member shared how intimidating it could be to manually search through a massive PDF document while trying to provide timely assistance to a customer. This insight drove us to focus on making the LLM interface and responses as intuitive and supportive as possible, by understanding the specific information staff would require related to their task.
Iteration played a vital role in our process. We started with low-fidelity mock-ups, gathered feedback, and eventually created something that could be placed in the hands of users, continually capturing insights, and applying those to improve the solution. This iterative cycle ensured that we were not only meeting the user’s needs but also adapting to new insights as they emerged.
Initially, we proposed a traditional text interface that supported natural language inputs. However, feedback showed that users wanted to reference the source document in addition to the LLM output, leading us to building a hybrid interface that allowed users to browse the source document while they interacted with the LLM.
Managing expectations and educating users
Introducing generative AI into any workflow comes with its set of challenges. One major hurdle is managing user expectations and providing adequate training given user’s comfort with the new technology. We know that while generative AI can significantly enhance efficiency, it doesn’t replace the need for human expertise and judgement. It’s about augmentation, not automation.
During user testing, we discovered that response times and accuracy were critical concerns. Users appreciated the generative AI’s ability to handle
misspellings and complex queries but were also sensitive to how long it took to generate responses. Balancing these factors required us to manage expectations and educate users on the capabilities and limitations of the technology through educational microcopy and clear loading states within the interface.
The power of prompting in generative AI design
One of the most interesting aspects of working with generative AI is the art of prompt engineering. For designers, this involves crafting prompts that guide the AI to produce useful and relevant outputs. In one concept, we were developing a solution to summarise long documents into different formats, such as X (Twitter) posts or presentation summaries. Through prompt engineering and user feedback, we were able to identify which prompts were generating the most accurate outputs, as well as uncover new prompt ideas to provide additional value to users. This fine-tuning process is akin to iterating on a wireframe to improve user experience.
Trust and transparency
Building trust in generative AI systems is paramount, especially in regulated industries. We found that users were more likely to embrace the technology when they understood how it worked and could see the provenance of the information provided. For example, displaying citations that link to the source of the AI-generated content below the LLM response helped users verify the accuracy and relevance of the information.
This transparency helped build trust and reassured users that the information was reliable.
Takeaways for designers
As experience designers, our role is to harness the potential of generative AI responsibly, always keeping the user at the centre of our solutions. Here are some explicit takeaways from my journey:
- Start with a clear problem statement. Before diving into generative AI, ensure that you have a well-defined problem that needs solving. This clarity will guide your design process and keep the project focused.
- Balance desirability, feasibility, and viability. Use the three-lens approach to ensure that your solution is desirable for users, feasible from a technical standpoint, and viable for the business. We constantly evaluated whether our generative AI solutions met user needs, could be technically implemented, and were sustainable for the business. Remember, new technology means new considerations.
- Empathise deeply with users. Spend time understanding the real-world challenges your users face. This empathy will inform your design decisions and help you create solutions that truly meet their needs, as well as address any concerns they may have.
- Iterate relentlessly. Not a new concept to designers, but gathering feedback on generative AI outputs is crucial to understand how accurate your solution is and helps inform potential guardrailing to keep users on task.
- Educate and manage expectations. Make sure your users understand what generative AI can and cannot do. Provide help text, informative UI states and guardrailing to ensure a smooth adoption process.
- Master prompt engineering. Learn the art (and science) of crafting prompts that elicit the best responses from your generative AI models. This skill provides the basis of using generative AI technology, and can significantly enhance the effectiveness of your AI solutions.
- Build trust through transparency. Ensure that users know how your generative AI solution works by providing an explanation of data sources or even document citations which allow for further discovery of the source material.. This transparency builds trust, comfort and encourages user adoption.
- Focus on augmentation, not automation. Remember that generative AI should enhance human capabilities, not replace them. Design your solutions to support and empower users in their roles.
- Prepare for technical limitations. Be ready to address technical limitations such as response times and output accuracy. Collaborate closely with your development team to refine the solution’s performance.
- Keep user experience at the forefront. Traditional user interface principles still apply. Ensure that your generative AI solutions help users achieve their goals, but also addresses any trust or transparency concerns they may have.
What’s next?
Designing generative AI experiences is both challenging and rewarding. It requires a deep understanding of user needs, continuous iteration, and clear articulation of success criteria. The fact that this space is evolving so quickly, my parting advice is to embrace the journey, stay curious, and keep experimenting. The future of design with Generative AI is just beginning, and there’s so much more to explore.