UX & AI: Successfully Designing Human-Centered (Agentic) AI Products

What is UX and why is it Important in AI Product Development?

Successful user adoption of an AI-powered product starts long before the product is actually used, and this is where user experience (UX) comes into play. UX encompasses the entire user journey, from familiarizing oneself with the product (product onboarding), to the usability during use, and finally, post-use interactions (such as recommendations). Product onboarding addresses questions like: How easy was it to acquire and set up the product? How simple was it to follow the instructions? Usability, on the other hand, measures how efficiently, effectively, and enjoyably we can use the product. When users are satisfied with their experience, they are more likely to recommend the product to others.

User experience is not a new concept. In fact, for software solutions, UX has been around since the 1950s, when human-machine interaction first emerged at the intersection of computer engineering and psychology. Every solution begins with a problem or challenge that we seek to address, either through a non-AI or AI-powered approach. When it comes to user experience for software products and solutions, the technology itself becomes secondary. Users don't care whether AI or agentic AI is integrated; they simply want solutions to the challenges they face. While AI and agentic AI certainly enhance our toolkit for solving these problems, users primarily engage with the product itself. Whether a product is successful can ultimately be measured by one key question: “Is this product adopted by users?”

In our projects, we’ve observed that, alongside many successful products, a significant number fail. The successful ones that make it past the proof-of-concept (PoC) or minimal viable product (MVP) stages and become widely adopted AI-driven products typically went through a phase where user experience was the primary focus. While these products may have had brilliant technological solutions initially, the breakthrough in adoption came when the user was placed at the center of the process before achieving global success.

Ideally, those AI-powered products offer benefits to them and are user-friendly. If our aim is to create successful AI products that are finally widely adopted by users, then our primary goal should be: creating human-centered products that are beneficial to the user.

How to create human-centered products?

Creating human-centered products considers the UX along the user journey and puts the human user in focus from the beginning of the project. One popular method for crafting the usability is called user-centered design (ISO 9241).

1. User-Centered Design: Starts in the product development phase by understanding the user. With the help of interviews, workshops or user research, user needs are identified and understood.

2. Derive Requirements: From this user understanding, requirements for software features and functionalities are derived. Those requirements translate user needs into technical features.

3. Create Prototypes: The requirements are visually mapped to a prototype, wireframe or design proposal. Those prototypes offer the benefit of sparking imagination. Both users, business owners and developers can see and understand the product vision and provide feedback. The prototype can come in various majority grades (fidelity): from a pen and paper-sketch to a demonstrator in the target system. User-centered design often prefers lower fidelity to gather early and honest feedback from users.

4. Collect User Feedback: Direct user feedback is essential for user-centered designs to evaluate the prototypes and iteratively adapt both requirements and prototypes. This feedback can also be incorporated into an AI-algorithm so that the artificial intelligence is adapted to and based on the user feedback. A well-known example is personalization, e.g. when listening to music. By providing feedback to our music favour, the algorithm can adapt to more often proposing music matching our taste.

We’ve been applying user-centered design for a long time in our BI projects, especially when developing dashboards. Feedback from our clients highlights how this approach sets us apart from many internal GenAI initiatives. By incorporating user-centered design into the development of agentic AI products, we ensure that we create solutions that truly address the needs of the end-user, ultimately leading to products that resonate with users.

What’s new with AI for UX?

UX remains highly relevant for AI products, as users seek solutions regardless of the underlying technology. They want AI products they have a positive user interaction with. However, there are some new aspects which are more relevant when using AI.

With AI-powered products, we've introduced a level of artificial intelligence by leveraging available data, bringing a new dimension to user experience. This intelligence enables various levels of autonomy for AI-powered products, ranging from fully automated AI solutions (complete autonomy) to augmented AI solutions. While automation replaces human tasks, augmentation enhances human work with AI, creating a collaborative dynamic between humans and AI. This partnership introduces new human-AI principles of trust and control. A key aspect of this revolution in user experience is the role of natural language, which enables more intuitive, human-like interactions with AI. By allowing users to communicate with AI in their natural language, the technology becomes even more accessible and aligned with how people interact, further enhancing the overall UX.

These various autonomy levels require significant human trust and confidence into the decisions made by the AI system. If the fully autonomous AI product decides on its own, humans need to trust its decisions. Therefore, we need to create trustworthy solutions. Additionally, AI-powered solutions are built on probability and uncertainty which needs to be explained to the user, e.g. if I press a generate button in a Chat Bot the answer to the same question might differ due to the probabilistic nature of large language models. Designing trustworthy systems is a new chapter on its own. It covers providing control to the user, making AI explainable, avoiding biases, protecting against potential harms and misuse and providing the right level of transparency of the user.

Depending on the level of trust, humans need certain control over the system. An AI system ideally is a learning system, adapting to the user feedback for a higher degree of personalization or accuracy. The user needs to understand that providing feedback is essential for better outcomes. Additionally, product designers need to understand where users need having control over an automated or augmented solution. Common tasks for keeping human control are enjoyable tasks, responsible tasks, critical situations with high influence on safety, health or financial risks or visionary tasks [7]. AI products need to ensure the right balance to user control in autonomous and augmented systems.

What’s new with Agentic AI for UX?

The same applies to more specific Agentic AI solutions that leverage generative AI technology, where our goal is to create artifacts (such as writing, generating images, sound, or videos) or facilitate collaboration between agents. With agentic AI, the notion of human-AI collaboration is becoming increasingly prominent, as agents can assist humans in handling more complex workflows. By utilizing natural language, we can design agents and assign them roles through carefully crafted prompts.

UX for generative AI proposes the unique challenge of generative variability. With generative AI we create artifacts – but those artifacts might differ due to probability resulting in a higher variability.

We summarized this in 6 human-AI principles for building human-centered agentic AI applications.

• User Experience: AI projects are heavily technology driven [8]. Hence, it is even more crucial to design agentic AI products that truly address the real needs of users. The key to success is creating human-centered products through a user-centered design approach. In our client projects where we developed agentic products, our experience showed that our approach stood out from other GenAI projects. We remained committed to the user-centered design process to meet actual user needs. Remember: users care about the products they use, not the technology behind them.
• User Control: In Agentic AI, agents assist with human work through human-AI collaboration. Since they can either automate or augment tasks, it's crucial to identify areas where humans should retain control and where they can entrust control to AI (see section “What’s new with AI for UX?”). This control has been established under the term human-in-the-loop. The research about UX for Agentic AI is still an ongoing field. However, to keep the user in the loop, the AI system should make the data it uses to improve its services (e.g. chat history) adjustable and controllable by the user. Imagine your agentic system learned that you work in a huge company as a team lead and now personalizes all answers to your position being a team lead. In case you got promoted into a higher management position, it would be important to you that 1) you are aware and have the transparency about this system assumption (being a team lead) and 2) that you can adapt the information to your new role. This leads us to another principle, the need for transparency.
• Transparency: Agentic AI is a new technology that we, as humans, need to learn how to use effectively. Users must grasp how to leverage this technology to get the best results (e.g., writing effective prompts). How can we assist in this? By guiding the user on how to use the system efficiently, such as offering an onboarding tour or a quick start guide for better comprehension. It's important that users understand the AI's output is based on probabilities and may vary. AI can also make mistakes, and these should be made transparent to the user, allowing them to correct them. This could involve visualizing uncertainty, giving the user control to decide when the agent is unsure about its response. Transparency is key to building trust.
• Collaboration: The previous example illustrates how AI and humans are increasingly collaborating in Agentic AI. By designing for co-creation, the strengths of both AI and humans are leveraged. This could involve teaching humans how to craft effective outcome specifications [5]. With Gen AI, a new human-AI interaction paradigm has emerged, where the human expresses intent-based outcomes in natural language. Teaching humans to do this effectively enables clearer communication with AI agents, leading to more effective results. At the same time, we should bridge the gap between the worlds of UX and AI by providing both domain-specific and generic input UX elements to the user (e.g., use case-relevant buttons), allowing for adaptation and supporting co-editing between humans and AI.
• Trust building: “A successful team of people is built on trust, so is a team of people and AI” [2]. When collaborating with Agentic AI, we need to trust the decisions made by the AI, such as by providing sources for the knowledge it uses. This requires the right level of transparency. It's important to explain how the system works: users need to understand the AI's capabilities and limitations to better comprehend its usage and output. By assigning the AI product a clear role, such as a tool, assistant, or copilot, within the collaborative human-AI team, it becomes easier for humans to understand and trust the AI [5].
• Ethical design: At the same time, we need to craft products that identify relevant user harms and provide mechanisms to mitigate them. One common example are hallucinations in GenAI which might be a danger for fake news. One solution is to minimize risks for those dangers or making them transparent to the user to that he/ she again has the control and permissions to correct them.

Our experiences with designing human-centered AI products

The following is a compilation of insights from our customer projects:

• Create human-centered products: Many agentic AI projects are driven by technology, but users expect carefully designed solutions that meet their needs to ensure adoption. We believe that with a user-centered design approach we can combine both technology and user acceptance.
• Make user information of AI-created results adaptable for better user control: We need new integrations that bridge simple prompt-based chatbot windows with user interfaces, combining natural language with user control. As a result, UX must develop additional solutions for UIs that account for human-AI collaboration in the future.
• Design ethical AI products: Using user feedback and monitoring to understand relevant user harms and protect users against them is important for building trust through competence and benevolence.
• Incorporate feedback for continuous product improvement: Only through regular user testing can we assess how successful our product design is and refine it further to better meet user needs.
• Focus on user benefits instead of technology: Users aim for well-working products. Communicate in an understandable way both limitations but further the benefits of an AI-product.

Conclusion

In summary, to create AI products that users actually adopt, it's highly recommended to invest sufficient time and resources into carefully designing the user experience. By taking a user-centered design approach from the start, we can save both time and money by getting it right from the get-go. One of our customer examples is a copilot that supports technical document creation, which we developed with a human-centered approach. How did it help? By saving development time and money, prioritizing the right features, and staying connected with the end users.

Remember our earlier example of grandparents and their smartphones? Let’s create Agentic AI products tailored to a wide range of people—like a copilot for grandparents to help them stay connected with family and friends.

Struggling to create solutions that resonate with your users? We can help. Contact us today.

Sources

• [1] Stracqualursi, Luisa, and Patrizia Agati. "Twitter users perceptions of AI-based e-learning technologies." Scientific Reports 14.1 (2024): 5927.
• [2] Kore, Akshay. Designing Human Centric AI Experiences. Berkeley: Apress, 2022.
• [3] Lew, Gavin, and Robert M. Schumacher. AI and UX: Why artificial intelligence needs user experience. USA: Apress, 2020.
• [4] Shneiderman, B. (2020). Human-Centered Artificial Intelligence: Reliable, Safe & Trustworthy. arxiv.org/abs/2002.04087v1
• [5] Weisz, J. D., He, J., Muller, M., Hoefer, G., Miles, R., & Geyer, W. (2024, May 11). Design Principles for Generative AI Applications. Conference on Human Factors in Computing Systems - Proceedings. https://doi.org/10.1145/3613904.3642466
• [6] https://www.forrester.com/report/the-state-of-generative-ai-2024/RES180458
• [7] https://pair.withgoogle.com/chapter/user-needs/#section1