AI + UX: design for intelligent interfaces
How design shapes, elevates, and (let’s be honest) saves AI
In just a few years, we’ve gone from laughing at it to fearing it. By now most of us have come to terms with the fact that, like it or not, it is here to stay. Designers in tech face a unique set of challenges in this evolving landscape. As design practitioners, we must be adaptable- to the needs of stakeholders and clients, to our teams and professional partners, and to rapid changes in the technological landscape in which we work. In this article I’ll be looking at how UI/UX and Product Designers can create better AI-enabled Products, the design strategies to consider when building intelligent interfaces, and how design-in-tech might prepare us for the future of this technology.
AI-enabled products? What are those?
By “AI-driven interfaces” and/or “AI-enabled products”, I refer to mobile and web apps that employ large language models (LLMs) or other ML techniques as their primary technology. “Intelligent Interfaces” refer to the UI of these products.
In the last couple of years “.ai” is everywhere and it’s not unexpected, given that LLMs are getting more efficacious and reliable with each passing week. AI-enabled products can serve users for productivity, design, sales, marketing, business development, and probably many more use cases. I don’t know them all, but maybe you do! If not, just ask your favourite AI.
Something that underpins all of these applications is data. Data is the lifeblood of AI, it is what allows the product in question to function effectively. In turn, AI-led tech products continuously amplify the value of data by unlocking insights. This makes data analytics not just a category in the list above, but also a fundamental component of AI-powered tools. [Edit after DeepSeek R1: AI now also amplifies the value of data by training itself!]
Where does UX get on (and off)?
AI without good UX is like a Ferrari on display with the steering wheel removed– Impressive, no doubt. A fine specimen. But sit in there and you’ll feel a mix of disappointment, confusion and anger.
AI systems are complex by default. All the fancy neural networks, machine learning models, and data pipelines are impressive under the hood, but quite intimidating to the average person even on the surface. A well-designed interface abstracts the stupendous complexity that is AI, presenting users with more familiar and intuitive options for input and interaction. UX anticipates and accommodates the need for accessibility, usability and empathy in these otherwise “cold” systems. Data and code make AI smart and clever, but it’s UX that makes you actually want to use it. It’s what turns algorithms into tangible tools, creepy predictions into comforting conveniences, and raw computing power into refined experiences. Without UX, AI is just a jumble of intimidating potential; with it, AI becomes a partner, effortlessly upgrading the way we work and live. Think about the next time you see “Your New Productivity Co-Pilot”, or hopefully one with a more original name.
Designers, pay attention
Best practices! They make the design world go round. They keep us designers on our toes and make us mind our Ps and Qs. If you’re a designer of any of the varieties mentioned earlier: Welcome, I shake you warmly by the hand. These are the things you’ll want to stay on top of to sharpen your Intelligent Interface, make users love using your product, and don’t get spooked and/or confused as soon as the AI does something unexpected. If you are a different variety of tech person: You’re equally welcome to continue reading, to the benefit of all the designers you may work with.
1. XAI for days
The average user is going to be pretty skeptical, disoriented even, when they first start using your product. We want to make sure our users trust the product and its capabilities, feel confident using it, understand how it works (to a reasonable extent), and are able to interpret the output effectively. We don’t want users to necessarily need a tutorial or crash course to use the product (unless it’s really that specialised). We also don’t want them to have to verify the results through Google, ChatGPT, or another “better” tool. Following design principles of explainable AI (XAI) is how we get there.
Explanation
AKA “So creepy, how did it know that?!”
The AI should be able to explain how it arrived at its output. The explanation should be accurate, comprehensive and easy to understand (Phillips et al., 2021). While creepiness isn’t as much of a concern today as it was 2 years ago, explanation is still beneficial in creating confidence in the product, the brand and the technology. Providing appropriate evidence and reasoning for output helps users trust the efficacy of the tool without worrying about more technically dense aspects of what it’s doing and how it’s doing it.
Interpretation
AKA “That’s cool, but WTF does it mean?!”
The answers and explanations provided should be easy to interpret, understand and act upon. Your user may be an expert in the given subject matter, but in most cases it’s best to assume that the user is not an expert and provide output that anyone can understand and move forward with.
Knowledge limits
AKA “Turns out it AIn’t all that smart (haha geddit?)”
When the AI reaches the edges of its domain of knowledge, it’s best to communicate it effectively to the user. Follow-up questions are welcome. Don’t be overconfident in your system, and don’t encourage users to be overconfident either. Instead, help them to quickly understand what it can and cannot do.
Fairness and bias prevention
AKA “I think it thinks I’m a 35 y/o straight white man”
AI is only as good as the data it’s trained on, and the data is often a hot mess, with biases and hallucinations just waiting to happen. Bias prevention and other ethical guidelines need to be baked into the system, taking the form of self checks and effective communication on potential ethical issues within the responses of your AI. It’s easy for these biases to get politicised, which is likely to be irrelevant and antithetical to the purpose of your product.
2. Error handling and management
Errors– whether from the system or the user– should be met with clear, helpful messages that guide users forward without friction. Your product likely won’t have perfect outputs every time, and your users should feel comfortable with learning through failure without their workflow being adversely affected. When the errors come creeping, it’s only polite to be accountable for them, communicate the error in a way that’s easy to understand, and provide options for resolution.
Error messaging may take the form of basic error indicators, tooltips with an action button, and/or manual control panels as a layer of redundancy (Cornianu, 2025). Subtle, well-placed animations can help bevel the experience. General UX best practices for error messages are important here — the error communication should be visible, proportional to the scale of the error, and follow accessibility guidelines. Effective UX writing also shines through. We don’t want users to feel like they did something wrong (even if they have), or experience confusion about how to resolve an error. Well-designed error management protocols help users trust the system, build confidence as they use the product, and reduce overall friction in the experience. In the words of Roald Dahl, “smooth and oily as hair cream” (The Hitchhiker, 1977).
3. Usability, prioritised
Don’t make your users think. You can achieve that by paying attention to usability, a cornerstone of the UI/UX practice. With AI-enabled products, we need to think about usability from both the engineering and the design side of things. An AI-enabled product is essentially a smart partner that proactively anticipates the needs of users and provides just the right amount of information and context where needed — that’s the engineering bit. The role of designers is to ensure that the system as a whole is intuitive, efficient, satisfying, and easy to use. Interfaces must work to bridge the gap between fickle minded humans and cold, data-driven systems. The gold standard can be achieved when designers work closely and iteratively with engineers, paying attention to visual consistency, adhering to accessibility guidelines, writing effective UX copy, and even providing neat third party integrations that can supercharge your users’ workflow.
4. Feedback and inputs
So far, proactively getting feedback and inputs from users allowed models to improve and adapt, gain credibility, and gave the user a sense of agency in the interaction (Chaturvedi, n.d.). Only a year ago, most AI-enabled products had a now obsolete mode of user feedback baked into their system in the form of ratings or a thumbs up/down button towards each answer. These products had largely been playing around in uncharted territory, but that territory is quickly becoming more charted. Very soon, models can be expected to move toward more self-sufficient operation, as demonstrated by DeepSeek’s reinforcement learning and built-in Chain-of-Thought capabilities. This doesn’t mean that user feedback and inputs are no longer needed. Far from it, they will continue to have an important place in interaction design for AI-enabled products for the foreseeable future. However, they will need to diversify and take different forms as we move toward more intuitive formats of interaction with AI enabled tools. This is something both designers and engineers will need to stay on top of.
5. Iteration and collaboration
Designing for AI-enabled products differs from designing websites and apps in a pretty big, fundamental way–the outcomes and user journeys are uncertain. Unlike traditional design, where every user flow can be meticulously planned if designers so wished, AI introduces a wildcard element that makes it impossible to map out every single scenario. Nobody can predict, list, or account for every potential type of output the model may give us. In fool proofing this situation, Designers and Engineers must closely collaborate to guide AI’s potential outputs, keep things aligned with what users actually need, and provide output in ways the user expects it to. We will need to learn each others’ disciplinary principles, needs and, to some extent, skills. Siloed operation is fairly detrimental to effective product design and development today. For the AI-enabled products we will be creating in the near future, it could spell failure.
Where are we headed with this?
Just as I was wrapping up this article, DeepSeek R1 was launched. In the span of about a week, the tech, AI and ML industries may have entered a new era– time will tell, but I suspect we won’t have long to wait. I’ve written about Designers working with AI in the past, in an article that aged well but also became obsolete in barely a year. I looked back at my draft after the launch of DeepSeek R1 and realised that in the few weeks since I started writing this article, it already sounded a little outdated. Depending on when you’re here, it may sound even more outdated.
“Vibe coding” is the new trend that’s cropped up in product design (Holmes, 2025). When I first heard about it, it sounded a lot like something that people from other disciplines can get their fingers stuck straight into product design and promote learning. My view is that product designers will need to pay attention to the use cases of user experience, now more than ever. In the short term we might see torrents of new AI-enabled products that come into existence in a “throw stuff at the wall and see what sticks” manner; user experience learning curves will be steep here. In the medium to long term, it provides insights on how we could refocus and refine the way we design AI-enabled products in a new age of tech.
A new layer of “multimodal”
Multimodal AI serves to allow AI-enabled products to communicate with users in more natural ways. Usually, this means you can type in a text input, record a voice message or video, or upload media files. Theoretically speaking, this enabled models to process all these forms of data in a way that seems more “human-like”. My personal view is that engineers and researchers cannot get complacent after developments they think will bring improvement to how easy a product is to use, when we just don’t know for sure that it will in fact bring said improvement. Processing data in a “human-like” manner may not ensure better usability of a product that already runs on non-human intelligence, is relatively unknown in its workings, and limited in its scope of interaction. Usability principles ask that the “multimodality” of these models extend beyond input possibilities, to also include output and interaction possibilities. There is a lot of scope for research in this area to understand what actually constitutes an intuitive AI interface, when AI is by default not an intuitive thing.
What I would like to see in the near future is a more visual and graphical mode of interaction that encourages more back-and-forth exchange with models before arriving at a comprehensive output that is self-initiated and self-evaluated by the model. Maybe it’s just me, but plain old chat-based interaction with an unknown entity leaves something wanting. It feels like a stopgap MSN Messenger-esque format that’s serving this purpose until we can figure out a better, more user-friendly way to get the best out of these powerful tools. And it’s time to start figuring that out.
The Third Paradigm
We’ve been in the early stages of AI-enabled products, and the people who have been making them are usually not UX specialists. Jakob Nielsen speaks of a new paradigm of UI design – Intent-based Outcome Specification– where we specify the outcome we want rather than specifying what we want the computer to do. In doing so, we hand over all control in the execution of the intended outcome to the computer, creating a somewhat one-sided “interaction” that is handicapped by the capabilities of the technology. So far, the most commonly employed interaction style has been the chat interface, since it helped account for the usability issues posed by these products. And so, the Prompt Engineers were born to paper over the cracks and “tickle ChatGPT in the right spot so it coughs up the right results” (Nielsen, 2024).
As AI continues to improve (itself?), we can expect a shift towards more complex non-command interactions — where systems don’t wait for you to click a button but anticipate your needs based on passive cues. We are right at the point where user research can step in, digging into existing user data to fine-tune these interactions, making AI-enabled products not just smarter but genuinely more usable.
Non-command interfaces mark a fundamental shift from the old-school “I click, you respond” paradigm to systems that sense and react to user behavior without explicit commands. Unlike traditional menus or GUIs that demand direct input, non-command interfaces are designed to quietly observe and adapt, blending seamlessly into the user’s natural flow (Nielsen, 2019). However, there’s still a lot to learn from the classic graphical user interface (GUI) playbook.
By combining the intent-based nature of AI interactions with the intuitive familiarity of GUIs (and some technical wizardry that I’m not qualified to elucidate), we can create systems that balance passive observation with user control. Imagine an interface that knows what you might want but still gives you the option to steer the ship when needed. It’s about merging the best of both worlds — AI that feels like it’s reading your mind but still respects your agency. I believe that a fusion of non-command design and GUI elements might be the key to pushing Intelligent Interfaces into their next evolutionary phase.
References
Chaturvedi, R. (n.d.). Design human-centered AI interfaces. Reforge. https://www.reforge.com/guides/design-human-centered-ai-interfaces#recap
Cornianu, B. (2025, January 30). Training AI with pure reinforcement Learning: Insights from DeepSeek R1. Victory Square Partners. https://victorysquarepartners.com/training-ai-with-pure-reinforcement-learning-insights-from-deepseek-r1/
Holmes, R. (2025, March 6). Rich Holmes on substack. Substack. Retrieved March 10, 2025, from https://substack.com/@richholmes/note/c-98472653
Kashyap, R. (2023, April). I Don’t Know A Great Answer OR I Don’t Know, A Great Answer. Qeios. https://doi.org/10.32388/DFE2XG.2 Figure 4
Lipiński, D. (2024, August 13). Designing AI interfaces: Challenges, trends & future prospects | Pragmatic Coders. Pragmatic Coders. https://www.pragmaticcoders.com/blog/designing-ai-interfaces-challenges-trends-and-future-prospects
Moon, M. (2024, September 27). Google launches Gemini’s contextual smart replies in Gmail. Engadget. https://www.engadget.com/ai/google-launches-geminis-contextual-smart-replies-in-gmail-140021232.html Thumbnail Image
Nielsen, J. (2019, May 4). Noncommand user interfaces. Nielsen Norman Group. https://www.nngroup.com/articles/noncommand/
Nielsen, J. (2024, October 24). AI: First new UI paradigm in 60 years. Nielsen Norman Group. https://www.nngroup.com/articles/ai-paradigm/
Phillips, P. J., Hahn, C. A., Fontana, P. C., Yates, A. N., Greene, K., Broniatowski, D. A., & Przybocki, M. A. (2021). Four principles of explainable artificial intelligence. https://doi.org/10.6028/nist.ir.8312
Sunwall, E., & Neusesser, T. (2024, August 2). Error-Message guidelines. Nielsen Norman Group. https://www.nngroup.com/articles/error-message-guidelines/ Clear change password.
