Episode highlights

00:01:56 – The confusing reality of 2026 

00:03:54 – What are we doing with AI? 

00:06:28 – Prompting and user experience

00:09:56 – AI is the next Internet

00:13:52 – Current state of AI 

00:17:57 – The main problem with designing AI products 

00:23:44 – How AI amplifies structural and data issues 

00:31:12 – Final advice from Vitaly  

 

About our guest

Vitaly loves beautiful content and does not give up easily. Born in Minsk, Belarus, he studied computer science and mathematics in Germany. While writing algebra proofs and preparing for software engineering at nights in the kitchen, he discovered his passion for typography, interface design and writing.

After working as a freelance designer and developer for 6 years, he co-founded Smashing Magazine back in 2006, a leading online magazine for designers and developers. His curiosity drove him from interface design to front-end to performance optimization to accessibility and back to user experience over all the years.

Vitaly is the author, co-author and editor of Smashing Books , and a curator of Smashing Conferences. He is the UX lead with the European Parliament and Smashing Magazine and front-end/UX consultant in Europe and abroad, working with large and small companies and organizations like Haufe-Lexware, Axel-Springer and others.

He also runs Measure UX and Smart Interface Design Patterns, friendly video courses on UX and design patterns, along with a live UX training for passionate UX and product designers.

 

Podcast summary

In this episode of UXR Geeks, Tina Lickova talks with Vitaly Friedman, UX lead with the European Parliament and co-founder of Smashing Magazine, about what it really means to design AI-powered products in 2026.

Rather than focusing on hype around models, automation, or whether AI will replace jobs, Vitaly explores a more urgent challenge: most AI problems are actually UX, workflow, and systems-design problems. Teams often rush to add AI features without asking whether those experiences are useful, trustworthy, or even integrated into how people already work.

The conversation explores why prompting is often weak UX, why chat interfaces shouldn’t replace proven interaction patterns like filters or structured inputs, and how AI often exposes deeper issues inside organizations-such as fragmented workflows, poor data quality, and inconsistent systems.

Vitaly also explains why companies should stop asking “Which model should we use?” and start asking better design questions: Which parts of a product should remain deterministic and predictable? Where can AI genuinely reduce friction? When is human judgment still essential?

The episode also looks at the future of AI agents, trust and transparency in AI experiences, the growing importance of design systems and content strategy, and why human decision-making remains critical even as automation becomes more capable.

What this means for UX and product teams

AI is changing how digital products are built – but this conversation makes it clear that successful AI adoption is less about adding intelligence and more about improving systems.

For UX designers

Prompting alone is not a complete interface. Designers still need to create discoverability, guidance, trust, and clarity. AI products need intentional UX, not just conversational layers.

For product teams

Instead of adding AI to every step, look at the full user journey. In some cases, the biggest opportunity is not optimization—but removing unnecessary steps entirely.

For research and UX teams

Understanding workflows, pain points, and user intent becomes even more critical. Task analysis and behavioral research should guide AI implementation, not follow it.

For organizations building AI features

AI will quickly reveal weak data structures, fragmented tools, and disconnected workflows. Teams that improve governance, consistency, and integration will create stronger AI experiences than teams focused only on models.

For the future of UX

As AI automates tactical execution, UX work becomes more strategic. Trust, transparency, system thinking, content clarity, and human judgment will likely define the next generation of product design.

Frequently asked questions

Will AI replace UX and design roles?

Vitaly’s view is more nuanced. Some tactical or repetitive tasks may become automated, but strategic design work becomes even more valuable. Areas like design systems, content design, workflow architecture, trust, accessibility, and decision-making remain deeply human responsibilities that shape whether AI products are actually useful.

Why does Vitaly Friedman say prompting is not always good UX?

Vitaly explains that prompting places too much burden on users. People often don’t know exactly what they need, struggle to express context clearly, or forget important constraints. While natural language feels intuitive, relying only on prompts can create friction. He argues that AI products should still use familiar UX patterns like structured inputs, filters, and guided flows where they reduce cognitive load.

What is the biggest mistake teams make when designing AI products?

One of the biggest mistakes is starting with the technology instead of the user problem. Teams often ask “Which AI model should we use?” too early. Vitaly argues the better approach is to first understand user goals, identify repetitive or high-friction tasks, and then decide whether AI is the right tool—and whether that part of the experience should be deterministic or probabilistic.

How can AI make organizational problems worse?

AI often amplifies problems that already exist. Poorly structured data, duplicated records, fragmented systems, and inconsistent workflows become more visible when AI relies on them. Instead of solving these issues automatically, AI can expose them and sometimes increase cleanup work unless teams improve their data and operational foundations first.

Podcast transcript

[00:00:00] Vitaly Friedman: A lot of teams go in and they plug in AI just to realize that if anything, AI just exposes the deficiencies that has been around in the company for years, but hidden behind the dust of work, and AI is really amplifying those deficiencies. A lot of designers are very confused and worried that they will be replaced and all of that.

Somebody has to clean up this mess!

[00:00:30] Tina Lickova: Welcome to UXR Geeks, where we geek out with researchers from all around the world on topics they’re passionate about. I’m your host, Tina Ličkova, a researcher and a product manager, and this podcast is brought to you by UXweek, the UX research platform for recruiting, conducting, analyzing and sharing insights all in one place.

This is UXR Geeks, and you’re listening to a conversation I had with Vitaly, who is a well-known designer, one of the co-authors and co-founders of Smashing Magazine. We were talking about AI and how to design for AI because our work and design shouldn’t be lost because of some big change, but design can actually help the adaptation of the big change in our society, in our products.

I am welcoming you in a very witty, funny conversation and the wisdom of Vitaly, who is just a very bright mind and a very bright designer. Tune in.

Hello again, Vitaly.

[00:01:42] Vitaly Friedman: Hello again, Tina. This is what, two months, three months, seven years, 22 years since the last time we spoke? How long was it now?

[00:01:51] Tina Lickova: I don’t know. I’m not counting. It’s always very refreshing, so I’m not counting.

[00:01:55] Vitaly Friedman: Yeah, it’s always a pleasure to see you, and pleasure to be exploring this wonderful world of UX and research and everything in between.

Thank you. It’s so confusing these days. Don’t you find it so confusing, Tina?

[00:02:06] Tina Lickova: It’s confusing. It’s exhausting. It, it’s super fast. Everybody wants to be even faster than we were before. Hell, I don’t know. I mean, and, and the world is crazy outside of our business as well, so…

[00:02:19] Vitaly Friedman: Yeah. But I would also say, like, I have this theory that we just discussed.

It feels like we’re all living in the world where everything is happening so fast and things are moving and things are evolving, and you look at the news, wow, things are moving. You look at AI, wow, things are moving. You look at the discussions about the workflows and the processes, things to do, things not to do, the pressure that comes in to be more productive and all those things, and it feels like a lot of people feel maybe that they have to be more productive, and they have to do more, and they have to be more active, and so as a result, we get just a lot of stuff happening at the same time.

So I think we all need a bit of a vacation or just calm down for a moment, maybe…

[00:02:59] Tina Lickova: Maybe I will start a meditation podcast after this. Just, you know, quiet for a while.

[00:03:05] Vitaly Friedman: Yeah, that makes sense. I think this could be a new groundbreaking hit of the year, so we’ll be waiting for it, you know.

[00:03:13] Tina Lickova: In the next episode, another round of quiet…

[00:03:15] Vitaly Friedman: Why not? Like scheduled quietness. Sounds about right.

[00:03:18] Tina Lickova: Yes. Great. We… Already in the first minute and we have a new great business idea. But talking about the devil, I mean, the world is crazy, AI is crazy. This is what we’re going to talk about. You brought up the topic of designing for AI products, which I am, of course, super interested because in my new role, I am exploring how to design for it, how to frame it, conceptualize it, how to research it.

So chip in what you have for the start, and then we will take it deeper and further.

[00:03:52] Vitaly Friedman: I’m just like everybody else, I guess. I’m just utterly confused. Not necessarily because AI is going to take our jobs or anything like that. This is not necessarily on the top of my head, to be perfectly honest. But then what I’m actually confused by is what we do with it.

Because when I look around and I look at how companies small and large try to implement AI in their work, very often there are three types of AI work that happens. One of them is let’s improve internal processes. For that we could do… Obviously we could use AI tools for research, we could use AI tools for prototyping, we could use AI tools for this and that and everything else, right?

But this is very internal. On the other hand, there is this notion of let’s bring in something customer facing like a conversational interface, chat, AI buttons stuff, things like that. And then what seems to be also gaining momentum as well is we need to make an AI space as a knowledge base for our internal organization to give access to our employees to work faster, better, be more productive, more accurate, and all of that.

All of those discussions, very often there is actually very little attention paid to UX. It just feels like sometimes in some conversations feels like a very technical thing. You just need to bring in that AI thing into your existing product and then you basically have all the benefits. But then, of course, when you bring an AI into your product or into your work or into anywhere, we need to have conversations.

We need to have conversations about data. We need to have conversations about the privacy and the sensitivity of that data. We need to have conversations about how do we help people articulate what they actually want better. How do we get them to actually make sense of the output better? How do we get them to refine whatever it is that they’re actually looking for better?

And all of this is on this side. That’s very confusing to me. Mm. Like, actually that notion of how people actually use that thing that we can bring in into the world of our digital products and services. I wouldn’t say it’s left on the side, it’s just not even considered.

[00:05:44] Tina Lickova: I had this idea because for me, and it’d be out of different reasons, and from the mental health side or neurodivergency side or whatever, but even the prompting thing, it’s not great UX.

It’s not necessarily bad, but if I have to write something, I have to structure my prompt to a machine thinking, and then if it’s not good, I have to re-prompt it in some, like, prompt cowboy and stuff like that. It starts there, the bad UX or the not good UX. Now we are at the point where y- I’m, like, switching and talking to AI tools, but then you have also AI agents which flows are complicated.

[00:06:27] Vitaly Friedman: Yeah. I would say in general, prompting on its own seems to be, like, almost natural, right? We … ‘Cause in the past we were using keywords, now we just ask questions. If anything, the questions we’re asking are lengthier and often miss important details that we just don’t think about because we just formulate a question, right?

And I think at this point it’s also worth noting, of course, that there is a very big difference between how we type a question versus how we speak a question. So the voice mode is very helpful and very useful as well in many scenarios. But very often we cannot just interchangeably say it’s the same. It isn’t.

Because also when we type, a lot of people are struggling with, I would say, three different type of issues. On the one hand, they might not know what exactly they want to know just yet. They know that they need to know something about something, but not exactly what in that space is relevant for them. So it’s very difficult for them to articulate what exactly they want.

But even if they know that they would want something and if they know what they want, they find it difficult to articulate that thing that they want in enough detail to engineer the right context for AI to give you the right output. And even if they think they know what they want, and even if they think that they know what they want to articulate and how to do that effectively because they’ve been taught on how to prompt better and stuff like that, people are very forgetful creatures.

And so every now and again, they write this meticulously written, designed, imagined prompt just to miss some very important things, details that really make or break that experience in the end, because they just ship something to AI, then realize like 10 seconds in, “Okay, this is not quite what I was thinking.”

And then they need to wait until AI comes back to them or stop it and then re-prompt again. Yeah. Copy-pasting the initial prompt, then editing it and doing all of that. Like with text we can do this, but with voice we cannot. Like editing with voice is a nightmare. You kind of need to really restart over and over again or kind of use commands.

And people are very inventive and imaginative, I think, as well, so they try to find ways. For example, in the very beginning, if you remember the times, Tina, maybe you remember the times before AI. We would just go to Google and type keywords. Keywords, remember that? And then we would get links, like top 10 blue links that we then would click, right?

What the times there were. And then we got into the world of AI, and we now write questions. But then in the end, what we get as output is basically the same. So if aliens ever came to the planet, they’d be very confused because we repeat the same problem, the same mistake that we did for 20 years now with AI now as well, because now when people get output, they get links to pages.

Sometimes those pages don’t exist, right? And so they get links to pages, and then they have to go to those pages and check where this information has come from. But of course, some tools are really creating this mapping. Like it would be very nice to be able to say this sentence, this statement, this phrase, this paragraph, this number, where is it coming from?

So I could click on it, right? And then I would get a reference directly to the segment of that page or that paragraph or that paper or that chart or whatever that actually where the information has come from. Some tools do that, but only few. So a lot of times when we even think that people know how to articulate things, i- it’s not really working well.

And sometimes you’ll find some remarkably strange behaviors that people have adopted because people are very good at bending the rules and tweaking things and making sure that things work for them. For example, there was a wonderful article by Norman Nielsen Group when they conducted research on how people prompt, and what they found was that it’s not uncommon, if you look at the logs, to see some very strange combinations of keywords and questions.

Something like, quotation, “Dishwasher- Two doors. Why? When? So what is this? It’s kind of a question, but it’s a really bad one, and it’s a keyword, but a bunch of keyword kind of plucked together in some way. And ironically, AI is typically very good at answering that question because it always is in answering any question that you give it.

But before answering to also, of course, tell you just how incredible that question is. But anyway, we get to the point where we have people who have to interact with that machine in some way by prompting. That prompting is not the most efficient way of ex-articulating and expressing intent. There are better ways.

Filters, sorting, buttons, things that people are used to and know for years already. Just because it’s AI, we don’t have to forget everything we learned in 20, 25 years of doing this digital stuff on the web.

[00:10:57] Tina Lickova: We’ll be right back after a short break with a commercial message from our sponsors. Hello, UXR geeks.

This is Tina speaking. You might already know that this podcast is brought to you by UXtweak, a research tool, and research tools are great until you need five of them to run a proper study. That’s why I actually use UXtweak. It handles everything from concept testing to usability checks, both moderated and unmoderated.

It also takes care of the analyzers with no messy spreadsheets, no frustrating dashboards, or endless reports. Just clear insights ready to share. So if you’re curious, go to uxtweak.com website and start for free. No credit card and no strings attached.

A year ago, I had a discussion with some friends about that visual part of internet is going to disappear. It’s, again, the same thing as we saw with the radio, with TV. They are still here. Radio is reinvented by podcasts, so I don’t think it’s going anywhere. It’s just reinventing itself, and why do we need the visual cues?

‘Cause it’s just somehow easier. Although it sounds, like, super stupid because the person can write it, of course, but that’s the effort, and the effort also inhales this thing of, like, I have to be able to adjust to the machine, which is exact the opposite of good UX, right?

[00:12:35] Vitaly Friedman: We often forget that sometimes people don’t know what they need until we show it to them, so they need to be exposed somehow to particular ways of what they can do in that product, in that system, in whatever that is.

So they might not think about creating an Excel sheet. They just don’t know if this is possible. So it’s almost like discoverability or learning. You can call it also onboarding, right? Those kind of things that people need to be exposed to at first to understand what they can do with the product. So we cannot just rely on a magical text box that says, “Ask me anything and I’ll give you the perfect answer.”

It will give you a perfect answer, but not necessarily a perfect answer for that user.

[00:13:14] Tina Lickova: Before we stop, quote unquote, complaining to maybe exploring constructively what are the ways how to design actually for AI, where do you think we actually are with AI? Because when I think about websites 30 years ago That’s the metaphor where I feel AI is now.

Like it weren’t useful things. There were some information, it was basic HTML. It was a start, and already people were amazed by it. Is it time-wise there, or do you feel that the adoption is already way bigger and we are still not there with the UX?

[00:13:52] Vitaly Friedman: No, I think when you look at the data, the adoption is actually not as high as we think it is.

There was data that around 60% of people on the internet have never experienced AI yet, at least not to a degree that might be considered fluent by any means. And so that’s a big number. But I would also say that it’s very difficult to even estimate where we are because just if you look at the iterations and the speed and evolution of how things have been evolving.

Me trying to pay attention to what’s happening because this is also a part of my work as I’m working with companies bringing AI into the products. I find it extremely difficult to even keep track because there has been a massive acceleration almost every six months or so. So it seems like almost every six to eight months, just like we know with Moore’s Law, there is, I wouldn’t say like a new groundbreaking development or anything, but there are things that really are changing where you can say that, “Okay, things are better now.”

So if you think about, let’s say, your experience with AI over the last three years, you probably will remember some things that were really an important milestones. If you maybe remember back in 2022 when you actually asked AI, “What’s the weather today in Berlin?” It wouldn’t give you an answer because it wasn’t connected to the internet, right?

Then we got to the point where we introduced RAG, then we introduced ways of how AI can then go to search on the internet and again come back to you and give you an answer. Then we got into the world of deep research, where all of a sudden you can have now AIs doing stuff for 10 or 15 minutes, like really extensive research and a summary of, I don’t know, dozens or even hundreds of academic papers.

That’s quite impressive. And then we got to the world where we started looking into agents, right? And agents being these autonomous things that can actually act in the real world and do things for us. And we can also define tasks and schedule tasks and give agents goals, and they would break down these goals and try to do something.

But in the beginning, they were very fragile, not reliable, kind of even annoyingly frustrating, to be honest. But now, if I look at 2026, we are in a pretty impressive place when it comes to agents, so these things are really evolving. And not to mention things that we usually don’t even think of. Like for example, there’s been an incredible development in generative voice.

But also if you want to speak to an AI agent in a way that it actually feels like it’s almost human or very close to being human, we are there now. When it comes to voice, tools like ElevenLabs, for example, absolutely incredible. Also, Sesame, just unbelievable how close it is, really mimicking human breathing, human mms and uhms, and all those things.

It’s ab-absolutely unbelievable. And then, of course, what we usually don’t get, because it doesn’t really get in our tech news, there has been some improvements around chemistry and around medicine, and it’s not groundbreaking. It’s not like we solved cancer, but things are moving. So all this just to say that it’s been remarkable just how fast things have been iterating and improving.

I think now with those agents around us, it might be one of those shifts where things all of a sudden really start accelerating even more. Because if agents then write code for themselves or for the tool that you’re working on, and do so more or less reliably, then that means that three, four, five months down the line, they’ll be even more reliable, hopefully, at least.

And then who knows what’s going to happen next. So the speed is unbelievable. It’s never been faster than now.

[00:17:15] Tina Lickova: And thinking back, I’m going to be super egoistic in this whole conversation.

[00:17:21] Vitaly Friedman: That’s okay. You are totally entitled to do that.

[00:17:24] Tina Lickova: I am crafting stories for AI products, which is also a very weird experience, ’cause I’m trying to look at so many user possible scenarios, figure it out then with tech people, like, how do we handle it?

Where do we handle it? And I have a first meeting with the designers that are building up the whole thing, and I feel that they feel the messiness heavily on their shoulders as well. So very easy question in the end, how are you handling messiness when you are helping your clients with AI products?

[00:17:57] Vitaly Friedman: I think that often the problem that I have is that we’re starting with not critical questions.

So like for example, very often the first question that emerges is, what model do we use? This is one of the first conversations that typically emerges. And I think that a much better question to start with would be, okay, so given the goals that we want our users to achieve, which parts of that AI feature or that AI product that we bring to the world must be deterministic, and which ones can be probabilistic?

Meaning, what part of that experience has to be a good old-fashioned software, and which part of it is allowed to be AI-ish, right? Could be conversation, could be also an agent doing something with his behalf in the background or anything like that. And of course, what you can also do is you don’t have to use large language models.

You can also use just AI, right? If you need to, let’s say, make sense of a particular data to give people slightly more accurate results and identify outliers and things like that, this is pretty much what cancer research has been doing for many decades now without large language models. So if you have, let’s say, a scan, you can look at everything that’s what happened there and then give users some sort of an output like chance or spots would require attention or anything that would be then assisting the human, right?

So we could also do this in our products. So if there are some things that require attention, we might want to identify them, maybe classify them, maybe cluster them, and do all those things which doesn’t necessarily have to do with conversations or generative AI. So which parts are we really sure should be predictable, reliable, repetitive?

If you want to do the same task, you need to get the same results. And where can we then plug in AI to make certain tasks that people are performing better, maybe more efficient, reduce the pains where they struggle? Whenever you have a lot of repetition, this is a great candidate for automation in general, with or without AI.

Now, that’s one of the questions that is really important to answer before you even go anywhere. And the other question would be, if we want to use agents in some way, because maybe we want to run some background tasks, maybe we want the users to be able to be more efficient without having to worry about the low-level tasks.

Give an agent a goal. An agent has the orchestrator. They break down that goal into sub-tasks, and then there are sub-agents doing the sub-tasks, and then you bring this all together. Good. So if we want to do this, for what kind of tasks do we want to give it to? And how do we want to make sure that it’s working within the guardrails that we’re establishing?

And the third question I would also ask is, at which point is the human judgment critical here? So that requires us to start thinking about what is the approval process if, again, AI is proactive. And also, if there are some decisions to be made, how can we then help a person in one way or another make sense of that messiness that they get?

So this is very important because ultimately, what we’re doing there is we’re designing for trust and confidence and reliability then as well. And designing for trust is really tricky because, of course, this requires transparency. This requires people to understand what’s going on. But sometimes we don’t know what’s going on, right?

And very often, we don’t know what’s going on. They also just need to be very clear about the implications of that. And I wish AI was really proactive about that. If you ask ChatGPT something today, it will give you an answer, but it will not necessarily point you to things that you might want to consider when taking decision, because ultimately, it’s up to you to make a decision.

So I would love it to be like a sparring partner to say, “If you take this decision, these are the negative and positive implications of that,” based on any data you have, based on research, based on academic papers, based on evidence from experts or anything like that. And these are the positive sides of that.

So you need to now weigh the, these risks, because otherwise, if you get just one answer, then you don’t necessarily know how to deal with it, right? And so trust also is a really critical part of it. So before we go into all the technical details, what kind of model we’re going to use, I need to understand the goals that we need to help users accomplish.

Then what is deterministic? What is non-deterministic? If something is, well, non-deterministic, which means then basically AI, then we need to think about capabilities that we need. What do we need here? Do we need clustering? Do we need classification? Do we need, I don’t know, aggregation? What do we actually need in terms of fancy AI capabilities?

It could also be just something very isolated, like face recognition or specific capabilities that you need in order to translate from one language to another. Something like that. You don’t need to load the whole OpenAI ChatGPT model into this because there are some decent tools around that you can use as well.

Some of them open weight that you can use for free. So there are so many other questions that have nothing to do with the technical selection. And this entire chain of questions then basically brings me to where do we actually add value for that user, for that customer, for that employee who’s going to end up using that?

Not to mention, like, a lot of other things, like, we need to talk about when we actually think about what is deterministic, what isn’t, and we think about capabilities, then there is a question about data. Is the data even clean? Because I think a lot of teams go in and they plug in AI just to realize that if anything, AI just exposes the deficiencies and the poor structure, poor organization that has been around in the company for years but hidden behind the dust of work that everybody’s doing.

And if anything, yeah … AI is really amplifying those deficiencies really well, and somebody will in the end have to deal with it.

[00:23:31] Tina Lickova: And amplifying data quality in the worst possible horror scenarios. Like, oh, this is how bad our data is.

[00:23:38] Vitaly Friedman: Yeah. The funny thing is that mostly it’s actually reflecting us, right?

This is how we are. Just to explain my perspective on this, I think that a lot of designers are very confused and worried that they will be replaced and all of that. Somebody has to clean up this mess, right? This organizational mess where you have a lot of data that’s totally inconsistent, has all weird, maybe even contradicting things living in different parts and islands of that organization.

And so who will clean up AI will not clean up on its own. It can identify that there are some things and outliers that kind of really not working well together, but somebody will have to go in and say, “We need to clear up.” The same goes for data and the way of how you structure it and organize, and if you have any duplications, if you have any wrong data, incomplete data, missing data, and all of that.

And then ultimately, somebody has to also design those experiences to make sure that people who actually use them will end up actually using them over and over again, and also value those experiences. Very often as this, as the time being, it feels as a very fragmented part of the whole journey that we’re considering.

What I mean by that is in a lot of organizations, you have data. That data can be very sensitive, very private. So there is a big ambition. Let’s bring in AI to make people more efficient here. But then you create this AI thing, like a knowledge base or something, and employees have access to it. The problem is that usually you’ll be very careful with giving AI access to private sensitive data because of legal compliance and all that.

So you don’t let AI touch that private sensitive data, but you let it get some data. So when people do the work though, AI should happen where the work is happening. It shouldn’t be a separate tool where they go to in order to do something and then go back to whatever tool they’re using to do something else.

Because what usually happens then, people then have two tools to choose. The old one, which they’ve been using for decades, and then the new one, which is then AI. Now, they need to go to AI to get some stuff done relatively quickly, and they can. But then the important bits that they’re really missing or may be missing will be in that old system.

So they go to both to double check things and to find things. And then they get something from that old system, they get something from that new system, and then they have to bring this together. So where do they go? And it’s sensitive data we’re talking about. Then they go to ChatGPT and say, “ChatGPT, summarize that for me, create a report,” and something like that.

That’s not a great workflow, and it’s also a big trouble because of data. So you’re just basically leaking the private data potentially into a third-party system or tool which has nothing to do with that. And a lot of tools and a lot of AI experiences are fragmented like that Literally fragmented. It’s like a thing we have here now, but you still have that other thing, and people are struggling because they have to combine both and they don’t know how.

[00:26:33] Tina Lickova: And this is exactly why productivity and, or the whole being more productive is such a big joke to me nowadays because we are actually not more productive, we are doing more intense work because of this fragmentation. Yeah.

[00:26:45] Vitaly Friedman: One thing I wanted to add to this, because I think it’s also quite important, is that when this starts happening, where do you have this fragmented experience, it really amplifies even more because you get this AI work slop.

You get something that somebody, your colleague or a colleague of your colleague or whoever, generated with ChatGPT, and then they send it to you, and you look at this, you’re like, “This is, this is not great.” So then you have to take over and do the work to clean it up. Mm-hmm. So it adds more work into your space.

You now have to clean up not only after your AI, but also after the work done by other AI, by other people. And so if you actually get those systems wrong, they’re fragmented, people don’t know how to use them well, and so on, it really, over time, compounds, so you end up with a pretty bad experience for everyone and definitely not an increase in productivity.

[00:27:37] Tina Lickova: The thing that strikes me, because I’ve been through all possible roles in this digital product business in the last 15 years, but I still consider myself a designer. Design mindset is something that you pinpointed that agents and even conversational products need, and it’s the guidance. Visual stuff can guide you on a website.

You probably know where to click because somebody thought about the flow and thought about, “Okay, how do I guide the person who wants to achieve that? These are the goals which should be achieved by the users.” We are probably thinking less of the guidance because we have this discussion, which model do we use?

And thinking of the capabilities, but it’s again about like, okay, what do we do with the capabilities when we have them? How do we show it to the users? That’s the big question.

[00:28:24] Vitaly Friedman: I think it also depends on how do we, again, expose what AI can do. Sometimes you would find companies trying to give people to flip it, and I think it’s actually a good idea.

What I mean by that is they don’t even think about what AI must be able to do for people, but instead they work the other way around. So they look at what people are doing, what kind of tasks they are working on. They try to do this task analysis or workflow analysis and understand just what is it that people do here, and understand both the goals and the needs, like the underlying needs that actually guide or create or establish those goals.

And then they try to break it down to see at which point can AI then help improve or make more efficient or even totally bypass some of those tasks that people are performing, especially the routine ones. Anything repetitive, that’s probably a good candidate for Being automated in one way or another.

And then you start thinking, okay, so which capabilities do we need from AI? So you basically orient yourself based on whatever it is that people have to do. You look at what people do, you look at where they need time, where they waste time, where they actually doing it well and relatively fast. And then I would say you do two things.

You either dial down failures and frustrations and annoyances and anything that just takes too much time and it’s very difficult maybe to get right. And on the other hand, you also try to dial up successes, whatever that is, faster, better. That also could mean higher accuracy, even although it might mean slower speed.

Often accuracy is more important than speed. So if you’re working across these two dimensions, like reducing frustrations and improving successes, then you will probably find a bunch of capabilities that AI brings to the table for your employees and also just for customers. So instead of looking at a model and say, “Where do we plug that capability in?”

Take the opposite approach. Look at where do we need to add some help, and then we add that. Just one thing to keep in mind here is that very often teams get lost in trying to improve every single step, like almost like optimizing every single step with AI. Sprinkle a bit of AI across every individual step of the journey.

Sometimes, with or without AI again, you might be better off by zooming out every now and again and looking if you can actually skip the entire bunch of steps altogether because, I don’t know, uh, somebody could just take a photo and then AI could recognize everything in it and map that structure for you in the dashboard without you going to step one, add your file, step two, add your details to it, step three, tagging or anything. AI can probably take care of all of that.

[00:31:05] Tina Lickova: Is there some advice that looking at younger folks in our business feel like, “oh, this has to be said”?

[00:31:12] Vitaly Friedman: This is actually quite a difficult question because I think it’s really quite difficult to see what’s going to happen next, especially because a lot of tasks that are usually done by junior designers can or will be automated for sure.

I think if anything, I would highly recommend to go in and explore things that are probably not going to go away. Think of everything that we are talking about today and in general, if you want, let’s say, a company to prototype better or faster in some way, there must be a good design system in place.

There must be a good design system where you have everything established, the documentation, the accessibility guidelines, the rules. Of course, it could be the one universal design system. It’s not like every company has to have w- its own, right? But you probably will be dealing in, especially when you want to get good prototypes, in a way to communicate clearly to AI what it needs to generate, but also to have existing components, flows, examples, patterns, voice and tone guidelines, accessibility guidelines, and all those things established in the system.

So I don’t think that design system is going to disappear at all. If anything, it’s probably going to be amplifying the need to really standardize and compartmentalize what we have in our product. So that’s one side. On the other side, what I also think, especially when we’re looking at the world of agents, I don’t know if you remember, maybe some of you who are listening to this will remember, we had this phase maybe one, two years ago or so, where we had systems that could actually go to websites and then mimic humans’ behavior by clicking on things, opening menus, and stuff like that.

And it seems like we’re moving away from this now. It seems like when we’re thinking about how agents would be communicating rather with websites or web applications, it will not be happening on the level of UI, most likely. It’ll probably be happening on the level of those APIs or frankly, markdown files.

There is a wonderful example from Atlassian. Atlassian has set up a design system And they had one for a long time now, and they actually have a view now for large language models to consume it. They basically translated their design system into a bunch of markdown files, which includes everything that AI should know about the design system, the tokens, the components, the guidelines, and all of that.

And of course, that sort of markdown files is not visual, it’s just text. Right. But what it gives them is almost like a skills file that gives AI everything it needs to know about the system in a very condensed, compressed, and LLM-friendly way. Because if it needs to go on the website and go through this mess of divs and spans and HTML, there is a chance it’s going to get noise as a result of that as well.

But if it’s really, like, a bunch of, very small bunch of files that you give to AI, it’s also much more efficient, right? Because in the end, you’re just burning less tokens as well. And so the question then is, and this is what’s interesting to me, and maybe I’m wrong, and I’ll be happy to be proven otherwise.

But it seems like if we are communicating to agents, we are communicating with, guess what, Tina? Words, because it’s markdown file. But if we’re communicating with words, we need to have somebody who is really good at communicating with words. Who is that, Tina, on most teams?

[00:34:29] Tina Lickova: Humans.

[00:34:30] Vitaly Friedman: Humans are very good with that indeed.

If you actually ask AI agents to communicate with other agents using text, they would be like, “Why would you even use human text? That’s so inefficient. That’s so strange.” Yeah, gotcha. But as we are communicating to AI, I think that the role of writers, I mean, very broadly speaking, content designers, UX writers and all, I don’t see it going away at all.

Because somebody has to articulate very clearly and very well different situations, different examples, different conditions and all of that. And yes, of course, AI can generate it for itself based on the design system that you have, but you probably still want to be the guiding force behind what kind of comes on the other side when somebody’s prompting something.

So you need to be shaping it in some way. And for example, a lot of roles, design-related roles, content-related roles, yeah, they will be probably transformed in one way or another, but I don’t think that they’re going to disappear altogether just because we have now AI. It’s just different work because it’s probably going to be much more strategic and much less tactical.

[00:35:32] Tina Lickova: Maybe just last thought from my side, because I have the hunch, looking at how we work in the last half a year, that communication skills and really the soft vibes, and I don’t mean it in the vibe coding type of term. But being able to read the soft vibes and being able to react on it, to name them, to make them clear, people will have to learn even more because the communication will become more important.

It’s not just the words written down with writers, but we have to be more clear and trying also to understand each other- More because the whole communication and everything in between

[00:36:13] Vitaly Friedman: In the end, I don’t know if a lot of people would feel comfortable leaving the fate of their company in the hands of just tons of agents that make decisions so fast that they cannot even follow.

I think that ultimately there will be a very important layer, which is the human layer, which is actually the most decisive layer of them all. Because ultimately, yes, you can automate a lot of stuff. You can be a bit more efficient here and there. You can also create prototypes, like, in 30 seconds. But ultimately you need to know where you’re going.

And so I don’t see people really being willing to give away the sense of navigation, compass, orientation, strategy, all the way to AI to say, “Eh, you know best.” I don’t think we’re there yet. It might come this time, but I don’t think it’s going to come soon, if at all.

[00:37:00] Tina Lickova: Vitaly, thank you for a fascinating, witty, and funny conversation. It was a pleasure.

[00:37:05] Vitaly Friedman: Thank you so much, Tina. Thank you so much for having me.

[00:37:07] Tina Lickova: Thank you for listening to UXR Geeks. If you enjoyed this episode, please follow our podcast and share it with your friends and colleagues. Your support is really what keeps us going. 

 

If you have any tips on fantastic speakers from across the globe, feedback, or any questions, we would love to hear from you, so reach out to geekspodcast@uxtweak.com.

Special thanks goes to my colleagues, to our podcast producer, Ekaterina Novikova, our social media specialist, Daria Krasovskaya, and our audio specialist, Melissa Danisova.

And to all of you, thank you for tuning in.

💡 This podcast was brought to you by UXtweak, an all-in-one UX research tool.