Yash Shah (00:00.352)
Okay, I think we are live. So it's time to start. Yes. Today, Koushik is like redder than usual. Very, very red. This is okay. So hello and welcome to Momentum of His Arms. My name is Yash and I'm joined by my co-founders Harsh, Jay and Koushik to discuss topic of the week, which is UX for AI led SaaS.
Our goal is to provide you with actionable insights and practical strategies that you can apply to your own products. Throughout the session, we encourage you to engage with us by asking questions and sharing your thoughts. This is a fantastic opportunity to learn from each other and give new insights that can help drive your tech initiatives forward. So let's get started. Jay, Harsh, Koushik, how are we doing today? Yeah, doing great. Great, doing perfect. Getting ready for the company offsite.
Yeah, we've been practicing some performances. Our whole company is going for an offside in this Friday. interesting. But Koushik, so tell us why is UX for AI led SaaS products different than UX for a SaaS product? So why does that need to be a separate conversation?
It is because of the recent, quite recent improvements that we have. So, two aspects here. We are seeing the amount of integrations with respect to funding in AI inside the products have been significantly increased and the demand for it also have been significantly increasing. That's one. And we are also seeing that no business is ready to wait for the AI to get perfect. So, with that in mind, everyone wants AI.
irrespective of whatever the potential, the model that they're using in is. So with that in mind, the whole aspect of how consumer friendly and how user friendly it is for using AI hasn't been defined very clearly yet. And it's something that major organizations are pouring a lot of money to do research upon like Google.
Yash Shah (02:28.334)
has started a couple of years ago itself, giving more focus into this, respectively with different organizations. Similarly, how, for example, when there is a specific UX pattern and specific rules for different, different fields like FinTech, for e-commerce, for SaaS, we have time to establish. Similarly, for AI also, with increasing demand for usage, it raises new problems.
So there is a need to define clarity within NASBA, which are the best practices, which patterns work out better. And here there's also a technical aspect. For example, let's say you're building an AI-based product. You can't expect the AI model to give you results instantly better.
We know it, we have seen, we have been building and we are seeing it quite often. Instantly you won't see better results coming in. At that point of the time, if you deal with user frustration, you need UX to come in hand over there. To make the legs better. that's the thing. So like you said a couple of things. The only thing that I had a, the only takeaway that I got was that technology is still not great and we need
UX to cover for the bad technology. So either the technology is bad or the expectations are very high. Good design to sort of bridge that gap. that be a fair? Yeah. I'm just getting there. This has always been the case with emerging tech. So it will always take some time to catch up and get there. at this...
during that transition period, how much of UX can help it to withstand it and make sure that it's easier for the users. Because for it to grow, more people should use it. More people should come in touch with it, so that they can share feedback, which will help to eventually make the tech better. So that's why it's very important for us to define UX per AI.
Yash Shah (04:48.748)
Yeah, yeah. Go ahead. problem. Yeah. So Kausi, I just want to understand. So like, if there is a SaaS product, right, then we have, we can say the one standard checklist, right? These are the things that require for a SaaS product, right? If, is there any checklist or we can say some standards that we should follow if we have AI products? But AI can be integrated or I have an AI SaaS product, right? That can be, both can be done.
What are the standards we can say or we can say the checklist that is to be need for the AI UX thing? So this can be divided into three sets. I'll give you three ways in which you can have a better way of having AI. So for example, we have automations, have augmentation, and then we have generation. Three segments I'll try to divide. So in automation, there are certain tasks within your product that you're trying to do.
who want to automate certain steps within this task using AI. So this could be, you want to probably skip, there is one, then you realize that two, three, four are reputable tasks. And these are causing a lot of friction in your product. So you want, you could automate that particular process and jump to the. So this is automation. Augmentation is basically you have certain data in your product. And then you're using AI to forecast or create an entire scenario which would
which would help them to use the product in much more valuable way. For example, your product, let's say recommendations, and for example, our one aspect of it, like AI could understand your previous usage and then recommend similar or something much better than what you were using previously within the application. Any particular aspect or the feature that's existing.
So it's trying to create a scenario based on the data of your usage or the data that is available within the product. Then we have generation. So generation is completely from scratch. That's generative AI. So you are trying to create from scratch. That's where the prompt basis comes in. In each one of these sets that I just explained, they all have their own onboarding, own set of, you know, understand,
Yash Shah (07:12.514)
making it easier for the user to use it. The aspect where the activation aspect of the product and then also the error states. To be very honest, one very important thing that I feel in UX pattern for AI that needs to be very clearly defined is how will the errors look? How will the feedback loop work out if the model doesn't respond properly?
And that if that process, it causes frustration. What happens is that the biggest plus point of having AI in your product is to build more trust and give more value. And that breaks instantly. So if the error is bad, so designing good errors and good feedback loops in the design is very important across all three of them.
these could be, you can literally divide them in these categories and then decide which one should plug in inside your product and what aspect at what point. Okay. So let's take one example. Okay. So I have two AI products. One is for like the prediction. Okay. They will use a prediction or it is a forecast. Right. And other AI product is like the generative AI, right. In which you can give the prompt and they will give the output. Right.
So we can say both AI models the chances are high to give some wrong output. So how the errors things can be differed for both AI model is respect to UX respect to UX how can we show the errors to user that according to different AI models. Is there a same thing that we should follow for all the AI models or that can be different error format or
How we can show the errors to the user for AI? Not different, it depends on product to product also, to be honest. But at the same time, I'm just saying that the error should just do the job of information delivery. It should also help them to do something else more than that. I'll give you an example. We are recently working on a... We are recently... One of the products that we have, we have created this...
Yash Shah (09:27.566)
design aspect where when someone searches for something and a set of results were supposed to show and the results doesn't show and a set of design templates or something that supposed to show that doesn't show up. We have created the error message in such a way that hey you have tried to look out for so and so it is not available with us as of now would you like to request for a design so when a design that you searching for is not available we are not just saying that it's not available
You're saying that it's not available, but you can place a request and we can actually send it to you. There's no AI aspect involved in this. I'm just talking purely from the sense of good error design. So you can think of this as a, I always give this analogy of error designs as the good librarian analogy. So when I was in college, the college had a very good library.
I used to say that know good design should be like the librarian because whenever I wanted a book and if it was not available without me telling him he used to already place the know order for the book beforehand in some other library or try to get it without me even asking for it so it's just being polite and being nice right so an error message that is nice and polite would be great that's the whole point
Great, so we talked a lot about trust factor that you know design has to build especially at the current situation where we are where know AI is a must component inside any of the SaaS products apart from this you know at times it happens that let's say before having an AI component inside a SaaS product it's already placed in such a way that all the buttons all the experience throughout all the wireframes are set in a way where
There is no more scope of adding any additional feature, but let's say now because of the need of AI component, which is to be implemented, there is no right space. So, did you share any of your examples based on the work or any such approaches where, you know, if this is the situation, what should be the approach? Because many a times I have talked with founders and they have said that this is a challenge with them. we'd to know that. The best thing is to make sure do not break the existing rules and patterns.
Yash Shah (11:52.078)
I'll give you a very good example. So for example, let's say you are in a language learning platform, like GoLingo or anything. And let's say you are trying to learn a particular phrase in a different language. It is in English and you are trying to learn it in Spanish or something like that. So if the phrase is there and below that it shows suggestions saying that, know, hey, this is...
This how this face has been spelled out in Spanish and this is a YouTube video related to that. Let's say your AI, it is still doing scraping and getting the data but still it is able to give such as a US display pattern I'm talking about and it is just showing us citations on hindsight just below the This is one. There are certain platforms that create a special tool just a special
character that just pops out of nowhere like a chatbot or something and then says that you what would you like to do? There is difference between showing these two things. This, in the second option where it specifically comes up and shows, the problem here is that it's breaking the existing pattern that users are used to. So do not do that. You rather work around the same patterns like this because people spend six
People have spent significant amount of time in other people's platforms and apps more than in your app. So they expect to see your platform to work the way they have experienced it previously. So it's better to follow the pattern. So your job is to make life simpler and not to make them think at all. they just, if a good AI, good AI additional should be in such a way that they don't even feel like, you know, it's an AI transition specific that's happening. So yeah.
So about the last piece, right? So I wholeheartedly agree that it should not feel like it was an AI piece that helped them. What I have also noticed and I don't really have an opinion on this but I'm sort of torn between these two. What I have also noticed is that the business will always come in and say that no, no, no, no, make them realize that AI did this, right? Because they want to charge.
Yash Shah (14:12.76)
for that AI, right? So as an example, like if you use AI in Notion or AI in ClickUp, there's a separate glow that appears, right? So it's not seamless. It like hammers down that, hey, you this output has been done using AI. Want to know your opinion on that, right? Where the business wants to make sure that the user or the customer realizes that this value that they got,
is because of the AI. But first principles UX is that you don't want to tell them more information that they need to know. So as a user, I want to accomplish a task, as long as that task is getting accomplished, I shouldn't get to know like what automations ran, what APIs were sent out. And it is equivalent to saying that, you know, in absence of AI, if I'm using a software,
And if I click on a click on sign up button on the next screen, it tells me that, you know, seven APIs were fired, you know, your sign up to to actually take place onto the software. And so which I don't need to know as to your from an engineering center. What's your opinion on on that? Talk about the I think so. Answering your first aspect of the question, if there is a definite need to make them understand about
what AI value is my product of. At least for now because when the clarity is it's very important for us to deliver that message initially so at least for next 5-10 years I'm expecting that to happen where it is going to be where people will preach about what literally about what AI could help them. Even with that like you have mentioned
I think the best way to deal with it is keep your messaging at the time of, so I mostly this is happening at the time of onboarding and mostly at the time of using a particular feature that has AI, right? So the affordance is like literally preaching about the whole aspect of AI that it has. So I would say talk about the benefit that is being offered rather than what is under the hood. What it is, no one cares what API model, I mean, at least when they use it.
Yash Shah (16:35.842)
What API model you used? What LLM is? I have seen platforms that used to show error messages that is completely in tech language. Where users would understand what it is. Yeah, it just says error 500. So yeah. So, that's why, rather talk about the benefits. think Notion has done a really good job with that. Like for example, when you
They talk only about what benefit it is offering. Let's say if it is something that is aiding you to summarize something or aiding you to shorten or make it, you might have seen this a lot of time in text based AI generation where to summarize or to make it sound more smarter or to make it sound more quirkier, funnier and all that is a benefit addition. So you are saying that this is the benefit that is there. you say, in fact,
Instead of putting it in text, you could make it more actionable, that is great. Instead of telling your user, make them do it. That is a better way to educate someone. In real life also, if people could approach that pattern with respect to, you know, from business point of view also,
it would be very helpful for the business also right like there's no better way than you know shouting it out loud then making them realize the value of it on their own so that would be a better way you should do it yes but but see in very we can say AI products where input is very important right like the prompt which we have tried to getting the output right
So for that it's must require to give some training or we can say onward the user in very good way so that they can use the AI thing very effectively. Right. So what do you think? Is there any effective ways we can say the best way to give training or we can say to good understanding of features how to use AI thing in a product? Right. Is there any good way to say that? Yeah.
Yash Shah (18:58.85)
So this particular problem is more significantly in generative aspect rather than in automation and augmentation. In automation and augmentation, there is a fixed set and you can understand it. In generative model, this is a huge problem. Even the most tech savvy of us are not so good with prompts because you have to literally convert your mental thought process into words and put it together, which is a separate skill set itself. And tech is demanding the skill set from
everyone on the planet. it is going to take, in fact, so the best practice is you could see platforms like Microsoft designer is one example. What they try to do is they give prompt itself as a solution. Right? So prompt solutions are one interesting ways in which currently we are solving that issue. So let's say I go into Microsoft designer, there are the entire dashboard is designed in such a way that there are just images and when I hover over an image.
It would showcase me the prompt that is associated with that image. Mid-Journey does the same. All those platforms are starting to do the same. This is very clearly to beat the mental block with respect to writing the prompt or the prompt issue that you addressed. And once the, I think what would be an additional level that would be great is that if I could edit that prompt. So if I could put, you know, the ability to change few words and, you know,
then use that prompt. then what happens is I come in, I see the prompt that is closest to my thought process. I take the prompt and I edit the prompt and it output. So if we could create a small step in between, that would be a great help for the user's mental model to work. You can also give something like a score, right? How much a prompt is similar to the actual prompt. That can be, we can say that.
that we can get on scoring type. Yeah, but for that I need to give an input for prompt first. Yeah, right. But yeah, we could write. There are many things like this. There's one very interesting, since you asked I'm trying to bring it up. Is that one very interesting problem that I noticed across all of them is that quantity versus quality of the data that's coming on. So for example, now this is something that has
Yash Shah (21:24.322)
as a company, as a product that one needs to make a decision about, how good is your data model? So with respect to that, especially in generative AI aspects, how much quality am I delivering and how much quantity should I deliver at the same time? So if your quality is less, it's a very tough decision to comment. When should I do what kind of which we are seeing, mid-journey themselves have faced this issue.
That's when they started giving full iterations initially and then you choose one and then you can iterate upon one. So this entire iterative aspect came from the conclusion itself. That you know, should I give quality? My data model is not there yet. So should I give the quality one? But from business perspective, I can't look like, you know, the one single image output that I gave looks bad. I, so that's when the quantity comes in. So.
These are UX aspects because I'm thinking from the user's perspective of what will make them happy with the product as of now. Eventually it will get better but now can my UX help to make it better? experience. So quantity versus quality is another aspect over there. Yeah, I agree. Great. Awesome. So Koushik, like we all know that AI would help in personalizing the user experience, right? In that particular aspect, what design, I mean,
Could you share a few examples on what design considerations are there that could be critical for delivering that personalization effectively? there any example? Yeah, so I think with respect to this aspect, AI, if there's one place that AI is really getting is good at and we can trust as of now, we can try build products around in a great way is personalization.
which sort of falls under the augmentation category where there is a fixed set of data models and based upon that is this trying to personalize things. So one is, in this the pattern says one is your own data or your own usage of the product upon which the personalization can happen. Then there is a other set of personalization where upon your peers who and again there are two categories in this. One is peers to if there is a community aspect inside the platform then
Yash Shah (23:48.11)
within your community, within your friends what sort of content they are consuming and based upon their personalization being suggested. Then the other aspect is you might have followed or liked a particular category or associated with a particular tag and the people who are also associated with the tag but they are not anyway in connection with their relation with you. What sort of content are they consuming or what sort of data are they consuming and that being suggested as a personalization.
which itself could be benchmarked again and you know could be used in multiple ways. Then the third aspect is more like my own data, my peers data and then the third comes with respect to from business aspect of anything that we want to highlight our show business. There also the personalization works and we are trying to as a business you're trying to sell something to a user based on their interest again.
Now again you need to know their interest and data with respect to that such that that gives you insights to upsell a particular aspect. So you would see that in Amazon where Amazon suggests Amazon made productions that are in line with my interests. So that is again happening around us.
There's one interesting example. So let's say it's a fitness app and I'm trying to, there are like five, six exercises that I try to do. Right now, and the fitness app has the data of respect to my BMI and my entire body. Then comparing these both, taking these both data, if there is a certain set of, you know, my exercise
sets that I need to do suggested by AI. That is again a combined personalization model where it's taking data from multiple points and trying to create a personalized or tailor-made set for you. So these are different kinds of, it depends on product to product but mostly around this is what it tries to play at.
Yash Shah (26:05.006)
So, what is the process of QA slash QC? I don't want to say QA because that is typically for bugs and all of that, but like, how do I know whether the... So we've largely spoken about generative AI, which is a very simple text based input and then output and so on and so forth. But in cases much like where you were talking about personalization, which is...
Not generative AI, but let's say in case of where AI is augmenting, where AI is offering some insights depending on data and stuff like that. How do I do QA of sorts, which is where how do I know that the output that was generated, how close was that output to the expectation that the user had?
Like what's the framework to do that, right? So I'll again come with those three broad sections that I tried to categorize and I'll try to address this from each section in the game's generality. In case of automation, the idea is that mapping your user's current journey and then identifying the friction points and then converting them into opportunities to be automated is much better.
One bigger mistake that lot of UX designers do with respect to automating AI is that they think every task is a chore. That's not true. Not every task is a problem and it doesn't mean that every task needs to be automated at the same time. So map, there is a precursor to this thing, especially in automation phase where while doing the UX research itself, you try to map it correctly. So here,
UX research plays a significant role. Whereas in respect to augmentation aspect, the suggestions should be trustworthy. I hear that the quantity quality aspect also comes in. Where the quality of, let's say your data model is not so good, then the quality of the suggestions are not going to be great. So then you might want to focus upon the quantity of the suggestions that is being okay, such that until your data model gets better.
Yash Shah (28:25.55)
The user doesn't feel like, know, even the ones that are shown to me. Let's think the other side. Let's say your data model is not so great and you didn't focus on the, you didn't make a decision from UX perspective that I need to show more quantity of the suggestions or the personalization items. the user would feel like even the ones that are shown to me are bad. Right. So that's another check that needs to
And especially in personalization, you also need to know about how much should you show, how much is not too overwhelming for the user. This is a very common, it is not just with AI, even in Netflix, with any sort of information delivering UX, we have seen is that how much is consumable information, how much is, and even the consumable information,
There is information that I take action upon such that you don't create the paradox of choice in UX over time. So that is another check that needs to be kept in mind with respect to the augmentation aspect as well as the personalization aspect. Generating we have spoken a lot but one is with in that the stages with respect to prompt how much better can you help people with giving the prompt because yeah you're saying. So let me
Let me, I think, clarify a little more, right? So, I will clarify this with an example. So, as an example, if I am looking at, if I am in Netflix, there are two separate experiences that I am having. One is how easy is it to navigate Netflix and how many steps do I need to go through to land on to the piece of content that I want to watch. So, that is the first.
part of the UX, the second part of the UX is the content itself. How good is the content piece in case if Netflix has originally produced? What I trying to get to with my question is that pre-AI software was only responsible for generating the output. The quality of output was completely dependent on me as the user. So if I am creating a chart or if I am creating like
Yash Shah (30:50.88)
spreadsheet or if I'm, let's say using Photoshop or using Figma, whatever software notion, right? So whatever software that I'm using, it's only responsible for generating the output, the quality of the output, I'm 100 % responsible. And so that's the UX for a SaaS product, right? Which is how easy is it for me to generate an output that I want to? With AI, I'm, one is basic UX, which is how easy is it for me to generate an output? And then
UX is also important in the quality of the output as well, right, which is as an example in Canva or in Figma if I am using some AI elements, in Photoshop if I am using AI to edit my images and let's say remove backgrounds or replace a mountain with a sky or whatever the case may be, right. So how easy is it for me to replace and what is the quality of the replacement that happened? What I am trying to ask you is in the second part of it,
Once technology has been deployed, we've established this already that there's a difference between the expectation that a user has and the output that technology can generate. What are the frameworks that a UX person needs to have to know whether the quality of the output is matching with the expectation and if it is then to what extent is it matching? And that's the piece that I'm trying to understand. What is the QAQC sort of framework around the
UX of the output that is being generated through AI.
So I think, I was trying to piece it together in my head. It's a very specific question. think one is, think understanding the data that is there itself is a problem here. Because UX designers, they don't know what data
Yash Shah (32:52.706)
They know what it is, they don't have complete idea about what data they are dealing with. And until and they don't have complete idea or they don't know what categorizations they have with respect to data whether they are dealing with it. Because once they know what categorizations there are present, they know which data to prioritize and which data not to prioritize. So that itself is a set of QA, QC that you need to do even from the designs. Because it's high time that know that
whenever you're designing AI, need to have a lot of clarity upon the entire data set that you're dealing with, categorizations within them, what goes, what priorities that you need. I'll address this with an example. Let's say Apple Music is there or your Spotify is there, right? So now there are lots of songs that is there and you're listening to a particular set of Without knowing all the categories of songs that is present.
You don't know which song to be prioritized and personalized to the individual person. So that level of understanding is necessary now for the UX designers to keep in mind. As well as for the product managers and for the entire team. Everyone should have clarity upon this otherwise it becomes very... Otherwise what happens is that it would be nothing different from it before AI.
It will still be the same pattern, same UX pattern, making sure that onboarding is great, in activation's perspective, making sure that we have spoken about this in our previous podcast also, where for better, how can you have better UX for activation rates in SaaS and all those things. Those will still be the same. The data, the idea about the data and also the clarity of the data is very important check here, which previously UX designers
didn't play a role upon and now it's very important that they have to otherwise it becomes very tough to personalize. can complete that for the rest. Got it. Awesome Koushik. That brings us to the end of the conversation that we have for today. I know that there are so many other things that we can discuss and I'm sure we have a lot more questions as well.
Yash Shah (35:17.67)
But that's the time that we have and we'll be for all the people who have been with us up until now, consider wherever you are, whether it is LinkedIn or YouTube, sharing this, consider following, subscribing, whatever is your jam. And we'll be seeing you again on next Wednesday. And thank you for joining in and see you soon. Yeah. Bye. Bye.
Bye.
.svg){kind=small}
.png)
