From Lip-Service AI to Architecting AI-Empowered Experiences: A Customer-Centric UX Framework

Muriel Naim at UXDX USA. Video: https://youtu.be/3-qD9C-1Vuc

Readable transcript: edited from the recording's captions for readability (fillers and false starts removed, punctuation and section headings added). Wording is the speaker's own. Timestamps are positions in the video. Names marked [?] could not be verified against the audio.

Knowing just enough to innovate

[00:00:12] Hi everyone. Can you hear me? Good. Cool. All right, where do I stand so you can see stuff? Here. I'm going to start really quickly. You got a great introduction here, so thank you, friend.

[00:00:25] The one thing I'll mention about my career is not that I was leading cross-functional teams with engineering and back-end and front-end at Disney and omnichannel in Walmart and all that. What I do want to mention is one of my very first gigs as a sole designer in Sisense, which is a big data analysis company. I was tasked by the CTO, who is this crazy brilliant guy in the best possible way. I came in and he's like, "Hey, do you have a month to try something out?" And I was like, "Well, you're the boss, you tell me."

[00:00:52] He's like, "You have a month to create a mechanism so that anyone, even your mom, can make a big data analysis dashboard in less than a minute. Now it takes 20 minutes to an hour. Good luck." I was like, "Okay." Those are data experts. They have doctorates and stuff. I'm not going to learn data analysis now in one month.

[00:01:16] But it was a great lesson in understanding that I need to know just enough to be able to innovate properly. Just enough, not too much, and it became my mantra for my entire career in UX. I quickly became the head of UX in this company, but that's not the point. The point is he taught me this really big lesson. The same thing goes for AI, and I want each and every one of you in the room that is not 100% sure to hear this. If you're OCD like me, you need to know everything to be feeling confident or comfortable about something. You don't need to know everything in order to speak and innovate in AI. You just need to know enough.

[00:01:46] Today I'm going to talk about the misconception of what AI has been the last few years, what we should framework to, and a very pragmatic flow of what you should do to have a bulletproof way to develop and come up with AI ideas for your UX features. You might even want to come up yourself with an AI framework that matches your UX feature design. I hope you can get this from today. And if not, catch me later and I'll make sure you do. All right, let's move on. I think I have the clicker. It's working.

AI is not a means to an end

[00:02:18] Anyone face that ask from a boss, board member, peer in the last three years? How many times? If more than one time, raise another hand. I faced it like seven times. What I mean by that is, "Hey, we need AI. Can you do AI?" For what? It's like saying, "Hey, we need the internet." "Yes." "Can you do internet? Can you do internet real good?" "Yeah." But what exactly do you want to do with the AI? Oh, I'm sorry. I'm being nice to AI for a second before I'm getting really mad.

[00:02:55] All of us face that, I would argue, in some way, shape, or form in the last four years. There's a big bang happening. Big bang, aka something big dropped. It existed before, but there was a big innovative move in November '22. And since November '22, it's been a really fast train. It can be a train wreck. It can be a train success. Depends how you treat it. But we face that a lot.

[00:03:15] That's a really big misconception, because AI is not a means to an end. It's a framework. Sounds obvious to maybe a lot of us, but we're not necessarily acting this way day-to-day. I'm going to quickly give you a good example of AI, some good data. And I'm going to rush a lot through this presentation. I over-practiced it again, so excuse me in advance. I'll send it later.

[00:03:33] It can save money, for sure. It can incredibly increase employees' productivity. This is taken for customer service representatives, specifically. So there's a lot of good it can do. It can definitely improve NPS scores if done correctly. And there are a lot of numbers about increased revenue for personalized AI, and many more numbers in many different industries that it helps. So if it's done well, it can be wonderful.

[00:03:59] However, what happens when it's not done well? I'm not going to focus on anything fancy on this slide except the very bottom line. Yeah, it can be a big money dumpster. Yeah, it can introduce friction, so it can make your experience worse. I faced a bunch of that recently within different startups. It can make it actually worse if it's not done well. It can reduce NPS scores. And the biggest thing, which as designers we know really, really well, the biggest thing you can lose is trust. There's nothing bigger you can lose than customer trust. Agreed? Makes sense?

[00:04:31] That's the thing that makes people buy things. That's the thing that makes people believe in things, and that's the thing that makes people walk away. And if you've ever been in a tough relationship and you lost trust in someone, you know how hard it is to regain that. So trust is really big. And the biggest problem, I would argue, with not good AI, and there's a lot of not good AI out there, is that it can actually lead to losing the customer's most important thing, which is the trust.

[00:04:57] I'm going to give really fast examples of good and bad takes on that before I jump into the framework. But again, I want to start with this little idea that we just talked about, this misconception that AI is a means to an end. And I have a zesty take here. Not to talk too much about global warming, because it's not the point of the conference, but AI is actually really heavy on the environment. It's really, really heavy. So if anything, please don't do it just for the sake of it, because it can kill trees and animals, and that's a bummer, because we live here as mammals as well.

[00:05:25] But most importantly, I would argue that the biggest problem is not really AI or the layoffs around AI and all the emotions going around. It's about not focusing on people. At the end of the day, it's supposed to help them do something. So why do we not start with a problem statement? Why do we start with just doing something in AI instead of looking into what we actually need to solve? Simple design exercise.

[00:05:47] And again, as we said, and I keep saying it, trust is the one thing you cannot completely quantify. Trust is not quantifiable. It's a qualitative metric. We have a lot of researchers here and research companies. That's the one thing you cannot put into a number. But if you lose a customer's trust, I would say it's maybe even impossible to regain if it's really a big loss, especially in the health care industries. Moving on.

Bad examples: A24, McDonald's and Figma

[00:06:11] Some bad examples. It's really hard to find examples. I wish you luck. It's really hard for me, because no one is publishing their bad examples. Surprise. "We really failed in this way." Forgive me if I'm a bit of a sassy take again. I'm being a little teasing here, and I'll tell you in a second why. But we're going to start with a company that I actually personally love as a film lover, A24.

[00:06:32] They released their to-date most expensive budget film, I think it was last year, Civil War. Anyone heard of Civil War? Not if you want to watch it. It had a $50 million budget, I think, 50 or 40, sorry. So the most expensive to date. For some reason they chose to use generative AI for their poster creation. It created a really problematic mayhem over the web, not because of the aesthetic, but because the $50 or $40 million budget was not a trillion dollar budget. This didn't happen in the film. And even more importantly, there are no swans.

[00:07:10] People looked for swans in the film. There are threads, if you're being a geek. Go inside, look for Civil War swans. There's a mayhem: why are there no swans in the film? But the point is there's a disconnect between the expectation setting with this generative AI motion and the film. And it created a bit of a mayhem online, and people lost some trust for A24, which is a very, very high quality production and distribution company.

[00:07:34] Another one is a bit funnier, I think. I was really laughing about that one personally, so I don't know why you guys did not, but I thought it's really funny to expect a swan and not see a swan in the big film. A gigantic swan. But another example, of a slightly bigger company: McDonald's. They've been doing it for a while. Their mechanism was failing in the sense of not doing enough iterative response on generative AI. They made their agents for drive-thrus AI early on. It was two years ago, two and a half years ago. And they recently cut it because there were so many problems, to the degree of potential losses.

[00:08:10] I'm going to just read a few. There are so many examples online. I'm going to publish it, but look up TikTok. There are so many videos. It's quite hilarious if you have some time later. Someone even without an accent, and it's obviously generative AI, voice recognition... I will not be able to buy fries at McDonald's in my accent. But this really American guy was like, "Mountain Dew! Mountain Dew!" And he could not get a Mountain Dew for the life of him. Or someone almost sued them because the machine added bacon to an ice cream as a topping.

[00:08:40] I'm going to give you just one other bug here. I hope it works.

[00:08:46] "Give me just a moment, please. Okay. Sorry about that. I'm back. What can I get for you?" "Do you have..." "Okay. Sorry about that. I'm back. What can I get for you?" "Do you have..." "Okay. Sorry about that. I'm back. What can I get for you?" "I don't know." "Okay. Sorry about that. Hello. I'm back. What can I get for you?" "Okay. I don't know what I'm talking about."

[00:09:11] I think Tori's on one right now. Tori was on something. Tori, the representative. And not to dive too much into this example of McDonald's, their failure was in not doing iterative takes enough to break down the generative AI that didn't work. So, soft touch. We'll talk about it later. Soft launch and soft touches are super important.

[00:09:31] And lastly, not really a failed example, it's not fair to say. Sorry, Figma. I know there's Config and really cool stuff, but for the life of me, I really want to generate some design system items, because it's on the web, it's open source, it's repeatable, it's reusable. Why don't we have that? So I asked for that. I got this. Which is cute. It told me what I need to do to create one. It was like, create a button, create a color, create another button. I was like, that's useful.

[00:09:59] I'm laughing a little bit. We all expect something that relates to what we do day in and day out. So to me, yes, I do diagrams in FigJam and all that. But what I really want is to erase the tedious work we all do in Figma. And that's just one example. We're the end users. There are a lot of designers here, I looked into the stats. And we're not necessarily getting exactly what we want. So is it useful to have a way to quickly create a flowchart or an information architecture? Yeah, it can be useful, but is it where it helps me the most? Absolutely not. It takes a second to create one. It's a button away. It's a template I click on.

[00:10:33] So focusing on the right area, I would say, is what I would like to see from Figma moving forward, on top of the things they're releasing today. And that's just an example, because we all know Figma.

Good examples: ChatGPT, Deutsche Telekom and Target

[00:10:43] Moving into some good examples, and some surprising good examples. I'm not going to dive into all of them. I'm going to run through, because I really want to get to the framework. Starting with a really simple one, which made a big bang: ChatGPT. I'm going to touch upon Deutsche Telekom. It looks like T-Mobile; it's the partner of T-Mobile in Germany. They did something that saved an enormous amount of money for them internally. And Target, on this kind of nautical on again jam [?] into this example.

[00:11:06] They essentially employed an in-store app for their employees, new and older employees, to find things quickly in the store. Simple. Scrolling through the aisles. It really helped onboarding new employees, and it helped existing employees find things faster. Super simple. But it was the biggest problem they faced internally. So sometimes the solution for your AI features, and the things you might want to propose as product, UX or engineers, might be inside and not outside. Keep that in mind. It can save you an enormous amount of money.

[00:11:35] Let's dive into ChatGPT. I don't know if you see the source here on the bottom. I asked ChatGPT why ChatGPT works. Not the most objective, but I thought it was funny. There's one problem here. ChatGPT works well because it does solve the problem of accessing complex knowledge quickly, aka the internet, "by providing natural conversational answers to broad or specific queries, like having a helpful expert on demand." I'm doing this.

[00:11:58] We all know the internet is not an expert in many cases. If you ever had a headache and you looked up what you have, you immediately knew you have cancer, which is what I think a lot when I have a headache. So please don't rely on the internet fully; it's not an expert. However, what did they do well? I broke down why it's actually good. There's clear expectation setting. It goes back to any product that you know works really well. You know what to expect, so you know what comes back.

[00:12:24] The more important thing here, I think, is actually the imperfect delivery. ChatGPT is not an obsessive perfectionist like some of us. I think it goes without saying that a lot of designers actually are, because you have to be tedious. It's already telling you in the experience, and it's really small. The experience of ChatGPT is very, very simple. It's like, "Hey, look for something. Is this what you were looking for? I'm not sure." It's like this confused agent. It's not 100% sure, but it's trying to help you out. There's something about this expectation setting that sets the bar a little bit lower and enables you to refine your search. It's actually quite smart. So that's simple.

[00:12:58] And continuous learning. That's the key in any AI thing you put out there, and it's what we don't do enough even for feature releases. Quick question. How many people are actually doing A/B testing for every single feature they release in their company today? Every single feature. Anyone here? A/B testing? No. It doesn't happen, because we forgot to do A/B testing and check that things actually work as expected. That's what they do really, really well.

[00:13:23] A quick example before I move forward: from two minutes of search time to 18 seconds and less for Deutsche Telekom. Agents in HR, people in finance, people in legal had these separated data systems for looking for information. And they realized after doing research that their employees spend a lot of time, and that's the key here. Why focus on this? Because it's a lot of time spent multiple times by multiple people.

[00:13:48] And they started making an assessment of how much money are we losing from our employees, our HR, finance, legal employees, and then other departments, just looking for things inside our systems. So they hired this company that created this Ask T, which reduced the time of searching for things in complex data queries. Again, a pretty simple solution that saved an enormous amount of money. So sometimes the best solution you can have is going to be internal.

[00:14:12] All right, I'm going to jump into the how. We already know why it's working. I'm repeating myself here because there's a repetition in good AI. The clearest problem statement, and the what. We're going to dive into the what together in a second. Start small and expand; that's really big. And continuous learning. This one will repeat itself in every single good AI thing: continuous learning. AI is a self-prophecy of sorts, but you have to keep guiding it in the right direction, because it cannot guess what you want. And the better the thing knows you... the better we know our customers, the better we give them the right products. It goes all the way back to good design. It's the same sentiment, the same thing for good AI.

Three elements: understand it, target the what, soft release

[00:14:50] As we know by now, AI is the framework. And what do I mean by framework? You can suggest to your company, hopefully with enough information from this presentation, that we can do something utilizing a certain AI framework. I'm going to flash some words you can use with your engineers, but you should study just a little bit to know just enough, and create a big lasting effect that will help a lot. Also, as designers and product people, we can really be the future of AI, because we define the what and we define the why. Or rather, we search for the why. So why aren't we leading the charge in that sense?

[00:15:29] We're going to jump into the do part. You might take some pictures, because I'm going to run through some of those best practices. Essentially, as the UXers among us like to do, I boil it down to three big things to look for. Where to start, and honestly maybe when to finish. Three elements I'd like you to remember here, and we're going to break down each of them.

[00:15:50] First of all, as I mentioned, you cannot be this innovative person if you know nothing about what you're talking about. I wish you luck. It's going to be hard. You need to know just enough to respect the people that are going to work on it, and also to communicate with them. When I just started doing UX, I learned a little bit of Java, a little bit of HTML. I just wanted to know. I learned it so that I could tell them it can be done, because I was told no many times, and I was like, "There's no way." I'm going to study just a little bit. And then I found out that it can be done. So I was a bit annoyed. But it could be done.

[00:16:20] It's about common languages. If you speak a common language, you can go really far. You can do okay if you don't speak a common language. We can get coffee together or something. But if you speak a common language, you can go really, really far. So just know the languages a little bit.

[00:16:37] Oh, and I put a little thing. If you have a Python engineer in your company, make a little Post-it or note on your phone. Just set up a time with them: "Hey, I'd love to meet you. I'd love to know a bit more about what you do." Just do a 20-minute call. Just start, just to get some more knowledge. If you don't have one, there are a lot of back-ends that are really getting into it. Or you can go into Coursera. Coursera has really good courses on basic Python engineering. Nothing fancy, just enough so you know the jargon.

[00:17:02] I would argue that's the biggest thing, and that's what so many companies are missing. It's not just targeting the what, but why is this what more important than other whats? There might be a problem in your company, but the percentage of people doing that, the percentage of people having that frustration, is super minimal. Maybe don't focus on that. Many times I think that's what happens in companies: "Oh, we can do this thing." The fact you can doesn't mean you should. Maybe quantify it a little bit and make sure it actually solves a bigger problem. So it's both understanding the what and also: is it big enough? Does it have enough volume to hold water for a while? I'll get into it in a second.

[00:17:38] And that's a really big one as well. Again, what you saw as a repeated chord in all of these good examples: don't skip soft release, iteration, coming back to it, soft release, validating, and keep going. Also, if you can create a model that does this, as simple as that, and get some input, it will be gigantic. Then you can use AI to analyze that. There are a lot of things you can do. But make sure you keep listening. It's good in general to listen, but especially if you're trying something new, like AI.

Understanding it: types of AI, active and passive

[00:18:07] I'm going to break those down as fast as I possibly can. Understanding it: I'm not going to dive too much into it. Python is a fascinating, fascinating language, but I'm not going to get into it too much. I started learning it a few years ago. But I wanted you to know the basics, and I also boiled it down a little bit into maybe a new language that can help us dumb some things down.

[00:18:29] First of all, that's not the same. I know some of you know that, but it's important to make the differentiation. Generative AI and classic AI are completely different thought processes, in the way they've been working. There are a lot of differences between them, especially because classic AI, or machine learning, has been around for freaking forever. Forever. It's everywhere. Maybe you're not aware of it, but it's everywhere. And I would argue that's the most interesting part. I don't think GenAI, which is super, super cool, is actually as useful for many, many companies and many products we work on.

[00:18:58] We're not all generating things all the time. For some things we need a service. We need to get food. We need to go somewhere. We need to do something. It's not always about generating. It's lovely to generate, I'm an artist too, but that's not always the case.

[00:19:11] Just to give you a bit of the jargon, again, you can take a screenshot of that, but those are largely the big types of AI you can dive into. And then I'm going to get into the languages in a second. Essentially, a lot of companies actually just do narrow AI. Those are simple, specific tasks, such as Google Translate or a spam filter. Those are narrow AI features, and they can be incredibly useful if done in the right place.

[00:19:33] But there are so many more things we can do with AI. Again, generative is one thing; ChatGPT is a great example, DALL-E. You're going to see a lot of Easter eggs at the end of this presentation if I have time. And predictive AI I personally love a lot, which is forecasting outcomes. That's been around forever. It's not been nailed by many companies, but when it's actually nailed, you're so excited. Think about the Netflix "next" that's actually interesting to you. They're getting better.

[00:19:56] This type of model, which is learning you as you go, is actually really fascinating in different companies, and we should think about it more. But those are the options. A lot of companies are using chatbots; that's conversational AI. There are conversational AI designers. If you don't know one, look them up online. They're fascinating people. You can learn how to do that; it's also a good next profession to look into. But those are the options.

[00:20:18] I like to boil things down and make them really stupid to understand for myself. I like to call it active, or proactive, AI and passive AI. What I'm doing now is just putting a high-level blanket over everything in how the user might approach AI. Let me give you an example. Active AI can be something like ChatGPT. I go into ChatGPT, I actively go somewhere, I actively search for something, and I actively get a response. Active. So I'm an active user doing something to get something else.

[00:20:49] I love the second one. Those are these quiet agents that make your life better. And I'd argue that we need to use them much, much, much more. We need to put them in our products, and I really hope everyone here can start thinking about a potential passive AI thing that can help your life, either in a product you're using or in your own company. Just be easier and faster. This thing that you do again and again and again repetitively, that can be a template that can be filled in, that can be adjusted, that can be run through. All of those are options for passive AI.

[00:21:17] I'm going to break down active AI really fast, because it's pretty straightforward. It's a deep learning mechanism. It's using some machine learning for neural networks. I'm not going to get into this too much. You can go into generative AI; go ahead and do that. I want to focus on the second one, passive AI.

[00:21:32] Machine learning is super cool. Read about it. I love just reading about it in my free time. But essentially there are four types of patterns that work with data. Now, is this familiar? Don't you have patterns somewhere else? Design systems? Design systems? Tokens? Design systems? So we already know a lot about AI in this room without even knowing we know about AI. We just need to put a different skin on it.

[00:21:58] If something can be tokenized, repeated and broken down easily, it can be AI. It can move faster. You can tell it things to do. You can have a rule and formulation around that. Makes sense? If you call something green 100 in your design system and then you adjust that one green 100, it will make a lasting change all over your green 100. If you suddenly make the green purple... it's a horrible example, but if you change the token, you change it across the board.

[00:22:25] The same thing goes with machine learning, a little bit. But you need to formulate the data. So the next thing to do today is to reach out to the data engineer in your company and have a little coffee with them. Twenty minutes. "Hey, can you tell me about our data structure a little bit?" The reason I want you to speak with this person is that, depending on your data structure and how the data is constructed, you can do much more or much less in your company.

[00:22:47] If you have a well-structured, templated, tokenized data framework, then you can start pulling things. Just start having conversations. Even if you're not there, fake it till you become it. Just set up the calls: "Hey, I know nothing. Can you tell me a little bit about our data? How is it structured? Speak like I'm a five-year-old." The really good ones can do that. I would say the ones that are not as powerful data engineers cannot do that, because it takes an expert to break things down and boil them down.

[00:23:17] Long story short, I think meeting our customers or our people where they're at is the most useful thing. It's really cool to have generative AI. I think it can save a lot of time, can save jobs. It's great. But we already have a lot of products, and some of them work well. Can we make them better? How do we make them better? Where do we have the most friction? Can we solve it?

Target the what

[00:23:35] Number two. This is the most fun I had in this presentation. Please target the what. It's the number one problem with weird AI things that I'm seeing across the board, and I'm sure you have seen them as well. It's really simple. You target a good what, you can get a good outcome. It goes back to a simple example. Who here, if you joined a startup or tried to help a friend with a new idea, asked, "Who is your audience? Who is it for?" And who of you faced the one saying, "Oh, everyone"? That's it? Okay, I thought so.

[00:24:12] "Everyone should use this product." Okay, that's great. That's the worst start to understanding what you need to solve for. And not everyone actually needs to be addressed. So the what starts with the W's. We all know them. But you really need to know who you're trying to solve for. Look at this really good example of Deutsche Telekom and Target. They're looking into who is the person they're trying to help. In this case, it wasn't outside people. It wasn't customers, and Target is a big, big seller. It was about employees. They want to solve their problems fast. Okay, that's useful. So let's go into that area.

[00:24:43] Second, again, you all know this framework, because that's how we frame things to work on in the future. That's a product mindset. Where do they spend too much time? Not the most time, too much time. If you speak with five people in a certain department in your company, you can find out really fast where they spend too much time. "Oh, this thing is so freaking annoying." And then you know where it's at. Simple. That's the key.

[00:25:10] Why do they spend so much time in there? Why? Please be the most annoying person in your company as a UXer, always. Please be annoying. Be annoying. But why? That's not enough. When you have that, now we need to aggregate it. What do I mean by that? Why did the Deutsche Telekom example and the Target one work really well? Because it was repeating across different people, different departments, and it happened a lot of times. It's not just about the problem. It's about it being repetitively annoying to multiple people over time.

[00:25:43] So you should be annoying and be curious, but look for what annoys people at scale. We have a lot of research companies here. We can use a lot of tools. I don't want to call out names, because no one was paying me to say that, so next time pay me and I'll say the name. But long story short, we can quantify really fast. There are so many tools. They're getting cheaper and better at quantifying what actually doesn't work and what works well. So, aggregating it.

[00:26:09] Every UXer I know is a bit of a head of research. We have to be. Someone talked about democratizing research, which I'm a big, big fan of. Make sure you collect information. Always repeat. The quantitative collection of information is just how many times, how often, how annoying.

[00:26:28] And this one you can do tomorrow or next week. Speak with support people. It's an incredible resource. They are so freaking frustrated. They're losing their minds. Speak with support people, and speak with the most verbal, annoyed, outspoken support people. They will know a lot about their users, especially if you do SaaS, software as a service. Not only that: if you're a user of a product and something annoys you, double down on that. Ask others if they have the same problem. Here you go. You just found a problematic one that's at scale.

[00:27:01] And then filter it. What's the most annoying? I personally like to get to the most annoying first. You might want to start simple, and be less useful for humans. I'm not kidding. But just look into the high frustration points. What is the thing that repeats itself? What is tokenized? What do I mean by that? What repeats itself being bad for many people?

[00:27:27] And if you want to examine it yourself, look into something like FullStory or Pendo, whatever that is, and see whether some information, for instance in a form, is entered in the same way. What are the fields that are entered in the same way? Is there any common denominator? If you have more than an 80% match, suggest having, not generative, suggest having simplified classic machine learning there to suggest a pre-filled form next time. It's easier to adjust than to start from scratch. Simple. This is really easy. Just a little thing in the code.

[00:28:00] So filtering it makes us know what we want to focus on. And simply speaking, just start thinking in tokens and formulas. You already have that. It's in your brain. Think about design systems. Anything that can be tokenized and repeated can utilize AI to move faster, be pre-filled, or get suggestions. Anything. And if someone argues in your company, you can find the Python person to talk with, or just talk to me and I'll bring you some data. It's easy to find.

Feature excellence and feedback loops

[00:28:31] I'm not going to dig in; there's no time. But there are best practices for doing AI well, and they're functioning based on feature excellence. So we go back into design practices. Feel free to look into that. I'm not going to double down on any of it; there's too much information here. But obviously, we're always user-centric. We start small. We're scalable. Hopefully you have user-friendly design. ChatGPT is super simple, as an example. Boil it down to what's actually necessary. If you know the what and you know the who, you know how to do that.

[00:29:03] And I want to actually focus on... well, interpretability is super important, and we'll talk about it later, but feedback loops. At the very least, when you release anything AI, please, please, please just make sure it works. Please make sure it works. Just put a simple stupid thing. Literally this, this, this. At least you get a gauge. If you're a bit nicer to your customers, you're also going to include a little bit of, "What didn't work?" And I'll give an example in a second of ChatGPT that really annoyed me recently.

[00:29:30] So make sure you have a feedback loop. But the point of feedback loops, as you know as designers, is also to follow up on the promise. If something doesn't work, please iterate and go back to the drawing board and then represent it. Please. Your customers will thank you.

[00:29:42] I like to have a framework, personally. I'm going to glide over that because you can create your own. But for me, what passes the threshold for my team to suggest AI ideas is that it has to follow these four elements. I think that's where they find success. It has to be insightful. If it's zero insightful, it's not actually helping, and I don't know why we'd do that. It has to save time. If it takes them more time to do something, are we actually helping people save time with AI or not? Probably not.

[00:30:10] Personalization is gigantic, and it can grow over time. Don't be crazy there. You can start really simply, just starting to target the people that you want. Who is the person you're following? And you can put in a plug-in to understand the repeated behavior. Again, when you speak with your Python engineer: when it repeats more than 80%-plus, you can make a prediction.

[00:30:30] And please make it contextual. That's the thing that bothers me the most. You guys have seen these little virtual bots. They're sitting everywhere, generally speaking, sitting in the world. It's really nice if you can have the suggestion in the right place. We as designers have the power to do that. So just doing AI, to me, equals there's a chatbot somewhere. But doing AI well means: where is the most frustrated journey? Where does it start? Where can I help? Can I help contextually? Can it save time, be fast, and be personalized? And if yes, you will not only not lose trust, you will actually gain trust.

Soft release, manual override and documentation

[00:31:05] Too much information here. I'm going to glide over that. I don't have much time, and I want to share the stupid little visual at the end. But soft release is the last part. And again, I'm not going to dive into the specifics. There are specifics here. But soft releasing, evaluating, continuing to learn, please don't forget that. And then monitoring, redoing that.

[00:31:29] But eventually, being transparent is what separates, I think, ChatGPT and other models from the rest. They're really transparent. They're like, "Hey, I suck a little bit. Did I help you? Did I not help you?" It's like a non-confident person that this represents. That's a really big deal, because it sets an expectation. And expectation equals result.

[00:31:52] There's a cheat sheet here. Feel free to take a screenshot. I'm not going to read through it. There's a lot. I just want to make sure I give you practical, practical help here. But again, soft release. Make sure you evaluate in context. We can talk about this later. Rule-based systems or manual override: whatever you do, make sure it's not only AI. That's the one thing I'll say from this whole slide. If you create something with AI, you help someone with AI, you have an idea for an AI framework, make sure they can also manually override it.

[00:32:18] That's really big. Just imagine someone giving you a perfect solution. They're giving you this amazing pen, and you have to do these few things to make it work, but they don't have any other pen. You can just click to make it work. Make sure you can click to make it work manually, even if it's a really magical solution.

[00:32:32] Monitor continuous learning loops, and make sure you have documentation. The biggest thing in AI data engineering etiquette is documentation. Documentation is gigantic. It's making a big flashy comeback. Documentation is not a sexy term, but if you have good documentation, you can ensure your next iteration will work. Really simple. You can say, "Oh, this didn't work, this now works." If you don't have that, you can't say that.

[00:32:58] And if there's anything I will leave you with here, and this is again a zesty take: AI just equals good design. What if you look at it like that? As designers, we iterate. We work, we do another one. We take feedback, we do another one. We take feedback, we do another one. Sometimes we do a four-up, sometimes we do an eight-up. If anyone studied design, they know we have to do it in school. And then we release, then we check we made the right thing, and then we finish. That's the same as AI. The same framework. So any good designer can be a good AI innovator, I think.

A swan, some fries and a closing Easter egg

[00:33:36] Lastly, just a zesty, fun thing. I was like, "Hey, can you create an image of a UX professional riding a swan in an apocalyptic lake while eating McDonald's fries?" Because I like fries, frites. I got that. He's all right, it's cute, but I'm not... Thanks, AI. UX equals a guy. Not a guy. So I left feedback.

[00:34:12] If I didn't mention it before, I use ChatGPT only for emails. I put in my response. I'm usually very direct; I'm Middle Eastern. And I'm just like, "Can you make it American?" I'm not kidding. And it's like, "Hey, this doesn't work. Let's make sure..." And it just makes it like, "Hi Lindsey, good morning. Thank you so much for this information." Anyways, I'm not kidding. That's the only thing I do.

[00:34:37] So anyways, I just asked ChatGPT, "Can you make him a girl?" I also added, "Can you make her cool?" So of course cool equals sunglasses, apparently, but it's close enough. So yeah, I don't think I have much time for questions, but I hope this was useful.