Conf42 Machine Learning 2024 - Online

Design Think before you create your next AI product

Abstract

Usually, the talks on AI are about technical issues & future predictions, resulting in products that are “cool” technologically but “unfit” in addressing the real needs of users. This session aims at breaking that trend & bringing AI/ML developers closer to the process of design thinking.

Summary

  • Welcome to this talk on design thinking before you develop your next AI product. In today's talk, we will dive deeper into the current state of AI. What are some of the issues that haunt AI application in the absence of approaches like design thinking? And then how can design thinking really help?
  • AI is to be a force multiplier for people and organizations taking over repetitive tasks. Biggest promise of AI is to ensure that the technology supporting our society keeps updating itself continuously with the changes that come with time. Over 80% of businesses expected to adopt AI by 2025.
  • One of the most prominent area of challenges in the current state of AI still haunt us. Currently, the kind of energy requirements that large scale AI applications are simply not sustainable and feasible for most of the businesses and people. The biggest thing, which is closely related to adoption and scaling, is the trust.
  • Bias is a big issue which is haunting AI right now. The second one is accuracy and real time intelligence. Third one is about explainability the AI system, the kind of recommendation it has given you. communicating them to the end customers are extremely important.
  • The next one is about the security, copyright and IP infringement. The last one is hallucinations. Users are not able to trust AI completely. The root cause is lack of explainability. Some of the problems that we are seeing right now will not exist in 2040.
  • Design thinking always keeps it's focused on user at the center of it. Everything that you design is for a specific Persona. There is also a focus on iterative development which can really help you to solve problems incrementally. There are definitely these application adoption and trust issues that could benefit from the approach of design thinking.
  • Design thinking is an approach to solving bigger problems by understanding users needs and developing insights to solve those needs. When the problems are really wicked, then that means they are extremely complex. And when you have, you need to create a unique experience. When should you use design thinking or not?
  • Traditional design thinking approach has been there for some time. IBM's designed for AI framework is looking at design thinking from the AI's point of view. How you can train the traditional approach and apply some AI considerations on it so that your learning curve is not as deep.
  • The next one is really about storytelling, journey mapping and Personas. You have to cross collaborate, not just with end user, but also with internal business team. Design thinking can greatly help AI solutions in maintaining their quality issues.
  • Coming to mode three idea in this mode, you basically try to get as much ideas for solution as possible. The next mode is prototype. Here, people create the physical form of the best ideas, the prioritized ideas, and then they allow people to experience and interact with them.
  • Again, thank you so much for your time today. Let me know if you have any questions. You can always shoot me a note at arushi dot shivastava mail. com. I would like to consider diverse data sets in my thinking as well.

Transcript

This transcript was autogenerated. To make changes, submit a PR.
Welcome to this talk on design thinking before you develop your next AI product. Before we go further, let me tell you a bit about myself and my background in this field. I have been in software consulting, design and development for the last 20 years. After doing some Java J two e development at the start of my career, I got several opportunities in entity data to lead emerging technologies, research, design development and now deployment. So right from big data around twelve years ago, I have been with these technologies like IoT, gamification, enterprise, social blockchain, data science, and now AI and ML. So now in my current role, I actually lead fantastic team of designers, developers and solution architects who design, deploy and scale these products for large enterprise grade customers. And that really gives me some of the real world experience in this field which I'm going to share with you all today. In today's talk, we will dive deeper into the current state of AI, where things are, what are some of the issues that haunt AI application in the absence of approaches like design thinking? And then how can design thinking really help? And how do you design things for AI applications? So with that, let's start the talk today. To start, let's get started from the big picture. What is the ultimate aim of AI products? In my opinion, it is to be a force multiplier for people and organizations taking over repetitive tasks or very complex tasks that take time for humans to do and that can complement human effort. Things that are right now based on subjective knowledge and experience and which often result in inconsistent outcomes, can be changed with AI. They can be made to be encompassing of all the possible data and all the possible context around that data to result in the best possible decisions with consistency and reliability. And the data that we are talking about changes with time. In my opinion, the biggest promise of AI is to ensure that the technology supporting our society keeps updating itself continuously with the changes that come with time, with place, with the surrounding context, and it keeps updating this knowledge and understanding with this continuous learning. So lofty goals for sure, but I would say not too far. And in fact these goals are not at all new. Over the last 70 or so years we have made incremental progress in AI. After coining the term and making some early experiments, there was a long gap in which we did not really do much and it is often called as AI winter. But since 1980s, focused academic research as well as development of surrounding ecosystem of technologies has led to tremendous investments. But nothing has been as big as the last two, three years where we saw the first big use case of generative AI through chat GPT and that has rightly deserved its place on every next LinkedIn post that you see or the conference talks, newspaper headlines, and more often than not on the minds of worried policymakers. That said, it is not just OpenAI and chat GPT. AI startups are sprouting in every major industry vertical cross industry applications such as vision, natural language processing and search are still ruling the charts, but use case focused development which is application of these cross industry applications in a given business context, such as manufacturing plant floor machine vision or say traffic condition analysis on a highway. Those are the ones that are actually resulting in real business value for businesses, but none of them are as feasible or even deployable without the supporting platforms and processes. Where we are seeing a lot of development recently, such as those that are needed for Mlops or AI Ops. But with around all this buzz that we are hearing about AI, there are a few things which it is really good at and some things where it is still very niche, very new things that it is good at. First I would say is the NLP based chatbots BI has really shined great in these areas. Or inculcating knowledge from heterogeneous data sources, detecting anomalies and fraud in say, insurance claims, etcetera, and then predicting based on historical trends and creating better personalized recommendations. Something that you see when you shop on Amazon or Instacart, etcetera. It is not just based on your historic information, but also the context, people like you and where could your interests lead you to spend your next big part? Right, but where are we heading with all these developments? Around 90% of the AI experts when surveys felt that AI advancements will lead to human like intelligence in the next 100 years. I don't know what will happen decades later from now, but one thing we can say for sure is that it is here to get better, with over 80% of businesses adopting it by 2025. But some of the most prominent area of challenges in the current state of AI still haunt us, and one of the biggest one is scaling of AI application. Currently, the kind of energy requirements that large scale AI applications have are simply not sustainable and feasible for most of the businesses and people. That is something that still needs to be worked out. And there are some very good advancements being made which which tell us that this is something which will be solved in next few years. The next is about adoption of AI. Yes, it is related closely to the cost of AI right now, but there are other issues that are stopping AI from being adopted. Generally they are related to the quality of the outcomes, as well as the biggest thing, which is again, closely related to adoption and scaling, is the trust. The users and the businesses are not able to put their complete trust in the AI solutions. We will look at all these issues in detail in a little bit, but having said that, I think there are some things that you can solve with approaches like design thinking, and there are of course, a whole gamut of other frameworks and approaches that need to be applied in order to solve the others. That said, some of the most prominent challenges in scaling adoption of AI this is a report from Everest group. I'm not going to go through all the numbers here, but the key point is that organizations do not really have talent and skill to make informed decisions about their AI investments. The way they manage their data and the computational infrastructure right now is just not suitable for enterprise grade implementation of AI, and everything that would be useful for them is extremely expensive right now, at least at the current state of technology, and the skills to maintain that is just not there. There is also inadequate policy and compliance infrastructure, so businesses do not know if what they are going to do is going to be acceptable in the regulatory framework or not. And then there's simply just unclear ROI of the new things that organizations do want to do. But the kind of costs that it takes, whether it is really going to result, the ROI that they are expecting or not, is unclear. But that happens with every new technology, isn't it? Having talked about scaling and adoption, let us look at the last pillar, which was really trusting AI and why businesses and users are having trust issues with it. So the first one that is most talked about is bias. It occurs when an algorithm produces results that are systemically prejudiced due to either erroneous assumptions in the machine learning process or the lack of context, lack of data sources, lack of all the possible parameters that can inform better decisions. And algorithms can usually have this built in biases because they're created by individuals who have conscious or unconscious preferences themselves that may go completely undiscovered until the algorithms are actually used. Bias is a big issue which is haunting AI right now. The second one is accuracy and real time intelligence. All the AI systems that have been created either have supervised or unsupervised learning in them. The data sources that they are considering keep changing all the time, and depending on how the AI system has been implemented, it may or may not be taking updating its learning sources continuously or not, and which greatly affects the accuracy of outcomes. The third one is a big one, as it is about explainability the AI system, the kind of recommendation it has given you, on what basis has it said so? Again, coming back to the example of doing online shopping, your new recommendations, have they been shown just based on your past history, or has it also considered people like you? Has it considered what is the next big fashion trend that is going to happen? Or has it considered the weather implications in your area in the next few months? So, there is a lot of factors that AI algorithms can consider, but communicating them to the end customers are extremely important, which is right now failing that is one part of it, things that can be easily explained. There is also inherent in explainability, or the lack of explainability within the AI models themselves, because they have been built upon the understanding that we do not understand our brain very much. We created these neural networks which mimic the outcomes that our brain create, but we do not know how the process itself works. The data scientists have a very hard time explaining why did their model choose outcome a versus outcome b in two different scenarios. And that is a little bit harder problem to solve than the first one, where there are parameters that just need to be explained to the users. The next one is about the security, copyright and IP infringement. And as we have all seen with some of the image generation or the video generation softwares, with the generative AI, it does collect millions of data. So data sets, right? And where has it take? What influence, if, uh, what has affected the result? It is just simply not humanly possible to cite all of them with every outcome that is created. And it is important for it to look at all the data sources. But that also heavily weighs in on some of the creative ideas that original creators have put in, and they are simply not getting credited for it. Forget the, the reimbursement for their well thought work put in. So that is, again, some of the things from which the creative community has a lot of problem with AI. The next one is the quality. So a lot of you who have, who may have used chat GPT for must have seen this, right? I have seen this myself. If I ask to create the script for this talk, I would see a lot of things being repeated, which is known as parroting. I would see the answer that I get today for a talk on design thinking would be very different from what I may have got a year ago and maybe very different the next year, depending on what data sources and sets it has considered. And sometimes it will be very outdated as well, creating these drifts. And the outcome. The last one is hallucinations. Sometimes when AI is not able to bridge the gap between what it knows, it puts in things which are totally out of the context. Right? Just the other day I was hearing on NPR an example where they said a newspaper ad agency or a product review company was using AI, and they were talking about the gym belts. And suddenly the AI hallucinated and started putting in text, which was talking about the belts, the waist belts that you use, or the fashion belts that you use. And it doesn't take a human to understand when the AI has done so, and it just creates that cringeworthy experience, where would you really want to consider it for any serious discussion or not? So, all these issues are prominent challenges, because of which users are not able to trust AI completely. But what is the root cause? We've talked about some of them, the lack of explainability, etcetera. But one of the key reasons this is happening is because while the development of this technology, people were extremely focused on the technology itself and how to make the algorithm work, they were not really taking the end user in consideration, or society in consideration, etcetera. But now, when the business applications are out there in the world, this tech focused tunnel vision is not working at all. So that is one of the key reasons. Another one is that when people think from business applications perspective, they think about people, process and technology. And now data is a new participant in it. It is more than the technology itself, it is more than the process, it is more than the people. It is also an active participant in an AI ecosystem, in an AI application. So now, thinking from the data's point of view is extremely important. What data sources should you consider? What should be the quality of those data sources is something that have to be rethought of when you are developing an AI application, which right now are not being considered by every product, and which is where, again, which creates some of these biases, drift, etcetera. Lack of explainability. I think we've talked about it enough. There is, of course, the two kinds. The one that is just not done by the developers and implementers, and informing the users enough. But then there is also an inherent explainability in the implementation of these models. The last, but something that is very true for any new technology, is that it is just simply new. We have not lived in an AI world before. Some of the problems that we are seeing right now will not exist in 2040. So there will be new problems. Of course. I think the problems that we would be talking ten or 15 years from now would be about the supervised learning, the sustainability of quantum computing, etcetera. But probably we as a society would have found out solutions to problems related to bias, crypt, etcetera. So remember these four reasons for the rest of the talk. Now that there is extremely tech focus, the new participant and data is not being considered in designing. When the user experience people look at AI systems and then there is lack of explainability, both from the technology side as well as from the design side. And overall, it is a new word for all, not only the implementers and designers, but also for the users and customers. So nobody really knows what they expect from the ultimate system. Why I think design thinking can really help because if you look at the key tenets of design thinking, it is, it always keeps it's focused on user at the center of it. Everything that you design is for a specific Persona, and that is extremely important for AI systems, because when you consider the feelings and emotions of a person who's going to interact with the system that you have implemented, you are always going to do a better job than by throwing technology at them, by perceiving what they might or might not like. And then the next is it also involves around. It very much relies on cut. It also relies on cross functional collaboration. You can avoid some of or many of the biases and tech only vision problems by involving various groups, diverse backgrounds, diverse data sets, diverse context. In your AI solution design, there is also a focus on iterative development which can really help you to solve problems incrementally and then reduce risks. The way it does it is by taking the solution back to the user sooner than what traditionally system implementations too. You cannot solve all the issues that plague AI currently, such as the talent and skills gap, or the hardware sustainability and inefficiencies, etcetera. But there are definitely these application adoption and trust issues that could benefit from the approach of design thinking. So before we go deeper into how, let's see what is design thinking? What design thinking actually does is that it makes sure that whatever product or service you're trying to implement is actually viable for your business and it is feasible with the technology at hand. And most important, it is desirable by the users that are going to use it. It is centered for most of the innovation that has happened recently. Many of the successful apps, companies, products and services have benefited from it because of these key qualities of design thinking approach. But again, it is not new at all. It is not something that has only recently come and no products or services have become widely popular because of this. Without design thinking at the center, it has always been there. Everything has been said about this. You take up any industry is going to take a successful book, take a successful movie, take a successful song. It has always had these coordinates at its center. But the reason why we need to talk about it now about this again is because AI is new and it is showing some of the problems that other products and services have shown in the past, which when they use design thinking, help them get better about themselves. That is why AI could really benefit some of the key tenants of design thinking. For those of you who go by the definition, here is the definition for you. It is an approach to solving bigger problems by understanding users needs and developing insights to solve those needs, resulting in an AHA experience for not only the users, but creators and stakeholders as well. Now before, let me go back before we move on to when should you use design thinking or not? If you just look at the highlighted words here, it would give you a very simple guide on when to use design thinking or why is it important? Right? Wicked problems. Problems which are not easily solvable by simple if then else in for loops, right? When you have to consider users needs again, think of trust and adoption issues. When you develop deeper insights, think about all the data and context that we have been talking about till now and create an AHA experience, something that people have not experienced before. And that is what your AI systems is supposed to do, which we all saw that aha experience when we use that GPD for the first time, right? So coming back to when to use it, right? Whenever you need to understand user needs and develop insights into those needs, that is when design thinking should be used. When the problems are really wicked, then that means they are extremely complex. The answer is not straightforward. And when you have, you need to create a unique experience. Now let me give an example here. For example, in healthcare industry vertical, there is a lot of talk about analyzing the electronic health records and creating these use cases which come up with recommended treatment plans. The basis of that is that there are certain conditions which only involve looking at certain parameters and depending on the context of the patient, the recommended treatment plans are a few and you cannot really go wrong. It is really just a few options that a provider has to consider. So here you need to understand the user needs. When you're creating the system, you have to think about the patient trust issues, right? What if you give the information of these recommended treatment plans to the customers or to the patients directly? Right? Are you ready for that kind of world where people are getting their treatment plans from an AI solution? Would they be able to trust it or would it create us? Could it create distrust in the physicians and providers themselves that oh, if these are so simple things, there is a possibility of misuse of that information by the patients themselves, right? If they have just considered two or three parameters and coming up with their own treatment plans and maybe ignored a larger health health condition which a physician would have been able to look at much more in detail, right? So maybe for this particular use case, patients are not the user group that you should be focusing on. Maybe it still needs physicians or the human supervisions of certified nurse practitioners or any other provider Persona type, depending on where one group is permitted versus other. For example, nurses should get only these aspects of recommendations. Probably doctors can look at, or the high level of recommendations, etcetera. So you probably need to rethink about your user groups and the needs of that user groups. Maybe patients are not at all a user group in this particular use case at this time, right? The next one is why is it a wicked problem? No one knows the right answer here. No long term study has been yet done on automated recommendations of treatment plans in various conditions. Maybe recommended test plan may be good for weight management type of issues, maybe not for a diabetic's weight management, maybe not a weight management in pregnant people, etcetera, right? The correctness of this has not yet been been studied in long term, it is fairly wicked problem, right, that you have also not studied it. And what happens to the experience of the end customers and the provider spaces as well? You have to create the unique experience here. No one has lived in the true AI world and you do not know the repercussions of this, right? So you have to create an experience that fits their requirements right now that increases their productivity without compromising on the quality of care that the end customer is getting. It is a perfect use case where design thinking should come in before any actual AI application is launched in this particular area. There are several school of thoughts in traditional design thinking now. Literally, they come from schools, many universities. The one that is most popular, which we will be going in detail today, has come from Stanford's school's framework for design thinking, and we would look at it in a bit more detail. But all these different frameworks that exist out there, their key goals remains the same. They all focus on understanding the user. They all rely on radical brainstorming in cross functional groups, they all promote rapid experimentation and going back to the users with the test of those. And they all believe in co creation and collaboration between user groups, different teams, different sets of data, etcetera. So doesn't matter which framework you are picking up until it is these coordinates in it. But we've talked about all what design thinking is. Let us also see what it is not. It is definitely not a quick fix or a band aid type of approach, right? Where you are seeing certain issues in production and there is a big user trust problem which you need to fix in a week or so. Design thinking is not the answer. You need to do something else about it. Design thinking is generally needed when you are starting a new product or service because it takes time to do the iterations. It takes time to come up with the final solution. It is also not an approach where you have the technology at hand and you are looking for where to apply it. Right? What is happening with most organizations right now? They have AI and machine learning. Oh, how can we use generative AI in insurance right now? How can we use generative AI? So there again, hammer looking for nails. Right? This is not where, again, design thinking would help you. Design thinking would really help you. If you have a problem space, if you have a user group in mind and you want to solve a specific problem, it is also not a quick response to competition. Again, enterprises these days are saying, oh, our competitor a is not yet using generative AI. Maybe if we do, we will have a better edge over them. Maybe not, right? Nobody knows the answer. But definitely design thinking is not a way to get that competitive edge. You probably need to look at the strategy a little differently. You need to look at your industry specific use cases. You need to look at where your business is differentiated, etcetera. Again, a whole side of business that cannot be solved by design thinking. And then it is definitely not a foolproof formula for sure. Short business success. Again, related to the third point here, there is, it's not that you can solve every problem that your business is currently facing or your use case has with just design thinking. It is really about creating new products or experiences or services for known user groups. So let's look at traditional design thinking as well as what is emerging for AI in the market space right now. So as we talked about this Stanford D school design thinking approach, there are five modes to this approach, right? You go from understanding the user, which is the empathize mode, and we will go into the depth of these modes again in a little bit. But there are these five modes where you understand the user, you define the problem space, you collect the ideas, then you try to solve the problem with rapid prototyping, and then you test it with the user. Again, this is how traditional design thinking works. Again, it may look like a waterfall approach, but there are several iterations that can happen between the modes themselves. For example, between empathize and define. You might go back to the user to understand the problem space. Again, you might want to refine the problem statement again by talking to the users again, and you can do several such cycle back and forth, similarly between prototype and test, or between ideate and prototype, or the whole cycle itself. It is a very iterative approach, although it may look like waterfall in this diagram. So this is what traditional design thinking approach look like. If you look at what is out there in the world right now, this is an example. It has come from IBM's designed for AI framework. Now, again, they are looking at design thinking from the AI's point of view. From what, how AI is making you think about the business, about data, about the understanding of the users. How should you prototype about it? What knowledge have you created? What knowledge would you continuously learn about, etcetera? Again, looking at the problem space from the AI's lens. Now I want to explain this for the benefit of those of you who have worked with design thinking approach, as well as for those who are doing it for the first time. The traditional design thinking approach has been there for some time now. People are comfortable with it, they have deployed it, you've seen it succeed multiple times. So it is very much possible for people to just fall back on the traditional approach without really considering these new frameworks. And then this new framework that you have seen, for example, that of IBM, requires you to again practice it, learn it, and without really implementing it in the real world, it is difficult to be an expert in these new frameworks and approaches. So that's why what I want to talk about today is how you can train the traditional approach and apply some AI considerations on it so that your learning curve is not as deep. So that is what approach I'm going to take for today's talk. If you all have different ideas, you have any discussion point about it, we can talk about it later through questions as well. But right now, for the purposes of this discussion, let's just take traditional approaches and then apply AI considerations on them. So again, I'm going back to the five step or the five mode framework, the Stanford schools, and for each of the steps that we would talk about, I would talk about the goal, the process that is generally taken, and the tools that are available in traditional approach. So first of all, mode one for empathize. The aim is to understand the users within the context of your design challenge. The process is basically to observe, engage, and immerse with these users. Some of the tools that are available for you are interviews, empathy, maps, and in context immersion. Now, let me give you a quick example of how this traditionally works. So, for example, if you are creating an application for kindergarteners, generally what the designers would do is be in the classroom at the level of kindergarteners, sit in that space, see what are their physical constraints, what do the kids need, what frustrates them, what excites them in the classroom. And then look at that, and then see all these concentrations are met when they are designing the final solution. So they would actually immerse themselves, they will interview their end users, and they would create the empathy maps of saying, okay, what do they say? What do they do, what do they feel, what do they like, what do they not like? Etcetera. And this in context immersion becomes extremely important, even for non AI applications. So let's now, having said that, let's look at the AI considerations that any of these modes need to take off. So first we'll go to empathize. So it is extremely important to observe the user in non AI word for, again, going deeper into it. If you are creating an AI application, it is always very much possible that you would just interview a user virtually and say, okay, this is how you are going to use a treatment plan, right? Maybe somebody who is considering using treatment plan, somebody who is going through the diagnosis right now, or somebody who is going through the treatment plan itself right now. Until, unless you sit with the user, when the diagnosis is communicated to them, what do they go through during that time? What kind of questions do come to their mind? What do they ask their provider? What does the provider explain to you? Until, unless you sit in that context and you observe all your user groups properly, it is often possible to leave out certain data, certain key consideration, the point of explainability and all this is extremely important for your AI applications to create the trust. Also, you have to select the data sources based on the authenticity and accuracy, right? Again, considering the patient's questions, considering what do the providers or the doctors look at most, and what do they believe in? Making sure you prioritize that in your AI application is important. And then you have to ally the AI solutions to users context, something that may work very well in, say, a generative AI chat bot that is general purpose may not be right for what a diabetic patient is looking at to consider its next treatment plan, right? You cannot give them the same kind of disclaimer that AI results may be wrong. Here, right? What you maybe again, as we said, patient may not be the right user group at all. If it is the doctors and physicians that are your writer user group, you might consider giving them a confidence score, saying that this is what algorithm thinks is 90% accurate or this is 60% accurate, etcetera, so that it increases their productivity, but also tells them that how much relevance they should give to a particular again, it is extremely important to align a general purpose or a cross function AI application to the user's context. Going to the second mode, the define generally in traditional design thinking, you capture the findings of your empathize mode and you create a deeper understanding by creating a Persona definition of your users. You craft a meaningful statement, an actionable problem statement really, which is generally said like this, right? A user needs something in a way that they are able to do something. For example, they may something like diabetes especially needs to provide treatment plans in a way that increases their efficiency and productivity. So that could be our high level problem statement. And then you might be able to come back with more how might we statements saying that how might we be able to increase their productivity through AI? How might we be able to increase the trust of patients in the solution through AI? So there can be several how might we statements that could be horrible. The next one is really about storytelling, journey mapping and Personas. These are the tools that are available that people often use to create the problem statements, or create these how. How might we statements. These are the supporting tools that they use now. These are the considerations that people should keep in mind. They should go back to the user for validation of these problem statements. Are we considering the right things for this solution? Is this what you would really like to see in the final how would you act if you were able to solve this particular problem for you, etcetera? And you have to develop these insight questions for non functional requirements also. And this is where cross collaboration happens between teams. You need to bring the security and privacy teams, the technology team, the infrastructure team, and you have to also bring in. So when we are talking about efficiency and productivity of the provider, you also have to look at what are the security and privacy concerns that you need to take in mind. What are some of the technology concerns, investment concerns, etcetera, that you need to take in mind. So again, you have to cross collaborate, not just with end user, but also with internal business team. You have to define the values of your solution that should be tested with each phase. So this is also extremely important that especially for AI, you have to put in the values, for example, quality. What is the targeted accuracy of the solution? How much drift should be allowed? What is the continuous learning approach to this, of your overall? What are some of the privacy guidelines that the solution should always follow? And you would see, if you do that in this particular mode, it's really going to help you when you are coming up with ideas. It is also going to help you when you are coming up with prototypes, the final solutions, etcetera. And this is where again, design thinking can greatly help AI solutions in maintaining their quality issues, as well as increasing the consumers trust in the overall solution. Coming to mode three idea in this mode, you basically try to get as much ideas for solution as possible. The main aim is to create both volume and variety. You are totally non judgmental about the ideas that you are collecting. Usually people use brainstorms, brainstorming sessions, and looking at existing solutions with cross functional teams and creating. They use mind maps and notes, clouds tools, collect as many ideas as possible. Now, again, from AI perspective, if you think if we really actively seek out alternative sources of data and perhaps even conflicting point of views to include in our models of the work, perhaps the algorithms will be less biased, right? They will be less open to manipulation. And this is all true for supervised learning. We are not yet talking about the future where AI is taking decision of its own. Remember, singularity is still not here and humans still have the unique ability to engage in the decision making. So this is where, from AI's point of view, you have to involve cross functional teams, get their context, their alternative views of the world. You do not have to really solve the problem here. You're just collecting the data sources, the data sets, the values that you should consider, things that are important for other teams and groups, and you're just collecting those ideas, those solutions. And here you have to surface all the AI opportunities and pitfalls. Again, this is important. Traditional design thinkers do not do this, but AI has almost made it most important that you consider the security aspects, the privacy aspects, the infrastructure constraints that you may have on your solution. And you involved technology, not just business people, in coming up with the ideas for the solutions. You also have to consider the pitfalls that your AI solutions may have, right, the risks that are associated with getting the solution in the hands of the end user. And then you have to draw, and this is really an example, this is really a tool, I would say, which works for anything, AI or not. When you're doing such brainstorming and just collecting ideas is to draw some inferences and some motivations from parallel universes, which in our context, we would say, for example, if you are developing something for healthcare, you may look at retail, or you may look at finance to get some ideas as well. From there, how people have used AI solutions in those contexts, there may be a few ideas hidden there that may be applicable in your context as well. The next mode is prototype. Here, people create the physical form of the best ideas, the prioritized ideas, and then they allow people to experience and interact with them. And that is where they again record their emotions of how people would react if a service or product, as hypothesized, would be presented to them. The process that people use, they generally learn and explore. They solve any disagreements that could be there between ideas, between teams. Here, this is a great opportunity for solving that. They would start conversations about things that have yet not been talked about between teams or inside teams yet. And for example, in AI's world, we can talk about policies that have not yet been considered, the fears of the users that have not been yet taken into account. And how could technologists look at solving that? Right. Breaking the larger problem into smaller components is again a key tenet here, which really means that if you are creating, say, generative AI based chatbot, what are the different aspects that it would have? What are the different modules that it is going to have? And you can create a prototype for each different subset of the problem and test it individually before actually bringing everything together to see whether. And again, this is to reduce your risk, reduce your investments of future. You can fail quickly if people do not like something at the prototype stage itself. Generally, people use catching physical mock ups, wireframes, they use interaction flows, storyboards, prototypes, which are again, something that you can very quickly create and test with your end users. For AI, you have to now think of your prototypes a little more advanced than they have been before. You have to focus on the technology a little bit more. You have to set clear test goals for each prototype. Now, these test goals, remember the define mode. You are deriving your test goals from that defined mode. Again, the values that your system should consider and are your prototypes. Considering those values, what and what not should be presented to user in the solution. Have you considered all that? So test for those goals again with each prototype. Think how the user will test this when you are presenting an AI solution to them, right? What are the things that they could break? What are the lines outside which they are going to color the solution and then test the values again and again. I would explain this. I would emphasize on this rather that you have to think of explainability, you have to think of bias. You have to test those values in your prototypes as well as in the real solution. Next is the test mode, the final mode of traditional design thinking where you solicit feedback on prototypes by putting them into the context of use. You define these prototypes and solution and learn more about the user. You continue to ask the why questions and refine your point of view. Sometimes you create things which and by the time you have created there is additional insight available to you which might ask you to pivot totally or change your initial goals again. So sometimes the iteration can happen after you reach this end state as well. So be ready for that. Continue to ask those why questions. They would really help you to do that. Some of the things that people use here are again, desirability testing, field studies, feasibility testing. They do cost analysis, they do swot analysis, etcetera. Again, some of your MBA friends can actually help you with that, but this is where you actually test the feasibility, desirability and viability of your solution alongside your competition, right? An extremely important stage. This is what you do before you actually go ahead and code in a real prototype. And then some of the AI considerations for this mode are again, show the user something that they could test. Don't just tell them an idea and try to gain their feedback from that verbal idea. Regard their reactions when they see something for the first time and how they use it, right? Not just yes, they like solution and yes, it is passed, right? Test with a new set of demographics. You may have collected your requirements or ideas, or even you may have studied a certain type of demographics and user group along with testing with them. Consider alternative sources if you have till now for your say again going back to our healthcare treatment plan recommendations use case, suppose you have tested it with older demographics till now. Think of what happens when you go to teens or when you go to young mothers, etcetera, right? So that is where this can really help when your test mode expands. And this is where again those trust issues could be avoided. Because till now you might have been considering a view of the world that was not encompassing of certain things that you have not considered. The last is test with newer versions of data sets. This is again just given how time boxed things generally are in our industry, you may not have the freedom and ability to test all the data sets at every stage, even at the prototype stage. Or before that, when you finally are in test mode with a real prototype, open it up to newer data source and then see how it fares against them. See whether it suffers from hallucinations, drift, or those parroting issues that we talked about. So this is a place where I've put all the considerations together for those of you taking screenshots, so that when you are practicing traditional design thinking, you have something to go back to and look at all the recommendations at one place beyond the design stages too. This work does not stop and you have to keep these things in mind when you are actually designing the user interface, the final solution for your customers. You have to keep transparency in mind, you have to keep explainability in mind, and you have to keep testing for alternates users more than ever, right users, data sets, how your different releases of your software are acting and you will keep doing that, not just during the development of your product or service, but even after it has gone into production. And when can you do this is a question I often get right when I introduce these concepts to senior leaders, cxos, etcetera. They are like okay, we have already started a pilot or we do not know anything about AI or we are already on this journey, are we too late? So you can start it before you start prototyping your next idea. You can also employ it in your current project as a parallel stream and you can always look at any weird wicked problem and apply it there. Start wherever you are and you will be fine. Again, thank you so much for your time today. I hope you liked this session. Let me know if you have any questions. You can always shoot me a note at arushi dot shivastava mail.com or arushi dot shivaswapntdata.com and I would be happy to discuss this further, especially if you have any alternate views about this because just like this approach, I would like to consider diverse data sets in my thinking as well. Thank you so much again and hope you enjoy the rest of the conference. Thank you.
...

Arushi Srivastava

Senior Director @ NTT Data

Arushi Srivastava's LinkedIn account



Awesome tech events for

Priority access to all content

Video hallway track

Community chat

Exclusive promotions and giveaways