"Creating value from AI led transformations" by Pranay Agrawal

Are you able to see the screen? Yes. Okay, excellent. Very good. Look, I know the format typically is that I'd probably spend about 25-30 minutes taking through some content, and then maybe it's open for Q&A. But if there are questions in the middle, you can send questions on the Q&A box or directly in this conference. I'm happy to make it interactive and address questions along the way. All right, so let's get started. I want to start with a small story here. My screen's not... okay, there it is. I want to start here with a story about a farmer in Japan whose mother was spending up to eight hours a day sorting cucumbers. Cucumbers apparently come in nine different varieties, at least in Japan. He realized that this was an extremely poor use of his mother's time, who could have been using that time to grow more cucumbers, sell them, or just relax and enjoy life. It was also hard to find labor to do that because this is a very seasonal activity, and one needed to train the labor because a certain amount of skill and experience was needed to actually sort the cucumber. Makoto then decided to apply AI to this problem. He purchased a Raspberry Pi, and with some software, he was able to get high accuracy in sorting and recognizing the cucumbers. Soon he was well into the 90s in accuracy. Huge amounts of time saved, a great outcome for his parents, and in a fairly simple way, it got applied and was highly useful and productive. Now, what really happened here? If you deconstruct this cucumber sorting problem, we started with him taking several thousand images of the cucumbers from several different angles. He probably took about eight to nine thousand different pictures. He then used a neural network to train the software so that when it saw a new image of a cucumber, it would be able to sort it with certain accuracy. He created a system, a mechanical conveyor belt, where the cucumbers would roll over through that conveyor belt, the image would see it, the software would categorize it, and then the physical system would drop it into the right box. As it would drop these cucumbers into the right box, he would subsequently get the information whether the sorting was accurate or inaccurate, and from that feedback, there would be learning for the software, which would further help it in improving its accuracy. This is very similar to how the human brain perceives things and how the human brain learns. We see objects, we learn, and over time we learn to act and get further feedback, and things continue to get better in terms of our perception. So that is really how AI is working. But more importantly, the idea that if AI can be applied on a farm to sort cucumbers, the applications of AI are limited only by our own imagination. There is a huge expectation that AI will deliver tremendous growth and competitive advantage. Some estimates say up to three trillion dollars in business value will be added, and several billion hours will be recovered from doing mundane tasks, giving us the ability to put human capacity and creativity to better use. Some applications just in terms of improving human conditions: in the area of road fatalities, we lose about 1.3 million people every year to fatal accidents on the road, and the expectation is that autonomous cars with their self-learning ability will dramatically reduce that number. At the same time, there are a lot of different implications as well, from a defense standpoint, cybersecurity standpoint, and very importantly, what's going to happen with the capital and labor dynamics. One of the big fears is that with AI, more and more power will shift to capital, and what will happen to labor, or how will human beings continue to contribute economically and be relevant. These are all the kinds of things we're speaking about at a macro level. There are huge expectations of massive changes, and we are starting to see a fair amount of change on the ground as well. What we want to talk about today is what does success look like in the traditional world of manufacturing or services, and what does it take to create that success within our respective organizations so that we can realize the value of AI within manufacturing, services, financial services, insurance, and so on. Let me take a couple of points over here first, and then we will speak of a few success stories. The first idea I want to share is that we need a combination of AI, engineering, and design to realize full value from AI. AI is the ability to build algorithms to do a huge range of cognitive tasks at human capacity or better. Cognitive tasks could be things such as analysis, predictions and forecasts, computer vision to recognize images, find defects in products, text analysis, and so on. But we need a vast amount of data for these algorithms to be accurate and to learn. Equally, the signals or insights that come out of these algorithms need to be fed into applications where action can be taken. To operationalize AI and generate value, we need a very strong, robust engineering platform. That engineering will create the data infrastructure to feed these algorithms and continue to feed them with new data so that those algorithms can continue to learn and get better. Action needs to be taken, decisions need to be made, so we need the engineering to take those signals into those systems. The third thing, probably the most important and probably the place where the least amount of attention is paid, is in the area of design. Design is the notion of embedding human empathy into the applications we create. These solutions are ultimately created for people who need to use them and benefit from them. Design means understanding what the problem is as seen from their lens, what success and value look like again from the lens of the human, and what they care about in terms of the decisions and outcomes. I want to share another idea: what are the dimensions of AI transformation? As we look into our organizations and ask how we apply AI, we believe there are five pillars where we can apply AI. Number one is to use data to drive executive decisions and operational decisions better, providing insights as needed in an easy-to-use form. Number two is to improve consumer engagement. Every business is ultimately serving consumers, and growth is a function of how well we engage those consumers. AI can be used to deliver better value in terms of customer servicing, customer experience management, acquisition, retention, sales, marketing, and such. Number three is to eliminate inefficiencies and increase productivity. Every organization has inefficiencies, and AI can help reduce and eliminate them over time. Number four is the notion of building better products and services. It takes a reasonably long time to assess whether products are going to be successful. R&D, product development, bringing product to market, and assessing success can take a long time. AI can help us get there faster, increase the success rate, and reduce the time from idea to the decision on whether the product is successful. The last concept is driving disruption and fighting disruption. Every business model is under attack from new business models. We've seen in our lifetimes, in a short period of ten years, the number of business models that have been completely upended and changed. Every business needs to find ways to disrupt their own business and to fight disruption. AI can help us get there. So all businesses need to be thinking along these five dimensions on how we bring AI, engineering, and design together to get better on all of these dimensions. I'm going to take a few success stories and then speak about what are the building blocks required within an organization to create success using AI, engineering, and design along these core dimensions.

Sure, yeah. Certainly, I think great question. There are a big range of things happening in the retail sector which could create a big impact. Let's take an example. One of the biggest drivers of sales in retail is availability at shelf. Many times retailers and manufacturers themselves do not know when products are not available on shelf. One of the ways AI is being used is through computer vision. By taking images of shelves and using AI, retailers can find out very quickly whether a product is on shelf or not. Imagine that today it may take 48 hours to figure out whether a product is on shelf, but now you're able to figure that out almost instantaneously. That can reduce your out-of-stock situation dramatically and help increase sales significantly. Retailers spend a lot of money on in-store promotions. AI and advanced analytics can help us figure out which kinds of promotions work best at the store level. If you're able to make better assessments, it will help manage the entire supply chain and ensure better product availability. I'll give you one live example. We were working with a big departmental store that wanted to drive growth in their DIY tools department. DIY being do-it-yourself home tools like drills and screwdrivers. One of the biggest drivers of what kind of products get bought is based on the expertise level of the customer. Some customers are novices, some are intermediate, and some are professionals. Based on the kind of customer, they'll buy one product or another. AI was able to classify these customers into their level of expertise, figure out which products these people buy, and then optimize the product mix, inventory, and promotions for each customer. Suddenly you have much better customer satisfaction because you've got relevant products in the store, relevant promotions, and great value for everyone across the board. So I think that answers Anantha's question very well.

Product launch decisions, yeah. Absolutely. If you look back, there's been a lot of work that has gone into understanding consumer segments. We know that's kind of infinite. Traditionally, analytics and market research have been helping us do that. Today, that's being taken to a completely different level. If you look at online streaming, someone like Netflix is essentially tracking everything we do on their platform. They know what parts of a particular program we forward, what we rewind, where they get the maximum audience or eyeballs. Just think of the amount of data they have on what features should they put in the show, what aspects of the stories people engage with. That's just one example of how products can be made and launched faster using AI. So I hope that answers your question.

Yeah, please. Let's keep the questions going through the sessions. It's a good way to keep us all engaged.

All right, so we spoke about these five dimensions of AI transformation. I'll take a few examples along each of these dimensions. If you think about the first dimension, using data to drive better executive decisions, consider a consumer products company that sells into millions of retail outlets serviced by thousands of salespeople. A couple of key issues they face: salespeople do not have a good reliable way of knowing how their performance is in a given month. Very often, a large amount of monthly sales gets done in the last week of the month because that's when quota pressure increases. When you have a disproportionate amount of sales in one week, it creates huge inefficiencies in the supply chain. Also, salespeople don't have a very good idea of what is the right product to sell in which store. There are hundreds of SKUs, and when they go into a particular store, they don't know the right product to offer. Very often, the products at the top of their sales booklets are the ones that get sold. We created an application called Cuddle, an AI-based BI assistant on your mobile phone that sits on top of enterprise sales data. It's a conversational interface where a salesperson can ask questions like: which products have I sold better, which geographies or trade areas am I doing better in, and get this information on a daily basis. When this company applied this to its sales force of thousands of people, they found two things: the people who used it were able to consistently deliver three to five percent better outcomes, and sales started getting better distributed across the entire month, which meant a big relief on the overall supply chain management and better realization. Because of the recommendation engine, they were able to sell the right products in the right stores. This is a great example of using AI to power decisions for salespeople who need to make decisions every day on whom to call on and what product to offer, with superior outcomes for the whole company and the individuals involved. I'll take another example, very recent and topical, about the COVID pandemic. When COVID came in, most governments and city councils had very little data on how to deal with it, very little understanding of the pandemic. We created a command center, or a war room, where data from all sources was engineered: data on travel, rate of spread of the pandemic, rate of hospitalizations, etc. Through that, we created various visual interfaces for government officials to know things like the rate of spread, where it is spreading, the rate of hospitalization, and available hospital capacity. The government was able to make very rapid decisions on a whole range of things, such as which areas are high risk, where to deploy resources, and how to ensure we never ran out of hospital capacity in a particular location. This is the first part of that five-pillar example: using data to drive better executive decisions. Two very different examples: one in public health and pandemic, the other in consumer goods sales. I want to move to the second area: improving customer engagement. Through this pandemic, but over the last 15 or 20 years, especially in the last year, the amount of digital interaction has increased tremendously. One of the biggest problems customers face when they come to digital platforms is digital friction. They are trying to achieve an outcome through the website, mobile platform, or call center, but they can't because there is some problem in the journey. Maybe some data is not available, certain links are not working, some part of the website is confusing, or the IVR has put them into loops. This friction is highly frustrating for customers and loss-making for the company because the consumer cannot complete their journey. It's very hard to detect this friction because there are thousands of pages and millions of journeys. We brought AI into this situation where all digital data is engineered, and we put in place a range of anomaly detection applications that can point out anomalous customer behavior and potential reasons for it. This is served up as a digital cockpit to the business. On a daily basis, they can figure out the places on their digital platform where friction is highest and fix them. For a hundred billion dollar global telecom major, we were able to see that by eliminating or reducing customer digital friction, digital revenues increased by up to 100 million dollars a month, along with a big jump in overall customer satisfaction. I'll move to the third notion: eliminating inefficiencies and increasing productivity. In petrochemicals and oil and gas, we deal with very large infrastructure all over the world. There are leakages in pipelines. For one particular company, this could translate to as much as two billion dollars in lost product. These leakages do get detected, but it takes a lot of time to find them because there are thousands of such assets. By using AI and anomaly detection techniques with live streaming data, detection time was reduced from 48 hours to near real time, allowing immediate action and saving hundreds of millions in lost productivity. I'm giving intentionally diverse examples. In healthcare, up to 25% of X-rays are misdiagnosed, not because of any fault of the radiologists, but because many things on X-rays are not discernible by the human eye. A product like Cure applies computer vision and deep learning algorithms to X-rays. The algorithms have been trained with millions of X-rays and can detect a range of diseases and anomalies. Today, we're using this in parts of the world where there are lots of X-rays but radiologists are not available. A product like Cure can provide diagnosis immediately. We've also been using it to look at COVID progression, detecting COVID and assessing whether it has reached the lungs. The last thing I'll give a couple of examples on is fighting disruption. One of the things AI enables us to do is to look at sources of value beyond the obvious. For example, these are satellite images of parking lots. Investors are using these images to make assessments on how a retailer is doing. Well before the retailer releases their earnings and volumes, this is a very good signal into the traffic flow into the retailer, and that can be used to assess how well a retailer is doing. Small signals and data can create big advantages. Then look at a completely different and disruptive application: Amazon's Go store. You walk into the store, the store recognizes you, your payment information is already with the store, you pick up products, put them in your cart, and leave the store, and you automatically get billed for it. Just an idea of tremendous disruption that can be created using AI, engineering, and design. Each of these opportunities and each of our industries have these opportunities. Someone is out there trying to disrupt the business model, and we need to fight that disruption using AI proactively.

Yeah, so look, some of the best technology products we are using are examples of great design. Great design makes the technology intuitive, easy to use, and takes all the complexity out of the problem. The best design example is Google. You go to Google, you type a question, and you get an answer. There is a deep insight there: consumers just want to type a question and get an answer. They are not concerned about any of the underlying technological complexity. That insight about the consumer translates into great design and into a product that becomes widely adopted and widely used. That is the power of design thinking: developing that deep insight and empathy into the consumer's mindset. Look at other great technology products like the iPhone. It has a lot of AI, but all of that is wrapped into very beautiful design. As Steve Jobs said, design is a funny word. We often think about it in terms of how things look, but it's really about how things work.

So he's asking how to ensure AI-led transformation integrates with the existing tech stack. Without getting into too much context, let's think about what we need for AI. We need to know what we're trying to achieve from AI in our respective organizations. You can think of it as use cases: I'm trying to do better pricing, better forecasting, better customer engagement, and so on. We need to know what data we need to feed those AI applications. We need to know how the insights coming out of those AI applications will be used in the decision-making process. Those systems may not be able to do the AI, and that's totally fine. We always have the option of taking data from those systems into more modern applications, whether on cloud or on-prem, applying the algorithms. People are using Python and all of that, but those don't reside in our traditional transaction systems, and that's not necessary. The real thing is that the insights or recommendations from AI applications need to be available in the interfaces that people use, whether that's my CRM system or Salesforce system. It's a complex topic. Application of AI does not mean we need to let go of all our existing technology systems, but we may need to add some systems to do some of the work that is needed.

Yeah, great question. There are applications where explainability is very important. In the world of credit, there is a big need to explain our decisions to ensure there is no bias. Companies still use highly explainable AI, which could be traditional logistic models and linear regression models, or even more advanced machine learning techniques like random forest, so that we can create explanations and know what factors are driving the decisions. But if I look at things like image recognition, there is really no explainability. Why does an application recognize one product versus another? It may not even be needed. All we need is that the AI gets better and better at recognizing those images. So there is plenty of ability to explain decisions, and there are plenty of algorithms where decisions cannot be explained. It's merely a function of what we want to do, and we can use the algorithms we want depending on how important explainability is to us.

Yeah, so cloud provides us with scalability, elasticity of usage of data, and easier ways to manage data and technology. That's the way the world is headed, and it is going to be an integral part of the world. Soon, being on the cloud will be the default mode of operating. We are finding that a lot of AI applications are being developed for the cloud first, and the most latest technologies are available there. So that's an integral part of the future.

Yeah, it is a very industry-specific question. Each industry has its own challenges and needs. But if we apply this lens to any business, we can figure out what is important in that business at that given point of time and what is going to create value. I would say though, if we go industry agnostic for a moment, there are a few things. Number one: e-commerce and digital interaction has gone up tremendously, leapfrogging several years ahead. Therefore, every company today is interested in how to improve digital interaction, online conversion, and digital supply chains, and how AI can help do all of these things better. Number two: because the economic landscape has changed so much, our old models of predicting and forecasting have become irrelevant, and everything needs to be done afresh. There is a lot of interest in building new consumer forecasting models. The second thing is that these changes will keep happening, so there is a lot of interest in auto ML, which is the ability to automatically refresh these models as fresh data signals come in. So broadly across industries, these are two big things we're seeing. The third thing is in the area of employee safety, welfare, and productivity in the world of remote working. Companies have used AI to ensure that there is no disruption in their delivery to their customers.

So he's concerned about the integrity. It's a great question, and it's certainly one of the big risks and big issues we face today. Yes, and I think there is tremendous potential for bias in AI. For example, Amazon started automating its resume screening and selection process, and they found that the algorithms were discriminating against certain sections of people, like women and potentially minorities. They quickly discovered that the reason was because the AI was merely mimicking what had been happening in the past. In the past, when we were screening resumes, we had biases, and the AI had simply taken that as the truth. There is also the possibility of intentional bias. A parallel example: people have been suing some companies because their employment advertisements have been discriminatory. In certain cases, people above a certain age were not shown ads for jobs, meaning they didn't have access to the knowledge that a job was available. That was intentional. So overall, there is a lot of potential for bias, and there is a need for a lot of public policy and intervention over here.

Yeah, yeah. This was a part of the story we were not able to get to. When you think of organizations, if you want to deploy AI within the organization, think of the why, what, and how. The why is the five dimensions to see what's relevant to your organization. But then the how becomes really important. The how is the architecture around this. Probably one step before that: what is the intention from the top? What is the support at the board level, at the CEO level? Is this critical to the company as seen by the leadership? What about the org structure? Is AI in a prominent place in the organization, or is it buried under six different layers where it has no visibility or power and authority? The kind of people we bring in, the hiring we do, the knowledge systems we create, processes, tools, and technology. All of these are important to make AI successful within the organization.

Yeah, and look, if people want to stay on for a few more minutes, I'm happy to do that and answer some questions. I just need a minute over here if you don't mind.

Assesses speed relative to the environment. So we came up with this idea of how do you communicate speed. The idea was to paint the railway tracks with yellow lines. You know that this is the point at which people cross. Up to half a kilometer before that, you paint yellow lines in the track, leave some space, paint yellow lines again, leave some space. As the train is approaching, you see those yellow lines disappearing, and they disappear faster. It communicates to the mind subconsciously the speed, and people stop crossing. This was a very interesting design thinking example.

Yeah, so I think why COVID has spread quite as much is a fairly complex topic. The medical science of it actually has been quite advanced. The genome of the virus was released very quickly. China released the DNA code of the virus, and the vaccine got designed in two days. It was not possible to do this 20 years back. The whole idea that there is so much AI and computational ability available to us, we are able to do that stuff. Just recently, about a week or two back, the really complex problem of protein folding was solved by AI, and it's a huge breakthrough because there are millions of ways in which proteins can fold. To have a vaccine in less than nine months is tremendous progress in science. I think we know scientifically that wearing masks helps, social distancing helps, but that's about the extent of the science. We can know through AI where there is likely to be bigger spread and more transmission. So really, the failure to do better is not of the science or the AI; the failure is in human communication and behavior.

Good question. There is some small body of work we're doing from a standpoint of content. There are some possibilities of teaching and courses, but really not much more beyond that. There is opportunity here to do more with IIMs, to do some research, work on live projects. IIM gets a lot of grants and live industry projects from both industrial and government projects. There is some opportunity for us to explore together, but the honest answer is today what we are doing is well below potential.

Yeah, great. Thank you, Brendan. Have a wonderful day. Bye.