The emergence of Generative AI (Gen AI) has sparked a significant transformation within the retail landscape, wielding an unprecedented impact on various facets of the industry. From revolutionizing customer experiences to optimizing inventory management, Gen AI’s integration has ushered in an era of predictive analytics and personalized shopping, fundamentally reshaping how businesses engage with consumers and drive growth.
To give us more insights into this, we have Navin Dhananjaya, the Chief Solutions Officer at Merkle, boasting an extensive 27-year career focused on steering and optimizing large-scale data, analytics, and technology operations for Fortune 1000 enterprises. With a pivotal role in conceptualizing and crafting cognitive learning platforms, Navin spearheaded the adoption of AI at Merkle. This groundbreaking innovation harnesses the prowess of generative AI to revolutionize customer experiences, exemplifying Navin’s commitment to transformative solutions in the realm of data and AI-driven strategies.
AIM: Can you explain Merkle’s interest in Gen AI? Until recently, Gen AI wasn’t a focal point in discussions about language models and chatbots. What sparked this interest in Gen AI, especially in the retail sector? Understanding this context will guide our discussion.
Navin Dhananjaya: In Merkle, we have always been at the cutting edge of technology, whether analytics or any other technology platforms, leveraging these technologies for greater customer impact. From a Gen AI perspective, it naturally lend towards what we’ve done on analytics, advanced analytics, AI, and now Gen AI. So, we’ve been constantly evolving, and that’s our next step. And it goes well for our business because a lot of our business involves consumer insights. Many of our stakeholders are from a marketing or product merchandising background, so many of these use cases work well for them. That’s one of the reasons why we’ve leaned into it much earlier than some of our other counterparts.
AIM: In the early stages of retail implementation, what immediate use cases or wins are you observing with Gen AI? What areas within the industry show immediate potential for leveraging Gen AI’s capabilities?
Navin Dhananjaya: Retail is a significantly large sector in which Gen AI can make a significant impact. But from where I come from, there are a few things with respect to, say, if you start with retail, and the retail products or the retail sku. So, right from how our retail sku is like, if you look at the digital world, how our retails are described for all of us to transact with – how the product image is there, what kind of content it is associated with, the entire video, the description, the classification. In terms of any product, at least 100 to 150 attributes go along with the product, helping discoverability of the product and better conversion.
With Gen AI, a lot of that can be maximized. So, if you have a million products online or if you have thousands of products online, you will be able to get the same kind of treatment for all those products at a much faster rate and change the treatment or the relevant content for all those products in a much faster manner. So, that’s the most simplistic and immediate use case that comes to my mind and is very apt for retail.
AIM: How does implementing Gen AI differ from traditional AI in existing initiatives? Beyond implementation, what drives the decision to choose Gen AI over conventional AI solutions? What exactly does this difference entail in practical terms?
Navin Dhananjaya: The example I just mentioned is about something other than Gen AI coming in and solving or creating a new paradigm. It’s reimagining how you represent the product online or reimagining all consumer touchpoints from a Virtual Assistant perspective. When I talk about reimagining product or product content, I mean you always used AI to generate product content. Now, generative AI will add to that, giving a more personalized feel or improving that experience per se. From a generative AI perspective, it can aggregate a lot of outside-in data, whether it is market data, and it can help you test concepts or the entire content we were going to produce. Unlike just AI, you get content that you can take in, or you’ll have many more ideas to choose from, augmented with Virtual Assistants assisting the category manager or the end consumer of the product, enabling conversational choices to experiment with. Those differences are coming together. So, it’s the ability to visualize, make it more personal, and have those conversations. That’s essentially taking whatever applications we had with AI to the next level. And that creates a significant competitive advantage, especially in an era where customers expect almost a singular interface to aid their final decision. So, generative AI can help accelerate that process.
AIM: Could you guide us through a specific use case and discuss the interplay between data science and Gen AI? How were these elements utilized in the process, allowing us to delve deeper into what implementing Gen AI for this use case entails?
Navin Dhananjaya: Sticking to the same example of retail products, any retailer who wants to be in the digital world wants their products to be appropriately discoverable by customers so that conversion happens faster. Let’s take that as a simplistic use case where data sciences or analytics help better classify product categories. So many classifications are available online, so you need to have the product in the right category. For example, running shoes can be in sportswear, fashion, or regular athletic wear. Which is the right classification where running shoes should be featured so that it is more discoverable? And what should it be, and at what price point should it be? And how many variants of the product should it be? What kind of elasticity does the price point offer? All of this analytics.
In terms of getting from here to AI, AI can help with respect to when you’re doing competitive analysis. You would want to compare the shoes with similar shoes in the market. So, to get to those shoes, you will need to do image or text matching. And that’s where AI helped us a lot. From a generative AI perspective, you can have conversations based on what you want the product catalogue to be. It is more interactive to say that it suggests the best form of the visual for a shoe compared to what I have already. That’s one; otherwise, you can interact with your models based on all its learnings. Based on when we talk about LLMs and when you talk about large knowledge models, it can aggregate all the knowledge you have internally as well as what is there in the marketplace to suggest what the right description is. And that’s where generative AI is going to take place. All these three processes are evolving. Many retailers are not even at step one. But the other significant advantage of Generative AI is that you could directly jump on to step three and leverage a lot of learning outside the market, plus get a jump start on this process.
AIM: In Gen AI’s application within retail, the emphasis on domain knowledge alongside AI and data science is crucial. How is structuring teams for Gen AI development in retail distinct from previous methods? How pivotal is domain expertise, and what skill sets are prioritized for these teams?
Navin Dhananjaya: We’ve always structured teams in three specific boxes. One is the expertise in technology, which could be analytics, advanced analytics, or AI. People are adept with models or model building or referring relevant global models and conducting transfer learning with our data for increased relevance. Then there’s a team for platforms. They gather all this expertise and put it together on a no-code, low-code interface platform. It’s a technological platform.
Following that are the operations teams that access this platform. Essentially, we’re making AI accessible or aiming to make generative AI accessible to a broader audience. Only some people need to have in-depth knowledge of technology or data science. This approach spreads knowledge to more people by having an interface. Similarly, with generative AI, if we have a lot of product catalogue intelligence previously used for image matching, in the generative AI construct, we might use that to build extensive language or knowledge models. People can interact with these models conversationally to assist in decision-making. Our focus is always on making AI or Gen AI more accessible. That’s been our mantra throughout. These are the three buckets in which the team is organized. When we talk about large knowledge models, that model could be retail-driven, consumer-driven, or based on various areas with extensive data to generate those knowledge models.
AIM: Have any Gen AI models entered production amidst this rapidly evolving landscape? Two months ago, none were in production—everything was a proof of concept. Have you managed to reach the production stage with any? What challenges does your company face in this process, and how are you overcoming them?
Navin Dhananjaya: A significant chunk of a business is learn on this edge of AI already, so we use that construct to deliver to business. It’s been in production. I cannot tell the numbers, but at least 20 – 25% of our business is AI-driven in that sense.
On top of that, we have already started deploying for at least many use cases within that constructor. We are exploring because Gen AI is also the one thing there is an opportunity to step back and discover as well: where else could you have used Gen AI? So much so that even our development for the AI models, our entire development environment, is something they’re using Gen AI for. So, somebody who writes code once or twice. I mean, the third or fourth time, the IDE – integrated development environment itself illustrates what course should be written in the build. From that level onwards, we are building Large Knowledge Models. We’ve identified these five or six areas where we’ve built Large Language Models. Now, they are in semi-production. And why I say that is because from a client context or a deployment context, there are two kinds of challenges that we want to overcome before we take fully into production; the first thing is, how much dilution is there from an external world perspective so that, it is more contextualized to the implementation that we have. Secondly, we also want to ensure that when we deliver services to a client, the sheer fact that these knowledge models can learn a lot more from environments. We need to ensure that these models can sift through cross-customer data. So we need to draw walls between them so they don’t learn that customer A doesn’t learn from customer B. We need to keep that isolated. So, the weaknesses are in the strengths as well. So, the strength is the capability to sift through large, vast bodies of text knowledge. But we need to ensure that these interactions are limited to the clients that we’re working with. So that’s a boon and a bane in the Gen AI model. So, some of the challenges we need to overcome before we look at these large language models.
AIM: What organizational changes do enterprises need to consider, considering the points you just mentioned?
Navin Dhananjaya: Enterprises need to ensure that their contracts are more tightened when they look for service providers regarding where the data exists, where the knowledge exists, or where the language interactions and the technology extension sit. So, at the same time, they should not make it a barrier for real interactions. Finally, it’s the conversational nature and the extension of decision-making that matters the most, so they should ensure that their line of business managers leverage this and can augment their current work along with what Gen AI can do.
AIM: Navin, any closing thoughts?
Navin Dhananjaya: We’re just beginning; I would say it’s about this path, a lot more unexplored. And it is on a complete step function jump, or paradigm shift of sorts, where we can now integrate Gen AI with many of what we have with AI tools available to the customers. So, there are a lot of tools; we work a lot with Salesforce as an example. Our integration with Salesforce and Einstein GPT can result in significant business outcomes. That’s one way to look at it. Then, we also have our identity solution, Merkury, which covers almost 95% of US demographics. From that perspective, we can build better, more contextual large knowledge models that can enable a much more granular understanding of customer behaviour, enabling human-like connections at scale. Those are some of the chapters that we have. Likewise, concerning consumer insights, just the entire thing about how you do primary research, how you do the hypothesis testing of a product in the primary research, and how you develop that questionnaire itself can be conversationally got out of a knowledge model, one or had via conversations itself rather than research. At least, all these three examples I took are areas where we are furthering our Gen AI development. That said, in any area you explore, there is probably an application for AI or Gen AI. We are just discovering a whole plethora of use cases.