Putting the AI in Retail: Exploring its full potential with EPAM and Aprimo

My name is Ana. I work for Aprimo, the leader in AI-led content operations. I head up our global strategic partnership program, which means I work closely with all our clients so the different solution implementers and technology vendors best support them."

“I’m Val with EPAM Systems. EPAM Systems is a global digital transformation services and product engineering company. I’m one of the leaders in our data science practice. I am responsible for thought leadership around the practical use of AI in consumer goods and the retail vertical.

Val: "When you think of industries pushing the boundaries of AI, retail may not be the first to come to mind. But it is a very mature industry. If you think about Macy's being 165 years old and Nordstroms being 120, you see they have been on the cutting edge and early adopters all along. Retail companies were among the first to use business intelligence and advanced analytics. Then they started using AI automation in almost every part of their value chain. Today you can find AI in intelligent demand forecasting, risk modeling and social listening. Retailers use AI to optimize distribution and pick where the next store location should be. They use AI for smart inventory management, and so on."

Ana: “Everything we do in marketing is driven by the cost and skill required to create high-quality content. You have to think about how many campaigns you can do, how many different channels those campaigns will operate on, or how many segments they will be personalized against. When you're putting together your marketing plan, all of those questions are governed by how many designers you have, how long it could take to create the content and how big your agency budget is. What's interesting about generative AI is that when you start incorporating it into your forecasting model, you know that you could have high-quality output from the get-go, which helps from a productivity standpoint. Still, it also might take far more tries to create the content. And what's interesting is that even though that is extremely nominal, like when you're using ChatGPT, the variability around what it takes to create high-value content will make your ability to budget and forecast more complex for marketing organizations.”

Val: "When Microsoft commercialized chat GPT making it available to the general public, what it did was it showed people a different way to search and consume content. A much better way. Gone are the days when it was okay for a consumer to spend 10 minutes crawling, playing with search keywords and hoping to find the perfect t-shirt. The expectations of people have changed, and it’s beyond the point of returning, and that is a good thing. Since there is no e-commerce without content in search, disruption is inevitable. Retailers and consumer brands are using generative AI for hyper-personalization. They also use it to improve social listening and demand sensing, which enables better segmentation of customers and their needs. That, in turn, helps them design tailored reward structures and loyalty programs, which is a big deal for retailers in 2023.

"Another thing that I see across the board is a plain English interface to discover products to buy. The proper term is semantic search, and it is one of the most popular use cases along with content localization and translation."

Ana: "There are two really interesting use cases that I've seen already Incorporated with some of my clients. One is in financial services, but it still translates to everything in retail. It illustrates a quick productivity win and the ability to personalize at scale. This client trained up a ChatGPT extension within their DAM on collections of high-performing email content. Email blasts that are already brand specific to them and on point for their tonality have been sent out in the past. They have their email marketing teams create the base copy and feed it into the ChatGPT extension to create 12 different segments for all the different personalities they're targeting. This is a quick use case; everything gets routed off for final approval, so you still have that extra set of eyes against it before anything gets sent out.

“The other interesting use case is with a large retail client. Val mentioned this one earlier. This retailer trained up a ChatGPT extension on a collection of product-specific content, and then they invited that chatbot into their campaign ideation and brainstorming sessions. The result is a scenario where their product marketing and e-comm teams can pose questions to the chatbot, and it will answer from the product's point of view as if the product has come to life.”

Val: "The work we do at EPAM with AI is not just with content. Depending on who the buyer is and what kind of business problem we're trying to solve, we can engage with the data science departments, the digital product leads, or somebody in charge of e-commerce. Pretty much every business person is trying to understand what AI is and what they can do with generative AI that will make life easier and deliver a better consumer experience."

Ana: “Even internally at Aprimo, our CEO mandated that we all start incorporating Gen AI into our everyday tasks and routines across every department. It's interesting to see the different use cases that crop up across our departments. The sky's the limit. Many people are focused on just those initial productivity gains that they can see by incorporating Gen AI into their day-to-day work. It does span every single different department.”

Val: "Companies are still trying to figure out economical ways to use GPT. Because it is so fresh, the value proposition isn't always clear. It all looks appealing, but customers still have lots of questions. I tell my clients to think about scale and think about data. Productivity is a no-brainer; by all means, get productive right away. Just be mindful about sharing confidential data and how you want to scale your use cases. You need to consider how you will apply responsible AI best practices from day one. AI can be biased, and you need to put some effort to ensure it's ethical and represents your brand values. My final advice is don't do it alone. Bring in partners, especially those that have subject matter expertise in natural language processing. Let them assist you when you work on solutions you will push at scale."

Ana: “I've accompanied our CEO during his Gen AI briefings to several of our major clients. Regardless of their industry, they all seem to adopt a similar approach. Firstly, they establish a Gen AI task force comprising various department members. This task force then formulates a policy, guiding employees on responsible Gen AI use in their daily roles. Secondly, companies pinpoint low-risk, high-opportunity Gen AI applications as proof of concepts in response to the frequent question of where to begin. This helps to showcase the ROI, such as in Val's examples which contrast the creation of a campaign brief with oversight against sending out an email to the entire customer base without any checks.

“Additionally, the client using the product chatbot has implemented a checkpoint system within their existing technology, Aprimo. This is used to flag content influenced by Gen AI, acting as an enforcer of their policy. It's not meant as a punitive measure for employees but rather as a guideline, emphasizing that Gen AI is not something to be feared.”

Val: "From a crystal ball standpoint, the priorities for retailers and consumer goods in 2023 and 2024 are loyalty programs, personalization, commerce and sustainability. They will be seeking ways to integrate gen AI to advance these objectives. From a wishing well perspective, I hope for disruption in chatbot automation, particularly in eliminating tedious phone menu navigation, especially with government agencies. I hope this transformation extends beyond retail and redefines our interaction with customer service across all sectors. Additionally, I look forward to gen AI automating mundane tasks."

Ana: “Interestingly, Val and I share the same enthusiasm for chatbot automation, particularly in call centers. My initial exposure to its potential was at Aprimo, where our Chief Information Security Officer implemented a chatbot trained on our security policies and compliance documents. This innovation reduced the time and labor spent on compliance requests and security reviews by 60% and 50%, respectively. Translated to retail, such chatbots can significantly decrease call center traffic and empower customer success representatives with instant access to FAQs. If these bots are client-facing, they enable brands to address customer queries directly, enhancing the overall experience. I'm eager to see how gen AI can be harnessed to innovate digital in-person experiences, as this domain has vast potential for creativity.”

Ana: "Generative AI for imagery has advanced with stable diffusion models like Midjourney, and DALL·E, surpassing the older generative adversarial networks. Initially developed for noise reduction in images, diffusion models were later adapted to generate images. While they can produce artwork reminiscent of Rembrandt, they currently struggle to create specific designs, such as a retail shelf with specified dimensions or generating clear text. DALL·E, for instance, can assist in ideating product designs, but it cannot perfectly replicate a detailed product design spec or convert text to 3D. However, many organizations are actively working to enhance these capabilities.

Regarding the GPT's tendency to "hallucinate" or create content that isn't factual, the reasons stem from three main factors. First is the data it's trained on, largely sourced from the vast information available on the internet; the data is not always accurate. Second, the mathematical method used by the language model in generating responses has its limitations. Third, the prompts users provide to the system can be difficult to understand.

What do you do with hallucinations? It's unavoidable because it's a feature that looks like a bug. But what you can do, and what you absolutely should be doing, is to work on minimizing it. You can't avoid it entirely. But you can minimize hallucination to less than 1% of responses."

Ana: "EPAM, Aprimo and Contentstack collaborated on an intriguing project. A detailed blog post delves into the technological aspects of it. The essence of the project involved utilizing Gen AI to enhance the output from our systems. Users could take a picture of themselves and upload the headshot to Automation Hub, which then routes it to Midjourney. Midjourney transforms the image into a personalized superhero or supervillain. After the transformation, the experience is saved in Aprimo, transferred to Contentstack, and the user receives a trading card, complete with a backstory based on their job responsibilities. This engaging experience not only showcases the capabilities of gen AI but also demonstrates the synergy between different parts of the tech stack. Interested readers can refer to the blog for more insights into the decision-making process, especially the choice of Midjourney over other image generators. We also have a link detailing the user setup, which has provided much amusement for our team."