From ‘conversational’ shopping to AI-generated images of concertgoers holding your beverage brand, generative AI is going to transform every aspect of the food industry, Amazon exec Justin Honaman told delegates at the virtual ‘Generative AI Food Pioneers’ summit hosted by Israeli startup Tastewise last week.
And businesses that don’t engage now risk being left behind, added Honaman, head, worldwide retail and consumer goods GTM at Amazon Web Services.
Peppering his slides with questions such as, ‘What are you doing today so you are not obsolete tomorrow?’ he went on to tell viewers that generative AI would be “as big as laptops, as big as the internet, as big as the light bulb, as big as the printing press.”
He added: “The reason why many food and beverage brands, CPG brands, and retail brands don’t make it is because they don’t innovate. They get stuck in their own ways and they can’t move quickly enough to stay relevant.”
Acknowledging that implementing projects may require new skillsets including design thinking, prompt engineering, critical thinking, data literacy, and change management, he added: “I’d just encourage you all to get going and get started on an experiment.”
And it can’t be left to the IT department, said Honaman, who was joined by speakers from Meta, NVIDIA, ADM, Conagra Brands, Kraft Heinz, Bell Flavors & Fragrances, Waitrose, and Tastewise. “If you’re a businessperson, and you’re like, ‘that’s IT’s job.’ No, it’s your job. It is not just IT’s job to figure this out.”
He added: “It’s said that 50% of employees will need reskilling in the next five years [due to this tech], I think it’s actually higher than that, in fact it’s probably closer to 70, 80 90%. Last year, every day I was on calls with CMOS, chief supply chain officers, and chief legal officers, not just the CIO and CTO, but board members of several major brands [keen to integrate Generative AI tools into their businesses].”
‘Traditional’ AI
At Amazon, said Honaman, “What I call traditional AI is already built into many parts of our business, so we use it for product recommendations, image analysis, managing fraud-free ecommerce, and contact center solutions.
“AI is built into the robots that move product around our fulfillment centers, machine learning models are running on Alexa devices, while in our ‘Just walk out’ technology [in Amazon stores] we’re using computer vision and sensor fusion to build virtual [shopping] carts. Same thing with Amazon One where you use your palm for payment; AI is built into the ML algorithms that enable that.”
AI is also helping shoppers to search using an image in order to find similar products online, while virtual try-on (VTO) experiences for eyewear enabling users to see what glasses look like on their own faces are significantly increasing conversion rates.
Generative AI: From Rufus to diffuse to choose
As for generative AI, he said, Amazon has been doing a lot of work with a text model to create concise summaries of customer reviews, highlighting key points and common themes to help shoppers quickly grasp the overall sentiment about a product without having to sift through scores of individual reviews.
“We have many customers that have hundreds of reviews on their own ecommerce platform,” he added. “As one of our customers, you can have a model run against that data today and get some insights from it.”
‘Diffuse to choose,’ meanwhile, is a new tool built by the amazon.com retail team that deploys a diffusion-based image-conditioned inpainting model to seamlessly blend product images (e.g., a tee shirt) into users’ personal photos, creating realistic visualizations allowing customers to virtually try on products in their own environments, said Honaman. “There’s lots of work happening on images and the ability to manipulate images.”
Stable diffusion open models—machine learning models designed to generate high-quality images from text descriptions—also enable users to select an item of furniture, for example, and then see what it looks like in different environments using written prompts (see below).
Another recent innovation deploying generative AI is Rufus. Launched in beta mode to a small subset of customers using Amazon’s mobile app in February, Rufus is a virtual shopping assistant trained on Amazon’s product catalog and information from across the web to provide a “conversational shopping experience,” said Honaman. This will help users find and buy products through natural language interactions, improving the online shopping experience with personalized recommendations and answering product-related questions.
“You’ll see this kind of thing continue to grow not only with us but I think more broadly. Personalization is going to evolve very quickly in the marketplace because of AI and generative AI will be a big part of that.”
Ditch the agency, try DIY with AI
According to Honaman: “Many of our customers are already well into testing with generative AI, so one of our customers wanted to build photos of people at a concert holding its beverage products, for example.
“If you’re in marketing, brand management, category planning, shopper marketing, shopper insights, and you’re [currently] using an agency to do this kind of thing, you could be doing a lot of it yourself with generative AI far more quickly.”
Likewise, he said, “Say I want to revise every one of my product descriptions, titles, bullets and ad words on my own ecommerce platform. I’m going to use a text model at Amazon to do that for me in seconds instead of hours and days waiting for an agency.”
How AI is impacting the bricks & mortar shopping experience
As for bricks and mortar retail, generative AI is also a useful tool to rapidly generate store-specific planograms [a diagram that specifies the placement of products on retail shelves to optimize sales], said Honaman. “You can take inputs from sales data, out of stock data, and sell through data and provide dynamic planograms individually to stores on the fly.”
He added: “If you’re in category management or shopper marketing or allocation planning on the store side you’ve got massive spreadsheets and pivot tables that you’re manipulating every day or dropping Nielsen IQ data into. You will be amazed at how quickly you can upload a spreadsheet and then run AI models against that and start mining it for information.”
In the last 18 or 24 months, added Honaman, “We’ve also seen significant uptick in interest in [using AI for] loyalty programs because of the availability of customer data and the ability to leverage that on ecommerce platforms and customer data platforms.”
According to Honaman, some other applications for generative AI in CPG include:
- Automating consumer segmentation at scale to tailor marketing initiatives
- Generating contextual marketing content based on unstructured data from consumer profiles and community insights
- Tailoring product return offers based on individual consumers
- Converting sketches, mood boards and descriptions into high-quality designs
- Customizing products for individual consumers
Conagra Brands: Using AI to gain consumer insights
At Conagra Brands, meanwhile, generative AI is helping the company better understand consumers’ unmet needs, said fellow speaker Megan Bullock, director, predictive science (emerging demand).
The company—which owns brands including Healthy Choice, Gardein, and Hunt’s ketchup—is working with Tastewise to generate insights from huge amounts of behavioral data rather than waste time and money on costly surveys that tell you what consumers say, not what they do, said Bullock.
“‘Survey’ is an off-limit word at Conagra, because we have transitioned over many years to more observable behavioral data. We’re now living in such a digital world that there are so many [online] breadcrumbs, if you will, of consumer behavior that can signal [what consumers want].
“They are searching on Google about what they want to eat for dinner. They’re creating recipes on social media. They’re communicating with their social networks. They’re shopping at the grocery store. And then we can [analyze this data and] ask what are the bundled sets of attributes of the products they’re buying? What are the shortcuts they’re looking for? What are their pain points?”
The GLP-1 consumer
She added: “We’re currently studying the GLP-1 consumer [taking appetite suppressing drugs such as Wegovy and Ozempic].
“We have data that looks at what they say they’re looking to get more of in their diet, more protein, more vegetables, more fruit, and less sugar. And then you look at the data tracking what they’re actually consuming in-home… and there is such a differential between what they say they want to eat more of and what they’re actually eating more or less of. And I think that is just a stark reminder of the importance of looking at behavioral data versus the reported consumer response.”
‘I hear people say, AI is going to take my job. No, AI is going to make my job easier’
She added: “Conagra has been embracing AI for quite a few years now because it can really help us synthesize datasets at scale and understand trends and patterns. But more recently, we developed an AI sub-function with our broader demand science team, which I think is going to help us build out more robust datasets and make smarter and faster decisions. Generative AI can quickly help us ideate off of this data and really bring it to life.
“Innovation cycles take a while but trends move faster and therefore the technology and capabilities to activate on the trends have to move faster. But I feel like we’re finally at a point where we have those processes in place to be able to move faster.”
It is also enabling staff to focus on more strategic, collaborative and rewarding tasks, claimed Bullock. “I hear people say, AI is going to take my job. I’m like, No, AI is going to make my job so much easier because it’s bringing an immense amount of efficiency to my day-to-day life.”
Using AI, she said, “I’m answering hundreds of business questions within one day versus if we were relying on survey data, we could commission a study which would maybe take two weeks, plus an eight to 10 week turnaround time, not to mention thousands of dollars. So this method is cost effective and efficient and fast.”
