Episode Summary: One of most fun parts about doing our geolocation pieces at TechEmeergence is that we are able to interview so many people within a given country or city. Recently we did a huge piece on AI in India. We got to interview folks from the government and the bigger existing businesses, as well as a handful of people at the unicorns in Bangalore.
One of those companies is Fractal Analytics. Fractal Analytics works in a number of spaces. One of them, consumer packaged goods, is an area on which we haven’t done much coverage. Many of our readers are in the retail space, but CPG has some pretty curious AI use cases.
This week, we interview Prashant Joshi, Head of AI and Machine Learning at Fractal Analytics, about the different applications of machine learning in the CPG sector: doing chemical tests or finding new buyer segments within existing groups of consumers to determine who is buying from a company and who is buying from competitors.
Hopefully, for those in retail, this interview will not only highlight some of the interesting use cases of AI in the CPG world but also provide some ideas about winning market share from what some of the bigger CPG firms are doing with Fractal Analytics.
Expertise: machine learning, computational neuroscience
Brief Recognition: Joshi holds a PhD in Computational Neuroscience and Machine Learning from Graz University of Technology. Previously, Joshi served as a research fellow and a computational neuroscientist at Frankfurt Insitute for Advanced Studies. He was Head of R&D and Senior Principal Data Scientist at [24/7].
Subscribe to our AI in Industry Podcast with your favorite podcast service:
(03:21) Where is AI making a difference in consumer packaged goods?
PJ: I think in CPG there are a bunch of factors that make things interesting as an industry. One thing is the maturity of the CPG industry in terms of not being tech savvy or ready to adopt AI. There are lots of low-hanging fruits that can be worked on using AI and machine learning to deliver transformation. From the CPG world perspective, the possibility of growth is phenomenal using machine learning.
I think some of the key areas where we are seeing a lot of traction using machine learning in CPG is, for example, something called shelf-share analytics. The idea is to figure out what percentage of a retail store’s shelf is occupied by a CPG company’s product. There are certain compliances that the CPG company agrees on with the retailer. The idea is to figure out if those compliances are being met or not. To do that the old fashioned way, one would go to the store, take a photo, come back, and measure it using a ruler. Now using deep learning and image analytics, we have been able to develop a SKU that allows you to not only figure out what percentage of a shelf is being occupied by the CPG company’s product, but also whether everything is in compliance or not.
The other interesting thing that comes to mind is the recent work that we did for another CPG company. The idea was when people suffer from dental sensitivity, only 3 out of 5 people reach out to a dentist for a medical intervention. What happens is usually, when you go for a medical intervention, you’re given a certain brand of toothpaste. [This brand] was owned by the competitor of our client. What we figured out is that the client’s competitor had complete market share of the people who are going to the dentist for interventions. How about we figure out a way to figure out who are the 2 out of 5 people who have dental sensitivity but don’t report it?
The idea is, “Can I use machine learning to figure out unconventional ways people are talking about dental sensitivity? If I can do that, I am going to be able to create my marketing campaigns around that.” So the idea was that we are going to find out [through Amazon reviews] if the customer is talking about something [in a] very alternative way of talking about dental sensitivity. We found that people who are going through chemotherapy have a much higher probability of having dental sensitivity…We consumed a hundred gigabytes of Amazon reviews. Based on that, we could figure out these nontrivial fringe topics that involve dental sensitivity. The big win was our client was able to launch a new product and capture this market.
(08:43) What do you do when you figure this out?
PJ: All of this allows you to create very targeted marketing campaigns so that your conversion rate goes up. If I’m able to find a microsegment segment that has a high propensity of taking a new product, I’m going to create and advertise one that is going to appeal to that microsegment. There’s a bunch of these interesting things happening in the CGP space. A lot of CPG majors are getting the idea that social media and product reviews might be a good data source for doing market research at scale.
Essentially, if I look at the last 4 to 5 years of machine learning transformations in every industry, they’ve been (1) the realization that we are leaving a lot of data on the table that are unstructured in nature—text data, speech data, image or videos—and, fortunately, this is the time when we had algorithms that could handle the unstructured landscape. Having the realization that we have to work with this data, and at the same time having new algorithms that can consume and leverage the data has lead to a lot of disruption in the last 4 to 5 years.
(13:43) How does share of shelf space application get applied?
PJ: This is not a win-win situation for the store or the CPG company. Essentially, the way it happens is the CPG company will get into an agreement with the retailer that they will need X-percentage of the area to be covered with their product. The objective of the retailer is not to sell the CPG company’s product, but to maximize its sales. So then the issue is, “To what extent are the compliances of self-share being met by the retailer?” This is a situation where we need to go out and take a look at it. The difference is huge. Right now it has been as simple as you upload a picture, and within a matter of seconds, you have statistics about what percentage of the shelf is being occupied by your product and whether your shelf is [meeting compliance].
For example, let’s say your product has to be stacked vertically, and the retailer has been stacking it horizontally—that can be figured out. This is phenomenally complex in terms of the algorithms that make it work. This is highly computationally intensive. To develop such a SKU requires a huge amount of investment and research, but the ROI is also phenomenal.
Whenever a CPG company has to launch a new product, let’s imagine a toothpaste. First, when you are done with creating the product, you have to pass it through a bunch of experiments where you have to figure out different aspects of the product: what is the durability, how long will the chemicals stay active, how will the packaging respond under different temperatures. Essentially, there are about 440 experiments that have to be done before launching a product. Each of these experiments lasts around 8 to 13 weeks. The problem here is many times it happens that by the time you finish the experiment your competitor launched the product.
The hypothesis that we started with is, “Do I really need to do all these 440 experiments?” The idea is that if I can use machine learning to tell you within 99.99% certainty the outcome of a particular experiment, then you don’t need to do that experiment. I have reduced the need of doing the experiments at scale. We looked at all the historical experiment results; then we used multiple algorithms to predict the outcome of new experiments. The big win here is we were able to reduce 12% of the 440 experiments; we simply knew the company didn’t need to do those experiments anymore.
Subscribe to our AI in Industry Podcast with your favorite podcast service:
Header Image Credit: Strategy