Embracing AI-Driven Solutions to Combat Rising Retail Theft

I'm Christine Russo, and you're listening to what just happened on market scale. Hi, everyone. I'm Christine Russo. And welcome to what just happened. Today, we welcome Benoit Koenig from Vision. Vision is using AI to stop retail theft. Welcome, Benoit. It's very nice to meet you. Thank you for having me. I'm really excited because Faftian retail is such a problem has always been a problem, but we're really at a crisis point right now. We've been on the market for five years, and the the the major difference from five years ago to now is that shrinkage has become a real a real pain for retailers because five, ten, fifteen years ago, it was it was okay to have zero point five, zero point eight percent of sales in shrinkage because the margin were comfortable. Now with the inflation and with the increase, the the fact that shrinkage is skyrocketing, now it has become a real priority for retailers. And so they're looking for for new solutions. They've been using for years, security agents, cameras, RFID tax, and all of this is not sufficient. You know, they still see their shrinkage increasing from zero point eight to one. Sometime two percent of sales. This is huge. So just for clarification, so we have shrinkage, and then we have organized retail crime. And those are very different. That's actual, you know, sort of not on the same playing field as what you're talking about because shrinkage has been a long standing metric for retailers, brick and mortar retailers. And as you said, there was like a standard, and that standard has been changing. So it's a bifurcated approach. I mean, the the smashing grabs, that's a separate issue. We're talking about typically shrinkage. A lot of it was caused by employee theft. Yeah. So it's very difficult to to know the origin of of shrinkage. It can be due to, employee theft, client theft, and administrative errors. And so you can see like, usually the the the numbers that we have in the in different studies is probably forty, forty, and twenty, forty percent external, forty percent internal, and twenty percent, I mean, mistakes. So, obviously, with the smash and grab cases that that you can see, more and more often, it decreases it actually increases the, the excellent theft proportion. So it's still related. Our technology relies on what we call deep learning. This is, a field in artificial intelligence, that has proven super successful in image analysis. And we're using those progress in in deep learning to apply to video content. So it's much more difficult But now our technology is able to understand human behavior, a human gesture, and our first use case is detecting any gesture relating to shoplifting. Okay. So let's get into the AI conversation. So you built your platform five years ago using AI, and you're using not generative AI. But deep learning, computer vision. Is that correct? Yeah. It's really it's really close. Actually, this is another phase that we're gonna enter in terms of r and d next year to getting to generative AI for video. The next step will be to prompt the AI and to get in return a proper video. So today, our AI is able to understand human behavior But in the future, our AI will be able to produce video content automatically because it actually learned a lot to be able to understand human behaviors. It will be able to generate human behaviors in the future. Sure. So how does that connect that back to the up retail operations manager of a retail store. So let's start with what you have now, the deep learning human behavior analysis, first because that's what's on the market. And so they can see flagged gestures hiding something, etcetera. So let's just talk about that for a second. And then what does that look like when you do kind of add on a generative AI layer? Yes. Awesome. Our AI relies on different intelligent bricks and all bricks together. They will provide you the profitability of the occurrence of a specific gesture every second. So the algorithm is analyzing what's happening and every second is able to tell you, okay, in the few seconds I've just seen, this is the gesture that is currently happening. So sometimes you're just grabbing an item and putting that item in a shopping cart. No problem. Sometimes the algorithm did text understand that a person took a product from the shelf and placed that product in an object that looks like a backpack. And this is when we will send the alert to the store owner so that they can intervene in real time. It doesn't mean it's necessarily a theft, but at least the store owners, he knows there is, I don't know, a salmon salmon steve in the backpack, and you just wanna see it on the checkout on the team. You know? Okay. So let's go functionally in the store space, is there like an an alarm, a a pinging that goes off that's a that is audible? Because are they sitting and staring at the screen, or are they notified audibly? Like, okay. And then you can identify the person in the location. No. Very good point. We go even further than that. So before yes, in super centers, you have a video man that was actually monitoring a hundred cameras at the same time, which is impossible. I could do a you could do a little test, but you can follow what's going on in three three four screens maximum if you have the experience. And so now I want to remove that guy, and put that guy on the field instead of in front of the screens, and he will just receive a mobile device, a short video centered on the suspicious behavior that has been detected. So he receives a notification, like, you receive a notification on WhatsApp. You have a new message you just open the app and boom, you can see a five second video clip showing you the suspicious behavior that has been detected. Okay. So that's how they become aware of it. And then I guess from there, store policy takes over, which is we approach, we don't approach, we Whatever it is, but they at least are aware. At least they're aware. And what you realize is that probably eighty percent of staff are committed by regular customers that are not professional shoplifters that are not part of a organization criminal organization. You know? It's just an opportunity. They just want to not pay for that product. And if you're aware that product is in the backpack or in a purse, can just ask, hey, maybe you forgot it, but just you need to pay for it. That's the main point. And indeed for harder cases, that's the store policy, either they wait and they just get a proof and they can use to, start a file with the police or sometimes we we equip a lot of independent store owners. It's their money, they're willing to, to confront the client in most of the cases. Right. Get by the merchandise. This is just what they want. They do. They wanna get back their items, and the guy can leave the store. Right. Okay. So logistically, it works within store cameras already. So it it goes into a store security system. Is that Remind me what the how the logistics of it works. Exactly. You don't have to change anything. You just have to plug a small box this size containing our technology. And that's really what makes the difference with alternative solutions is that we made a huge effort in R and D to fit this technology on a small device. And you just have to plug this device on the same network as your CCTV, and the device will analyze the stream live continuously and try to detect gestures associated with shoplifting. We have around two hundred clients in the US already. And on your development map road map, you said you're adding a generative AI layer. What will that do for the, product? It takes time to collect and to annotate this data to train your AI. If tomorrow you have, a generative AI able to produce millions of cases of people placing items in backpack you have millions of data to train on, and you don't have to wait to collect this data in in stores or to collect to connect it from wherever you want. You can create your training data set. What I wanna know specifically is how it will change as a client, as a customer, as a retailer, how will it change what you're already doing? Today, our algorithm has been trained on one to one point five million of examples. It will change a lot in terms of detection rates. Our product like any human being, any security agent, you cannot detect one hundred percent of shoplifting cases happening in a store. Obviously, there will be some thefts, some really smart people that will manage to, yeah, overpass the the the algorithm tomorrow if in a matter of seconds, you can train the same algorithm on one billion videos, then, yes, you will be really, really close to a hundred percent detection rate. And so you will basically detect anything you want. And if a client wants to add another scenario, For example, hey, I want to detect when someone puts an item under the baby's troller. This is a new gesture that we need to train on. Well, tomorrow, we used it this generated AI to generate one billion video of people placing items below the baby's troller. In a matter of minutes, you have a new algorithm able to detect that new gesture, and you don't wait for months to collect this data to annotate this data and to retrain your algorithm. Got it. Okay. So In terms of the insights that come from the product, your product. So, for example, will will retailers be able to manipulate the information and the insights so that they can see that there's a higher rate of theft or attempted theft on certain products versus others. And let's say, you know, kind of an aisle scenario versus when there's, like, you know, a case with a human behind it? Like, are they able to analyze this information and then make strategic decisions. I'm assuming they can, but why don't you walk me through it? Exactly. So our clients, they have access to a back office where they can check all the statistics from the app. They already know what products are particularly subject to shoplifting because, again, they have this stock difference per product so they know beverages, they are, stolen a lot. What they don't know is where the shoplifting is happening, you know, because a lot of shoplifters, they can save the product somewhere and go to another aisle to conceal it. So what they managed to do with our technology is to know what are sensitive areas physically in the store not only the product that are particularly subject to shoplifting. Of course, they can, get some statistics after a few months of use of what are the days of the week that are particularly sensitive, what are the hours during the day that are particularly sensitive So what I'm what I'm gathering is a company is monitoring their shrink. They're at a certain number, and then they use the data from vision to qualify the data. So in other words, my shrink is here, and here's where it's And now I can it used to be just this black hole, one single number, one and none. But now you can back into where it's coming from, internal, external, blind spots, etcetera. And then sort of action item and and address it to reduce the shrink. Do you have data on how effective the technology is to reduce the shrink number from your from your clients? Yeah. Yeah. Very good question. So it's it varies a lot from store to store, of course, but I have an example of this organic chain here in France, two hundred stores after one year of use, they manage to decrease. I have their shrinkage after one year of use. But it's not necessarily, a hundred percent related to vision. They put in place a lot of different processes while using vision to change the schedule of their security agents to ask the security agents to be much more cautious at this time of the day, this time of the, of the week. Thanks to vision. And so all in all, they managed to decrease it by by half. Right. Well, that yeah. Exactly. So it gives you that it it's a road map. Then whether or not they act on it is really kind of an internal, as we said, operations decision. If if I may have picked up on that, an interesting another pilot we did with a major retailer in Europe. They wanted to first test our solution on a graph of ten stores because they had a huge, like, it was twice as high as the average they had in other stores. They shrink. And so they wanted to know where it comes from. They have no idea, low clue. So we put the solution in ten stores. We detected a few a few theft And we actually managed to send an alert on an employee that at the end of his shift, he would leave the store. And every day, every night, at the end of of his shift, it would just take a bubble on the on the way. And this was just an alert who would send because we do gesture recognition. We don't care who who actually does gesture. It's just a bow concealed in a purse. It generates an alert. And so what they realized is that most of the thefts were not coming from clients, but they were coming from internal inter from employees, and they got this answer they got this insight thanks to vision because we analyzed all the cameras. We detected a few, shoplifting from clients but they still had a a huge, a huge shrinkage, and they actually deducted that it was coming from employees And so they put in place a lot of different processes plus the added cameras in the back office. And suddenly after six months, of course, the shrinkage plunged because now employees would be much more much more cautious about about shoplifting. Tell me about are you finding Is the space competitive? If I may just to to have a a a view of of of the future of AI, you can see, like, how generative AI is is disturbing a lot of existing, businesses and a lot of existing technologies. The the the most important point today is that is the fact that it could significantly accelerate the training phase of an algorithm to take data driven decisions is gonna be the next the next revolution for retailers. You know, they already tried to get some data on the hot and cold zone, warm and cold zone in in in in the shopping floor to know where people go, where people don't actually go to know where to place the products, in the future with gesture analysis, gesture recognition, you will be able to go into much more details to understand exactly what they did at what what time, what product they interacted with, how they interacted with this product, and how we can change the placement or the packaging or whatever to make sure engage the customer and and drive more sales. That's a very competitive space, and it's also still very vague. Like, someone may touch a product, but it might have too much sugar, so they put it back. There's no way, like, a touch data would indicate that. But It's good to know how many times it's touched and put back and then just stop buying it because it's not turning. Yeah. Or you wanna know they touched the product, did they read the note, the the notice at the back, or did they did it did they just look at it and not even touch it, or did they miss the problem is at the the the the lower part of the of the shelf, because nobody actually looked down to to to that part and maybe you need to to change, to change to change the layout. Those are the the really the the the very detailed analysis that that might be lacking today. Yeah. Today, you just have a warm zone, cold zone. It's it's a bit vague. Indeed, you don't you don't know what to do with this. But if you go into much more details, you might have, some actions to take right away. For sure. It is the holy grail to be able to have the equivalent of what online data produces inside brick and mortar. It's it's still surprisingly fairy behind with a lot of different solutions available out there that to varying levels retailers have embraced. So having yours serve two functions could be very interesting and a good tipping point to get them to that that additional dataset. Exactly. Well, thank you so much. It's been really interesting to hear about vision and to understand the roadmap in this space in general role and how generative AI will be contributing to expanding the offering. Thank you so much, Christine. It was a pleasure.