How AI Will Power Our Vision of Adaptive Retail

Speaker 1:
Welcome to TD Cowen Insights, a space that brings leading thinkers together to share insights and ideas shaping the world around us. Join us as we converse with the top minds who are influencing our global sectors.
Oliver Chen:
AI, it's the future of retail, lots of exciting innovation happening here, and I'm here with an expert, a friend, a colleague, and a co-professor who comes help teach my class at Columbia Business School as well. Welcome to the Retail Visionaries podcast series, a podcast about visionary ideas and people. My name's Oliver Chen. I'm TD Cowen's new platforms, retail, and luxury analyst.
In this episode, we're thrilled to host Francois Chaubard. Francois is currently a partner at Y Combinator, and he's an accomplished entrepreneur and technologist with a strong background in engineering and AI. He was the founder and CEO of Focal Systems, a retail automation platform that automates and optimizes retail through deep learning computer vision. He's also the famous author of AI for Retail: A Practical Guide to Modernizing Your Business Through Automation. Francois, it's a pleasure to have you here. It was also great to learn from you when I was at Stanford's program in AI too. Nice to see you.
Francois Chaubard:
Always a pleasure. Thanks, Oliver.
Oliver Chen:
Francois, so the AI landscape is rapidly changing almost daily. For this podcast, what are your top tips you would share with retailers to navigate the landscape and use AI efficiently?
Francois Chaubard:
Yeah. I mean, I guess to pull a Bezos on this, I would say that focus on what's not changing. The AI is always getting better. Sam is constantly investing in the model, making the models better. All the competitors are following suit. So if any claim that, well, AI, I mean, I remember 10 years ago when I started getting serious about AI, they would always say, "Well, AI can't draw a picture," would be the argument? And it's like, it can do so better than me for sure now. And I used to hear all the time, "Why can't AI set planograms better than a human?" And then there was all these arguments why they thought that was going to be true forever. It's just not true.
And so AI is going to be able to do more and more things every single day. And it's focused on the slope, not the y intercept. And I actually think that AI is already able to do these things. And if the argument is that it can't yet, then focus on the yet part, is that it's going to be able to do that in short order, even if it can't just yet. And so that'd be my advice is focus on what's not changing.
Oliver Chen:
One of your perspectives is a driverless store. What do you think's happening with the store doing more? And your book also goes through profound models too. Would just love some highlights about which parts of retail should have more AI and what you're seeing now.
Francois Chaubard:
Yeah. I mean, in retail, it's all about the customer. And that's the most important thing. Sam Walton was passionate about that. And I think, walking backwards from the customer experience, what does the customer want? They want great products that are in stock at low prices. And they want to get in and out really easily.
So you have to make that happen and do better at that every single day. AI absolutely has a role in that. I mean, I was just at a Target, for example, and 20% of the store was out of stock. I'm just like, this can't be happening anymore. It's kind of crazy that we're living in 2025. We have driverless cars all throughout San Francisco here, and 20% of the store is out of stock. And it's like how is this still happening? And even Instacart, when you order on Instacart, and they sub you out 10, 20% of your order, you're just like, how is this really still a problem? The technology exists. It's an adoption issue. And that's like the part that where retailers aren't the first ones to adopt new technologies typically across all industries. They're sometimes the last. And they need to move on that. That needs to change.
Oliver Chen:
Yeah. Well, could you dive into that? What is the technology that's most exciting to you? I think you've been a pioneer with computer vision. And secondly, on the adoption highlights, what should we know? What have been some of the challenges that are worth sharing?
Francois Chaubard:
Yeah. I mean, I worked on inventory for a long time. I worked on checkout for a long time. The checkout part, I would say, is everyone was very impressed with Amazon Go. Everyone thought that was going to happen in all stores and it didn't. And the reason why is because of cost. And so it's important, getting AI to do the thing is one thing, then getting the AI to do the thing cost effectively is a different thing. And there's some things that AI can do, but not cost effectively. And that's the Amazon Go stuff.
So I would ignore things that are really expensive, like in the hundreds of millions of dollars, billions of dollars in investment. There's going to be robotics in AI and AI in retail in the future, but those applications, if they're specific to retail, I would probably not look at those. I like the humanoid general purpose robots and applying them into retail because you don't have the build out cost for that.
And then for inventory management, I mean, it's already here. I mean, Morrisons has us in 600 stores, taking an image every hour, detecting out of stocks, ordering based on that, updating the action tool for all the stockers to make sure they're working the most prioritized outs. It's already working. And it's mostly an adoption issue. I think that American retailers in particular have a bias towards building, not buying. That is absolutely not going to work in the future when AI specialization just becomes harder and harder to get, and AI talent is more sought after.
So building your own solution just doesn't necessarily... Some retailers built their own routers when Cisco was coming out. It's like now you look back, you're like, why did you build your own routers? Cisco's routers are great. You can buy those. It's not really differentiating you as a retailer by having your own custom routers. Similar with barcode scanners, similar with AI, especially if it's back office. If it's front of house, maybe you want to own that a little bit better, at least tweak it so that the customer interface is a little bit your own. But for back office stuff like ordering inventory, there's no benefit to build, whether it's your own or it's the same that everyone else has, it's not going to be that much different. Just do the best you can and take the best solution out there.
Oliver Chen:
Francois, as you know, retail's a low margin business. It can be high gross margin, low gross margin return on inventory, but high turnover. How can retailers afford this, or what does a typical model look like with pricing, et cetera?
Francois Chaubard:
Yeah. I mean, because the cost of AI is going down by 4 to 10X every single year, you look at these cost curves of what cost per tokens are, when they came out to what they are now because of innovations around GPU compute, but also innovations inside the architectures, these things called sliding attention and things like that, that really bring down the cost of inference. It just keeps getting cheaper and cheaper and cheaper until it's a no-brainer.
And now, I mean, Focal, we sign up for a 10X ROI. And so if you spend a dollar, we promise you we're going to return you back 10. And if you don't, don't pay the bill. And that was literally something that we would put in the contract. And honestly, 10X wouldn't be enough for me. I want to get retailers to a 50X ROI. And so I'm no longer at Focal, so I'm not pushing that, but it's definitely possible, and it's things that we put in contracts all the time and we actually deliver on.
Oliver Chen:
Yeah. Well, it's been tough because I've met with many CEOs and lots of AI doesn't really work. And so why does yours work and/or how does somebody figure out what works? Because I get spoken to a lot about these ideas that actually don't work for returns, but it's more [inaudible 00:08:57].
Francois Chaubard:
I mean, we've had a lot of bake offs against a lot of companies that tout that they're doing this or that. And then when push comes to shove, it's just obvious who's telling the truth or not. I mean, the dumb answer is go try it. And then you're going to find out and you're going to kiss a lot of frogs and you're going to see which ones actually stand up and which ones don't.
The better answer that might save you time is look at the credentials. If these are Stanford PhDs and that have worked at top frontier labs, that's credible. They're going to be able to deliver on that, and they're not going to oversell and do the vaporware thing. And if these are a bunch of MBAs that are building this technology and they're saying that they can do it and they don't have very strong AI people building it, I wouldn't believe that. And that's the case for a lot of these technologies. And so, I'm not going to name names, but there's companies out there that have raised lots of money, and that's not a good signal that it actually works. I mean, you have to look at the credentials of these people and see if there's any credibility to it.
Oliver Chen:
And then your product and what you've done with retail, which technology excites you the most? It's been computer vision focused. And I think reinforcement learning is also a key model. But could you walk us through a deeper dive on the tech that we should focus on for retail more practically?
Francois Chaubard:
Yeah. I mean, this is like definitely my research area right now. So I'm back at Stanford doing research 80% of my time. And computer vision worked starting in 2012 with the AlexNet paper. RL did not work and RL is just starting to work and getting to a really good place. If you saw this paper that came out last year called Diffusion Policy, we went from 20% human level performance to like 88% human level performance in one paper. And it's a huge deal. And that's given rise to Figure AI has raised billions of dollars to build humanoid robots, which those algorithms didn't exist more than a year ago. Chelsea Finn, physical intelligence, and we're going to have Rosie the Robot in our homes and be able to teach a robot to kind of do anything now. And that's mostly because of this innovation in RL.
And we've seen this also in DeepSeq. They had this R1 paper that was a big famous paper where they invented this GRPO loss function and got the RL to really work. That direction is also very promising and I think it is meaningful to what's going to happen in the future.
Oliver Chen:
Francois, in your book, AI in Retail, you also lay out some financial parameters in terms of what can happen. What are some highlights from your book and top findings on AI and automation for retailers, as this can apply to valuation models too?
Francois Chaubard:
Yeah. I mean, I say a lot in the book, so I don't want to give too much of the ending, but I would say the big highlights are if you look at the P&L, and this is what I did before I started Focal, just look at the P&L of major retailers. Like take Walmart, for example. These numbers might be a rough approximation. But they do like 600 billion in revenue. They have like 130 billion in gross profit. And 100 billion of that goes to SG&A. And so that leaves like 30 billion approximation into operating profit.
And so if you automate half of that, and improve the customer experience, like just ignoring the improvement in sales because you've improved the customer experience, you can double or triple operating profit. That's a huge, huge deal. And most retailers are not equipped to completely reinvent the entire stack and how they planogram, how they select SKUs, how they price, how they order inventory, how they stock, all those things, but someone will. And I think that that person will have not 3% net margin, but maybe like 15% net margin. I think that that's actually possible now. And that takes a lot of work and people are going to have to figure that out.
But I would say most retailers are extremely AI light. They're like doing the simplest little thing that like... My favorite one is the AI based robotic cake decorator that people were excited about. I'm like you're decorating a cake with this thing. I'm giving you like this thing that is smarter than all of us. It only gets more intelligent, and you're decorating a cake with it. That's the thing that you wanted to do.
And it's like Moneyball, of all the things I can do with this algorithm, instead of like selecting the team, running the team, and winning more games, I'm like using AI to like mop the floor after the stadium clears out. That's not the thing. And the thing is, go change the fundamental way that you run your business. And that takes a Billy Bean to really take a swing and change the way that the game is played. And that takes courage. And I'm not seeing someone really do that yet.
Oliver Chen:
Yeah. Our book and our model that we wrote includes MLI, meaning thinking about fuller price selling on merchandising, as well as labor reallocation and inventory management and reducing safety stocks. So those are algorithms to really induce your return on invested capital if you get it right.
But AI is very, Francois, pervasive in terms of magic and logic and pre-purchase and post-purchase. How do you break it down? Is there a framework you think about as you compare and contrast? Because there's still surprise and delight with the customer experience, as well as fraud detection, as well as inventory management.
Francois Chaubard:
Yeah. I mean, in AI, the framework that's applied, that is applied for all optimization, not just AI, is called standard form. And so to put problems into standard form, the first step is to recognize a single objective function. And as per Jeff Bezos' 2004 shareholder letter, the objective function in retailers is to maximize intrinsic value, future discounted free cash flow. And people get confused on this. It's like, no, it's to maximize... If you ask 10 executives in retail, what's the objective function of retail? And it's to make customers happy. I'm like, okay, give everyone a Ferrari. That'll make customers happy. You'll go bankrupt. So clearly that's not the objective. It might be a nice KPI, but it surely is not the objective function.
It's to maximize sales. I was like, okay, again, give everyone a Ferrari every single time they spend 10 bucks. You'll get more sales. Trust me. You'll go bankrupt. So again, it's not the objective function. If you talk to a COO, what's the objective function? Keep my shelf stock. Okay, close the store. Great way to keep the shelf stocked. You don't have these pesky customers ruining your shelf conditions. That's not the objective function.
The objective function is future discount of free cash flow. And it's just a normalizing, amazing thing. Once you have that in your head, and you can get the entire executive team to think like that, which Jeff Bezos did in 2004, you'll get Jeff Bezos levels of ACSI, you'll get Jeff Bezos levels of market cap. And most retailers can't not think that way. It's very difficult for them to think that way.
And then the other part to it is subject to. And so I want to maximize some objectives subject to some constraints. Well, I want to make sure I never fall below 70% or 20% gross margin or something like that. You'll have something like that that you can put in there. I never want to fall below ACSI below some score. And so you can put things like that in there if you want. It's completely fine. That's the way that the game is played.
And this is actually the same exact structure is how SpaceX lands rockets on moving boats in the Pacific. And it's I want to minimize some objective, which is like getting the distance between the rocket and the landing pad to be minimized, subject to some glide path. So a valid glide path will not be land on San Francisco and go two feet over someone's house and burning to their house and then sliding over. You want to like follow some glide cone, they call it, where I'm making sure I'm not coming down, even though it might be a valid solution. So that's the subject to.
Oliver Chen:
Yeah. In terms of maximization or minimization, and also that aspect of gradient descent is like such a principle of machine learning. How do you juxtapose what you just said in terms of the standard form relative to like the cake exercise or even creative uses of AI across like product design relative to conversational commerce, relative to inventory management? Because a lot of your product was focused on inventory management and optimization. What about all the other stuff? And cake's not bad, like decorating cake could actually help you maximize some happiness and extract pricing.
Francois Chaubard:
You get prettier cakes.
Oliver Chen:
Yeah. That might be the return on equity for a certain aspect.
Francois Chaubard:
Yeah. Yeah. It's a fine thing to do. It's just like there's only so many things to do, and you have limited amount of things to be done. It actually, it might be gross beneficial, but net might be not beneficial because the amount of time and effort and the foregone opportunities that adopting this thing versus something else is more impactful might be not be there.
And so I have this framework in the book about how to adopt AI in the third section, which talks about setting up these three Es. And if I can remember here, so the first E is education. And this is what Jeff Bezos did a great job of is educating. He recognized in 2014, I think, really early, that AI was going to be transformational, and he forced every single manager to have an AI OKR on educating their people about AI.
And so once it's inside your organization, then there's the second E, which is exploration. And you want to explore and find all these opportunities, like the cake one, it's a fine one. And I'm sure like the mopping the floor might be another one. And those are all great things. And maybe you're going to put them into a giant spreadsheet, and then you should rank sort by the opportunity. And you should pop off the top most important opportunity you can go after. And trust me, it's probably shelf conditions, especially at Target, having just been over there. How do I improve the shelf conditions for the customer? And you can do that by free cash flow. I think that's probably the right way to think about it is future discounted free cash flow. What's the impact? I think the cake decorator won't rank very high. I think that increasing shelf conditions from 80% OSA to 90% on shelf availability is probably going to be much higher.
And then you have the third E, which is execution. And then you give this to a bunch of program managers to go execute on those things. And here's the trick that I strongly advise. Give those program managers a piece of the pie. If they execute and they are successful, and they can bring back $8 billion of sales or something like that, give them a $10 million raise. And you're like, why not? It happens in sales, it happens in finance, it happens in entrepreneurship, but it does not happen in these major firms. And that to me, it's the reason why most of these programs are not successful.
You asked why most AI doesn't work. Part of it is like on the AI people. I think that there's a lot of snake oil salesmen out there. The other part is the retailer. There's perverse incentives to not make the thing successful. So if you're a really smart program manager, and you get handed a project that's massively impactful, you get none of the upside, and if it doesn't work, you get all the downside. And so if you roll it out to all the stores and it doesn't work for some reason, you're fired.
So I get none of the upside, I get all the downside. And if I don't do the thing, I still have my job. So what am I going to do? What would a rational person do? They just don't do it. And they say, "No, it doesn't really work. It's not ready yet." They'll say things like that. And I've actually had PMs sandbag the project completely because of this issue. And so they're smart. These are smart people, and they'll do the thing that's optimal for them. It's not optimal for the organization, but it's optimal for them inside that organization.
Oliver Chen:
Yeah. It's a real leadership question in terms of, A, prioritization, and B, how do you structure fair goals that incentivize stakeholders, including employees? Okay. ChatGPT checkout, hot topic, daily announcements here. And Walmart, Skims, Glossier, probably your portfolio company's moving here, what will OpenAI's marketplace look like in your best guess? What's happening here?
Francois Chaubard:
I think Sam was just at YC, I think not this past Monday, but Monday ago, and was talking a lot about what's going to be the new way that people find out about products, buy products. Starting from the, I mean, the YC ethos is build something people want. And if I'm chatting with ChatGPT, and I'm like, "Hey, I don't know. What's a good present to buy for my wife for... " This is like literally a use case I just had. We have Christmas coming up, and I wanted to buy a present for her, and it's giving me some ideas. And they just send me some links. And then I don't want to go to your junkie website that takes forever to load and has all this stuff and then I create a login. I just want to like buy it. Of course I want to do that.
And so of course that's going to be the user interface and that's the right thing to do. And you better hope that the AI likes your products is all I have to say. So if you're not recommended, it's almost like this is going to be the new Google for sure. Everyone's been talking about that. But the whole user experience is going to be like coming through the ChatGPT experience. That will be a huge piece of the economy and purchases in the next 10 years.
Oliver Chen:
What do you think looking back in time, is the analogy Google Shopping?
Francois Chaubard:
Like Google Shopping Express?
Oliver Chen:
Yeah. Or the way you search and you see Google Shopping. It didn't destroy all malls and there's still physical retail.
Francois Chaubard:
Yeah. I would say so. My understanding though is different why Google Shopping did it and why ChatGPT is doing it. So my understanding of Google Shopping was that Google knows a lot about you, but they don't actually know what you buy because they hand you off to the website, and then you buy or not on the website. And if you have Gmail, then they know that you bought it. If you don't have Gmail, they don't know if you bought it or not. So there's no closed loop. It was a leaky bucket and that's why they did Google Shopping, and that's my understanding.
The ChatGPT reason why they're doing this is because it's the right thing for the customer. And like it's truly what... Sam is the most YC person in the world. And YC propaganda and what we're trying to make sure every startup knows is do the right thing by the customer. It works out. And that's what Sam is doing, I think. He's just following what's the best thing to do for the customer, and it's for sure, it's like, I don't want to go to your website. I just want to buy it.
Oliver Chen:
Yeah. Just thinking about friction, thinking about customer centricity. And maybe he has a standard form too, because you got to maximize customer's interest of these subject [inaudible 00:25:39].
Francois Chaubard:
Exactly right.
Oliver Chen:
But yeah, we want Ferraris too maybe. Okay. Francois, it was really great to be with you and talking about all about AI and retail. Thanks for the helpful advice and also conceptual thoughts around what's happening. You're a real leader in the space.
Francois Chaubard:
Thanks, Oliver. Always a pleasure. Thanks.
Speaker 1:
Thanks for joining us. Stay tuned for the next episode of TD Cowen Insights.