Agentic AI For Retail — ALDO Group Vice President of Data & AI Fatih Nayebi

Fred Schonenberg
Hello everyone and welcome to the VentureFuel Visionaries podcast. I'm your host Fred Schonenberg. I'm so excited today to welcome to the show Fatih Nayebi. Fatih is the vice president of data and AI at ALDO Group. He has over 20 years of experience leading AI initiatives that transform retail operations and customer experiences. 

Fatih is a faculty lecturer at McGill University and he is the author of Foundations of Egentic AI for Retailers. This is the first book on autonomous AI systems in retail and he's going to share his insights on how AI is reshaping retail overall and the challenges of scaling these technologies, something I know all of our listeners are very interested in. So please join me in welcoming Fatih to the show. Fatih, welcome to the show. Thank you for being here.

Fatih Nayebi
Thank you very much. Very happy to be here and also sharing some of the insights and looking forward to having a great discussion with you.

Fred Schonenberg
Well, I would love to kind of set a level with maybe your journey like what led you to specialize in data analytics and AI.

Fatih Nayebi
Yeah, so I'm coming from comp sci, computer science and engineering background and I've been into programming, computer science and also artificial intelligence and I did my master and PhD in the field and back then that AI wasn't as popular or as sexy. I was doing some AI, let's say, mostly like predictive and then prescriptive AI. So yeah, so that brought me to doing any different types of AIs in different types of industries including financial institutions, a bit in the healthcare and most recently in retail, which is very interesting of course because of all the trends and because of all the velocity of specific to fashion retail and a lot of data that we have.

Fred Schonenberg
Yeah, it's really interesting even just thinking of those three industries coming together for you to be doing that work in all three different times must be fascinating. I know at ALDO you're leading the initiatives related to AI. I'm wondering if you could maybe help us think through maybe some of the listeners that aren't familiar with ALDO or are not familiar with some of the work they're doing in AI. Can you give a couple examples of maybe some of the projects you've worked on and some of the results of that work?

Fatih Nayebi
Of course, let me just do a very quick introduction about the ALDO group itself because ALDO is a global retailer. We have multiple brands including ALDO shoes, that's the most known version. And then we have the Cold Spring that operates in Canada. We own all the stores in the United States and Canada. Then we have the franchisee system all across the globe. So we operate in more than 100 countries. We have more than 1500 stores across the globe and we do everything.

So we start from the ideation and the design of the products, manufacturing them, doing everything, allocation, logistics. And we have our own e-com, we have our own stores. So we have the whole spectrum. When I joined ALDO almost three years ago, we had a lot of things related to reporting and dashboarding and to leverage the data to make business decisions. But then we were lacking doing AI, really. And in one of the discussions that I had with our CEO, David Ben-Seddon, he told me that he regrets that we don't use AI enough. I said, OK, that's no problem at all. Let's actually go for it. I started to look around to see what are the opportunities that we could really work on AI and then to bring business value.

I started to talk to my stakeholders, with our business partners to see what are the pain points, what are the opportunities. And we identified quite a few of the challenges that we could solve with AI and with automation in a sense. Then we said, OK, can we actually afford doing all of that or should we be able to do a bit more research and maybe to bring in a bit of the talent from universities, from academia and then maybe some funding? So what we did, we formulated a project that's called Revenue Growth Management for retail and supply chain with AI. 

And then we applied for funding to a super cluster that from the government of Canada is called Scale AI. I'm not talking about Scale AI, the company that's in Silicon Valley, but we have a super cluster in Canada that provides funding for projects that are in the intersection of AI and supply chain. In Canada, the project was accepted and then we started to work with one of our partners, Ivado Labs, and we started to build it. So within that project, we have multiple different AI models that include demand forecasting, which is really the holy grail of retail. 

You want to ensure that you have a good demand forecasting model such that you can use it to do any type of planning or any type of, you know, decision making. So an in-season demand forecasting, and of course, demand forecasting by itself does not bring a lot of value. It's not actionable at that moment. We paired it up with two impactful use cases. 

One is the markdown optimization. So within the season, we want to ensure that we have a good pricing approach, not marking up, but just marking down, but to ensure that we can have a good sale through, we can have a good gross margin. Then the last piece is all about order fulfillment optimization, because we ship from the stores to the clients. And for that, we want the model to pick the right store to ship from, but not just based on location or the distances, but also considers the demand tries to avoid out of stock situations and the excess inventory and such. So these are more of the traditional AI models that we've been working on. Besides that, we have some gen AI and agentic AI that I think I'm going to have for another question, I guess. 

Fred Schonenberg
Yeah, I'm excited to dive into those. It's very interesting to see where you all kind of began that journey. I think one of the things that's interesting to me, obviously, is that you have your PhD in computer engineering, and your research was focused on machine learning and human computer interaction. And obviously, you've been in the multiple markets that we talked about kind of doing some of this work. How has your research impacted your real world business impact? And what I mean by that is oftentimes, we see things in academia that don't necessarily translate into real world business outputs. I'm curious how you've sort of connected those dots.

Fatih Nayebi
Great question. So I finished my PhD in 2015. And then back then, AI wasn't as popular, maybe in industry at least, but in academia, we've been trying to see how we can apply AI and machine learning specifically to different software problems. There was a field that was called mining software repositories. So we would build machine learning models that would identify defects in applications, they would decide which part of the application needs to be tested. They would do effort estimation of knowing that, okay, how long does it take for this project to be completed, and things related to usability and the human computer interaction of applications. 

And mobile was very popular. And I was also building a lot of mobile applications. I did research around how I can leverage machine learning and predictive approaches to measure and then to predict the success of machines, like mobile applications in the market. And then I actually use this for many multiple reasons, because I created my own questionnaires, I did some of the measurements, as well as some of the calculations around the code base, and then leverage that in a machine learning and predictive model to predict that.

I worked a little on mobile and the intersection of the mobile and the AI after that, and applied this in some of the examples of applications that we've been building. And just want to say that the human computer interaction field by itself is hugely disrupted right now with the large language models and the foundation models. So many of the things that we've been doing back then may not super apply today because of the foundation models.

And the human computer interaction is, as the name suggests, to be able to find out what are the best ways to interact with computers and to be able to have bidirectional communication and accordingly to be able to learn from each other. But now, the foundation models are, in fact, maybe the only user interface, or they will become the only user interface between humans and the computers. So that is like, I'm really looking forward to seeing how this is going to fold. And then maybe we will have some of the hardwares and different types of applications coming along. But yeah.

Fred Schonenberg
Yeah, very, very interesting. You mentioned Ivado Labs, and then through our research, we saw you've embarked on a number of academic hackathons and capstone projects. I'm curious, in either case, can you share a little bit about the commercialization efforts from university partnerships into this future of AI and retail that we're all kind of experiencing that's moving so fast? 

Fatih Nayebi
Amazing, yes. So we do many things. I will start with the Ivado Labs part, which is already partially funded by the Government of Canada, Scale AI. And definitely doing these types of things and being able to work with researchers from multiple universities, that is definitely really helpful. Ivado Labs is the execution partner of Ivado, which is a consortium of multiple universities across Montreal and then one in Quebec City. And then we have access to multiple different researchers in different areas. So some of them are specialized in, let's say, supply chain optimization. The others are really into pricing and some others just purely machine learning and technical.

Within this project, within the Scale AI project, we were able to create a scientific board, a scientific approach, and then to bring in multiple of the different professors from different universities, including McGill University that I also work at. We have weekly and sometimes biweekly meetings to talk about the scientific approaches. And all of this translates to better results. And we can definitely see that with those discussions and some of the solutions that we find during those discussions, we can improve the models. 

Every time that we improve machine learning or AI models, they translate to something very tangible because these are real stuff, right? We are optimizing markdowns. And so far, we've seen that we are able to increase the gross margin by 3%, which is very significant for us and I think for many of the organizations. And these are because we are leveraging sophisticated machine learning and optimization techniques. 

Besides that, it's very important to have access to talent because we're talking about machine learning and AI. Of course, we have good projects and many of the talents are interested to work in these domains. But having the right connection is also important. And so we organize hackathons and within those hackathons, we formulate some of our problems in terms of use cases. And we provide sample datasets and the students work on that real problem with the real data. And we hired some of them already at the outdoor group. And this definitely helps in terms of retention. It helps also in terms of bringing the right talent.

Fred Schonenberg
Yes, super interesting. I almost feel like we could dive into that for a long time. But let me flip over to your book. So The Foundations of Agentic AI For Retail. There are so many, just the word agentic AI is probably, or that phrase is something that I'm hearing multiple times a day. And I'd love to tie into retail, right? I'd love that it's actionable and it's an emerging, fast-moving field. Let's spend some time on agentic. Could you maybe start with someone that's listening, that's new to the concept, what is agentic? And why is it a paradigm shift for retail?

Fatih Nayebi
Yes, definitely. That's a great question. And I think it's one of the most important questions to ask as of today. You know, everyone is talking about it, but let's demystify. I will start with traditional AI and mostly with machine learning as well. Because AI is not new. We all know that AI has been here for many years. Neural networks, they've been there since the 1950s. But we never get the chance to build agents that they can act by themselves, right? 

So we need to predefine things to decide what are the exceptions, what type of things that they should do. We needed to implement rules. Then we said, okay, this is not trivial. We want these agents to learn from the data. And at that moment, let's say we call them agents, but they did not have a full agency. It was really artificial. Then we said, let's do machine learning.

Of course, using the data, learning from the data, they could find out the patterns and they could revamp themselves. They could achieve better results. But human supervision in most cases was required, such that we needed to feed the data into a piece of code, a system. We need to also say, okay, what the outcome is and what can be. A model generates something. The AI generates something. As humans, we needed to go ahead and evaluate it. So this supervision needed to be there all the time. Of course, we had unsupervised models as well, but they were not as business oriented in many cases.

Now we get to the point that we can have an intelligent entity that has agency and can use other types of AI as just a tool. So I just want you to imagine very simply a loop that has a reasoning, a brain. That is the foundational model, like large language models. Sometimes smaller language models are going to be sufficient. This tool, this entity is able to use tools. So it can call APIs. It can write code by itself. It can read files. It can search the internet. And once you give them an objective to say, okay, I want to do this specific type of thing, then it can do things by itself. It can refine things by itself. It can make things happen. It can decide. 

So when I wrote The Foundations of Agentic AI For Retail, I could write it only for foundations of agentic AI. But then I was like, okay, can I actually make this a bit more practical to provide examples only in one domain such that we can really see what is going to happen behind the scenes? Otherwise, the foundations are true. Like the foundations we've been talking about for many years. Now addition to the foundation models, the large language models are exceptional and it's really helpful. But there have been different types of approaches to build these agents. 

Then if you're building these things for one specific domain, such as retail, can I build multiple of these agents so that they can work together and they can make things happen? Not only am I going to build a machine learning model that does demand forecasting, but then the output of that demand forecasting needs to be presented to some of the end users, right? Or some of the business users in a dashboard. And they would just look into that and they would try to make some decisions. Otherwise, sometimes these models make mistakes. Sometimes the data is not there. There's always some sort of a drift. And the users would provide you feedback. And then you need to go back and update these models. 

Now imagine that you have an agent, which is a system instead of sitting in front of a screen that can call and coordinate with another agent just doing the demand forecasting. And look at the result and say, OK, wait a minute. This doesn't make sense, right? I just need to adjust this. And it has access to the tools to be able to validate, evaluate these results, and then at times to go call another API or another agent to do the corrections. This is a big paradigm shift. This is just not static. This is not predetermined. We don't need to go ahead and create a lot of different rules or UIs, user interfaces. Instead, we need to build these agents that they have the sense of agency and they can decide. They have the reasoning behind the scene. And I could talk about this forever, so please stop me at some point.

Fred Schonenberg
No, no. It's fascinating. I actually want you to talk more about it. My biggest question is maybe because we're seeing we run Comcast's AI accelerator. We're in our fifth cohort. And so we've looked over 3,000 AI startups. We have I think 46 pilots running currently throughout their organization. So we're seeing not only this, but the application of it across multiple challenges and business units live at the moment. And it's truly transformative.

I want to stay in retail because one of the things I shared with one of our clients maybe a year and a half ago was I said, you have to be prepared for a world in which bots are selling to bots. That agentic is coming and the consumers are going to be using it. And you as a retailer are going to be using it. That's going to be working with your supply chain. And it seemed almost like a futuristic comment at the time. I thought it was several years away. And agentic has changed that, in my opinion. I feel like that is coming. What do you think about how agentic is going to transform let's say omni-channel retail as we know it?

Fatih Nayebi
There are many answers to that but like the shortest one is it will transform everything, any system, any application, any role that you see today will change. And the main reason for that is that this is the first time in human history that we will have a system that is more intelligent than us. Something that can decide properly knows pretty much all the domains.

So an example for that is that we have a benchmark for the large language models that is called humans last exam. There are 4,000 PhD level questions from different domains and if you go find very genius people that have maybe more than one even PhDs and ask them these questions they will not be able to answer more than 3 let's say to 5% of these questions. But with the pace of AI we are about like 40% right now and these models get better and better.

What I'm trying to say is that they know any type of business. They also know how to code, they know technicalities and that. And in the future, instead of having static UIs, applications for omni-channel or for anything in any industry or any organization, instead of having these static things that people sit in front of and try to interact with. As you said, we can have networks of agents, chains of agents, we can have an internet of agents that they can do things with and they can work together.

So of course, this is not gonna happen in one day, but we see and we build some of these already, right? So like with the co-pilots assistance and that we've been trying to build like single agents, not multiple agents, but they can still follow very interesting goal-oriented tasks and then achieve things. Now we're talking about different types of protocols. They needed to have protocols to be able to make this thing work. So Anthropic came up with MCP, Model Context Protocol, that makes it very easy and standard for the agents to call tools. In the future, instead of just writing so many APIs or different ways to interact with that, we can just use MCPs. 

Then Google came out and said, okay, let's build a protocol to make it easier for agents to talk to each other. And then that's A2A, agents to agents protocol. And we started to adapt that to say, okay, if I'm building an agent, I want this agent to be in this agent's network or internet of agents so that they can work together. And that's the key, not having, you don't need someone to just supervise them. They can work together, but this is not sufficient. 

So now we move forward with how we share memory and context between the agents? And some of the protocols are coming up, like some such as mem0 to say that if I have multiple agents, they should be aware of what is happening in that context. And then finally, at times that we need human supervision, that we will definitely, we need to implement human in the loop type of approaches. Then what should be the protocol around the UI itself? So there is a protocol called AGUI that makes it simpler for the agents and the UI to communicate with each other so that they can bring humans to the loop. 

So like the not so short answer was this, that actually everything is gonna change. I think this is bigger than the internet or even bigger than electricity to some degree. And I don't know if everyone, many of us actually realize that or not, but those opportunities and also to some degree, the risks are just limitless.

Fred Schonenberg
Can I ask you, and I'm going off script here, so feel free to tell me no. One of the things I had a conversation earlier this week with a loyal listener, my mom from the show. And I was sharing a very similar point of view that I think while a lot of people feel that AI and agentic are overhyped, I think they're wildly underhyped. I think people are not understanding the possibilities, the challenges and the disruption. And I do think it's on the scale of the internet and even more so electricity. That scares a lot of people, rightfully. You mentioned humans in the loop, which I think might be a piece that people would be really excited to know more of. How do you see this?

And I don't know, this is the part where I was like, I don't know if you'd want to use this as an example, but to me, like a simple example is like, okay, I need a pair of shoes from ALDO. And like, I have a very specific experience set for how I would go about that, right? And maybe I wouldn't think ALDO first, I would think I just need a pair of shoes and then I would go through some mental loops of my own and maybe there's not a store nearby or I'm in a rush and I go online and I find the site and I look to see what you have available in my size and I buy it or I don't. And a bunch of stuff happens in the background, which I don't care about, but you guys care a lot about at ALDO, right? So I'm curious, how does that change with agentic? And where's the human still in the loop?

Fatih Nayebi
One of the key points, as you also mentioned, is the search that is going to change drastically and it's already changing. So what we do today is we go to a website and then we put some keywords and that website doesn't know much about us. If we log in and then we provide information, of course it knows to some degree, but we still need to provide keywords. It doesn't have the context. It doesn't have much memory. It just gives you some products or lists or links. And then you need to refine them, right? You need to go ahead and use some filters or this type of thing. So be able to get something that you want, but you are still in the loop, maybe too much in the loop just to decide what you want.

And this is gonna change because the Gen AI models and not even agentic models, just Gen AI is capable of understanding and then generating things that are accordingly fine for you because they have the context. They can… you can actually talk to them. You can talk about the occasion, you could talk about the style or the look, or if it's for a rainy day, if it's for a wedding, if it's for this or that, and then they would provide you something. You could further refine the results by just keeping… keep talking to them, saying that, okay, but I just want this in black, not blue or something like that. And people are going to get used to this. 

So one of the things that we've done in our, one of the hackathons, one of the winning themes, we implemented this, but then we got to the point that, wait a minute, everyone is going to just implement this, right? I mean, this is going to be pretty much the interface moving forward. So I do see that many of these applications and systems are going to be having this, like the iOS, Android, the web is going to have this. In our case, it was Shopify, that Shopify is going to just integrate this thing inside.

And then it will bring you back something, but then you need to train this thing because it's going to have the memory and the context such that it can continue to serve you well. That's the part that human and the loop definitely works. How do we interact with it? What do we provide? How do we prompt it? And all of these types of things are going to play an important role.

On the other hand, when we get into more of the backend stuff, like supply chain related things, sometimes the data is not perfect, you know, things can go wrong. And sometimes the agents and the AI do not have any clue about those types of things because the data is not fed. And in those cases, of course, there should be some human supervision such that we can refine the results and make the right decisions for it.

Fred Schonenberg
Yeah. Yeah, it's interesting. I mean, I get so torn with the excitement of what is possible here. And also the flip side of like, well, wait a second, how does anyone make money to have the agent tell them what shoes they can have for their feet? Because the job they had of finding the shoes before is gone. And so that's probably the doomsday side of me that is just thinking through where that human in the loop is, what the value is there, that exchange, because agent to agent is such a powerful concept in so many different ways.

Fatih Nayebi
I think as a society, as humans, it's hard to answer that at this moment because so many things are gonna change. But I do see that we are going to be working differently and we're just going to be faster in doing everything. There's of course, a lot of oversight that is going to be required, governance, regulations, and the guardrails that needs to also be implemented to ensure that these agents, they're gonna act for good, right? Not for bad. So there's all of these things. 

On top of that, anyone that does anything related to STEM, I think they're gonna need to do more. They're gonna do more. And then we are going to be using these technologies and the agentic approaches to do faster research, maybe to solve some of humanity's problems, right? Healthcare issues and so many of the things. So recently Google published an article around Alpha Evolve, which is using reinforcement learning and some of the evolutionary approaches to improve itself. So we see that AI can improve itself and can improve its performance, but also can solve some of the existing problems that it's very hard for us to solve. And this will hopefully translate to many of the breakthroughs that we require as human beings. 

That's the positive side of it that I'm also looking forward to. And I do think that with the vibe coding and the agentic coding, we always talk about, okay, there will be no job for the software engineers or for the scientists or like for the data scientists. But I tend to disagree because we can do more and we should do more using this type of tooling.

Fred Schonenberg
Yeah, I think it's really interesting. Well, this has been such an exciting conversation. All sorts of other questions I'd love to ask you, but before we get out of here, is there anything that you wanna share that we didn't cover that you just wanna talk about before we kind of tell people where to go to learn more about you? 

Fatih Nayebi
If you want to stay a bit longer, I'm fine. If you want to ask more questions, it's fine for me. 

Fred Schonenberg
Okay, well, in that case, I have two in particular. One is, there's all these things that are happening. We've talked about the multi-agent systems and causal reasoning, federated learning. What are you most excited about and how are you and ALDO preparing to leverage these advancements? 

Fatih Nayebi
Yeah, so there's a lot to unpack over there. So the federated learning is really interesting. And of course, we are talking about learning on the edges. We are talking about offline learning. So these are very interesting because if we want to have our own personal assistants, we don't want this to be all over the place. We just want to have the right control over it, right access and the right security. So that's, I think, a very important piece. And the better the GPUs and the computation becomes, we will be able to create these types of things. 

Now, privacy-preserved federated learning is another area that is going to be key to ensure that we do share these information accordingly and then we can train these models and make them sophisticated for us. Before going for the ALDO examples, I would like to say that anything that touches humans in terms of either education and training or healthcare, I'm really looking forward to being kind of disrupted or improved.

So today, education is particular. You go through a very specific curriculum or programs and then some generalization happens. But with agentic AI and specific foundation models, we will have personalized, truly personalized education that is going to be accessible like 24 seven for us. It can adapt, it can make things happen accordingly. So it's going to be super easy to learn anything and to have access to any type of information. 

The other part of it is that many of the researchers that we have today, that they're untapped. Let's say not every paper that is published gets enough attention. And sometimes some of them are just hidden somewhere, but they could in fact have a profound impact on everything, on research, but also on humanity. I would say by just looking at the type of researchers that they failed, let's say, or the hypothesis was rejected, we could get into some very meaningful results and we could use it in the future.

AI is perfect for doing these types of things, to crunch a lot of different documents, data and that, and then to be able to find very interesting patterns in them. So I'm looking forward to having very interesting results in terms of the research itself and the education, because the better educated we are, the better the world becomes and then the more that we can achieve. That is my hopeful sense.

Then at the ALDO Group, we try to use AI everywhere, like agentic AI. So we see this as two types. One type is that low code and the no code for everyday essential types of things that using, let's say, Copilot, you can have a lot of productivity gains. You don't need to summarize the emails or like the meetings or some of the documentations and that is a lot easier. On the other hand, like the multi-agent approaches that require coding, in my opinion, they're a little more interesting because now you can achieve things.

An example that we have done is we have a data privacy officer that we receive data privacy related requests and then the agent passes through them, quantifies them, qualifies them and groups them and then creates like tickets for them to be followed. Or in the IT organization, we have a role in Azure that is called product owner. What it does, it captures all the requirements and then creates stories, epics and such to ensure that we can implement the right products from the data or application point of view. 

My team created an agent to automate the product ownership approach. And you just give a bit of the text blurb, it does everything, it writes the stories, it writes the success criteria. It also looks at the workloads of everyone, defines their roles, makes suggestions in terms of architectural dependencies and that. And of course, like the coding itself. What is important to emphasize is that coding is not just coding. Coding is the way that computers solve things. They can write code to achieve things.

They can write code to solve mathematical problems or physics or this and that. And anything that we do that somehow uses code or requires code is getting a lot more productive because now we can analyze the data, we can create applications or systems a lot faster and this is going to have a profound impact on everything.

Fred Schonenberg
No doubt. Well, Fatih, let me get you out of here on this. What, where can listeners go to find your book and learn more about the work you're doing at ALDO and McGill?

Fatih Nayebi
Thank you very much. So the book is in the print format right now. I don't have the electronic version yet. One reason is that there's a lot of diagrams and a lot of codes and such. They can find it on Amazon already and it's going to be published in Barnes and Nobles and different places soon. And then if you just search it for me, like they're gonna find some code GitHub repositories that they can go explore the code and there's some code examples. 

And I try to share this at different conferences and hopefully I will bump into some of your people that are going to be listening to that in some of those conferences. And I would love to have their inputs and their feedback.

Fred Schonenberg
Wonderful. Well, Fatih, thank you for sharing your time and your insights today. I can't wait to read the book. And also thank you for all you're doing to spark change in this really exciting space. 

Fatih Nayebi
I appreciate it. Thank you very much for the great discussion. And I'm very happy to be here.