How AI Is Reshaping Retail

Samuel Boulanger
Hello, Fâtih. Thank you so much for joining us on the show today. I'm really happy to have you.
Fâtih Nâyebi
Hi, Sam. Very happy to be here. Looking forward to our great discussion.
Samuel Boulanger
I know you've been in AI for almost 20 years now, if I'm not mistaken. So what brought you to this field, and specifically in retail?
Fâtih Nâyebi
Yeah, so that's a great question. Of course, I have a lot of experience with different domains, such as financial institutions and healthcare and some consulting as well. And I was looking forward to work on an area of business that is really innovative and is open to try different technologies. And I had a great introduction with ALDO's CIO. At the time, we had a couple of discussions, and then we figured that we had a very good synergy, and there were a lot of things that could be done at ALDO Group. And I started to analyze and also to review and research the type of things that can be done with retail in terms of AI. And I've seen a lot of opportunities. There's a lot of data, good data, because you run business based on that data. And there are also a lot of challenges. There are a lot of opportunities in a sense that, like, from manufacturing the products to selling the products, from ideation to actually logistics, you have many different processes that you can improve, and AI is there to help you. So, yeah.
Samuel Boulanger
I will assume that from all of the fields, retail is one that is really impacted by AI. I'm just thinking about recommendation for the next best product on the website, or even on the supply chain side of things.
Fâtih Nâyebi
Yeah. So, like, I think all the different domains are going to be impacted because of the AI, what it can bring to the table. Retail is specifically impacted because they have a lot of problems and they have a lot of data too, right? So when you think about it, just processing that much data and getting insights out of it is going to take a lot of time. And in most cases, analysts, planners, allocators, different business lines that work on that data, they will not have enough time to look at all the aspects. And AI is good at that. It can just go through all the data. It can find very interesting stuff over there in a lot shorter time. So that's one of the reasons that AI specifically for retail is becoming very popular.
Samuel Boulanger
And you just released a book on the subject, actually called Foundations of Agentic AI for Retail. I haven't had the chance to read it yet, but you showed me a bit the table of contents and what's inside of it. So I'm really excited to. It will be my next book on AI for sure. But any favorite use case? I know there's a lot of use cases in the book. There's a lot of diagrams, architecture diagrams as well. Which one is your favorite? Which one do you think has the most impact on AI in retail?
Fâtih Nâyebi
Yeah, I try to cover many different use cases from ideation, creation of a product, then going through assortment planning, allocation, and then actually getting it to the stores or on e-comm and being able to provide recommendations and actually selling it to the clients. So long story short, the key pillar of everything AI in retail is going to be demand forecasting. Because if you know what you're going to be selling, what is going to be demanded, then you can use it for many different types of things. So that's something that you should build. But then demand forecasting by itself is not something that's going to bring value right off the bat, because, yes, forecasting this many products you are going to be selling at this store or e-comm at this time, at this week of the season. But so what now? How do you use it? How do you gain any business value out of it? Then you would pair it with pricing or with allocation or assortment and those types of things to say that if I can predict how many products I'm going to sell, then I can also look at the price elasticity to say, okay, if I sell the product at this price or the other price, what would be the demand? And based on that, I can adjust my pricing strategy. So that is one of the most important use cases that you want to tackle: pricing, markdown optimization. But for that you need to have the demand forecasting. So I would say a combination of demand forecasting and pricing is going to be the one to look after for the first comers, I would say.
Samuel Boulanger
Interesting. I mean, it's predictive AI, right? So it's been around for some time. I will assume that those models are on point and powerful at this point in time. I've been hearing about demand forecasting for decades, I think. I will assume that as of today, as 2025, they're really at a state where they're not fully mature, but they bring more clarity than it was. You've seen this evolve. You've been in the field for some time, right? How much better are they today than they were 10 years ago?
Fâtih Nâyebi
Yeah. So, like, let's go through the evolution of maybe just AI itself. Because when we started to build AI, let's say many years ago, there were different approaches. One approach was, in most cases, to build rule-based systems or expert systems to say, okay, can I find out how to calculate or estimate or, you know, to somehow forecast, let's say, the demand? I'm just talking about the demand forecasting in this case. And you need to adjust this all the time because there are so many different types of trends depending on the time horizon, depending on your specific industry, you will have challenges that you need to tackle. Then we started with the statistical type of models, you know, ARIMAs, ARIMAX, Prophet, and those type of models, like GARCH and such, that in most cases, they're univariate models, meaning that they're just going to be looking at the trends. They're going to be looking at also the seasonality. They're going to be looking at also what we call residuals and that, to find out, okay, what would be the demand just by itself, but they're not going to be looking at some of the external factors. And that's okay, but is it sufficient to just look at the behavior, seasonal and trend behaviors of just one product? No. Then we started to have these exogenous variables and that. Then moving forward, we realized that in many cases, the relationship between some of that data, some of those variables, or in terms of machine learning, we call them features, is going to be nonlinear. What I mean is that you cannot just easily formulate if the weather is like that, and then let's say the season is this and I'm in, like, Montreal, then I'm going to be selling this many. It's not as simple as that. Then we started to build machine learning models, specifically the ones that are used for the time series forecasting, but they require a lot of data preparation. Data, we will call them feature engineering, to build, like, some specific information, indicators that are important for AI to learn the seasonality, the trends, and some of the other patterns. So we've seen that machine learning models or ensembles of machine learning models and statistical models, they perform better than, like, just statistical ones or just the rule-based ones, right? In many cases, many of the retailers would look at the last year, and they would just project some adjustments to that last year. And that would become a kind of forecasting. So from there to get to the point that we have the statistical models, and now we have machine learning models and a combination of machine learning models and statistical models ensembled together, they achieve a lot better results. So we started by just the rules and then statistics and then machine learning. Ensembles or gradient boosting type of algorithms are pretty much the state of the art in terms of tabular data. And then we got to now this AI, right? So what I'm talking, and when I'm saying AI, just to simplify, is the large language models or the foundation models or the frontier models that are pre-trained. They know many of these concepts. They're not good at prediction, but they have reasoning abilities. So if you combine large language models with machine learning or, like, ensembles and those types of things, you get the best possible results. So there's always something and there's always some different granularity. There are always some challenges in terms of the trends, you know, in terms of you will introduce a new product and you don't have the data for it. So how do I find out if I'm going to sell this product if I don't have any data about it, so I cannot really use machine learning? Then I will need to go find out, hey, this product looks like the other product that I had last year, but can I use the data from that product for this product? So these type of discussions.
Samuel Boulanger
Great. And you just introduced the large language model in the equation, right? So how do you use machine learning in collaboration with large language models? Is it what we call agentic AI, basically? So for our non-technical audience, can you just explain in simple terms? Because there's a lot of noise on the market right now. Everybody's talking about...
Fâtih Nâyebi
Yeah, exactly.
Samuel Boulanger
Agents, everybody's talking about GenAI. Now, what's the difference between simple, simple, it's not that simple, but GenAI that everybody knows, like ChatGPT, traditional automation, because I think there's a confusion between automation and agentic AI. So from GenAI to automation to agentic AI. So can you explain in simple terms, like, what's the differences?
Fâtih Nâyebi
Of course. And I think this is a very important question because I think many of the people in the industry are kind of confused with all of these things. And is it just a buzzword? Is it, like, a hype? Or is it what exactly? Let's actually tackle that. We used to do a lot of automation with robotic process automation, right? So, like, the RPA systems, what they would do was to go through the screens and use applications, really, to say that if you click on this, enter this information here, and this and that, you could automate this process. But they were kind of dumb because they wouldn't understand the context. They wouldn't know what they're doing and they would just follow some sort of a protocol. And that would apply also for pre-ChatGPT era chatbots as well, because they would just follow a sort of a protocol, right? Then we said, okay, can we make them a lot more interesting and then a lot more intelligent? With the dawn of these frontier models, we started to have generative AI, meaning that the AI models were able to generate content. Okay, so now you're asking a question and then there's a model right there that can understand what you're asking and can generate the text to answer. But this is just a generation, or it could be, I mean, image generation or it could be also, like, video generation, but just a generation. Then, okay, is it actually sufficient to make things happen? Because, yes, it gives me some answers. So what do I do with it? It does not execute on anything. Then we said, okay, this is just maybe an assistant. I can talk to it. Then let's go find out if we can build Copilots. And while I'm doing something, then let's use these large language models, foundation models, to actually achieve things. Part of the process would be to generate a text to send an email to someone. But we have, let's say, Copilot Studio, or we have, like, Power Automate, that they can follow instructions, and you can actually build a workflow or process. And part of it, you plug in large language models and it can read an email, understand that, it can generate something, and then your system is going to send that email to the people. By itself, it doesn't do anything. It just generates something. Now we get to the point that can we use this intelligence that is behind even GenAI to make it actionable? So it understands that it needs to do something and then uses tools to make it happen. In an example, I want to have, like, a multi-agent system that can optimize the inventory. So what I mean by that, you know, you want to know how many of these products are going to be, let's say, in this store. I'm just going to oversimplify it. And then for that, you're going to need to know how many of the products you're going to sell, right? And how many is going to be demanded, really. On that side, you will have a demand forecasting to say in the course of next 10 weeks, I will be selling this many products or these many number of product units is going to be demanded. Then if I know that these many units are going to be demanded and I know how many products I have in my inventory, I can predict how many of the products are going to be in my inventory or when I'm going to be running out of it. So if I look at this from the agentic point of view, not just maybe generative point of view, I can have an agent that uses demand forecasting as a tool, because it can use tools. It can call APIs. It can write code or whatever. But let's say that I created the demand forecasting model and exposed it as just an API. And then I gave this API to my agent, which is, like, an inventory optimization agent. Then I ask this agent, what are your suggestions in terms of the inventory and what should I do, right? Then I'm going to be needing to replenish, to bring in more products. The first thing that it needs to do is to go check the demand forecasting results, maybe call it to say, hey, go do demand forecasting for me so that I can know how many of these products I'm going to be selling in the next, like, four weeks. And based on that, it will receive back the results because I can call the API, will have the result, and then use that information to say, I would say, because you're going to be running out of these number of products in these many weeks, then because of the lead time, because of the cost of delivery and this and that, you should be maybe replenished or go order these many products. So that is a way that agents are to work. Just to say it in a different way, imagine an AI that can listen, can understand, but it can also use tools. The tools can be anything, but most probably can be, like, searching a website, searching Google, let's say. They can go read files from your laptop or wherever you actually run them. And they can call APIs. And if they can call APIs, they can use any other form of the computer program as a tool. And in this case, our machine learning demand forecasting model becomes just another tool for our agentic system that knows how to call APIs to use tools and then to use the results to do other things.
Samuel Boulanger
So traditional automation will be part of the tools of the agent as well. So it can retrieve information, but it can take actions, right?
Fâtih Nâyebi
It can be that, like, your agentic system runs existing workflows, right? You can kick the existing workflow, actually make things happen. The key ingredient of these agentic systems is the reasoning engine. The reasoning engine, what I mean by the reasoning engine, is really the frontier large language model that knows the concepts. Then, when you say, okay, go do inventory optimization, it knows what is inventory, right? It knows that, okay, for the inventory, you need to understand the demand. And because it understands the business as well as being capable of calling APIs or using tools, it can make things happen. Now, whatever application or system that you have, you can put it in these agents' disposal and then they can use them to make things happen, to execute on things. Maybe they're going to be calling an API to send information to an existing ERP system, to SAP, but they're capable of doing these type of things.
Samuel Boulanger
I will assume that you already built some agents for ALDO. Which use case right now, if we have to pick one, which one already delivered measurable value, like good ROI? What's your top use case for an agent in the retail business?
Fâtih Nâyebi
So we've been working on a program that is called Revenue Growth Management. And this is partially funded by Scale AI, which is a supercluster from the Government of Canada, not the Scale AI that Meta bought, not that one. So this one provides grants for the projects in the intersection of supply chain and AI. This project includes demand forecasting at very different granularities and different time horizons, as well as markdown optimization, basically adjusting the prices and the discounts, and order fulfillment, a model that decides from where to ship to the client. If you buy a product from e-comm or the website, we ship it from a store, not from a distribution center or fulfillment center. You can just ship it based on distance to say, Sam's home is here, so what is the closest store? And then let's just ship it. But that's not what we want to do. We want to look at the demand for that specific product at that store that is closest to Sam, but also to the other stores that are good candidates for delivery, and pick the ones that we know the demand for that specific product is going to be lower, and then ship it from there. In this case, again, we're using demand forecasting to identify the demand in upcoming weeks and then to say, okay, from which one, from where to send it. The other one was the markdown optimization. And each of these models, they're working in a sense that the markdown optimization is going to be using demand forecasting and then is going to be doing also price elasticity to know if I'm adjusting price like this, what's going to happen with the demand. And based on that, it's going to say, that's the final price that we suggest to be stickered on the product.
Samuel Boulanger
Is it an agent doing the orchestration, or is it just basic foundational model that you use with automations?
Fâtih Nâyebi
What we have right now for these are not really agentic. They are like AI applications that are multiple models that work together, more so of a workflow. But in the future, we're looking forward to have, like, let's say, demand forecasting model to be used as a tool for this markdown optimization, let's say. Why this is interesting and important: when you do the demand forecasting, you have these numbers and, like, you just feed it into the optimization model. But before feeding it, you're going to need to look into the results to understand if there are any troubles and any problems with the forecast by itself. And in most cases, you need humans to look into that and say, hey, this makes sense. This doesn't make sense. Maybe the model made some errors here or there. With an agentic approach, this orchestrator agent can review, because it has the right reasoning abilities and it knows the business, can review the results of demand forecasting and say, hey, wait a minute, this doesn't make sense. There's an issue. Let's just go maybe rerun or maybe adjust. And then I'm going to be using it for the markdown optimization. But we did not implement this piece yet. We worked on some other agentic AI use cases, more so to improve our productivity so that we can do more of these things. I can give you two examples. One is that my team built an agent product owner. Like, it's like an agile role that needs to manage the backlog, right? To write stories and all of those types of things. Now we have an agent that you provide, like, a blurb of text and it goes through the whole text, creates all the stories, puts the success criteria, definition of done, definition of ready and everything. Looks at existing workload on Jira, who does what, and assigns tasks already, tasks and the stories to the right people by itself. So that's one example. The other example is in our privacy office, data privacy office, that it goes through the data privacy requirements and requests from the clients and then cleans them up, groups them together, and creates also tickets in the ticketing system to be dealt with. So these are two examples, but we had some interesting sessions with you guys, right? So to look into some more opportunities. And the team is working on many different use cases at the moment.
Samuel Boulanger
Yeah. That's really two interesting use cases. I mean, I think you're just limited by the imagination when it comes to building useful agents. What I found with working with different organizations out there is that sometimes the use cases that are identified are not necessarily the ones that have a strong ROI. I mean, people want agents to do what they don't like to do, but it doesn't mean that it takes a lot of time, or it doesn't mean that a lot of people are doing it. I mean, if you're building an agent to save an hour a month and you're the only person doing that task, that might not be the good use case, right? What are the key metrics or signals that will show that an agentic AI system is really bringing value and will deliver an ROI? And actually, I don't like the term ROI because it's not always a return on investment, right? It could be improved quality, it could be the quality of life improvement for the employee as well. But what are the key metrics you're looking at?
Fâtih Nâyebi
So there are different types of objectives that you have. Some objectives are related to productivity. In a sense, you want to ensure that you're more productive and then you can do time saving. And then some are related to cost saving and cost optimization, that you can do automation for that. And some of the use cases are going to be actually like ROI or, like, gross margin related, to say that I want to actually be able to do more. The way that you will calculate is that, can I do some time saving? So that time saving is of course money by itself. Because if you're wasting your time on something, that's for sure, it's like the money that you can recuperate. And then, because I'm wasting too much time on this, then I'm not able to do this and this and that. It's like an opportunity cost as well, that I know that I cannot go do more. And then you can maybe calculate that in some cases. For instance, if I'm not able to go check all the products, like the demand for all the products, I may have some products in my collection that are not performing very well, but I'm just overwhelmed with so many other things that I cannot actually look into the right analysis and provide better solutions for them. So that's another thing. Then you have some of the use cases, like the markdown optimization, that they have the direct impact on your profit. If you can reduce the, let's say, the markdown percentage, like some numbers, some digits even, it will have a huge impact on your bottom line. But how that is possible is really based on if I can do better demand forecasting and then a better price elasticity model and a combination of these that can lead to better prices, then I will get more gross margin and I will have a very good sell-through. So the challenge with agentic systems is that you're going to need to give them the environment to do things. And once they do, then you will need to calculate what has been done and then compare it to some, like, B groups. Let's say you should do some sort of a testing, for instance, like you need to do A/B testing to say, if I have an agent that does X on some of the stores, then I should have some of the other stores that I don't run agents on, and then you just get some results and then you compare after a while to see how the impact will be. And in this case, just, like, on the retail side, you can definitely go and look at the real numbers, the sales and those types of things, to see how much improvement has been done in the one that, like, agent is running and against the ones that humans are running. This is pretty calculatable, but forecasting it, estimating it, is not as easy. You can say I'm doing some testing on historical data, but to see these things in action, you need to let them run and then calculate that. Then you can know what exactly has been done.
Samuel Boulanger
Interesting. You mentioned the opportunity cost as well. That's not something I was using when talking about ROI around agents. This is definitely a good one. I'm thinking about a seller who spends so much time doing administrative tasks. This is an opportunity cost of not being in front of their customer. So it's totally a good one to look at as well. And the A/B testing also, it's a good tip. I think that everybody should apply it to just see what's the impact of these agents, if they should continue with the development of these agents.
Fâtih Nâyebi
Yes.
Samuel Boulanger
For any retailer out there that is just beginning their agentic AI journey, which is the case of most organizations out there, it's so new, what will you say is the foundational step that they should take first if they want to go as far as ALDO is doing right now?
Fâtih Nâyebi
The first thing that I would start to look at is going to be the tooling around the AI itself. What exactly I can use and what exactly I can squeeze out of existing tools that I have access to. Let's say that if I have Copilot, let's ensure that I know how to use it and everyone is actually using it. Because there is a lot of potential there. The more that you use it, the more productivity gain, the more time saving, and then the more things, the better way that you can do. You need to do better coding. If you're an IT organization that has coders, programmers, just go ahead and use agentic coding. Because of course you can be a lot more productive. You can do many things with the agentic coding systems. You can do the data analysis. You can write code. You can debug. You can document and that. This is going to drastically improve existing systems. You can do a lot of unit testing. Who wants to do unit testing, right? Maybe some people do, but you know.
Samuel Boulanger
I don't.
Fâtih Nâyebi
And AI is pretty good at doing unit testing and writing all this unit testing and running it. So you're going to have a lot more robust system. So you're not going to have downtime, and because you don't have downtime, business can continue actually to do that. So that would be the first thing to do. Okay, now that I'm getting better and better actually using these AI tools, everyday AI tools, and then getting more productive, then let's go do some of the agentic stuff, really. I mean, automating processes, making them more intelligent using agentic systems, LLMs behind the scene. Tap into, like, OpenAI APIs. Use, like, Agents SDK or those type of things that behind the scene they use OpenAI APIs, let's say GPT-5, which is very new. Then, trying to adjust to understand what's out there. And then...
Samuel Boulanger
A few hours ago.
Fâtih Nâyebi
They can call tools, provide some of these systems to that, and then see moving forward how these systems can help you in doing anything. One key area that I'm really looking forward to is to democratize data engineering. Cleaning data, preparing data, engineering data is not an easy one. And every system has its own, like, schemas, and you need to do a lot of mapping. You need to prepare a lot of data for AI. If you can automate that task by itself to, I mean, not fully automate, but improve the velocity over there, productivity over there, then you can feed a lot more data into these AI systems. So this is going to have a compound effect in the future.
Samuel Boulanger
It will, and I think it will affect how people work as well. As these agents take more place in the workplace and take more operational decisions and are more broadly used, how do you see the role of the employees or the retail teams evolve? I know a lot of people out there are worried about their job. You can see this all over LinkedIn, like an AI will take your job, or somebody using an AI will take your job. But how do you see people working in retail, the daily job evolve?
Fâtih Nâyebi
The AI is very exciting. I understand many people would think that AI can take their job because it's very capable. And it's the first time in human history that we have something that could become more intelligent than humans. Imagine, like, an AI, they talk about, like, having PhDs in your pocket or something that knows many different domains at some good level that you can always consult and you can use for different things. And some people, they're just, if they're not going to be using this type of tooling, they're not going to be as efficient, as productive. And at some point, they're not going to perform in their job, right? And as organizations, we should provide them with the right training, education, learning, so to ensure that they can use this type of tooling, because I'm sure that they would want to use it because they're going to be having an easier life for sure. And this doesn't mean that, like, AI is going to replace you. AI is going to enable you to do more, a lot more, instead of just dealing with very small, small, small things that are not super interesting. You can automate many of those types of things. You can make more intelligent decisions, and then you can achieve more, but you need to be using it. So anyone, I'm talking about not just retail, like any industry. And this is going to be huge. Let's admit it. We don't exactly know where are we going, but the implications of this AI revolution is going to be huge. Like Demis Hassabis, the CEO of DeepMind, is saying that, like, this AI revolution is going to be 10 times bigger than the industrial revolution, happening 10 times faster. Like in 10 years, we're going to get to the point that we're going to innovate for 100 years.
Samuel Boulanger
If you look at the graph and you're seeing the curve of innovation, it's like it's a wall right now.
Fâtih Nâyebi
And we see that. We see self-improving AI systems. And I just want to emphasize this. If AI can write code, it can do everything because AI itself is written as a code. And if it can get better at writing code, which it does, that's not news. You know, everyone knows this now, then it can improve itself. It can become more efficient. And in fact, there are systems doing that. You know, there's AlphaEvolve from DeepMind. There was a publication just around that. GPT-5 is the same thing there. They say that it's just self-evolving and self-improving. And every time that they get better, they become more intelligent, they can do more and more and more. So how this is going to have an impact on us, maybe we're going to solve all the humans' problems. Maybe someone is going to say, okay, I'm going to be using AI to solve, like, global warming, or I'm going to be solving, like, all the different disease, like, the healthcare issues. And, like, no one is going to be poor anymore in the world. Maybe that. Or maybe some of us will say, hey, okay, this is a great time to go actually do some bad things. Because now you can do pretty much everything.
Samuel Boulanger
Yeah, I was about to say, in some cases it might sound worrying that it's self-improving and there's not a lot of guardrails right now around AI development from all the big tech companies out there. And I fundamentally think that AI will solve a lot of our problems, will probably create new ones, but we're talking about disease or global warming. I am convinced that we need AI to solve these issues. However, it might create other issues. I don't want to go down the societal side of things, but from an organization standpoint, how can you build guardrails or kill switch, or make sure that the AI you're using and you're developing is ethical and won't go over its boundaries?
Fâtih Nâyebi
There are techniques to review the results of AIs by humans, but also by AI, and to look at the confidence levels, to implement guardrails to ensure that they don't get out of what is accepted. There are quite a few of these systems being implemented, but are they sufficient? You know, the more intelligent that AI becomes, can it actually trick those systems? It can, right? So there are so many of these things. The guardrails, we can implement them, but then the governance of AI at the governmental, institution level needs to be also sorted out. Let's say that if I'm building an AI, I should be responsible for whatever that I'm building. And if it's harmful, I should be responsible. I cannot just say, hey, AI did that. No, I'm building it. If I'm faking people, I'm, like, abusing something, and you know, it's like someone is actually faking your money, right? If someone is faking money, they're responsible. It cannot say, okay, fake money is bad. Okay, fake money is bad, but you're worse because you're actually printing some money, whatever. If you're building bad AI, you should be responsible. And what that bad AI is, is not super clear. That's one part of it. The other part of it is that AI as of today is very sophisticated, and as human beings, we are not able to really understand what's happening inside very sophisticated ones. So there are two things. One is we can look at the results and then start to explain why AI said this or that, or why it created this or that. That's called explainability. But if you want to really open the box and start to look inside to interpret how this AI works, it's not as easy. In fact, in some of the most sophisticated neural networks, it's impossible because there are so many dimensions. It's like, okay, we cannot really understand. For those, we need to start to build some techniques that make it simpler to really understand, to better understand, to better audit these type of systems until it's too late. It becomes too sophisticated that we cannot do anything. Honestly, if you're at that point, it's, like, too sophisticated for our brains, but maybe we can use AI to solve this issue by itself.
Samuel Boulanger
I think it's called the black box concept, right? We know what's coming in, we know what's coming out, but we have no clue what's happening in the middle.
Fâtih Nâyebi
It's like our brains as human beings. We don't even know our brains, and they try to mimic the brain. There are some people saying that, yes, this is really like some of the concepts that we have in neural networks, let's say, are really parallel with what we have in our brains. And some of them say, hey, how, because of this and that? So there are a lot of discussions around that. But as you said, it's a black box. You know what comes in, what comes out. And then based on that, you try to say, okay, well, how this is built. But we're talking about many things that can happen in that black box. Yeah.
Samuel Boulanger
So as a business using, creating agents, or using AI tools, like existing ML model, I always assume that the best course of action right now is to make sure you use an internal model or you use an internal platform that already has some of those guardrails embedded.
Fâtih Nâyebi
So it's important to pick the tools correctly, to be very mindful of the type of tools that an agent can actually use, to provide the secure and private environment, and also the right privileges as well. So it cannot do this and this and that. It can just do the things that it needs to do, the type of APIs that it can call, the type of functions that it can call. So once you limit that, and then you go for the agents that are like vertical agents, that they're very good at doing some specific stuff but not doing everything, then it's a lot easier to control these things. And they can perform a lot better. And of course, you should provide some observability, transparency, like audit logs and those type of things around these systems. And then maybe use AI also to analyze them constantly. So the AI can be a problem but can be the solution too. Many of these things that we are talking about can be done with AI itself. You have the AI auditor, you can have the AI reviewer, you can have the conflict resolution AI agent. You can have all of these types of things that work together to ensure that whatever that they do is okay. Let's say you have five agents that are going to review one agent, and then they're going to calculate some confidence intervals. And based on that, they're going to say, okay, raise the flag or not, or just do this or that. That's all possible.
Samuel Boulanger
This is funny. It looks like human beings, right? Where agents are looking after other agents, make sure that they don't go over their boundaries. That's the same thing you're doing with humans.
Fâtih Nâyebi
It's the same thing. Now we talk about multi-agent systems, but we also talk about hierarchical multi-agent systems. So there are going to be agents that will, some of them are going to be the boss or, like, the reviewer or team leader, whatever. So, like, you've maybe seen examples of agentic software companies that every one of the roles in a software company is built as an agent, but then you still have a boss, you have a product owner, you have this and that. And you have an architect that reviews the code, and yeah, it's pretty much the same concept.
Samuel Boulanger
This is fascinating. I can't wait to see where it evolves. We're almost at the end of our time, and there's two questions I'm asking all my guests. This is new, actually. You're the second guest I'm asking those questions. But what's your number one productivity tip when it comes to using AI in your own personal work? And the second question will be looking at 10 years, how do you think AI will change the way we live and work? So let's start with the first one, obviously. What's your number one productivity tip?
Fâtih Nâyebi
The biggest productivity tip for me is agentic coding. Because if you have an agentic coding environment, you can do everything and anything. And you can make everything happen a lot faster. And even the simple stuff, like reading emails and all those type of things, they can be a lot faster with that approach. But maybe not everyone wants to code, of course. If you're not using the coding and agentic coding and those type of environments, then of course use ChatGPT or Copilot, whatever that is allowed at your organization. And, like, 10 years ahead is honestly, it's really hard to know. It's like the way that the AI evolution and revolution continues, I see that is going to be beyond our comprehension. Because they get better and better. They're going to get, like, a lot more intelligent and my mini brain is not going to be able to catch up to really understand the implications. That's what I would say. There are some people saying that, like, in maybe a year or that, like, the coding is going to be just fully automated. And that for me by itself is huge because if the coding is automated or we're talking about software on demand, then it means that AI do pretty much everything because they can write a piece of code to do anything in digital mode. I'm not talking about, like, whatever that requires dexterity or, like, the physical environment. But then you may say that, okay, what happens with robots and that, so maybe the next frontier is really going to be having the robots that understand better physics and they understand their environments, the world better. And then they have the same level of intelligence as, like, existing frontier models, and they can achieve many interesting things. I'm on the optimistic side of things. So I hope that many of the researchers, many of us, we're going to be solving pretty much many of the problems that do exist. And then we can have fun. We can be nice. Yeah, I'm going to stay there. It's just that's kind of wishful thinking, but yeah, I'm just going to stay there.
Samuel Boulanger
Yeah. I love the concept of software on demand, and let's see how it unfolds. But I do agree that it's not an easy question to answer, specifically when you're in the middle of the storm and all what's happening and at which speed it's going. Thank you so much, Fâtih. That was a pleasure. It was very insightful. And I think our audience will really gather a lot of knowledge and best practices from our discussion. So thank you so much for joining me today.
Fâtih Nâyebi
Exactly. My pleasure. Thank you very much.