Doron Hazan: 0:00

The concept of agents is alive. And you can see how tremendous they are at being assistants and being autonomous, helping us doing things. And if we think about it's, it's like a very strong brain but the brain is not enough The brain needs more organs. It needs a nervous system.

Tom Raftery: 0:29

AI has become remarkably good at thinking. But what if the real limitation isn't intelligence at all? What if it's the fact that AI still can't see what's happening in the physical world? Good morning, good afternoon, or good evening, wherever you are in the world. Welcome to episode 134 of Resilient Supply Chain Stories and strategies that Keep business moving. I'm your host, Tom Raftery. In today's episode, I'm joined by Doron Hazan who leads AI at Wiliot. We explore why the next leap in supply chain AI won't come from a better model, but from giving AI continuous awareness of what's happening across products, pallets, and warehouses. If you are responsible for supply chain, operations, or logistics, this conversation challenges one of the biggest assumptions in AI today. After listening to this episode, you'll come away understanding why the companies that get the most value from AI may not be the ones buying the smartest models, but the ones building the strongest data foundation beneath them. Let's get into it. Doron. Welcome to the podcast. Would you like to introduce yourself?

Doron Hazan: 1:46

Yes, thank you very much. Pleasure to be here. So I'm Doron at Wiliot, three years based in New York, and excited about AI and the weather is super sweet. So we have great mood here, great day.

Tom Raftery: 2:00

Doron, for listeners who don't know you, what do you do at Wiliot and what problem are you trying to solve?

Doron Hazan: 2:09

Yeah. So at Wiliot I guess I should start with what Wiliot is trying to do. We're in the field of ambient IoT, really trying to give the ability of every little thing in the real world the ability to speak. In the very abstract terms, that's what we're trying to do. So we're, we have the technology and we have the platform to enable that so our customers can actually talk and interact with their things. And what I'm doing is I am I'm responsible of the AI component of it, which we're gonna talk about, how do we integrate AI into this system, into this technology to make this work?

Tom Raftery: 2:54

And what pulled you from the likes of machine learning and responsible AI into supply chain operations?

Doron Hazan: 3:03

Yeah, no, that's that's a great question. I had very little knowledge about supply chain before joining Wiliot. Didn't really know much about even, sensors, RFID, active BLE, all of that. What I was passionate about, coming from, from doing more not hardcore, but more traditional machine learning research is, the question of where AI will go. And I started with this before ChatGPT was actually pretty cool. I am a big big believer in artificial intelligence. So what people call artificial general intelligence, some people call that now, or singularity, whatever you want to call it. I am a genuine believer in our ability to get there. You can call it in many ways, consciousness, the brain, that, that kind of stuff. Or I am a believer in that. And I think after seeing all the developments and in the digital world, we're gonna talk about, I think I think that we are much closing the gap in that realm. But I think there's so much to the real world around us that is mostly about physical things, which is completely related to supply chain that AI is, not there. It's a little stranger there. That's what actually attracted me because I feel like if, if, if we really wanna get AI to where it should be, and I genuinely believe it will be, we need to put the effort in that domain supply chains in the physical world and, and all this.

Tom Raftery: 4:32

Okay, and why has this become urgent now AI in supply chain rather than just another supply chain tech discussion?

Doron Hazan: 4:44

I think it's urgent now because, I mean, especially now as, as you can see, that the agents are, the concept of agents is, is, is alive. And you can see how tremendous they are at being assistants and being autonomous, helping us doing things. And if we think about it's, it's like a very strong brain that we have developed and we're continuously gonna develop. I think everyone now is certain about the cap capacity capability, but the brain without, I would say a nervous system, or something like this is not enough to operate. So I think especially now when you can see how the brain is actually here, some people were maybe a little sceptic. We can see that the brain is here. The brain needs more organs. It needs a nervous system. That's why I think it's especially relevant now.

Tom Raftery: 5:36

And so if we're talking about brains and sensory systems feeding into brains, what happens when, system record and physical reality disagree? what breaks?

Doron Hazan: 5:54

When a system record and a physical reality, disagree. The power of having many votes. I think to answer your question, if you can look at that, an analysis may be coming from traditional machine learning algorithms. When you have algorithms like decision trees, for example, or yeah, that's just a traditional machine learning algorithm where in that predictor you have many, many little voters that predict something. And the way you actually combine an overall prediction is that you do a vote, some sort of a vote between those little decision makers or predictors similar to the physical world where you have little sensors or little, predictors I would say that disagree. I think the power of the, of the many is actually what, makes a real good vote, but because they will disagree, this is a, and it's a real thing in nature. The things, it's, it's never a consensus.

Tom Raftery: 6:55

Can you gimme an example of where an AI system looks smart on paper but fails?'cause it can't see what's happening around it?

Doron Hazan: 7:04

Yeah. Many times, let's think about a robot, for example. A robot or more the things I call about. More of the physical AI. The, I I, I'm especially talking about AI that is in the physical realm. Not just talking about the digital realm, like the agents, but let's think about a robot or an an electric car. Maybe let's pick the robot situation. Robot has maybe three modes of operation. It senses, maybe it has some sensors, then it ticks, it's computes, and then it acts based on these things. That's the way I see it. Robot has a task maybe to do some one thing, maybe to take this box and put it in this conveyor belt. So that's what robot does. It has sensing, it looks at what's around it, and it takes the box and maybe wants to put it in the conveyor belt. But the AI, that robot is not aware of, let's say the distribution centre is gonna go through a shutdown or something like that. So the conveyor belt is not gonna run. AI is not aware of that situation. So AI is trying to do something and maybe it just failing because it's just simply not aware of what's happening beyond the scope of, of its sensing.

Tom Raftery: 8:15

And if we then start putting sensors on things around the place, what changes when a pallet, a case, or an item starts to become a source of live data?

Doron Hazan: 8:30

Yeah. And that's where the real magic is I think. I think that those two actually complete each other. It's not like this changes versus that. But if you think about this one unit, that one brain the, that robot that has everything in it, so the, the sensing is in it, the compute is in it, and taking the actions using its tools is, is in one unit. It has some advantages, of course. It's fast, it's reliable within its scope. As we said, you know, it's not, everywhere. It's not spread out. So once you go into more cheap, not super good, it's like the same analogy to the decision tree predictor which actually is one of the best, I think, classical models till this day. Lots of researchers and, you know, machine learning practitioners use it. When you go to these cheap, not super reliable predictors, I would say, which can be sensors or other things. You can allow for real time snapshot of the environment, not necessarily a hundred percent accurate. It will never be accurate, but it will give the brain that would work with it a better understanding of the environment. So that's kind of the change.

Tom Raftery: 9:43

And. We haven't mentioned it yet, but physical AI is what we're talking about. Can you explain in one sentence that a supply chain leader listening to this would understand and remember what is physical AI?

Doron Hazan: 9:58

Yes, of course. Physical AI or physical intelligence, and it's, it's a term like if, the simplest way I think for me to understand it or to explain it, is what we refer to as AI now is mostly digital. The AI that we're used to which are, the models, the agents, all that they were trained, and interacted with information, data coming from the digital world, mostly texts, images, videos, all this. Physical AI pertains to this AI system, AI agents and, and, and models that interact, learn, and train, act on data coming from the physical world, it's more dynamic and, and it's not necessarily the, the same types of data that what we're used to the, the AI that consume the form, yeah.

Tom Raftery: 10:52

And where does the likes of this continuous sensing add real value, and where is it overkill?

Doron Hazan: 11:01

Yeah, great question. I think if we need to break it down, I guess, in decision making I, I, I like to think about this, especially in, and what I do in Wiliot about, how does AI help? If we look at this, if I break it down into four steps of, this decision making framework. First step is you want to understand what happened. It's more of a descriptive analysis of what actually happened before, which is I think where, where we are with most things Then what you want to understand is what is happening right now, which is related to the sensing and putting a lot of senses there. Then the next steps are predicting what's gonna be the next move, and then, okay, I predicted what's the next move? What do I do with the next move? So this is kind of like, if I have to break it down the way I see it into four steps of decision making. I think the sensing, putting a lot of sensors there, doing this maybe distributed physical AI. So I like to separate it into, distributed is more of the sensing versus the integrated, which is one brain. I saw it somewhere and I really like it. The sensing comes in the, the first two parts. You need to understand what happened before, but you also need to understand what's happening right now, to make a prediction and to act on it. Now, where it's an overkill, I would say in very simple scenarios. Some of our clients that their use case is very simple, I would say, and simple. It means, let's say for example, instead of a big supply chain and you have downstream flows from supplier that move things to to a distributor, and it goes through massive distances. You have just a warehouse and the warehouse can be very small. Let's measure a simple room just for simplicity. In that case, if you wanna do something like just track your inventory and you have a simple room, sophistacated snapshots sometimes can be needed, depends on your use case, but it may be an over overkill for a very simple use case.

Tom Raftery: 13:11

Okay. Yeah, that makes sense. And is the industry at risk of buying AI before fixing the data foundation that's underneath it?

Doron Hazan: 13:21

Yeah, no, absolutely. Absolutely. I, I strongly think so. If you have that AI without enabling that sensing capabilities, you, you pretty much go nowhere. I think the AI is, it's, it's a false kind of conception that we have right now, or perception that AI can do so many great things, and AI is incredible, and I'm the biggest proponent of it. You see it in the digital world as, as I said before, can do so many things. But in the physical world, if you don't give the data foundation to that AI, it's pretty much useless. So, enabling, it's giving a really, imagine a really strong brain, as I said before, without being able to see or to touch or to move. What, it's a, it's a, it's a useless thing to have brain.

Tom Raftery: 14:13

I mean, that takes us from the idea to the messy bit, which is deployment. And what then separates the likes of a useless pilot from one that never escapes the innovation lab?

Doron Hazan: 14:29

I think it's a lot of repetition. It's what you need to do. It's a lot of, it's not just that you deploy things and you, hope for the best. I think there is it's, it's a, it's mandatory for, especially for supply chains that there's gonna be human involved, especially for decision making. I have an interesting, an analogy to I am pretty sure it's a Jeff Bezos quote that says that as a CEO, I have, I make very few decisions throughout the year. So big decisions, but very few of them. In supply chain, I think the situation is a bit different where you have to make decisions every, every now and then, every short periods of time.'cause a big supply chain is just to give, to give kind of an illustration. Imagine millions and maybe billions of items flowing around, and its a chaotic system. For decision making, we really have to make for a really good decision making framework. With AI, we really have to make decisions every now and then. So it has to be continuous. So going back to my point is where you deploy things, you need to have the tools to, to evaluate fast enough to see, okay, are we going where we want it to go? Do we monitor it? We act on it. We see the results. We do. We don't just leave it and hope for the best because you can't just make these decisions once in a while. You have to be involved and you have to have the right tool set to be in control because the systems are, are very dynamic and very chaotic in a way.

Tom Raftery: 16:09

Sure, sure. And particularly in a dynamic environment like a supply chain. I mean, what, for example, looks sensible on a dashboard but falls apart on a dock or a truck or a shelf?

Doron Hazan: 16:20

It's a good question. I think that if the data foundation is good enough and everything flows well, then the dashboard which is sits there on the top of the pyramid will be good. The dashboard would show you good things if all, everything flows well, and then the information that feeds that dashboard would work well. But if something breaks in between, then a dashboard would show you just static or irrelevant things or, and then you, the, the, the interesting thing is maybe to give insight to your question is you won't be able to know, 'cause the dashboard can say, Hey, you have I don't know these, anomalies or, or this is your inventory. And you would say, oh, okay, but you won't be able to know actually what broke in between, because maybe, you didn't see that part. Something fell through there. But a dashboard can say, okay, this is your snapshot. Okay, it looks good. But in fact it's, it's not the case.

Tom Raftery: 17:14

And what surprised you most when you moved into this live supply chain environment?

Doron Hazan: 17:18

That is so chaotic. Yeah, it's, so, I think that's the, the, the biggest thing.'cause I'm used to work with data sets and, and online data, data that you can find in the digital world where you can have more control. In the physical world no one has control. That's why I think this is one of the main challenges of, you know, our clients at Wiliot and, I think in general. You would be surprised, I was so surprised, so surprised to see how little control and visibility anyone has towards big sub supply chains. It's just so hard, so hard to understand this. And you think that you, you think behave a certain way, maybe behave a certain way 98% of the time, but those 2% of the time they don't in a massive scale. It's huge. So these, these, these things, these chaotic environments is something that I, I, did not, I mean, maybe I expected but not imagined to be so, so significant.

Tom Raftery: 18:23

Sure. Sure, sure. And is there anything that with this experience now in supply chain, that you've changed your mind about AI in supply chains?

Doron Hazan: 18:33

I don't think I changed my mind. I think I just got more intuition into why the data foundation is important. And I think most AI experts would tell you that. I mean, I think it's a common sense that the data is more important. I mean, the model is important, but the data, it's garbage in, garbage out. So the data is so important. And I think, even before this revolution of gen generative AI, I would associate that, atribute that, let's say ChatGPT. But it's not just that, of course. Lots of major steps in the data, collecting data, data labelling, all that kind of, of movements. Actually, actually I think were big factors into enabling this amazing things we see today. So I think it didn't really change about my way of AI, but I think it's similar that to do that in the physical world, enabling that data, collecting the data, visibility, all, that will actually enable an equivalent of what we see now with agents. But in the physical world.

Tom Raftery: 19:40

Right. And what about human oversight? Where should humans absolutely stay in the loop?

Doron Hazan: 19:48

Oh yeah. in, in decision making. So if we're going back to the decision management steps that I see where you have, first you have to have to make a decision to have, to make an informative decision, especially at scale or fast decisions. You have to, first understand the descriptive part, like what happened, what is happening right now, what will happen, with some confidence, and then you make a decision. I think that's, that's what I see it. AI, even if we have all the context in the world, we have sensors and we have everything spread out. AI cannot understand things that are being said in meetings and in a room between you and I. We're still, we're still not there. Maybe we can imagine a world where AI is integrated anywhere and can integrate it in our eyes and our ears, and, we can talk about that, but I, I definitely think that this will happen in the future. I don't know how long, but I don't think that that long. But AI will be, it's, part of us, but it's not the case, and we're not talking about it now. Hence, a human has to be in the loop, especially in supply chain. If you want to make, big big decisions such as, changing how much you ship or changing a distributor or, or, you can imagine on scale, you have to be informed of everything that's happening. The sensors and, and the nervous system is a, is a must, but it's not enough. We have to have the brain, which is being developed. All the AI stuff is tremendous, and it's just gonna get better. We have to have the nervous system, but we have to have also the, the context, the person that has all the context, which again, I think that in the far or not so far future, maybe that will be autonomous as well. My intuition is that it will happen. But the question is when.

Tom Raftery: 21:36

Okay. And speaking of the future, but maybe a little closer, not the far future, but let's say over the next three to five years, where does this go? Are we looking at better alerts, automated decisions, something bigger, something else entirely?

Doron Hazan: 21:54

Yeah, I think it's, better alerts and better decisions, or better insights reports or, or, or all that are just a part of a partnership, I guess between the AI and the humans in that domain of supply chain. I think in the next few years, once data foundation is more established and it's more present then I think we will be more mature into developing this partnership between the humans and the AI. Sounds like I'm coming, coming from the Matrix, but you know, the humans and the AI with managing physical things.

Tom Raftery: 22:38

Okay. And what worries you most as companies push towards more autonomous operations? And on the flip side of that, what becomes possible once supply chains have continuous physical data at scale?

Doron Hazan: 22:57

Yeah. I think this comes back to the informed decision making that a human can make. This is something that we see today in the digital world where AI is great and software engineers and sometimes big decision makers, that command and Claude dangerously skip permissions. You can do that. And then that thing is great sometimes because AI can be autonomous. If we kind of, make that similarity in physical world, I think once you have continuous visibility, if you don't have that discretion, as I said, the last step of decision making, and you do dangerously skip permission to make decisions in the physical world without being informed, without really having the, the, the, the continuous visibility that's a risk because you need to be very sure that you have control of your supply chain visibility, good predictions of where your things are, good alerting system, good insight system, all, all of this to, to really let AI do more and more of that. I think the risk is once people have that, have that control, they can say, okay, let's automate that. Let's optimise the system. Let's optimise my supply chain. Let's do it. Let's AI and do it needs to be very careful.

Tom Raftery: 24:14

And what then becomes possible once supply chains have that continuous data at scale?

Doron Hazan: 24:23

That. I think that once they have that and they have a good decision making in the process and the process itself is good, then you can, achieve tremendous things, I think. You can first of all have better, better control of your supply chain. You can understand better your gaps where things go wrong. Just, from the simplest examples of of you ship something you put, so you put a TV in a truck and it's supposed to go to location X and it goes to location y, just be able to know it right when it got to the truck. Or if something is, you have a case of, of fruits and it goes, bad 'cause of temperature, you get an alert immediately. And just as, simple examples. But you'll be able, decision makers in the supply chain industry will, will be able to act fast and act fast and scale, which I think is something that we clearly do not have. And I think once we have more and more spread of that data foundation and the system itself from the descriptive all the way to the action decision making is, is more mature, decision makers will be able to look at lower resolutions, let's say. If I have a distribution centre, how does this store did compared to that store? And how did this storage within that store did? And it's more lower resolution. So I think are very, very hard to do in right now in supply chain.

Tom Raftery: 25:47

There's that case of a truckload of Kit Kat bars going missing this summer or this, this last few months ago. Is that the kind of thing that, shouldn't be able to happen?

Doron Hazan: 26:00

Yes, Kit Kats are so great. Yeah, so that's very unfortunate. You see, that's exactly an example. Yeah.

Tom Raftery: 26:07

Okay. And for listeners who want to start implementing something like this, where should they begin?

Doron Hazan: 26:15

Yes. It depends on who who is the listener. Cause there are many scales to it. I think there are many ways to have that data foundation, to have that nervous system. There are many ways, different types of sensors. We, here at Wiliot, we, we have our own technology. The, the Pixel that's already gen generation number three is out working on generation number four. Part of active BLE, there's passive RFID, there's computer vision. There are many ways to get sensing from the environment. Just a simple, if you're a user at home, just a, just a ordinary person. Even taking computer vision like a camera and getting some input from the environment and connecting that into a brain. So connecting the brain into physical, the nervous system. Just a simple exercise to get a sense of what this physical AI thing can actually do.'cause you can act on real things from the real world. But I think on massive scales, you need to think that the user needs to think about the very little item that they want to track that is important to them. Then let that item and ability to speak. It's like giving it a mouth and whatever thing that needs to, that it is suitable. If it's a, if it's a sticker, if it's an active BLE, if it's a camera, whatever. Yeah.

Tom Raftery: 27:43

Okay. And after hearing this conversation, what should supply chain leaders think differently about when they hear the phrase AI in supply chains?

Doron Hazan: 27:54

Two things. I think first of all, that AI in my in my opinion will not be able to replace them. At least not now. So, humans have to be in the loop always, especially when the in decision making happens. That's one. And then the second thing is the double down on the data foundation. Data foundation, as we said this entire talk is probably the most important thing for AI. So to invest most of the time in the data foundation into a real time visibility, into dynamic snapshots of the, of the environment and integrate that well with the AI. those two things are probably the most important.

Tom Raftery: 28:33

And a left field question for you, Doron. If you could have any person or character, alive or dead, real or fictional as a champion for closing AI's, physical world, blind spots, who would it be and why?

Doron Hazan: 28:54

That's a great question. I, I think, to be honest with you, the instinct that I have is someone like completely unrelated, completely unrelated to the industry. But as someone is Winston Churchill. Maybe I'm a big fan. Someone that is so adamant about what to push, I think, to answer your question more seriously we're in a big intersection or a big, I think most we, we know, we now understand it where the world goes and AI, most people, if not all, we don't understand where that would go in terms of capability. Especially now we integrated into physical things. We integrate it into our bodies. The AI just gets better. The most crucial thing now is not is not the technology, is the leadership that actually pivots this and decides, make big decisions that will affect, I think, society affect us, affect the world, the universe, hence my instinct was to go with Churchill.'cause I believe as a good leader, to make decisions, not be afraid to convince, to just go with their way. That's kind of my intuition.

Tom Raftery: 29:59

Cool. We're coming towards the end of the podcast now, Doron. Is there any question that I didn't ask that you wish I had or any aspect of this we haven't touched on that you think it's important for people to think about?

Doron Hazan: 30:14

No. There's one thing maybe I, I know, I think I make the, the distinction with physical AI, that physical AI is a broad term, but we talked, I think, yes, we did cover the integrated versus distributed mostly like the nervous and the, and the brain. I think we talked about that. So I think we're good. Yeah. Yeah.

Tom Raftery: 30:34

Okay, super. And if people would like to know more about yourself or any of the things we discussed on the podcast today, where would you have me direct them?

Doron Hazan: 30:44

Yeah, they can reach me reach out to me in LinkedIn. I have my LinkedIn there and probably the best way yeah.

Tom Raftery: 30:50

Okay, fantastic. Doron, that's been fascinating. Thanks a million for coming on the podcast today.

Doron Hazan: 30:55

Course. Thank you very much. It was a pleasure.