Inside The Rise of Physical AI — With Amir Khoshniyati

[00:00:00]Can AI go beyond chatbots and agents and actually change the way our physical world works? Let's talk about it with Willot VP Amir Kosniotti in a conversation brought to you by Willot and Amir is here with us in studio today. Amir, good to see you.

[00:00:14]>> Great to see you. Thank you for having me.

[00:00:15]>> You bet. Great to have you here. So, let let me set the stage. Um, at the late 2010s, I wrote a book called Always Day One. And in it, I wrote about a company called Amazon. You might have heard of it. that was making use of AI and machine learning within the company. And they had this program called Hands Off the Wheel, where they basically took past purchase data, things like what people bought in what zip codes, in what season, and predicted how to stock their fulfillment centers with what products when, and at what price. And they even had these AI, this AI front end uh effectively do the negotiation with the vendors because they knew the prices that they needed in the ranges and became a portal instead of a person. So the vendor manager's uh job actually was handed over to AI. Now that was mid2010s to late 2010s. We're here in 2026.

[00:01:04]Obviously AI has advanced tremendously and we're entering this era of physical AI. So could you just take me and all of us on a journey of what's happened since then and how AI in the supply chain and physical AI itself has advanced from that moment until today.

[00:01:22]>> Absolutely. We're in a transformational time. I would say it's exciting on all fronts. We've we've evolved quite a bit and I I think I would take a step back before we even get into physical AI and kind of paint the picture around Internet of Things, this coin coin term of IoT, what it really meant and then how that's kind of given the foundation to where we're at with physical AI and what that means. Um, Internet of Things when it launched and it was before the COVID times, it excited everybody, but truly the understanding of it really wasn't stabilized yet. So, everyone was using the coin term, but the reality was we were just moving from every type of device actually having some access to the cloud. Even hosting information on premise, it was moving to the cloud. And you may have seen in Microsoft's pitches, in some of the phones they were launching and tablets they were launching, it was mobile first, cloud first. So it wasn't really internet of things yet. It was internet of devices.

[00:02:21]All these items were actually getting out into the market and we were getting multiple tablets, multiple devices, uh multiple phones in your everyday use.

[00:02:30]We fast forward a couple years, these devices have applications. They're built on top of it. These devices started to give rise to being intertwined and then we hit COVID and everything started to go contactless and then you had more reliability on the cloud than ever before and your devices interacting with everyday things. So these everyday things may have had QR codes. They may have had what we call a subset of RFID NFC nearfield communication for contactless payment and engagement for the consumer side. That's the chip in your phone that when you like go into the subway for instance, >> it's exactly it's standardized now. And a lot of these in the Asia market, they were picking up, but Europe and US were were behind in adopting it. All the credit cards were transitioning to these contactless chips that you could then go to these areas and then just basically be frictionless and and move forward. So the that's where internet of things really started to take rise because the devices gave rise to a level of things but those things essentially were not intelligent yet they were just a means of summoning data and what what we're now stepping into what we really feel is a definition of what physical AI is is the next frontier from this. So I I kind of took a step back to answer the question with internet of devices, internet of things. But now physical AI is our ability to take every physical asset in the world and give it a digital identity. And you go beyond just having recognition of what information from that asset is associated to a tag around it, but the condition that that asset is sitting in.

[00:04:08]>> Can I can I drill down for a second?

[00:04:09]Absolutely. So this is really interesting because um when I giving that Amazon example, the way that the company um would try to do forecasting of demand and you know figure out how many items they needed in the fulfillment center when was honestly it seems like they were looking at spreadsheets, right? You had your numbers and you know you would be typing those numbers in or they would be populated after you scan items in and stuff like that. Uh but but what you're saying is right now talking about the internet of things is that for it's almost like forget the spreadsheet. You can have effectively a real time idea of where inventory is at any given moment.

[00:04:51]>> Absolutely. And it's and it's beyond just knowing the location and the traceability of that item, but the condition that it's in. And that's where physical AI starts to pick up value beyond what historically was in place.

[00:05:04]So with condition, you know the temperature of that item. You might know even humidity of that item. You know if it's been exposed to light. So through the supply chain, if it's in a package and there's light detection through that package, you know that it's been tampered with. You may even be able to locate in on movement based on triangulation of the item being picked up in different areas. So physical AI is giving rise to physical items that are being digitized. So they have a digital identity and as they sit in the cloud you have AI capabilities that can drill into that data and then make sense on it in a predictive manner and that's what we do with Willot.

[00:05:45]>> Okay. So I have two questions coming up.

[00:05:47]One about why this is AI but let me just ask the more basic question first. Um so let's say I'm a warehouse manager and I have like items that are coming in and going out. How do I make first of all how do I collect this data and then how do I make use of it?

[00:06:01]>> Sure. So, so a starting point is understanding what the assets are that we need to to track and trace and it it might be in some cases that we we uh interact on a on a engagement level. It might be in the middle of the supply chain. So, we might start with a distribution center. We might go upstream. We start with a manufacturer um some some packaging organization that then gives rise through the supply chain. So, we have to start and see what is the asset that we're looking to track. We have a large retailer that we've went uh public with where we we started to tag the pallets as >> So this is Williet that did this.

[00:06:38]>> Yes.

[00:06:39]>> So explain a little bit about the Williet product and how you're able to track this and then you can go through the example.

[00:06:43]>> Absolutely. So so with Williet what we do, we're a physical AI platform. We have proprietary IP around IoT pixels.

[00:06:52]So these are uh reference designs that all RFID manufacturers could produce.

[00:06:57]These tags go on physical items.

[00:07:00]They're battery free, so they harvest energy off of ambient waves.

[00:07:05]>> So, we have a capacitor that acts like a sponge. It powers up and then it transmits a signal via Bluetooth low energy. And that signal that it transmits can be everything from the location, the temperature, the humidity, exposure to light, movement, all these capabilities all in a passive format.

[00:07:23]And when it's transmitted, we take it in the cloud and we ingest that data and have actionable insights behind it. And it's a small sensor >> size of a postage stamp.

[00:07:32]>> Okay. And so what what you're able to do is put it on top of packages or pallets and get this information that you're discussing.

[00:07:40]>> Yeah. So we could be at the pal pallet level, we could be at the case crate level, we can be even at the item level in in some cases.

[00:07:48]>> So and that that will give somebody now let's go back to our warehouse manager example. that will give them a real-time picture of where every single item or every pallet that they have is and that additional data that you're discussing.

[00:08:00]>> Exactly. So, it's value for the warehouse manager. Then it's also value for the brands on a larger scale because you can go beyond your warehouse, you can go into the full visibility of the supply chain. So, everything that leaves that warehouse or everything entering that warehouse, you get the value upstream and downstream through the process.

[00:08:17]>> Yeah. Okay. So, so obviously the question now is where's the AI in that?

[00:08:21]I mean that's that's data >> but how does it go from data to something that's artificial intelligence related?

[00:08:27]>> Absolutely. So from a from a platform uh level we pride ourselves on being able to synthesize and decrypt the data as it comes into the platform. So we're a source of truth all these capabilities that I went through these are all different data signals and AI is only as good as data that comes in. So you have some sense of truth that you're mining into and making logic out of. When we have data coming in, we have to unpack that data. So it might be repetitive data. It might be data in a sense of location, temperature, humidity, light.

[00:09:01]Um there's variance as items move through supply chain. We take all that data, we ingest it, we make sense of it.

[00:09:09]Once we uncover it in the platform, then the AI component layers in. So then you can ask it questions and then you'll know more around your assets and products. So this could be in a predictive format. You have assets that are reusable. They leave your facility.

[00:09:24]You can ask it when is it going to return based on trends. And then you'll get the answers.

[00:09:28]>> You may ask it simple questions like is my pallet within compliance? If your compliance levels are between 32 degrees Fahrenheit and 35 degrees Fahrenheit and any kind of variance outside of that temperature doesn't allow you to sell the item that's located inside of it, we would be able to then tell you, yes, it had a twoderee variance outside of the threshold. You can't move this to the front of the store because the truck went through some something during transit and the temperature had some volatility. so that the AI can unpack the data that that it ingests.

[00:10:04]>> Interesting. All right. So, so let's make this real for people and kind of talk it through an example. Um, so let's say I'm I'm running a store and I and I have access to this platform. Is there like a chatbot or LLM style front end where like typically I'm there's like a number of questions I need to ask that I probably can't get answers to um in in an old version of this world that I can start asking. Um, when is my inventory going to be in? Is my inventory within the temperature range like the milkshakes or the the protein shakes? Is that in the in the right range uh that I could I could be in? I mean, where is is that the type of questions that you think are or that currently are being asked by people using the platform?

[00:10:46]>> Absolutely. I I think there's a macro level and there's kind of a micro level uh at the item level. So, we we could start with the with the macro level. So this could be as easy as an example of a pallet that has meat and vegetables on it and it starts at a source at a distribution center and is packed in.

[00:11:02]Let's say one of the pallets is meat, one of them is vegetables. They're they're refrigerated and frozen at different temperatures. As it goes through the transit on the uh on the back of the trailer, ends up at the dock door at the back of the store. You have a certain window that that pallet gets unloaded. it can dwell and then it has to go directly to refrigeration or some level of a freezer.

[00:11:27]We calculate with our customers what that ideal dwell time is and we send them proactive triggers via events to let them know if this dwells another 5 minutes, you can't sell it because it's out of compliance.

[00:11:40]>> So that's kind of the proactive way at the at the macro level.

[00:11:45]>> But it can go one step further. It could go through the item level journey that let's say it is a package of meat and then it ends up in the front of the store. Some reason it was moved out of the freezer a couple times outside of that dock door >> and it went through some variance changes that you can actually engage with that product and say, "Can I eat you? Is this safe to eat?" Really?

[00:12:07]>> And then you go to that step and it says, "No, it's went out of variance once or twice. You can't consume this.

[00:12:13]There there's there's a high risk of you you contaminated you. It's been contaminated through the process.

[00:12:20]>> This is important because I think when we have discussions all the time on this show about where's the value of AI and is AI going to have a real return and um it often times it is in the nitty-gritty of business, right? It's in these areas of business where you're going to find, you know, that true uplift, right? It's like um a lot of people when they think about is this AI moment sustainable, they're thinking about like people paying for $20 subscriptions to chat.

[00:12:52]And we've come to you back to this topic again and again here. Um industrial AI, physical AI, AI and retail because when there are companies that can better serve their customers and see so just for example, right, the the meat has been out of the freezer or out of the refrigerator for long enough, you're going to get that customer sick.

[00:13:10]Absolutely.

[00:13:11]>> If they're able to retain somebody who, you know, doesn't get food poisoning because, you know, an agent pushed them this notification uh that if that if that meat is out of the freezer for that much longer, you're not going to be able to sell it. And either they don't sell or they get it back in the in the freezer or the fridge, that's really where you start to see business changing outcomes. It seems >> it's it's transformational. So, it it could be at the item level that protects one person. It could be >> it could be pallet. So it could be something went wrong and somehow something was missed and then you trace it back to your point to the pallet and then you discard everything that was part of that pallet. So you're preventative on that front and then also it it creates a frictionless handshake between food processors and the folks that are receiving the items. So there's no fingerpointing. You can actually go to the source of the truth where something went wrong and then fix the compliance between those two parties. So there's value on the consumer side and then also on the on the business side.

[00:14:10]>> Are you at a point where the AI is making some of the decisions? So for instance, um what we've talked about is a notification, hey, you probably want to do this. Um could it be that like an agent for instance, I'm just going to think of an example, you know, sees that a shipment is uh is coming in late and it's not let's say it's a holiday thing, right? So your Halloween shipment is coming in. It'll be there in time for Christmas. Mhm.

[00:14:36]>> Um and proactively can say you're going to be late enough that we don't need this anymore. And instead of having the the item go through the whole supply chain, actually kind of pause it and send it back to prevent those steps from being taken.

[00:14:50]>> Absolutely. You can flag it. I mean, we've had these examples with uh everyday items like strawberries that go through some level of, you know, refrigeration. and they come out, there's condensation, it goes through another temperature variance that it might be too cold and then there's freezing and what you're getting through these cycles actually levels of bacteria because there's water accumulating and then going through temperature changes onto the um onto the strawberries. So we've we've seen these examples and to your point we can flag them so it you just stop the process through the supply chain and then you fix it with some uh replacement before it's too late all the way at the store level and then you deal with stockouts and and uh going through the process again.

[00:15:30]>> Now these are missionritical type of decisions that are being made right. So, um, we're talking in our short time talking together, we've talked about health, health related things, you know, potential spoiled meat if it's, you know, the sensor picks it up outside of the refrigerator. You kind of can't guess at this. You can't really make mistakes here.

[00:15:51]>> Mhm.

[00:15:51]>> But then again, uh, something that we encounter often is that when you're dealing with AI systems, they are probabilistic. They are, you know, they're they're not, you know, if A then B systems. They are systems that tend to freelance a little bit for lack of a better term. So how can you then trust an AI system to act accurately in these environments?

[00:16:13]>> It's a great great questions. You you have to start with the source of the data. You know, I've I've been a victim of chat GPT where I've asked it questions that I knew maybe 60% of the answer and then I get an answer that's completely off base. And then you look at some of the sources that the data is being pulled from, you say, okay, this is not credible or this was a opinion article and it pulled the information from there. So tying it back to the real world of um asset tracking and what that means from a data source is that you have to have a source of truth behind the data and validity and trust that that data coming in is factual data and then you can rely on the AI to do the the leg work behind it. And and for us from a platform perspective, we really pride ourselves that the assets that we're tagging and the data that comes from it, we really synthesize that data in a right format so that when it is sitting in a platform and AI is now layered on top providing the insights that you're working off of really clean, decrypted, trusted data at the end of the day >> and it works like 100% of the time, 90.

[00:17:23]>> Absolutely. for for us right now we we pride ourselves on the readability side >> uh first so that you're getting the visibility because if you can get the visibility and we have a basically a three-step process we have a PC we have a pilot and we have a deployment and what we tend to do is through these stages when we come in with some level of infrastructure our intent the way we qualify deals is that that that hardware stays on site we don't rip and replace it so the hardware that you use in a PC or pilot is only going to exponentially expand for the volume and the the scale that that organization is going to go through.

[00:18:02]>> So while while we go through the evaluation process and the use cases, >> we really pride ourselves that we're making the process really bulletproof and it's something that has validity.

[00:18:13]you are getting the scans because if you are getting the scans in the read range then you are getting the data that's consistent out of it and then you have the assurance that the AI is going to do what it's supposed to do with with factual data points >> could you see a world where AI agents just run the supply chain overall >> with with the right infrastructure behind it and process uh I don't think we're far from it >> really and I can give you know maybe a couple examples around it is that you have you know the fortune twos out there that are retailers, e-commerce brands.

[00:18:47]They have control over the 3PL cycles and their supply chain. They control a good chunk of what we get on an everyday basis. You know, when you look at the packages you go online and you order, they're they're running through these organizations. And a lot of times when they are getting suppliers in, they're the same suppliers that work across the board. So being able to be a part of that supply chain with the larger organizations, a byproduct is that any organization that works with them is going to be a part of the technology eventually that is enabling them to make these insights and it adds value upstream as well. So we aren't far from that day. It's just a matter of adoption and how quickly we get there. Okay, let me give the counterpoint u which is that we've seen attempts to let AI manage or run even elements of the supply chain and it hasn't worked well. Uh there was a very high-profile instance of Starbucks recently that attempted uh to use some AI system where I think I'm going to get this wrong but directionally accurate right where they folks within the store would take pictures of what was on the shelves and that would feed into an inventory system and they just couldn't get it to work.

[00:19:59]Um, so when you think about that side of things, how do you go from that point to an area where AI is more trusted and actually is able to do the actions that you're talking about?

[00:20:13]>> From from Willot's perspective, we we want to make the process frictionless.

[00:20:18]As you add nodes of manual work through the process, what you're creating is more friction and areas where things can't be consistent and adoption is more difficult. We have um handful of use cases that we see that are common between customers. And just in a quick summary, we have you know items that are shipped out, items that are received.

[00:20:39]Automating the process as it leaves your facility and enters it. Can you do that in a in a in a way that's frictionless?

[00:20:46]So we we pride ourselves on being able to do that with the technology. If an item is within your facility, like the warehouse manager example you had, being able to recognize where it is in your facility and if if it's sitting on a pallet or some kind of reusable plastic crate or rolling cage, how many assets are sitting in there? Can you do the cycle counts? Mhm.

[00:21:07]>> We also have the temperature side of it that is more popular from condition monitoring and even sometimes tracking these assets that are reusable that you own as an organization that move products around being able to predict where those are. For us, those five create some level of frictionless adoption. And then when we get engaged, we try to rinse and repeat the foundation of these use cases in and then we land and expand on the innovation side. when you start to heavily customize something early up front, you face that uh friction where you have to then tailor it completely brand new and then adoption isn't easy. So some of these use cases that I went through around the freshness coupled with these use cases we have seen that it's uh the adoption is much easier to digest and also it's something that builds on technologies that have existed over the years. So you have barcodes, QR codes, anything with machine vision, line of sight, it's been there over 25 years.

[00:22:09]RFID, it's been there over 15 plus years. So we're building on those processes and workflows, but we're doing it in an additive way, not with the capabilities. Yeah, I think that I guess one of the things that could happen with AI buildouts is um there could be I'm not saying this is necessarily the case for Starbucks. Um but there could be a directive to just go build something >> and uh and do it now and teams just kind of run to it and then it fails.

[00:22:37]>> Uh or like there's not enough time maybe ahead of time to like think of the proper technology because there's aren't there so many options? There's computer vision, there's data that you're talking about. Um there's using uh you know rag uh with databases you know that you're you're keying information into and trying to build off of that. I mean it's amazing because we're not even that far away from the release of chatt and there are already so many different ways to build this.

[00:23:04]>> Absolutely. Absolutely.

[00:23:07]>> Um multimodal AI. Um so AI is mostly trained on publicly available data. Can you tell us a little bit about what happens when uh LLMs are trained on the physical world itself?

[00:23:21]>> They they give a good foundation. So I think they that level of training and it's only getting better as more data gets ingested in and these models get stronger. So I think it's foundationally very very important but it is a additive benefit to what we're doing on a day-to-day basis. You still need the foundation of the data being ingested in. It's kind of a theme of this discussion is that if you don't have a reliable data source >> and the assets slashproducts that you're tagging, they're not providing the right data triggers behind it and the right insights into those assets and products when they're through the supply chain.

[00:24:02]None of this matters from an AI perspective or any of the language models that are out there. You first need to start with the source and the consistency that these assets and products they actually have a foundation that they're consistently tagged. To your point earlier, they have read ranges that we can have certainty that they're actually being picked up and that we're not missing things. Once that's in, sky's the limit with all these different models and I think they also are learning from each other which is a additive benefit as well.

[00:24:34]>> What do you mean by that? So, so as it gets more intelligent from a supply chain perspective, there's predictive trends that are picked up, one model can actually learn from another model and share that data because they're they're they're scrambling for data when you do a search.

[00:24:50]>> So, there's there's heavy benefits in the in the areas of the supply chain that actually get publicized. Um, when you look at handshakes between organizations, a lot of that data is proprietary, but in many cases, there might be consistency from a food processor to three or four different types of distributors.

[00:25:09]If that information to some degree is shared of who they're working with, you know, you have many prominent brands out there, Yum Brands of the world and all that, they have multiple sources that they work with. any one of these that have consistency through the supply chain, they open the door for more data sharing and then these language models can work together which also is going to be added a benefit for uh predictability down the line.

[00:25:32]>> Right? So look, I'm sure you've heard of the token maxing moment. Uh we've had various debates about it uh on this show. The idea that like spending tokens for the sake of spending tokens has been a thing. Um, and I'm curious to hear your perspective on um, how you and the companies you work with measure the ROI of these AI implementations um, and whether there's been a a an emphasis on cost efficiency recently.

[00:26:04]>> It's a great another great question. So when we look at right now any type of ROI, we first start with the problem. We always have to start there before we start to quantify anything. is what is the problem that we're solving? Is it have to do with shrink? Does it have to do with manual labor that's in the process? Is it maybe even a combination many cases of both? And then you do the the backwards engineering. Okay, quantify that problem for me. And we have our solution sets that I went through. When we approach the problem with one, two, maybe even all five of them in one one case. We then work backwards into what that ROI could look like based on the cost and and and what the problem is. We have a major retailer out there. They factored in that every time they had a pallet and it was sitting idle more than 37 minutes there was loss. And then we multiplied that out on an annual basis and that 37 minutes translated to north of $270 million annually that they were losing.

[00:27:05]>> So we we knew the problem. We were able to quantify what it meant from a loss perspective. By the way, manual numbers behind the manual work of scanning those pallets, all that not factored in. So, it's incremental to that 270. And then we knew the solution set that we had to put forward to solve the problem. And in many cases, that's the process you follow. And then the variability that you get on the ROI is the complexity of the implementation across all your sites. um and and what that means from a time to actually get the the ROI back.

[00:27:40]On average, we like to say anywhere between 7 to 13 months is what you're going to get on an ROI perspective with our technology coupled with um the problems around these use cases. And in that case, it was a 13-month uh ROI.

[00:27:54]>> And how about the way that you're wait sorry what does 13-month ROI mean? It means that you will make your money back within 13 months.

[00:28:00]>> Yes. Behind the solution. Yes. Are you spending more as these AI uh models become uh bigger and more token hungry?

[00:28:08]Like for instance, if you were just in an information synthesis versus of the more agentic type of use cases like that you seem to be putting putting forward in into production in your product, you might be spending less in that more information synthesis versus the agentic side.

[00:28:24]>> Yeah, I I would say right now it's it's less around spend for us on the on the AI side. It's more around solving the problems with the technology set.

[00:28:33]>> We're letting a lot of these agents themselves get integrated in so that they have data to work from.

[00:28:39]>> They're getting more and more intelligent dayto-day with the way that they're processing information. I mean, you look at even like a a cloud and and what it's able to do on a on a weekly basis with some of the rollouts they have across all these product uh suites.

[00:28:54]We focus on the data side and then we let this just be layered in. So if they're intertwined with any major ERP WMS system out there, if we relay the data over there, we'll let some of these modern agentic AIs just go in and and do their own um data mining from there.

[00:29:12]>> Finally, I'd love to ask you um what is a data foundry for the physical world and what happens when you have a program or uh one of these foundaries that can collect, synthesize, and then act on real world data?

[00:29:26]>> It's major. It's it's it's uh it's transformational when you look at all the data sitting together in a in a basically a pool and what you can do out of it. Um there's a couple subsets. So when we work with any type of organization and where we're looking at where they are in their supply chain and what we're tagging and the data coming in, that's one part of the pool that we're getting into and we're mining and then we're using AI to get some insights around. There's also a larger pool when you look at the total supply chain and what you're doing with it. Most of the organizations we work with, they're not publicizing all that information. So, it's up to us to be able to use the power of the Willia platform to decrypt that data and then give them the actionable insights. But as we get macro level and this becomes standardized, you will start to pick up the trends very similar to the to the example earlier that food processors that work with distributors are going to have commonalities. And then some of these more public domains around AI agents can actually go in there and give you some level of analysis on what that means.

[00:30:34]And so if you're a smaller mom and pop shop and you want to start up something and get involved in the supply chain, you'll have a baseline to work from and you'll have your own KPIs to be a part of the process and evaluation.

[00:30:46]>> So the foundry just like uh typical real world foundry, right? To go with this metaphor, the raw materials come in, the refined materials come out.

[00:30:54]>> Mhm. Inputs and outputs.

[00:30:56]>> Okay, great. Amir, thank you so much for coming on the show. Great to see you.

[00:30:59]>> Absolutely. Thank you for having me.

[00:31:01]>> All right, everybody. Thanks so much for watching and we'll see you next time here on Big Technology.