Name: [Expert Panel] AI in Parcel Shipping: Approved, Denied, or Proceed with Caution?
Uploaded: 2026-09-01T18:30:45.043Z
Duration: 1 h 16 s
Description: [Expert Panel] AI in Parcel Shipping: Approved, Denied, or Proceed with Caution?

Transcript for "[Expert Panel] AI in Parcel Shipping: Approved, Denied, or Proceed with Caution?": Alright. Welcome, everyone. Thanks for joining us. Thanks for joining us for today's pros who knows session, AI and parcel shipping, approved, denied, or proceed with caution. I'm Justin Kramer, cofounder of ProShip, and I'll be serving as both the moderator and, I suppose, our our our judge for today's proceedings. If you spend any time in the shipping industry lately, notice that, AI is everywhere. Every conference, every webinar, every demo seems to involve AI in some way. The challenge isn't deciding whether AI matters. The challenge is figuring out where it creates real operational value, where traditional automation still wins, and where human experts remains absolutely essential. Before we get, before we get started, though, I'd like to do a couple of housekeeping things. Please make sure we've got a q and a section. Submit those questions. The panel line were just talking about this right before we came forward here. We'd love to answer some questions. We'd love to see where, our audience is thinking about AI. So make sure you enter those. Also, this webinar is being recorded, so you'll receive your on demand recording within the next week. You can also download additional resources from the related content section. So you'll also find things there like links to surveys that we're running. And most importantly, we are coming out with the with the annual pro ship report. Please, we'd love to get more people responding to that. Eligibles respond eligible people who are eligible will receive an Amazon gift card if you fill it out. Finally, please complete the survey at the end of the session. Provide feedback so we know how to create better pros who know sessions in the future for you. As I said, I'm Justin Kramer, cofounder of ProShip. Joining me today are are four supply chains experts. I've had the pleasure to work with with many of these guys over the years. But would each mind introducing yourself to our audience? Adam, let's start with you. Yeah. Sure. Thanks, Justin. My name is Adam Klein. I'm a senior director of product management for Manhattan. We've, I've been with the company for a long time, twenty nine years, all working in the warehouse management space, and have gotten the chance to work with, Justin and ProShip and, some of the other panelists, as well for, you know, quite quite a long time, and I look forward to the conversation today. Excellent. Thank you. And, Tyler, could you introduce yourself? Hey, everyone. Tyler Linderman. I am VP of Strategy and Growth at Bricks. We're a supply chain technology, strategy and implementation partner. Excited to talk about AI with, this great group of panelists. A guy many of you know from his amazing post on LinkedIn, Matthew Chapa, can you introduce yours? Yeah. Justin, thank you so much. My name is Matthew Chop. I'm a senior account manager at, Flexport. I'm happy to be on the three p l side after spending a good good bit of time last decade over on the shipper side. So I'm very familiar with, all the parcel. And, yeah, I'm excited. This is my second go around on the pros who know tour. Excellent. And and last but not least, the guy I ask all my analytics questions to, Quintin Nelson, can you introduce yourself? Yes. My name is Quinn Nelson. I'm a senior program manager with Reveal. I've been with Reveal for about a year and a half, but I've spent about the past twelve years in the space. My background is primarily in everything related to transportation, analytics, business intelligence, customer reporting. And my, my role and my role has been very uniquely impacted by AI. So looking forward to the conversation today. Excellent. Thank you. And thank you to all my panelists. Thanks for introducing yourself so the audience gets to know you. But before we dive in, I'd like to set the stage for today's session. The goal of today's session isn't to debate whether AI matters. That's that's already been decided. The real question is whether it should whether it should be making decisions, where it should assist, and where it should be completely out of the loop entirely. So as as we go forward with this, let's talk about some rules, what we're gonna be playing by here as we decide, whether or not they, you know, to approve, to kind, or or proceed with caution. We'll present a a real world shipping scenario. You'll vote first, you as our as our, audience, and then our panelists will argue each case challenging each other's assumptions and ultimately will deliver a verdict. Your options are simple, approve, proceed with caution, or denied. Without any further ado, let's start with the first case, delivery promise engine. So as we look at this, you've got all of this ability to determine estimated delivery dates during normal operations. Customers are are see faster, more accurate delivery estimates. And if set up right during peak season, as performance decline, we will prevent that that miss that communication, which could be an error. So, we should be presenting this on the screen. So if everybody could, could vote, let us know your thoughts. Do you feel that, yes, use AI to determine delivery dates, proceed with caution, or no? Let's stick with the tried and true dates that the carriers give us. K. Lots of different things that we can look at there. Alright. I'll give it another another few seconds here. Does look like, so far, nobody is denying this. No contrarians out there. Alright. And with that, we do have our our final votes in here, and it is overwhelmingly 80% say proceed with caution with a handful of people saying approved and nobody, denying. So let's let's, let's bring this to our panel. I wanna hear your thoughts on on a promise engine. How can AI help and thereby affect what we show customers as well as for, close rates for retailers, directories and manufacturers? Matt, let me start with you. Yeah. So from my time on the on the shipper side, it it was incredible to work with the warehouse, and that's where I'm gonna kind of lean on is that expertise there. AI is only as good as the information you put into it. So delivery promise raise is only gonna be about halfway accurate if you're just looking at what happens after you ship, from the warehouse. So there's a lot of information that it may not be tied into, and you may be giving customers the wrong information, right? So, if you are in like a peak holiday season, where your warehouse might be delayed, that gets a little bit murky. So I would be very cautious with allowing AI to promise something to my customers that I haven't vetted or don't doesn't may not have the right context. Yeah. Agreed. And and you said the right word there at the very end, context. Right? Sometimes our context scope is just a little too narrow. The OMS may understand labor capacity fulfillment lead times, and that's gotta be combined with the shipping data to provide if you are gonna do this, it's gotta be combined so that you can start the transit time after all that fulfillment time. If AI is gonna make any predictions, it needs to be visible or or the visibility needs to be broader than the fulfillment operation itself. So, Justin, just a thought here, you know, on on this on this point. So, obviously, my warehouse is in the you know, my my background is in the warehouse. Right? And so I think about things kind of after the promise is already made, because the warehouse sort of has to fulfill the promise that that maybe another system upstream has has ultimately made. But but I would say that a lot of, opportunities for AI ultimately occur after that promise. Right? So I I I think Matt's right, you know, in in everything he said. But, you know, as I think about, you know, exception handling, communication with the customers when a disruption is inevitably gonna happen, you know, tracking inquiries, issue resolution, etcetera, etcetera. Right? A lot of that stuff kinda flows through a customer service organization, you know, oftentimes a call center, and so forth. And so, you know, one of the things that we've really, looked at is is how do we reduce the impact and the cost of all those, exceptions and, you know, and and so forth. I think AI is a a really effective way of, of doing that. So, you know, the the the the promise itself, you know, we we can debate whether that's best handled by AI or or or more deterministic approaches. But I would certainly argue that, you know, kinda after the fact as we're addressing, you know, change and disruption, etcetera, that AI can step in and, potentially be kind of that first interface, for a customer to find out what's, going on. Adam, you raise a good point because I I think it is it's not an all or nothing. Right? It's not all AI or all deterministic. The best solution is is probably a blend of the two using deterministic where things are fixed or where things are well known or or need to be calculated and using AI where, well, it's probability. And and and so applying each of them to where they will provide the greatest value. Yeah. Adam, I like the word disruption that you used there. I was actually having a conversation earlier today around disruption management across supply chain systems, and I think that's an area where more and more we're seeing AI take an approach. I I think the other thing I would emphasize here is, when it comes to, delivery speed and expected delivery dates and promises, I think it's more about accuracy than speed as well. And, you know, we went through a time in supply chain to where shipping was about speed, how do I get it quicker to the customer, and those sorts of things. But now I see AI as a way to be more reliable. Customers care more about receiving a package in, you know, on the day that they were promised it or earlier than they were promised it rather than getting it tomorrow morning for the most part. And so I think finding ways to leverage AI to focus on making commitments more reliable, is really the way that I would look at this approach as it comes to expected delivery dates and and these delivery timelines. I think the other interesting thing here that, Justin, you started to pull the string on a little bit is and, Adam, you added to it. These decisions are not in a single system. Right? They are across multiple systems. They can touch OMS, WMS transportation, a parcel shipping management system, the website, the integration systems that come in between. And so the ability to have AI that can look across those systems and make decisions also played a role here. Yeah. Good good points, you know, Tyler. I I think from the accuracy side, I agree with you a 100%. I mean, you know, it's it's really critical because as soon as you broke break a promise, you start to erode customer trust. And then, you know, what what are they gonna do the next time they're ready to order something? Are they gonna order from you or for someone another vendor that they that they do trust. Right? So that's, I think, a a big deal. You know, I I think couple other things here, you know, if you think about these predictive models, you know, to use your words, Justin, you know, order management, warehouse management, transit data, etcetera. I mean, all the things you mentioned, Tyler, you know, coming coming together to improve, you know, deliver delivery reliability. Sharing that data, is is part of the the key. You know, and I think with, you know, with with enterprise solutions going into the cloud, that that certainly helps, you know, with with sharing the data, you know, as as well. And then, you know, look, at the end of the day, you know, we've talked a bit about AI thus far, and I'll just say, you know, different AI technologies ultimately serve different purposes. You know, you can look at something like machine learning and say, hey. This is really well well suited for, you know, forecasting delivery outcomes and and, potential disruptions before the order is placed, while something like a large language model or an agent or some a an AI assistant, typically better for those post purchase monitoring. Some of the things I mentioned before, alerts, communication, exception management, all those things can maybe be better handled by, you know, something like an agent. Yeah. You know, that that's a good point, Adam. Many many of us remember a couple years ago when machine learning was the buzzword. Right? And now we've got AI. Can can you can you help define the difference for those of us who are trying to understand what is the difference between machine learning and AI? Yeah. I'll I'll try, Justin. You know, AI is sort of an umbrella term. Right? Artificial intelligence. It's the the broad goal of building systems that act intelligently, right, in some way. Right? And, I I think that machine learning, and I mentioned large language models, those both sit underneath the umbrella of AI. And and there actually is a lot of similarity between the two. So machine learning takes past results. It looks at what's happening right now, you know, current conditions, and ultimately provides a prediction of what's gonna happen next. Usually, that prediction is gonna feed something like an optimization engine, some sort of deter deterministic engine, you know, etcetera. Right? It's it's usually not just used in isolation. Okay. So that's machine learning. Large language models are actually almost like a subset of machine learning if you think about it. You know, and what I mean by that is they're looking very specifically at at text, and they're really looking to to look they're they're called Mars language models for a reason. It's looking at massive amount of text that's out there. And what it's trying to do is predict the next bit of language or the next bit of text to ultimately go and answer a question or generate content, hold some kind of conversation, etcetera. Right? So it I think their distinction isn't really ML versus AI. It's really how are these things kind of similar underneath the broad, you know, spectrum of AI. So yeah. Awesome. You know what? Hopefully, our our audience came here to learn because, that's a very educational statement. Quinn, let me throw it over to you. What are your thoughts on on this, this case? Yeah. Absolutely. I'll take kind of an analytical view of this one. So what I'm hearing kind of over and over is that confidence as a whole is a big concern here, whether it's with AI in general, whether it's with this specific use case. But I think if if confidence is a concern, all we need to do is account for that in the models. So, ultimately, you kinda need to treat AI the same way you would a human. So, hypothetically speaking, say you're pretty confident with these AI models to to understand time in transit for most of the year, but maybe in peak season when things really go haywire, you don't really know if AI is up to the task to put trustworthy information out there. Ultimately, we just need to build in those confidence thresholds. So if you just give AI last year's peak season time in transit actuals, it's probably just gonna approach it and go, okay. What was the average time in transit on this day? But I think there's so much opportunity to add context and add confidence controls to that. Maybe you tell it, okay. You need to look at more statistical stuff, like what is the standard deviation of some of the averages you're seeing or what sample size are you looking at and get to a statistical level of confidence that AI has in the outputs that it's putting out in different situations. Because, ultimately, AI is undefeated when it comes to basic math. And if you can frame things in mathematical terms, it really sets AI up to succeed. An analogy I like to give a lot of times is that I think of AI like the best, most capable employee in the world who is an expert across all disciplines, but, critically, it's their first day working at your company. So while they have all the knowledge, they have everything in their toolkit, they still need to get the context to your business. How does your business operate? How does your industry operate? And what, parameters does it need to work within in order to provide an output for a given scenario? Excellent. I swear I've said that to to to many people as I've trained them on AI. Very good. Alright. I think we we we're the consensus is pretty clear here amongst our panelists. It's a proceed with caution or maybe a proceed with, intelligence about where the data is coming from and make sure you get all the data. Everybody agrees. Right? Proceed with caution? Definitely. Alright. Let's let's go ahead and move on to case two. One of the trade offs we're starting to uncover here is that predictive models may not always be as accurate as the carrier transit times used today, but our machine learning can identify potential problems before it happens. Does that change the value of the equation? That brings us to our next case. Should shippers trust AI to predict disruptions and take action before a failure actually occurs? So let's go ahead and bring this up for our our, bring up this poll for our audience here. Let's see what you all think. I think this is, a very interesting one because we were just talking in case one about the actual, you know, what to do before while we're in the cart or in the catalog. Now what we're talking about is this package has left our warehouse. It's on its way to the, to the to our recipient. Do we try to use AI to predict if we're gonna have a problem? Alright. Looks like the votes are coming in. Give it another second here. Alright. So with this one, proceed with caution was definitely the the lead vote with denied becoming in second. A lot of people saying, hey. No. And if the smallest one, is approved. So with this in mind k. Interesting. Okay. So the the, the audience says proceed with caution, denied, and then approved. This one's interesting because we start to move beyond prediction and into action. Tyler made a great point earlier. Once we've actually made a promise to our customer, our responsibility is to keep it. Right? So, let's talk about whether AI can help us by identifying problems before they impact customers. So, Tyler, since you're kind of the you know, I gave you a lead in here, let me throw it over to you to see your thoughts. Yeah. Yeah. I think you worded that question in a in a unique way, Justin. I I think this is the classic human in the loop, human out of the loop conversation is what it boils down to. And so what what I've seen with our customers and as we, you know, really dive into this space is a really valuable use of AI in the supply chain systems, shipping systems, etcetera, is about proactive notification and providing optionality. So I would say, I can bet that almost every company can find a use case to where it would make sense to act proactively and and automatically. That being said, I don't think any of us on this panel would say all of your decisions should be automatically made and and acted upon by AI. And so, really, what I would look at is rather than automatically making a bunch of decisions, we want our AI tools to be able to surface potential disruptions, kinda going back to that disruption management topic, and provide recommended actions. Right? And so it's not it's not as simple as, you know, an if then statement. If this happens, then do that. Right? There's there needs to be optionality to it. And and as we talk with more users, and try to understand how they wanna use AI over the past year, a lot of times what we've learned is, you know, it's not all science. It's not purely science. There's a lot of art to some of these jobs as we talk with merchandisers and how they're making decisions on how much product and quantity to order and am I ordering it for econ fulfillment or store fulfillment. Yes, there's some forecasting allocation and science behind that, but a lot of it's art too. Right? It's I've been ordering shoes for men for the past four years. Like, I'd I kinda know how this fall is gonna go even if the numbers don't don't say it directly. The the same thing Adam mentioned some warehousing stuff earlier. My background is also mostly in warehousing, but I think of the the companies that we work with that have really great wave planners. Right? They they know the data. They know the orders. They know the flow. And so even if even if it's not quite there, there's a little art that that helps it helps it make sure the day runs smoothly. And so, I think AI can do an exceptional job of monitoring signals and flagging issues and giving optionality and and scenario planning so that highly intelligent users can then act on that before customers are impacted. Yeah. So, Matt, you had a bit of a of a different take on this. I I just wanna say, like, real quick. I'm so glad there's more deniers on the on the poll, because that's kind of where I'm gonna stand. Like, I'm gonna stick my claim on that that for the entirety of this poll. So, yeah, I think there Tyler, you made a great point, that it's about it's about half science and a lot of art. And I, you know, I look at a lot of AI art, and it is uniquely terrible. So I think there are some really practical limitations, right? So, customers are generally unpredictable, especially when they know information. So I think my use of AI to proactively, you know, surface some problems or provide, you know, prediction dis about disruptions, it's gonna be incredibly limited because I know, like, having been in the game for a long time and having seen, you know, customers who are literally in a hurricane zone asking, where is my package? You know, I think that there are there are some really strong limitations to to AI, and that's where I'm gonna trust, like, a good human being to, like, absolutely do the job, that I'm, like, I'm hiring them for. The other thing is, like, being able to to change something that happens, it it really it really depends on where the life cycle of that order something might be. So up until the point that the warehouse, like, applies a label and hands it to to the carrier, you can do a lot of stuff. You can change things, you can change labels, you can change service levels. But once you put that label on it, like, wave goodbye, kiss it, and off it's going, and you can't touch anything else other than communicating to the customer, Hey. There was a disruption. Look outside your window. You might see a lot of snow. You might see a lot of rain. Those things are gonna impact your delivery. So my use is gonna be incredibly limited, regarding AI as the, disruption predictor. Yeah. I completely agree. Even from working in warehouses, once that label is printed, it is a pain to to have to go change it for for any reason, whether it's a a network disruption or just someone clicked the wrong button as they were running the way. Right? And so yeah. The earlier you can flag it and potentially change a path, the more likely you are to have a positive result out of it. And I I I wanna pull a point in here a little different than what we planned. Matt, you you allude to something, which is we've paid the carrier to for that to be their problem. Yes. But let me let me let me throw this back here for just a second. Tyler, before I interrupted you there, you were you were saying basically something to the effect of we've made that promise. We're supposed to keep that promise. If we're gonna have to break the promise, it sounds like you were thinking that AI can help us at least, communicate when that delivery promise has to be missed. Yeah. Absolutely. And so in in terms taking action doesn't always have to be corrective action either. Right? It's not like we're necessarily gonna reroute the package through a different carrier or we may not even have the ability. But but even something as small as communication, if we can proactively trigger the communication to the customer that, hey. A delay is gonna happen. That's a lot better than just not notifying at all or notifying late. Right? I I know I've had a few examples over the the years to where it's, hey. Your product's gonna be there on Friday, and then product doesn't show up. You it's crickets. Hey. Is it coming? Did you lose it completely? Is it just delayed? Like, do I need to go run down to the local store and get it? What what does it look like? And so I think organizations that can notify customers early, adjust expectations, maybe even offer options, like discounts or incentives to maintain trust and loyalty. That's how you can use AI in a network disruption situation with automatic action and, deliver a better end result for the customer to help maintain that trust. Tyler, you you almost just described an agent that we that we built that goes along with, you know, some of our order management solutions and so forth. We call it Maven. And, you know, as it's connected to the all the order data and all the you know, it understands all the exceptions that that that are happening. You know, there's a lot of sort of customer inquiries that come in and typically get routed to that customer service, agent that, you know, that that that we that I talked about a little while ago. But but some of the things you just talked about, where's my order? Like, pretty simple question to answer. Also, very simple to automate. Right? You don't need a human to answer that question. We just need to look at the data. Where is it? How far is it away? Is it late? Is it, you know, on time? And if it's gonna be late, you know, the person who's asking that question, you you typically through a a chat window on the on the website they ordered from, you know, maybe they're gonna ask, well, when it when is it gonna get here? Right? And, you know, if it's gonna get here late, well, well, what are you gonna do about it? Right? I'm not happy now because it's now a day late than later than I expected it. Maybe I needed it for a reason. Right? Maybe it was a gift. Maybe it was for an event, you know, etcetera. So now all of a sudden we get into, appeasements, right, and and customer management. And, you know, you mentioned, like, discounts and incentives, etcetera. You know, you can set up workflows and let an agent, you know, kinda work through those up to a point. Right? Now at some point, you know, maybe you do have to kick it over to a, you know, to a human, but but the goal here is to is to automate as much of that as possible, but give it a a very human like experience. And that's where large language model come in back into play. Because if you can if you can have a conversational and even friendly tone, as a, you know, as a chat agent, you know, sometimes if done right, you know, there's times where you're speaking to an agent or chatting with an agent online, you don't even know it. And, you know, sometimes you'll ask I'll even ask sometimes. Right? Hey. Are you, are you AI? And it'll it'll usually come clean, you know, if you ask it directly. But, yeah, I know. I think that that is a great example, a great use case, for using, you know, an LLM or what we call an agentic, you know, an an assistant or role based assistant, to to step in and, you know, at this point, you know, not not only raise some customer satisfaction, hopefully, and and and fix some issues, but but on the other side of it, lower cost. Right? Because if we can if we can take some of those conversations that would typically go to a human and take it through a a chatbot, we're we're gonna be saving money as an organization as well. So Yeah. So I think I think it I guess real quick, Justin. I think it also makes. sure that nothing falls through the cracks either. Right? Even if you have the best intentions with, hey, our SOP and policy is that every time there's a disruption, we reach out to the customer proactively. Sometimes that doesn't happen. Right? We're humans. We make mistakes. Things we get busy. We go on vacation. We have a sick day, whatever it might be. And so if there's at least a a backup mechanism there to to set that expectation correctly with the customer, it can go a long ways towards maintaining that trust long term. So we've kind of heard the entire spectrum here, almost like the poll itself. Absolutely not. Here's how you warm the humans, and here's how you take the humans out of the loop. I think I'm gonna have to average this one out and go proceed with caution and apply humans where required. So very interesting, how this one turned out. Let's move on. We're we're, right on time here right now. Let's move on to our third case, weather triggered dynamic rerouting, which is very similar to this one. AI can detect that a weather event is coming. You know, as a Floridian here, I'm I'm, at every morning checking out NOAA and seeing where the tropical depressions are. And if you're in, Louisiana or parts of parts of Texas today, you are experiencing a lot of weather, all that weather out there. So as AI detects these things before they could occur, should AI be allowed to automatically reroute shipments when it predicts a network disruption? Do you approve automatic rerouting? Do you proceed with caution or just say no, this one shouldn't happen? Okay, it looks like our poll is up. K. This one's getting a much different voting solution. Alright. Let's, keep this going here for a second. K. And give it just a ten more seconds here, but it looks like we have the biggest split we've had so far with the majority of people saying denied. The the the next set saying approved and a small amount of people saying proceed with caution. So interesting poll, process here. It feels like, this is different than the last one because now we're talking about AI taking action, taking action on a prediction, not just identifying, what we had in the last one. Adam, let's start with you. How much confidence should we really have in AI making these kinds of weather driven decisions? Well, look. I I I'm weather is, I think, one of the most difficult things to predict. You know, it is it is inherently unpredictable. Right? I mean, and just look at what's going on, you know, in the, you know, on the globe today. You know, we've got an unprecedented weather event happening in The Pacific right now. They're saying it's the strongest El Nino on record. Look. Most people don't even know what an El Nino is, let alone what happens when the strongest one on record hits. Right? But whenever you hear something like, hey. It's the strongest on record, it means we don't know what's about to happen. Right? So I I think that statement alone should, give us pause when we're, you know, predicting any sort of weather, for the balance of the of of the year as long as that's going on. Kind of point two, even in the best of conditions, your local forecaster, if they're 30% accurate, 35% accurate, they're do probably doing a pretty good job. Right? So I I I I'd say that just in general, weather is is is pretty unpredictable. AI certainly can help. It it can help improve the predictions, but I would just say the jury's out on whether, you know, how how much it can. Right? Because, again, it's it's looking at past events. And as long as there's some level of pattern and and so forth, it it's gonna be it's gonna be reasonably, you know, helpful. But, you know, as I mentioned, we're in a we're in a state right now, this very moment, and and it seems like every year, there's something else. It's like largest deviation of ocean temperatures ever or whatever it is. Right? I mean, I'm not here, you you know, here to preach on climate change or anything like that, but there's there's unprecedented events happening. All those things are gonna add a level of uncertainty. And so I'd say any weather based prediction definitely should be treated with caution, and a healthy dose of realism. Yeah. And I I've given this one a a a fair amount of thought, so I'm gonna take this a little bit further. This is one of the most difficult problems because it's one of the things that our profession logistics folks don't actually have captured data on. Right? And when we look at the people that do, I mean, in many cases, we are talking government level computers that are actually attempting to do this prediction as you were alluding to, Adam, not having data, not having the processing power, this puts us in a really difficult position, to actually provide the right data with the right context, whether it be to an LM, to an to an ML, something of that nature. So the challenge today is between, weather for the weather forecast and the operational decisions, there's a huge gap. So because this is a lot about data, Quinn, let me throw it over to you. What is your thoughts about this? Absolutely. Yeah. I think with this from a personal level, something I've discovered time and time again when working with AI is just we need to get real about what AI is actually good at. I think the marketing and the advertising that we're fed around AI is that it's just going to do absolutely everything in every situation. Kinda grab your cup of coffee and sit back and watch and watch what it does. And I think that's way too much of an all or nothing view of AI. I I think we just really need to focus on what it can do well. It's let's be honest. It's probably not sophisticated enough by itself right off the bat to forecast whether to an accuracy that's going to be usable for us. But let's look at what it can potentially help with and what it very realistically could provide a lot of value for. So as we get into situations where things are very much in flux or changing, weather's a good example. The way that, just the the shipping market as a whole right now is another good example is just being able to make decisions quickly and having the right information to make decisions quickly. So say, for example, there's a weather pattern. Maybe we're in the winter and there's a a snowstorm setting up over the South. To go out and start rerouting your packages, to me, seems like a bit of a a tall ask. But what AI could very much do is help you scenario plan. So if that situation is setting up, it could go out and use the context and information available to it to help you understand what the cost of diverting shipments out of the Atlanta area might be for that week or what would be the cost to upgrade shipments that are going to the Southeast that week, or what would be the cost implications if I wanted to set up my own line haul and direct inject my FedEx or UPS shipments into a completely different hub than I normally do to try and get around the weather? Those are decision points, and those are data fueled decisions that AI can not only make but can make quickly and really present to you in a usable way when time is of the essence while you're trying to make decisions. I'm gonna throw this one over to Adam because, Adam, I know you're in a space that is famous for not dealing with certain types of weather well and other types extremely well. Well, we I I'm in Atlanta, and, clearly, we deal with ice and snow incredibly well. Absolute. We have all of two snow plows in the city, and we have a shed of salt up, north of the city. I think one, maybe maybe two. But, yeah, I I think that, you know, you mentioned a snowstorm setting up over the south. And, you know, the only thing I think that you can predict with any level of reliability is that when you visit your, you know, grocery store, local grocery store, there's gonna be no bread or milk, as soon as there's any level of prediction. But generally speaking, are are we gonna get anything? Look. We might get rain. We might get some ice. We might get a foot of snow. Like, who knows? It really is is is a very, very difficult thing to to predict. There's so many things that have to come, together to make it happen. So, you know, if you look at a a a forecast for a foot of snow in Buffalo, New York versus a forecast for a half an inch of snow in Atlanta, I'm gonna be way more confident in the foot of snow in Buffalo every time. Right? So, I think the confidence level, you know, has to be part of the decision making process. So, you know, I I would think that, you know, the AI models are gonna get better and better at at at understanding what a confidence level is and how to how to, you know, ensure that that's a leveraged as part of the, you know, part of the prediction. You know, some things deserve more trust than others. And at the end of the day, organizations need to understand we're not only what the prediction is, but how confident is the model. Right? You think about the spaghetti, you know, in Florida. Right? You think about the spaghetti models of hurricanes. You know, the further out you are, they're all over the place. Right? Which it could be going to Louisiana or it could be going to the South Of Florida. Like, who knows? Right? But there's models that are trying to predict it. As you get closer, you know, sometimes they converge a little bit more. But even as you get closer, it's it's still, you know, it it it based on data, there's a lot of variability. Yeah. Yeah. I so I the one thing I wanna point out there is that we're talking about making major financial decisions around this, and we all know everybody on this panel knows that that the the shipping costs associated with, with logistics now has become so impactful, that even for us now, we're now seeing, medical firms, pharmaceuticals, rate shopping. When in the past, it was next day, highest quality, early AM, all that every time. And now we are seeing that level of rate shopping go there. But as as we look forward to this, I I think what I'm hearing as well is, hey. Let's keep people in the loop here. It's a it's a this is a very big judgment call. And like Adam just said, I'm probably going to agree with a foot of snow in in, Buffalo, and I'm going to be very, very cautious to pull the trigger on even a quarter inch of snow in Atlanta because I have history. Anybody wanna add anything on this one? Yeah. I think one thing just with this both the previous case and this one, I think in all different respects, holistically, don't forget that you are ultimately already a subject matter expert. Maybe AI will help in certain circumstances. Maybe it won't. Maybe you can do a better job than AI. But I find so often that when AI comes up, people wanna go all the way back to square one and say, oh, if we use AI, let's give it access to these raw data sources way over there. Where in reality, you have handled these situations for years. You have the industry context. You have the industry knowledge. Okay. Use that to get your AI models to a certain point of usability. Don't just throw everything you know out the window and let it come up with its own parameters. Use what you've learned to really accelerate your starting point when you set up these AI models. Because I think a lot of the benefit to using AI is not using AI as your subject matter expert, but using your subject matter expertise. to turbocharge where you can take AI in the long run. Yep. Alright. I'm gonna we go a little off script here. I wanna jump, to to Matthew. Matt, do you, you you've had some very good conversation in the beginning here. Let's jump to that. You just wanna hear me say no. You that's why I know that's why you brought me on this panel. Love the contrarian. No. I think, Justin, you had some really good foreshadowing earlier that I I hope our our very esteemed, you know, audience caught, where you're going with it. But, yeah, like, I'm paying I'm paying the carriers to deal with this for me. Right? Like, I have the sat across negotiation tables for people that, like and I'm doing it because I know and trust them or because I work with them and because I believe in their system. So I'm I'm never going to allow an AI model to, like, even tell me, like, what I should and shouldn't be doing. In a lot of cases, I'm seeing, you know, our our comments coming in, like, oh, well, ask the customer if they want it. I don't wanna ask the customer because the customer, again, does not know the context of, like, what my warehouse is able to do. So I find it best to just put a label on it, kiss that package goodbye, off it goes. So in the case of, like, a UPS or FedEx, those individuals, they know their network better than I do. They know what shifts they're gonna be able to to accomplish. They know what, like, which trucks are running where, and so I'm gonna just let them handle it. And, like, again, it's because I have trust and faith in them. That's the reason why we're doing business to begin with. And I think every time that I have, have tried to do something that might cost a little bit more, my companies have been very cost sensitive. So to your point earlier, typically, transportation and TSE company is the third highest cost, for the entire company as a whole behind HR, marketing, and then you come with transportation. So it's always a big line item. It's always a big, you know, discussion point, to explain your your case. And so I I wanna make, like, I wanna make good decisions upfront with people that I trust, to get packages off even in some kind of weather disruption. Excellent. Great great points there. Tyler, let me give you the last word on this case. Yeah. I'll, Matt, I'll agree with you on this one. I'm staunch staunchly denied on on this one. Weather's just so tough. Like, it like, it's it's rough. I I think as it relates to this this question is I I see too many organizations just start with a top down mandate of use AI. Like, just just use it without first really defining what the problem is they're trying to solve and, and how we actually wanna use it from a from a company context. And, like, I I think, Matt, to your point, we we still have human intelligence and it matters. How do we use AI to supercharge them and turbocharge them in the right way versus just like, okay. You think the weather's gonna be bad and it's snowing, so tell everyone else just to go do something different. So I I think while clean data and mature processes are important, like, AI also needs a clear strategy and direction from the business on where we're gonna use it, why it's gonna bring more efficiency, and ultimately, how does that align to our business goals and objectives? So I I see it as the goal of AI should be used to, like, intentionally solve meaningful business problems, not just apply it because there's some use case that maybe it works to. So I think finding what that strategy and structure is around to apply it versus just saying, well, hey. Wouldn't it be cool if AI could tell us that it's gonna snow and we can direct a package? Like, maybe. But does that does does that fit with your broader business strategy, and does that make your team more efficient, and does it make them make them more viable to the organization? And so that that's my take on on this end result here. Yeah. Good good points, Sean, all the way around. And and I I wanna one of our commenters, Jonathan mentioned that it's very costly to reroute packages. And I think we're all we're all focusing on definitely not approved, maybe not even proceed with caution, but leaning a lot more towards denied. And maybe if anything, we look at the weather forecast, things of that nature. There's a lot better places for us to spend our AI dollars rather than attempting to predict the weather and and reroute packages because of it. And I wanna go back to something that that Matt said because that just really resonates with me, and that is we pay the carriers a lot of money for that to be their problem. So we'll let them, earn that money. Alright. Moving on. Let's talk about what AI actually needs to succeed. Quinn, let's start with with data because it's a theme that's come up over and over and over. Right? Context, capability, tools. Context is really another word for data. So let's go to you. Yeah. Absolutely. I think for so many years, AI was the buzzword that you just hear everywhere. Everything was AI, AI, AI. Lately, I feel like the new buzzword is more better data, Better data, better data. But what does that actually mean? And I feel like it's it's not always just better raw data. It's not, oh, we need to give it of access to more and more raw data. And this kinda throws back to what I was talking about earlier about don't forget that you're a subject matter expert. So that prerefined data really helps drive better decision making by your AI models. And that context, whether it's industry terms, whether it's known KPIs that you track. I mean, it it makes a big difference if you're giving it data and you can immediately say, hey, AI. Here's this dataset. You should immediately be looking at these five KPIs before you look at anything else. What that allows AI to do is start from the correct place using the correct parameters instead of trying to just invent those parameters itself. So going back to what I said, use your subject matter expertise to, excuse me, to turbocharge what you're doing with AI. I think a good example here is this is just a hypothetical, but I think it puts it into perspective is, say, you wanted to compare two carrier agreements to see who is gonna be cheaper to ship, package. Theoretically, you could just go out to Claude or ChatGPT and drop in some raw shipping detail, a PDF copy of your agreement, and the PDF copy of the carrier service guides and say, okay. Go out and tell me which one's cheaper. And would it do a decent job? Sure. It would. But it certainly is nowhere close to the result that you would get if you went in and said, okay. Here's the context. Here's how you normalize zones. This zone is this with the other carrier. This is how the service levels compare between the two carriers. Give it all that additional context. All of a sudden now, you've put what you know into it, and the output you're gonna get is gonna be so much stronger, having pre refined that data and having given it the additional context it needs. So if I had to leave with anything, it's that better data doesn't always just mean more raw data. Agreed. Yeah. Context isn't just data. Context is instructions and expectations as well. Tyler, you've been talking a lot about strategy. What does that foundation look like from your perspective? Yeah. I kinda hit on it with my last answer. I think it's like intentionally building a framework and not just letting everyone run run ragged. I know, Adam, you and I have had a lot of conversations about AI, specifically in the warehouse and WMS, and I I love some of the stuff you guys are doing around. Like so on our agentic side, like, everything has a role and a purpose. Right? It's it's built around the roles in the warehouse. And so there's there's strategy and there's intent behind it. It's not, well, can we what what could we do with with AI here, there, and everywhere. Right? And so I I think when I think about foundation, it's it you don't have to bowl the ocean. You don't have to apply AI to everything. Pick a strategy for how you wanna apply it and pick a specific use case to dive deep on and learn and understand what the the pros and cons and what it means, and then find ways to enable your team to expand and beyond that. Excellent. Excellent. Adam, let's go ahead and give you the last word in this one. Yeah. So I I I agree with Tyler. I mean, the the look. Clean data is important, but but I I think that one thing that we've not really talked too much about is trust. And, you know, when you talk about those operators in the warehouse, you know, folks that have been doing the job for twenty years, thirty years, or longer sometimes, you know, they know how they think, you know, decisions should be made. And when all of a sudden, you know, you plug in a new, AI agent and it starts recommending to do this or that, and it's different from what the operator would normally do, how do we think that's gonna, you know, what's the what's the reaction there? So I I think that, you know, we've gotta really think hard about how do we present the recommendation to a user. Right? So you're right. Having an agent that's like an assistant at a role level makes all the sense to the world, but at the same time, we've gotta be able to kinda back up, you know, what the recommendation is. Here's why it did it and, you know, ultimately help the operator understand the reasoning behind the the recommendation and and and you'll get them on board. Right? You'll get them, you know, not not just on board, but super excited about using, the agents because it's gonna make them do a better job and make them even that much more valuable, to their organization. So, I think that's really important is to is to drive a sense of of trust and, you know, make sure that this becomes a tool that's relied upon, moving forward. Yeah. Absolutely. I I love that, the trust. And if you if it's not working out, it's not gonna you're not gonna build that trust. You're actually gonna build distrust. You know, we have seen a lot of, systems in the past that had that exact same problem. And AI is no different than those old WMS's shipping software analytics tools. If you can't have trust, people won't use it. People will not, go through it. So that being said, we're gonna move to our last section before QA. I'm gonna go around the panel, and I'm gonna have everybody give their final advice. Let's start with our contrarian, Matt. I'm gonna echo what Tyler said because I realized I that, when you're doing anything in AI, you have to, like, know your final business case. Like, know what you're trying to solve with it. So going going into, you know, some conversation with AI with a very clear end goal, will be the most helpful thing for you. K. And next, Tyler. Yeah. I I think my biggest piece of practical advice is, like, pick a small use case even if it's like a personal use case and experiment with it. Hey. If if I could do this, what would I do? And I think I'd I've learned the most about how AI can apply to my broader business role by by doing little small personal projects with it. And Adam. Yeah. I I think that, you know, as we think about, how to how to, you you know, best leverage these these tools, I mean, you know, we we do think that that applying, you know, kind of an assistant or an agent at a role level is is is gonna be a helpful starting point. And we we can't jump straight to full on automation and autonomous agents and so forth. It's a crawl, walk, run. You know, give somebody a a framework where maybe it's recommending and they can you know, this human in the loop, I think is a really great starting point. And then maybe there's some things that we start to see. Hey. I'm doing the same thing over and over again. Now I can make that autonomous. Right? Because I'm I'm trusting more what it's what it's recommending. So now I with that trust, I can now allow it to to to go autonomous. But you can't go straight to that. Right? You gotta have have some, checkpoints in between. Alright. And, Quinn, you've got the final word. Alright. Yeah. So what I would really say is just whether you like it or not, I think AI is here to stay. And obviously proceed with caution with AI. But I think ultimately, if you're too averse to it, you do kind of risk in the broader context of business. Once again, this is kind of that 10,000 foot philosophical view. You kinda start to risk falling behind a little bit. And I think that people, as they're using AI, need to also really ask themselves a really specific question and say, are you using AI to do something new that you've never been able to do before, or are you just doing the same things faster? And I think, companies are really gonna, at a certain point, need to start taking some risks as it's related to AI because, ultimately, with the AI models that are out there, everybody in the world now has the ability to do the same things faster. But true competitive advantage and the companies that are truly gonna start to pull ahead in the long run are the ones who start to use AI to do things that were previously, out of reach completely. Excellent. Excellent. And and I I I agree with a lot of things that were said here. I wanna throw out a word that I that, I stole from a guy off YouTube whose name just eluded me, but taste. Right? And, I see a lot of AI pundits talking about the good use of taste, and taste is what we as humans have. It's judgment. But the one thing that that, we don't always apply that to is where we're gonna start with AI. Right? We talked about things where they don't fit, like predicting, or rearranging things because of weather. That it's not good. It it doesn't fit right to do that, at least by the panel and by the audience decision here. But there are areas where where our judgment will allow us to go ahead and move down those areas where we can find where we can learn those lessons and start using it. Okay. We're gonna move on to QA here. We've had some great conversation today. I really wish we had a lot more time to actually, discuss this, but, we had a really good question today. K? So with the improvement of open source models and the improvement of consumer hardware, I don't know how many of you saw the Apple releases, this week, last week. I forget when they were. But where do we see the battle shaping up between the, the large language model companies like your Anthropics, your OpenAI, Grock, you know, etcetera, versus those local models, which it kinda feels like that's where we were three three or four years ago. A lot of ML was done on local models, not these cloud based models. So to my panelists, and and I'm literally just, you know, gonna throw the ball up here. What are your thoughts on on where this is gonna go? I'll I'll start. I know this is something, Justin, you and I at least briefly tell us back and forth. I I think it provides an interesting viewpoint, especially for larger enterprise customers. And so when we look at the the largest level, I think there is a conversation to be had at a leadership level around, should we invest in the hardware, in some of these open source options so that we can keep any of that, proprietary information that we learn or build to ourselves. Right? I think there's questions about privacy. I know some of the off the shelf models have really good controls around them. Quite frankly, I use all off the shelf models. I haven't explored, some internal models even on a personal level, but I do think it's more of a question now. And if I was running a multibillion dollar enterprise at a leadership level, then that would be a question I would start to ask my team to evaluate and and come back with some ideas on. Yeah. I I find this so fascinating because to me, as I look at it and as we as Adam's so so well defined, ML versus AI or LLMs, if you will, ML really seems like the type of thing you might wanna keep local. Keep that customer data local. Keep your personal or your company data local. Whereas when it comes to actually taking the results of ML, determining how to communicate that, determining how to interpret it maybe, that's where an LLM comes into play, and that's where maybe these frontier models, that cost, will start to actually provide the value associated with them. You know, sometimes it's what what do you need? A a small hammer or a sledgehammer? So use the right tool for the job. Right? Anybody wanna add to any of this? Yeah. Justin, the only thing I would add there is that I I I think I I love the analogy you just made using the right tool for the job. I I I think that's I think that's a brilliant, you know, analogy. And, you know, when you're thinking about, like, how how am I what what is the output I wanna get from this model? Right? And then what is the input that I really am required to put into it that's going to I I think somewhat inform, you know, the the answer there. Like, if I have some sensitive financial data from my company, I'm probably not gonna load that into Claude or, you know, one of the other models out there. I might wanna use something a little more local for that. Right? But if I'm just looking for some, you know, some help with with with reverbalizing an email or or how do I wanna, you know, phrase something, you know, asking some questions about, you know, general things, I mean, you have to go outside. Right? Because you need all the data you can get, you know, to to to do the best, to to do the best. So I I I think you gotta be thoughtful about the, you know, sensitivity of the data that you need to feed it. That's a that's a big piece of, you know, what I think about every day when I decide whether I'm gonna use our internal model, we use Glean. Right? Or, you know, sometimes I I do use Cloud quite a lot, but but carefully. Right? Responsibly. Mhmm. Mhmm. That's that's a good point. That's a good point. I mean, I I I will say that that, especially if it's, if it's HIPAA related, financial related, those local models are probably gonna continue to gain traction. But what I what I the other side of this is if you're attempting to do a frontier level model on local hardware, that's a lot of hardware that you're gonna have to amortize and and get a lot of, of use out a lot of tokens out of in order to get a true ROI. And that's where I think that, though the floor for that, ROI will continue to to go up, it is still, or will continue and improve, I should say. It is still to the point where if you're attempting to run a full Kimmy k three or or something that's gonna compete with a Frontier model, you are putting so much cost into hardware that, it is going to be years for you to get an ROI. And given, Moore's law, the rate which, that hardware is going to, depreciate, I really, really worry that if you're trying to do a Frontier model, that you are going to be, never receiving that ROI. But on your smaller models, on your MLs, on your, TTLs, on on all these other other things that that don't require, you know, seven sparks put side by side to run, all of a sudden that ROI model starts to change, and the value per token also starts to change. So I think it's very fascinating, and I think it is one of the most exciting times, since 2000, since I've been in the industry here. So it's very exciting to see where we are with this, but it's also very risky. Just because you think it will be best to pull everything local, if you don't have the needs, financial, HIPAA, privacy requirements, you may find yourself spending more money than necessary if you're pushing everything to a frontier, cloud based model. Alright. Alright. Time for one more. Alright. For shippers using legacy systems, have you seen challenges with layering AI on top of legacy OMSs, WMSs, and TMSs? It's a great question. In some cases, like, Adam, I know that you guys just built it into your new solution. Right? So so no need to go for the legacy part there. Yeah. No. I think it's a great question because, look, when we when we built our solutions in the cloud, I I would say that, you know look, I I I think we did a a great job of building a a a platform for the cloud, you know, born and born and bred and so forth. I'll be honest. When we when we built it, we were not thinking a lot about AI back in 2017. So maybe a little bit of dumb luck that, the architecture, you know, really is highly supportive of AI now with our API first technology, etcetera. So I I I, you know, sometimes luck is a is a good thing. But but it is it is really well suited for that and, you know, legacy solutions. If I look back at, like, our older, you know, kind of versions of, warehouse and so forth, not API based, very difficult to do any sort of actionability. Right? Maybe you can extract data out externally, put it in a data lake, and then throw AI on top of that. That's really your solution for a legacy, you know, thing. But but you're not gonna be able to do things. Right? It's gonna be very limited in terms of its ability to recommend, you know, and and actually go do something and take action. too think that's the big difference. hard. Alright. And and, yeah, lots of stuff there. I think I think you mentioned it really comes down to how how you can integrate to those legacy components. Not every single legacy component is created equal. Some of them are more difficult to get into a data lake. Some may have APIs with with varied capability. Alright. With that, I think that's about all the time we have. It's great. Please make sure you fill out the survey. Ask questions. If you've got anything for our panelists, please let us know. We've got a bunch of emails out here. Communicate as you need to. Let me be the last to say thank you to all my panelists. It's been a great conversation. I really wish we had an extra hour to keep talking about this. And thank you as our audience for attending.