AI Is Eating Logistics

Logistics is a multi-trillion-dollar industry that quietly powers the entire global economy — and it's shockingly manual. Ryan Petersen, founder & CEO of Flexport, joins the Lightcone to break down how AI is finally touching the physical world: making shipping cheaper, speeding up global trade, and automating work that used to live inside emails, spreadsheets, and phone calls.

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:00Speaker 00

  Logistics is a very scale-driven industry. And so the bigger you get, the cheaper you get. Our take is that we can make the price of shipping anything by ocean container shipping between 8% and 10% cheaper over the next few years. And AI is a big part of that. So our AI for that saved us 2% of our ocean freight spend while improving transit time 20%. Usually that's a trade-off. It's like you're either faster or cheaper, but not both.

• 0:25Garry Tan?

  And you're at $2 billion a year revenue and just getting started.

• 0:35Garry Tan?

  Welcome back to another episode of The Light Cone. We've got a real treat today. We have Ryan Peterson of Flexport with us. He went through YC in 2014 and he is easily one of the most awesome founders I've ever met. Ryan, thanks a lot for joining.

• 0:51Speaker 00

  Thank you.

• 0:52Garry Tan?

  To start, Ryan, what is Flexport? And what are some of the things in AI you're actually implementing right now?

• 0:58Speaker 00

  So Flexport is a global logistics company built around a modern tech stack. And that means we help companies ship cargo from point A to point B across any mode of transport, so air, ocean, truck and rail, and get that cargo delivered hopefully on time and in full at a lower cost, thanks to the tech. What we're doing with AI is, man, I had to make it exhaustive. We were able to extend the length of the podcast to pull that off. But it starts with customer user experience. What can we do with their data? Getting them better access. How do we load containers in the optimal way? How do we put that container onto the right ship at the lowest cost while maintaining or beating transit time expectations? Automating just tons of work that's done in email or that you, or phone, or work that

• 1:43Speaker 00

  you wouldn't even do because the cost is too high for a human, but actually does create some value that's worth it with AI. So most contracts in logistics come in giant Excel files, thousands of rows and dozen tabs. You can't just feed that to OpenAI and get an answer, get a structured JSON file back. It needs intelligence. But writing code and then having AI write the code, you write a parser that ingests it and then have AI that can write. Those parsers for you, learning, it's an endless list and we feel like we don't even know all the things that they can do. It's still pretty new.

• 2:16Garry Tan?

  So basically, one of the most human intensive things now can be streamlined to the point where actually it might affect GDP in the world.

• 2:25Speaker 00

  Our take is that we can make the price of shipping anything by ocean container shipping cheaper by between eight and 10% cheaper over the next few years. And AI is a big, not the only part of that, but a big part of that. As our, our business model, the way we think about it is as, I call it scale economies shared. Which is the bigger you get, the cheaper you get, the more automation is a form of scale. And the bigger you get or the cheaper you get, you lower your, the lower your costs, you give that, share that with your customer, which will make them do even more volume with you. There are scale benefits that come, logistics are very scale driven industry. Uh, and so the bigger you get, the cheaper you get, like the Costco model, I like, I like, I love Costco, even though I don't shop there, I just love the business.

• 3:08Speaker 00

  You keep driving down the price, that makes you more attractive, more competitive and, and just keep going, yeah.

• 3:12Garry Tan?

  And you're at $2 billion of your revenue and just getting started.

• 3:15Speaker 00

  Just getting started, yes.

• 3:16Harj Taggar?

  Something I'm curious about, so from our perspective, we work all the startups and we've seen like AI over the last couple of years go from like when ChatGPT launched. And then like some startups in the batch start playing around with it and it's become like progressively more serious. I think you're the first person we've had on the show who's like running a company at scale that was founded pre-AI. What's like, what have the last few years been like for you from that perspective, from like ChatGPT launch? Like, at what point did it start becoming like a thing you were paying more attention to?

• 3:44Speaker 00

  Like so many other people is on November of 20, was it 2022? It's already been a few years, uh, since the ChatGPT launch, been personally obsessed ever, ever since then. It's interesting to watch it take hold at the company and in some cases not take hold. And you then saying like, come on, guys, we can't be this boomer company, like everybody needs to be using this. We're trying to drive that sense of paranoia from, from the top, from me, but, but many others in the company, maybe even more paranoid than me or more enthusiastic, excited than me as well. To say that story that we say of like, well, we're the only large logistics company. He founded since the web browser and I know there's a kid in the next Y C batch who say, hey,

• 4:24Speaker 00

  we're the only freight border founded since chat GPT November 2022 like it's got a point. So we we have to be leading on this. This is true of all incumbents in an industry. They have some real advantages when it comes to AI and benefiting from it. And one is the scale of the data. Two is the domain experience to know, okay, which problems should we be solving? And some of those problems are small enough that you shouldn't start a whole company around the problem. It's maybe a feature, not a company. But for them, it's great. It's a valuable feature that they could add. And third is distribution, like when we build or any large company builds a great AI product, the next day it can be used by thousands of companies.

• 5:07Speaker 00

  Whereas a startup doing that has to go beg people for their data to train the model and trust, earn their trust to have that data from a security compliance standpoint. And then third, get the customer, get, you know, so that's the huge advantage that any incumbent will have. And we think we definitely feel that we have that advantage at our scale. But the flip side where I think we also have an advantage is that we are still a young company relative to our industry. But young in terms of our tech stack, like we build our own tech, therefore we can implement, integrate AI and just add it wherever we want. Most of our competitors treat AI, treat technology period as IT as a service that they pay for. Many cases like desktop app or like Windows remote desktop is very common in our industry.

• 5:51Speaker 00

  But still, it's something they buy and therefore you don't control the code base. If you wanted to add AI to automate something or do something, you're like, it's not really, it's hard.

• 5:59Harj Taggar?

  Has there been a specific moment since chat GPT launch where you started as a company taking it more seriously? Or because my impression, like the first version was a toy, even within the Y C batches, like we sort of see some founders playing around with things, but it wasn't clear that they'd actually be like companies founded on it. So I'm just curious what's like founder running large scale company. You start out, you're like, this is really interesting to me personally. Like, was there some moment where you're like, oh, like, we should probably try and like, build something or do something internally with this?

• 6:27Speaker 00

  Yeah, I think a lot of it has come through in our hackathons. But there could be an interesting metric here is like, what percentage of hackathon projects, first of all, used AI? Like, we're building something with large language models. And which percent of the projects actually you decided to fund and push into like, let's actually make this thing real. It's not just a hack.

• 6:44Harj Taggar?

  Is hackathon something you've done for a while?

• 6:46Speaker 00

  Yeah, we usually do two a year.

• 6:48Harj Taggar?

  Okay.

• 6:48Speaker 00

  Oh, I think now we're like, kind of religious about two every year. But, but one to two a year where, and for us, it's very much a free for all. You can build anything you want. If you look now at the last two hackathons we've done, it'd been like 90% LLM-based projects. In my, I haven't studied it, but I was just like feeling my feeling, my gut. Whereas probably 18 months ago, there were like four or five.

• 7:11Harj Taggar?

  Interesting.

• 7:11Speaker 00

  We, there's probably 50, 60 teams that do a hackathon project each time. In the beginning of Flexport, I was very much of this idea that like, you just let, let the people, get smart people and get out of their way and go execute. And

• 7:25Garry Tan?

  Oh, that sounds like manager mode.

• 7:26Speaker 00

  Yeah. It was, I, I had way too much manager mode. I had this idea of like, human beings are going to flourish. If only they could be set free. They don't want to be told what to do by the man. That's why I started the company. I don't want to be told what to do. And I went through my own Chesky moment of founder mode and recognizing, oh, you've got to be way more tops down and directive and tell people what to do and get people aligned and rowing in the right direction. And that's been my evolution the last two years at Flexport. I've been pretty, way more hands-on and hardcore in directing the business. But then, as I see these hackathons, I'm like, I never would have come up with that idea in a million years. And I gotta let these guys build what they want to build and flourish. And so I'm starting to now come back on myself and say, where's the room in our product roadmap for bottoms up innovation?

• 8:11Speaker 00

  Certainly you see it in these hackathons and Trying to maybe even start making sure I do the hackathon timing before we say we do our kind of roadmap exercise every six months or so. We should probably do the hackathon right before that, so that when you see a great idea, you can budget it instead of after.

• 8:26Garry Tan?

  I mean, there's a noteworthy change here that's happening for you that, I mean, I think most companies might throw a hackathon and then in most hackathons, 90% of the projects are just like toys and you never return to them again. Like, you know, someone gets a nice participation trophy and that's it.

• 8:43Speaker 00

  Yeah.

• 8:43Garry Tan?

  But like, it sounds like the difference right now in the age of LLMs and age of intelligence is that these hackathon things are actually turning into real product lines and features for you.

• 8:54Speaker 00

  Yes, and at the very least into debates in my head of being like, man, I've got to do that. But also, I lost, we're gonna, we're gonna crush everybody with just our regular roadmap. But I had this after the very, I think our last, our next hackathon's in two weeks. The last one was six months ago. I remember thinking afterwards, I'm like, you know what, we could just only do that stuff and we'll also win.

• 9:13Garry Tan?

  Yeah, maybe win faster, maybe.

• 9:16Speaker 00

  It's highly unlikely that the person at the top now knows best what the best implications are, applications are, that it's, it's just as likely that someone on the front line is closer to the problem is gonna go, hey, look, watch, it can do this. You go, oh, man, I wouldn't have guessed it could do that.

• 9:30Harj Taggar?

  You kind of need engineers who are just really into it and have been playing around with it and just like, understand how to like, build the products in the first place to come up with the ideas.

• 9:37Speaker 00

  Yeah, engineers and engineers being really close to the business is something we've always prided ourselves on, like really being in the weeds. And one of the other things that we've done is create a program for non-engineers to learn AI skills and if kind of formalized programs. So your manager has to agree, but you get one day a week for 90 days. It's a 90 day program, one day a week where we teach you kind of a AI bootcamp, vibe coding and different ways to apply. And it's a new program. So we're only about six months into this. We'll see how it works out. But people love it and you are seeing gains. But the promise of the leader who created this and convinced the managers to give up someone for 20% of their time to go into it was, I will return

• 10:20Speaker 00

  them to you as 10 times more productive than their peers. I'm sure we haven't have achieved that or it would show up in the metrics. But that is the, you know, that's the idea.

• 10:28Diana Hu?

  How are you training all these folks to up level skill in AI? What are, what are the sorts of things they're learning?

• 10:34Speaker 00

  Certainly it's, it's cursor and a set of related products like that. That like, I think we're using something called Streamlit, but probably there's Y C company. I don't know, maybe she's rep lit or something with similar ideas. You can spin up, uh, build your own little apps. Um, build workflow automation tools to say, okay. Cause a lot of what Flexport is, we call it freight forwarding. Uh, I've often joked it should be called freight email forwarding. You're like taking docs and sending it on. So how do you look at a person's job? And there's no one better to look at it than the person doing the job and saying, oh man, I'm doing the same thing over and over again. What if I instead, it's like if everybody was an engineer, they would. And I've thought about this in the past and saying, hey, what if I took one group of engineers

• 11:16Speaker 00

  and, and hire them as engineers as a big bait and switch? And then tell them, actually you're just moving freight. Sorry. And watch them automate their way out of the job. Right? And you sort of say, okay, I, I never really wanted to do that to an engineer. Cause I feel like I'd just have a revolt of, but now you're kind of like, well, I could do it to a non-engineer who's already doing that job and turn them into a, you know, a lightweight, no low code.

• 11:39Diana Hu?

  Which is cool.

• 11:40Speaker 00

  Yeah.

• 11:40Diana Hu?

  This is going the other direction where you're taking really all these super domain experts and now they can finally build and they can automate themselves out of it instead of getting the engineer to do it.

• 11:49Speaker 00

  Yeah. And that, that program started on our Amsterdam. We have an engineering office in Amsterdam. It started there. I think they did it without me knowing about it for the first six, a few months. And then now we're like, oh, this is great. Everybody loves it. So we're, we're starting to bring it global to other offices.

• 12:03Jared Friedman?

  I wonder if you could share some examples of the AI projects that you have rolled out that have been most impactful over the last couple of years, both customer facing features, but also like any internal operational things that you guys have automated that maybe the customers have no idea about.

• 12:17Speaker 00

  Yeah. Um, the customer facing one, probably the most impactful, like a lot of what you care about from your logistics company is your data. What's going on with my supply chain? Where, uh, the, the types of data that people are looking at. So the way Flexport works, you, you place orders to your factories through Flexport. I'm replenishing my inventory. I'm buying things, I'm placing purchase orders. So those flow out to the factory. Factory becomes a user. There's a nice network effect there. Once the cargo is ready, they place a booking. Uh, and then we execute that, a booking to move the freight to come pick it up on this date, uh, and we'll execute it. By air, ocean, truck, rail, whatever, and move it across the world for you. So that's kind of the the loop that we're trying to run. So you care a lot about the data for on time performance, SKU level, performance, cost.

• 13:02Speaker 00

  You care a lot about that. There's customs attributes here that are super important with tariffs and everything that's happening. So being able to get that data is one of the core areas that Flexport shines already historically. With AI, and this did start as a hackathon project, we just built like natural language ability so that you don't need to know SQL. You don't need to build dashboards. You just type your question and it generates those graphs, charts, tables. Don't think it does maps yet, but it should. And it works and that is done. One, one, customers love it, but two is it's about 25% of our account management time is spent helping people generate reports. That's the

• 13:41Jared Friedman?

  Got it.

• 13:41Speaker 00

  Another huge metric for us is we're cheaper, we'll more people will choose us. It's not that we just started using AI with L L M's. We've had a machine learning model for doing planning for and planning in the sense of logistics means, let's say on a containerized basis, I've got a container, which ship should it go on? Therefore, you need all the contracts with their price. You need the sailing schedules, like how long is it going to take? Which route variability all the around those, both those things. So our AI for that saved us 2% of our ocean freight spend while improving transit time 20%.

• 14:19Speaker 00

  Usually that's a trade off. It's like you're either faster or cheaper.

• 14:23Jared Friedman?

  Yeah.

• 14:23Speaker 00

  But not both. So huge win there. Customers don't, they care a lot about those metrics. They don't care how we did it.

• 14:29Jared Friedman?

  And for that one, was the unlock parsing a bunch of unstructured like emails and data that you get from the shipping companies that have this? But it's all like in like a big paragraph where like you couldn't just like run a simple query on it before.

• 14:40Speaker 00

  Sort of, yeah. The way to think about it is you've got, if you just put a container on the cheapest contract, you you made it's an optimization. Okay, which one's the cheapest, but also the fastest, you know, I'm trading off. So that that's one thing that machines are better at. And then it's the scale of that. So on a given week, we have about 2,000 containers that get canceled by our customers. They place the booking and then they say, oh, actually the cargo is not ready. The factory is late. It's just inevitable that's going to happen. What software does that humans could never do is go through 10 times a day and taking each one of those containers and say, okay, I lost this container. It's been canceled. Is there another container that was meant to depart one week from now? And I'll grab that and move it forward. That's how you get the 20% transit time increase.

• 15:22Speaker 00

  And then the optimization piece of is just find the cheapest contract, like a solver out, you know, algorithm to go find.

• 15:27Harj Taggar?

  And the humans can't do that because it has to happen really quickly.

• 15:29Speaker 00

  Is this happening 10 times a day for every container in the system?

• 15:32Harj Taggar?

  Okay.

• 15:32Speaker 00

  You know, it's just like you wouldn't, maybe you could, but you wouldn't.

• 15:34Harj Taggar?

  You have to have like an.

• 15:35Speaker 00

  The labor cost would be crazy.

• 15:37Diana Hu?

  If you calculate this first principle, sounded like that first version was using classical optimization problems. And you had certain data about all these shipments, inputs, outputs and schedules. What do you think is the delta that you could get with AI now that you could harness all the instructor data? What kind of efficiencies could you get?

• 15:54Speaker 00

  You may be able to get a lot more now that you're starting to see tool use because the tool itself is incredibly powerful. And I don't think an LLM will outperform that, but the LLM can use that tool. And it could do other things outside of that. So you can, we'll see, we haven't started to do that yet. So we're actually still using that.

• 16:09Harj Taggar?

  I can actually email people or call them up and.

• 16:12Speaker 00

  Yeah, but it could, you assign the LLM the same solver problem, but it is good at default to use this tool. And then it'll also say, yes, maybe this container, I'm not sure if I could move it forward. I should ask the customer would be a good idea, actually email the, hey, is it okay if I bring you this container early?

• 16:27Garry Tan?

  Like the solver is still there, but then basically the agent is the user.

• 16:32Speaker 00

  Yes. Instead of right now, there's not really a user or there's a, there's someone is approving the plan. And so you could make that person upstream of the solver, choose the solver as one of many tools. So that'd be interesting. We haven't done that yet, but. And then the other thing is to just routine work. For example, you've got a lot of email communication with your customer base. So how do you take this? You say, hey, I want to place a booking for a container. Translate that into a booking. LLMs are quite good at that. A big use case today is verifying warehouse addresses and other information and getting appointments. I've got to deliver to a warehouse. Quite costly to call the warehouse and be like, do I have the right address? You're not going to do it every time.

• 17:15Speaker 00

  And then you have a lot of misses where your address data was bad. And your truck got lost, you know, it's a pain in the ass. So LLMs now, before we deliver, if we haven't delivered to the site in the last three months, there's an LLM agent that does email and voice.

• 17:31Jared Friedman?

  Interesting. Wow. So if necessary, it'll actually call the warehouse and be like, hey, can you confirm the 2pm tomorrow is a okay time to deliver this?

• 17:39Speaker 00

  Yes, yes.

• 17:40Diana Hu?

  Which is great because you're turning this previous communication protocol, which is very much, I suppose, very lossy to work sort of like the internet, like TCP, fully acknowledge and you can get guarantees.

• 17:54Speaker 00

  And sometimes it's not replacing work, although I'm very happy to do so. But like, in some cases, the work would have been too expensive, so you just didn't do the work.

• 18:00Diana Hu?

  To do this phone call.

• 18:01Speaker 00

  And even if a human could do it, it's like, not worth it. Another good one is just messages. So like, the way we communicate with our customers, some of it's email, but a lot, we try to drive as much as possible through our messaging applications inside the Flexport platform. There's this huge amount of signal in that, just customer sentiment. If a customer, in AI, we've trained the model to detect unhappy customers in the way that they message us. And then that creates an automatic escalation to the manager of the frontline person. Hey, this person seems upset.

• 18:35Speaker 00

  There's a lot of emotion in logistics, you know? It's your stuff, your business is on the line, you need to get it delivered. Infinite list, in fact, we measured at the beginning of the year, we had automated 20% of the work. It was pretty low scale of automation. We're going to finish this year at 50%. And we had set a goal for ourselves next year of 80. We thought 80 was actually the upper limit of what could be automated.

• 18:58Speaker 00

  It's not scientific, but now we feel like, oh, it's probably closer to 90 to 95 current. And then that'll get way more so as LLMs keep progressing.

• 19:07Jared Friedman?

  How will that affect like the total cost of ocean freight? Like if all the human work gets automated, does stuff actually get materially cheaper?

• 19:15Speaker 00

  Yeah, it's 10% of the end cost that the by the importer exporter pays for their freight 10%. If you look at the full P and L, about 10% is the labor cost in the freight forwarding layer of logistics.

• 19:29Jared Friedman?

  Wow. So, so when AI is like fully rolled out, like stuff will actually get 10% cheaper.

• 19:34Speaker 00

  Well, the freight of moving the stuff, the cost of moving it, right? The stuff itself depends on what the ratio is, but yeah.

• 19:40Jared Friedman?

  But like the transportation costs of like international freight is actually like 10%.

• 19:45Speaker 00

  On containerized ocean freight, that's our, that's our view is that we can drop the price of everything by around 8%. And maybe it goes to 9% over the next few years by doing this.

• 19:55Diana Hu?

  That has some last big economic ripples in terms of if it's becoming cheaper to ship things across the ocean. Is it going to create just more trade? I mean, there's also trade wars, but.

• 20:04Speaker 00

  Exactly. It's very hard to control for that in the world where tariffs just made everything like 10 times more expensive. But we're doing our part.

• 20:11Garry Tan?

  I mean, the white pill on AI right now is, you know, this hope and sort of possibility that AI rolled out properly across society would increase GDP 7% a year. So this would be maybe.

• 20:24Speaker 00

  7% a year will double you in 10 years is the law of 72. That is the hope, right? And I think more people should talk about that. And, uh, and everyone's so worried about automating away the jobs. And I just think that misunderstands the role of companies in the, in society. Like the role of companies is not to employ people. It's to deliver goods and services. And in fact, whoever employs the least number of people will have the lowest cost and win. And that's how they benefit society is lowering costs and making things more available for us to buy and sell. And then there's this idea, well, how are people going to make money if AI is doing all the work? And I, I think that that very much misunderstands human nature that we'll, we'll just want more things. Like there's an infinite desire inside the, the human soul can never be satisfied.

• 21:08Speaker 00

  Without God, we need more stuff like we gotta have more, we gotta have more. And so

• 21:14Garry Tan?

  We're trying to return to the garden.

• 21:16Speaker 00

  We may get a return to some, I think that actually the internet first, we haven't quite reconciled this on like a spiritual philosophical level, the emergence of these technologies and AI were not even beginning to of what it means for us. But there's a period in history called the axial age. It's about 500 years B.C. And that's when coins really started to spread. What you had with, if you think about it with coins is taking transactions between two people and make and really making them very impersonal. You no longer care who you're doing business with. I don't need to have a ledger. Does this guy owe me money? What's my relationship? Do I trust him? Just like, here, take this thing. So and it actually

• 21:56Speaker 00

  led to this breakdown in societies because we just stopped being so knowing your neighbor, like you used to only do business with your neighbors. Now you could just do business with any old person. The internet kind of does that at scale. What happened in the axial age, you had this breakdown of ability for of trust and you started to get generosity and all kinds of like things that start to break down in society. And simultaneously across the world, you had four major prophets that emerged. Yeah, well, prophets is of sort. You had Buddha, you had Lao Tzu, Confucius and Socrates. They all lived at the exact same moment in time, right? As coins were taking hold.

• 22:33Garry Tan?

  Fascinating.

• 22:34Speaker 00

  As like a, hey, we need to, we need to kind of like get our hands around how do we behave in this new world. So I do think there's an opportunity here. Maybe it could be you, Gary at Y C to be the next, the next Socrates.

• 22:48Garry Tan?

  I'm in, but I might not be the right person. I mean, I particularly like this idea that like the idea that what are humans going to do is a little bit invalid in that, you know, that's a little bit like going back five, 800 years and saying like, oh my God, all of us are farmers. And then what are we going to do when modern agriculture comes?

• 23:10Speaker 00

  Yeah.

• 23:10Garry Tan?

  And it's like, we figured it out. Like we Yeah.

• 23:13Speaker 00

  Right? What are the monks gonna do? They're transcribing words all day. There's no more jobs for transcript.

• 23:18Garry Tan?

  So there will be implications for society and morality and how people relate to one another. And obviously, like

• 23:25Speaker 00

  And we have no idea because it's early days, but history does kind of repeat and there's lessons there and figure out, okay, how does this? But the human nature doesn't change much, right? You can't satisfy humans. You're just gonna want more stuff. The more money you have, the more, classically, right? Cliche, like the more you have, the more you want. That's not gonna go away. So you give people a lot more stuff. It's not like, oh, I'm gonna quit working. Most people aren't like that. I'm gonna get a lot of stuff. I'll just quit working. You find out you're miserable. You want to keep producing, keep contributing.

• 23:51Garry Tan?

  One of the interesting things that has been percolating around the YC community among young founders and like AI researchers that we've been talking to is this idea that like they're going to be humans in the loop. The humans in the loop may well be, uh, some, some might be like government mandated, right? You know, in fintech, there's a lot around, uh, you cannot have an AI algorithm, like approve loans, for instance. They're like requirements from the government in these highly regulated industries to have humans in the loop.

• 24:21Speaker 00

  And then customs brokerage as well. We have to have a human that's approving the transaction before, before we clear customs.

• 24:25Garry Tan?

  Yeah. And so vibe coding is happening. There's this idea of you enter a prompt. It comes back with a bunch of stuff and then you just click accept all changes without reading any of it, right? Do you think this might happen? Would this happen at Flexport? Would this happen, uh, more broadly across all businesses? Like what if businesses are at the core, like hyper intelligent AI that, uh, has access to all your systems of record, knows what to do, optimizes constantly. And then you have sort of like government mandated liability sinks that are humans in the loop. Ideally, the organization still actually serve human needs, in which case, like the decision to use, you know, vendor A or vendor B

• 25:09Garry Tan?

  sometimes boils down to who brought me to the nicest steakhouse last. So then like the model for companies ends up being ASI of some sort, like some sort of AI process at the core of each company. But then, you know, humans attach to it as either like decision makers in like, you know, accepting or preventing liability and or holding relationships with other relationship holders at other companies.

• 25:34Speaker 00

  Yeah. And presumably you're still relating with You're still here to serve humans. You know, once we get to where AI is serving AI, then fair enough. You don't need to learn that much from the record of human history because there's no more humans involved in the loop. And I don't have a lot to say about that. But as long as there's humans there, there's going to be humans are going to want to relate with other humans and have a relationship. And I think we're pretty, pretty, pretty far from humans preferring to work with AI than to work with other humans. We're seeing where AI is doing more and more work. You know, another good example is, and just that you made me think of with your bank. You know, you have to have an approver, approver is that even our, our humans in customs brokerage across the industry,

• 26:19Speaker 00

  we benchmark and make about 2% mistakes and they file the entry with 2%. And we built this sort of like AI spell checker, the two-digit code for Australia versus Austria. You could easily get that wrong and AI will figure out like, oh, this thing is not made in Australia. It's made in Austria.

• 26:38Diana Hu?

  I guess one question for you, Ryan, is if you were to start Flexport today, how would the company be different?

• 26:44Speaker 00

  Not that different, I hope. The things that Flexport did really well compared to all the other tech companies who have tried and failed in our space, both before we came along and in parallel, is we didn't look at ourselves as a pure technology company. We're willing to pick up the phone and solve problems with humans, drive down to the port. Still to this day, like I, we've got a new customer who's asking us to do something really weird. We need a crane on the truck to unload this thing and we don't have that. It's not typical what we do. And I just said, take the customer and I need you to drive there and follow the truck and make sure this goes well. So I would not change that at all. And I think that's, that's the mistake that a lot of tech people will, in traditional markets will fail at.

• 27:28Speaker 00

  Cause they're like, oh, if there's no API. I can't do it. If my agent is unable to do this task, I guess the task can't be done.

• 27:36Diana Hu?

  No tool use for cranes.

• 27:37Speaker 00

  Yeah, and it might take you a long time and you should not try to automate that last, that tale of things.

• 27:43Harj Taggar?

  You started in grew flexport, like especially in the first few years during the era where just like more money coming into like, there's more venture capital each year coming into startups. And like you, you had like multiple fundraising rounds. Like in what ways was like that capital like an advantage? And I feel like now we're sort of somewhat back there in the AI world now, like the rounds are heating up. There's more money flowing in it. It's sort of like post like the 2022 crash. What's your advice to the founders now who are in these like companies that are growing and have options to raise like huge funding rounds? Like, how should they think about it?

• 28:13Speaker 00

  Every company is super unique. So don't listen to advice on the podcast, like get someone who's like paying attention, you know, knows the details of your business, which no one will know better than you. Generally, capital is a beautiful thing. Having it in your bank account gives you a lot of advantages. All you really need to care about at the end of the day is. Price per share, because if you issue more stock, like the number, as long as your price per share goes up, you you are richer. Doesn't matter how many, what percent you own until it comes to control. So there's two things that matter. Do you control your company legally or otherwise culturally it also works? But do you really have control over what's going to, the decisions that are getting made? And do you have a job and and price per share? And that all, that's all that matters. I think, I still think that's true.

• 28:53Speaker 00

  That's always how I thought about is there's been a lot of dilution to our investors, but the price for sure went up and everybody's made better off. I didn't take away anybody's shares, so you're better off. The part that I underappreciated and that I now take very, very seriously is the degree to which money just wants to spend itself. And you will end up making a lot of mistakes where, and the biggest mistake is believing you, for sure, every company has a lot of problems. And you start to default to like, oh, we'll just use money to solve this problem. And the way that that manifests itself is, oh, I got this thing that we need to do. Okay, hire someone to do it. And you feel like you just end up very bloated. We had too many people. You start to really slow down. And it's just a super bad cultural approach to problem solving.

• 29:35Speaker 00

  Like you're going to solve the problems, not the new people that are you're going to hire. So I I give this advice. Only one founder's ever listened to me, but I tell founders who who are friends of mine who raised a large round. Sure, go raise a big round as long as you're up round, like you're doing good, great, raise a large round.

• 29:51Speaker 00

  Then do a hiring freeze for 90 days.

• 29:52Garry Tan?

  Hmm.

• 29:53Speaker 00

  The next day to tell your team culturally like, no, the money's not going to solve our problems. We're going to solve our problems and keep that. And then sure, go higher, but it's because it's super. It happened to us over and over again, where you just like headcount got out of control. So all the plans look good. I want to fund all the, we're doing budgeting for next year. And I'm like, God, it's so painful not to.

• 30:14Garry Tan?

  Yeah.

• 30:14Speaker 00

  Add OpEx, add engineers, whatever, but you gotta stay disciplined and the money will easily make that stop.

• 30:21Garry Tan?

  So I'm really psyched to hear about this idea that AI is actually transforming your business in pretty fundamental ways. It's like coming bottom up. What does Flexport look like in 2035?

• 30:32Speaker 00

  Well, a cool thing about Flexport is the way our vision has evolved. I mentioned, we started as a customer's broker. Now we do all end to end, all the way from factory floor to consumers' doors. Like we have an e-commerce business that does fulfillment, retail store, distribution, et cetera. So we want to take that globally to where you can really ship anything, anywhere, by any means, any mode and any quantity, and do it all via code, like all available via APIs or voice, or it's just like easy to execute transactions at the lowest cost, automate away the costs. And so that brands, companies of all kinds, don't spend time thinking about logistics. Logistics should be this utility that just works, just like you don't spend time thinking about the electrical grid. You flip the light switch, you get power.

• 31:14Speaker 00

  Uh, you go back to doing your thing, which is making something people want and talking to users, right? That's what I think companies should do. Our customers should be doing that all day, like make great products, make a great brand to sell those products. And we'll take care of everything in between in the most automated, efficient, reliable ways possible, uh, on a global basis. So today, and we have a long ways to go to actually make all that true. First off, the automation stuff I talked about, making progress, but I gotta keep going. And then the global aspect, so we have employee, we shipped cargo to and from 147 countries last year, but we only have employees in 22 countries. And therefore, people on the ground that can do the work. Yes, we are automating that work.

• 31:56Speaker 00

  And in fact, it's easier for us to automate our own employees' work than it is some third-party company that's doing work. You know, even though they're in our software, it's very hard to automate. We don't know what they're doing. So we want to be in every country by 2035, certainly. In fact, our roadmap has us covering 95% of all container trade with our own people doing all the work in the country in 2028. So by 2035, I think we could realistically say, look, we'll be everywhere that's legal. And that is a big extension of our original vision. And I didn't have all that in mind when I did YC Demo Day. My pitch was like, we'll do customs and then we'll add some other stuff. But it wasn't like, we will cover the earth. Any two plates on earth, whatever you want to move, we'll move it.

• 32:41Speaker 00

  So yeah, it's a very ambitious goal. The good thing is I really genuinely, we're going to win on tech. We're winning. We're going to extend our lead there relative to our peers, our competitors. We're behind them on the global side. And that's super fun. If you told 25-year-old me, they're like, oh, Ryan, your job this year is we got to launch Flexport in Indonesia, Australia, Japan, Philippines, Turkey, and Poland, and France. I'd be like, oh my, really? I get to go to all those places and talk to the locals and stuff? So it's a, it's a pretty fun moment in our history, but also really challenging, but fun kind of challenging, yeah.

• 33:19Garry Tan?

  No better kind. Ryan, thank you so much for joining us, man. It's always a pleasure. All right, we'll see you guys next time.