Transcript: Can Data-Driven Insurance Make Robots Safer and More Affordable?

[00:00:00] Sergey Litvinenko: Guess how much our payout is so far?

[00:00:03] Audrow Nash: 20 cents.

[00:00:05] Sergey Litvinenko: Five.

[00:00:06] Audrow Nash: Holy cow.

[00:00:08] Sergey Litvinenko: that's, how much better it is. Even like we're an early stage company, we don't have a ton of data yet, but from the data that we have, it's already that much better. And it's, literally printing, underwriting profits. And what we're going to do is that we're going to take that and we're going to, do the iteration of the models that we have and we're going to start bringing the rates down.

So it's going to be even cheaper to get our insurance. that's the magnitude that I'm talking about, and that gets me really excited

[00:00:36] Audrow Nash: Hey everyone, Audrow here. Today we're diving into a fascinating intersection of robotics and finance you might not have thought about before. I'm talking to Sergey Litvinenko, a co founder and the CEO of Koop AI, a company that's revolutionizing insurance for robotics companies.

Now, you might wonder, why should I care about insurance for robots? Well, it turns out that as robots become more and more common in our world, figuring out how to insure them is a huge challenge and opportunity. Koop AI, which is backed by Hyundai, is using cutting edge data analysis to solve this problem in a way that could accelerate the adoption of robotics across industries. If you're into robotics, AI or you're just curious about innovative business models in tech, you're going to love this conversation.

We get into some really interesting stuff about how they assess risk for robots from autonomous vehicles to industrial robots and even submarines. We also discuss the future of the industry and some surprising insights about robot behavior versus human behavior. So, let's jump right into it with Sergey Litvinenko from Koop ai.

[00:01:46] Introduction and Background

[00:01:46] Audrow Nash: Hi, Sergey. Would you introduce yourself?

[00:01:49] Sergey Litvinenko: Hi Audrow. thanks for having me. yep, sure. my name is Sergey Litvinenko. I'm a co founder and CEO at Koop. We are an insurance technology company that's focused on everything automation, robotics, self driving cars, drones, and we underwrite all kinds of risks in that space and, have a lot of things going on that we'll talk about today.

We're an early stage company, Series A, raised 7 million to date. one of our backers is Hyundai and Kia, a global automotive conglomerate. 15 employees in two locations, New York area, and also Dubai, UAE, and, making progress step by step. So, yeah.

[00:02:35] Audrow Nash: Okay. Hell yeah.

[00:02:37] Understanding Insurance and Underwriting

[00:02:37] Audrow Nash: so can you tell me why you, like, why is this important and what is underwriting?

[00:02:46] Sergey Litvinenko: Yeah. Let's start first with, what is underwriting or what is insurance in general? insurance is a financial product that people and companies buy in order to protect themselves when something bad happens. For example, you have a car accident and, you have to pay a hundred thousand dollars for liability because you damaged somebody's car or there was a medical injury or something like that.

You don't have that money, but because you pay a little bit of money to the insurance company every single month, they're going to take care of that. Okay, and what the insurance company does, they pull a lot of money from a lot of people and then they redistribute this capital to when the claims happen.

Same applies to businesses. As a business you pay money every month or every single year and when a big claim happens, big accident happens where you are liable for millions of dollars, insurance companies will pay for that. So what does it do for you? In exchange for a little bit of money that you pay on a month of a year basis, you potentially get more money.

You're self protected against millions of dollars of liability, and that will make sure you don't go bankrupt. That also satisfies a lot of contractual requirements. If you try to work with other people, because it's the safety net that people prefer to have in place. And in general, as a business owner, it's a peace of mind.

if something bad happens, I know somebody got me. So that's what it is fundamentally.

[00:04:07] Koop's Focus on Robotics

[00:04:07] Sergey Litvinenko: So what we do, this, but for robotics. Companies and for the robotics industry in general. And we ensure all kinds of companies that, sell robots or at least robots or provide robotics as a service. So think of them, anywhere between like day one operations that are about to do their first pilot all the way to companies that already sell industrial robots to many customers all over the world.

So that's who we're dealing with. And, Okay, that's step one, just to set the stage. Step two, why we're doing this? we believe that the robotics industry is the next great frontier in terms of what's going to happen with our world. And it's going to be a physical makeover over the next couple of decades, hopefully sooner rather than later.

And when you think about this, there are a lot of great companies and great people that build those robots and the systems and the services. And there is this layer that goes on top, which has to do with. Financial products and services, specifically equipment financing or robot financing, warranties, and also insurance.

Okay. Insurance is one of the financial products. And if you think about it, it's hard to adopt robots if they cost a fortune or if they don't come with a warranty and they break and you lose your money or if they make a mistake and then you're liable for millions of dollars. So those financial products, they are really important for the adoption and that's why we do it.

So insurance is one of the things that you need to have in place in order to make sure that People buy your robots and the businesses installed more systems and that when people ask you questions what if your robot makes a mistake and then I cannot operate my business you tell them that We have the insurance in place.

They're going to give you the payout and That's why you shouldn't be it shouldn't be a problem for you to go and start adopting My particular system or similar systems. So that's the why, that's the peace of mind. That's that level of friction that we're trying to, take care of. And, in general, we think that insurance at the end of the day will save a lot of companies from having a bad taste about robotics.

especially as robots become more and more ubiquitous. So that's the why. The why is at the end of the day. Get more robots into the world, help them commercialize and do this at the best cost possible to both robotics companies and their customers.

[00:06:38] Audrow Nash: Yeah, that makes sense.

[00:06:41] The Journey to Koop Insurance

[00:06:41] Audrow Nash: Now, how did you come to this problem? this seems like a super necessary thing, but as so I am primarily involved in the technology. I, I just recently bought a house and that involves like insurance and things like this umbrella insurance, but it's not like a thing I often think about.

so I would like, how, what is your path been? and how did you realize this is such a problem?

[00:07:09] Sergey Litvinenko: The path has been very non trivial, to say the least. so yeah, we originally started with an extreme focus on the autonomous vehicle space. And our original, the original problem that we're trying to solve had, didn't have to do anything with insurance. So when we started, we said, Hey, We made an assumption that autonomous vehicles are going to happen in the near future.

And, back in 2020, it was late 2019, early 2020. it seemed like it was going to happen. We had a five year horizon. And we said, look, autonomous vehicles, like the companies behind the autonomous vehicles, they operate, in like in a closed loop. Okay, so whatever

[00:07:51] Audrow Nash: do you mean? Oh,

[00:07:53] Sergey Litvinenko: So yeah, it's like whatever they data they produce, they use it for training and then they ship the updates. It doesn't really go anywhere else. Okay. And when we thought about, they have to interact with a lot of stakeholders, with the government, with insurance companies, with public, policy making organizations and the like.

And we're like, at some point there has to be a company that will make it very easy. To take the data from autonomous vehicle companies and put it into the hands of external stakeholders. So for example, let's say if you are an autonomous vehicle company and you try to report to NHTSA, the way it's done right

[00:08:29] Audrow Nash: What is NHTSA?

[00:08:31] Sergey Litvinenko: the national highway safety organization.

So it's the comp, it's the, it's the governmental body that, pretty much overlooks anything that has to do with traffic safety

[00:08:43] Audrow Nash: Yeah. Okay.

[00:08:45] Sergey Litvinenko: and, so for example, Autonomous vehicle companies, be those pure play level four robotaxis or partially automated, systems, they have to report to NHTSA. And there are certain requirements, there are certain fields of data that they have to show them.

And right now it's done in a very manual way. you've literally just pulled the data, a bunch of, Excel word files, you put it into the cloud and then you ship it to them. And we thought, what if there were a streamlined way where you can just, define the fields, click the button, and then we're going to build the API and pull all the fields, of course, with the permission of the company and do this like seamlessly.

That was the original idea. And one of the use cases that we had was, insurance and we said, wouldn't it be awesome for the insurance company who are trying to insure autonomous vehicles to just go to, a tool that we're going to sell them, click one button, and they're going to see all their data, the need for underwriting.

that was correct. But when we tried to sell that, it didn't work out. Why? Because insurance companies had two problems. Number one, they had no idea which data to ask, and number two, there were not enough autonomous vehicles on the road for them to make, commercial sense out of it. that's why they dismissed this.

And for us, it was a signal like, okay, what are we going to do? And then we made a decision, that, Hey, if we know which data we can pull and we know how to pull this data, and we, like our specialty was working with data in the first place, why don't we become an underwriter or an insurance company ourselves?

And that's, we'll make a decision to go from Koop to Koop insurance. And we went through all the hoops that you have to go over, like getting licenses, getting, relationships with like reinsurance companies and stuff like that, and we started underwriting autonomous vehicles ourselves. And shortly after we've done that.

We, we said, Hey, we can do the similar thing to other robotics use cases. we can do mobile robots. We can do drones. We can do automated vessels. we even insure submarines today, believe it or not, autonomous

[00:10:48] Audrow Nash: Super cool.

[00:10:49] Sergey Litvinenko: And, that's how we got to where we are today. And, when I say the path was non trivial is that when we originally started, we didn't envision that we would become like a I would like to brag about it, but I think we are the number one robotics underwriter in the US today, maybe in the world.

And,

[00:11:08] Audrow Nash: know other ones. So hi. Are there other ones?

[00:11:11] Sergey Litvinenko: we typically compete with like bigger insurance companies that

[00:11:16] Audrow Nash: Yeah, but they're so slow

[00:11:18] Sergey Litvinenko: robotics companies. But yeah, it's, not a systemic competition. Yeah. so yeah, that's how we got here. And then, we are a function of the robotics industry. So as long, as the robotics industry thrives, we're going to thrive.

And we're trying to do a really good job to make the life of robotics companies as easy as possible. Yes.

[00:11:38] Risk Assessment and Data Collection

[00:11:38] Audrow Nash: thing to be explicit about with the business model if I understand correctly, so you are underwriting which means that you are estimating the risk Overall for these companies so because basically the way that it works if I understand is you say for this many companies that we have, or this many robots deployed of this type, there's a risk of this much, like we have a risk of this much, and the payout for that is this much.

And so then what you have to do is you have to multiply all that together, basically, so you can try to get the expected risk for the whole fleet of robots. And then you have to charge just a little bit more than that, so that you probably need a buffer too. and then you have to have enough to pay yourselves.

And that's, the business model, correct?

[00:12:34] Sergey Litvinenko: You are correct. And you pretty much with what you described, you can possibly qualify to be, to be an insurance person. So think about it this way. It's what you just described is called frequency and severity. And that's how, that's what insurance comes down to. It doesn't matter whether it's robotics or anything else.

Frequency, what's the probability a bad thing is going to happen. Severity, if a bad thing happens, how bad is it going to be? Okay. Of course, a lot goes into both in terms of how much data and statistical analysis, but that's what it comes down to. And then there are a bunch of rules like underwriting rules.

It's okay, we're not going to touch companies that are too big or that are too small or that do international operations. But all of that is more like Rule based system that details exactly. yep, you're correct, and that's exactly what we do for robots. For every single company, when we evaluate them, we evaluate, potential frequency, potential severity, multiply one by the another, that gives you a base rate, and then on top, you make, the, percentage for the company to, to operate.

yep, you're 100 percent on the spot. And we, we, also buy reinsurance. what does it mean? for example, let's say you take frequency multiplied by severity. It ends up being, let's say, 10, 000 for a year. Okay? And then we, it includes some commercial considerations for Koop. And then what we would do, we would collect that premium.

And then we would go to reinsurance and say, Hey, we're going to give you a certain. percentage of what we, collected. So you can re insure us. And that means that, if, we, let's say if Koop is wrong with its underwriting model, and let's say we have a big loss, then they would cover the

[00:14:23] Audrow Nash: for insurance. They back you up.

[00:14:25] Sergey Litvinenko: And we do that because it's a, in general, just a good practice until you have a lot of money on balance sheet, you typically do reinsurance. So that's what we do. And then we're left with, whatever percentage is remaining as our, revenue, and then, it goes down the P and L from

[00:14:43] Audrow Nash: Yeah. I feel like it would be ironic if you didn't do reinsurance because it's like you don't leave in the process in some sense.

[00:14:50] Sergey Litvinenko: Yeah. Yeah.

[00:14:53] Audrow Nash: Okay, yeah, and so reinsurance in the event that something huge happens, and that payout is much higher than the severity you expected, so then it's too much cash from your perspective, you guys are insured by a bigger insurance company for this kind of thing.

Okay,

[00:15:13] Sergey Litvinenko: Yep. So that's how, it works. And, it's just a, it's a, it's an industry standard until you become too big to fail,

[00:15:22] Audrow Nash: yeah. Yeah, too big to fail. And then but one thing that is clever, I think is, and I remember from talking earlier, I think I remember from talking earlier, you guys are effectively a brokerage for insurance. Is it true? Or is it like underwriting? You're not, you're estimating the risk for some other guys who are insuring, them, or are you guys doing the full insurance yourself?

Or how does all that work?

[00:15:51] Sergey Litvinenko: Yep. So here's how it works. We act as an underwriter. So that means that when we have a customer, we'll look at their data. we do the pricing, we do the, rules, we apply rules to them, underwriting rules, and then, we also do the whole transaction from start to finish. We also are responsible for acquiring the customer and maintaining the relationship.

We are the underwriter. Now we not underwriting everything. We're underwriting only a part of the company. So what does it mean? We do a general liability, or it's also known as product liability, just for the company. We'll also do robotics errors and emissions for your fleet of robots that you operate as a service.

And we'll also do cyber liability. In case you, your robots get hacked or your company gets hacked. That's what we under, oh, and we also of course do commercial auto. So commercial auto it's, if you have, fully, or partially automated vehicles on the road in a commercial fleet, we underwrite that as well.

So that's what we underwrite. We also have customers who, let's say, come to Koop with other needs. We call them non core, non core needs. And that could be like directors and officers to protect your management. it could be employment, practices liability. It could be workers compensation liability.

This is something that we do not underwrite because it's a pretty straightforward how it's done. That's what we're going to go and partner with somebody else to take care of that. But it's going to be done through Koop. Okay, so in that sense, we are brokerage for whatever is not our specialty, and we underwrite the specialty lines of, lines of insurance, which are product liability, errors and omissions, cyber liability, and auto liability.

[00:17:37] Audrow Nash: awesome. Okay. And so with all this, so that, to me, it all is very clear now, which is great. How the main challenge that I see is now you need a very accurate way of assessing risk because you could be super conservative and you could charge your customers an arm and a leg.

[00:18:00] Challenges in New Robotics Domains

[00:18:00] Audrow Nash: but that's not helping the companies and you're not as competitive as you could be or you could be super minimal and if you're super minimal then you guys go out of business real soon because you don't have enough to do the payouts so how do you figure out a good risk and the payouts especially in these novel spaces like submarines and autonomous vehicles and all these kinds of things how do you figure this out

[00:18:26] Sergey Litvinenko: Yep. So we're pretty much built a model and, that gives us an estimate of how much we charge. And to check ourselves, we compare that model to the market benchmarks and we see how much, how different we are. And I can tell you that sometimes we are lower than the market, significantly lower than the market.

Because the data shows that it's a really good risk. And there is a good, both on a quantitative side in terms of what we collect from the fleet, and also on the qualitative side. Sometimes we have customers who we charge over the market because we've figured something out that they're not as great of a risk and we need to charge them more.

And, it's, what's. It's this, it's, how they are and it's up to them whether they want to accept that or not, so our model, there are two components to our model. number one, the quantitative component is that the API that I told you about, we use that API to collect data directly from different kinds of robots.

And when we collect the data, our methodology is based on understanding A behavior of a robot in a particular environment, and then we'll use the behavior as a proxy for risk, as a proxy for an accident that can lead to a financial loss, an accident that can lead to a claim. And the idea is very simple, actually.

The better the performance of the robot, the better the behavior, the lower the probability that something bad is going to happen.

And probably the most startling example that we have would be, Auto, so automobiles, we have a ocean of data when it comes to, human driving performance on the road and how it correlates to accidents, which I believe in 2023, there was 40, 000 deaths.

On, public roads in the U S and there were a million more accidents, that led to some injury or damage. And there was a lot of bad behavior on the road in general. So we have an ocean of data. So from the data that we collected from autonomous vehicle companies that we work with, and that includes different use cases on public roads, robo taxis, vans, trucks, shuttles, we're already seeing that autonomous vehicles are behaving up to 70 percent better than humans.

So what does it mean? What does it mean? That means that, there are, there is like a five to six common bad behavior examples that tend to lead to accidents. For example, excessive speeding, hard stopping, risky lane maneuvers, running the red light and stuff like that. we see that autonomous vehicles just don't do that. Even from the data that Koop has, 70 percent reduction in risky maneuvers, okay? I think it's going to be 99 percent eventually. And that's what's driving, the quantitative assessment that we have. That's what's driving the frequency. we have a way how we say, based on that. We can estimate that the frequency of an accident is going to be lower.

now that's on the qualitative side, on the qualitative side, every single company that we work with, we of course evaluate how good they are from the risk standpoint, and we have a certain, Technical questionnaire that we developed. We asked them about their software practices, what kind of hardware they have, what kind of safety methodologies that they use, what kind of compliance they have in place, and the companies that did more stuff, we assume that they're going to have, less of a chance that something bad is going to happen to the fleet or to the organization.

And that also goes into our underwriting. So once we do both quantitative and qualitative, combine it together and everything comes down to frequency and severity, and we have the rate. We look, okay. This is how much we're going to charge them and this is the market benchmark and we have the access to market benchmarks.

And most of the time we are charging less than the market because we have the methodology that says that this is a better risk. But most of the time, by most of the time, 85 percent of the time. There are also 15 percent of the time we see companies that could do a better job and we also provide them with recommendations what they can do.

But unfortunately we have to overcharge them because they are. A riskier company than the market thinks.

[00:22:39] Audrow Nash: Yep. And you have to do the smart risk estimation. And so that makes sense.

[00:22:43] Sergey Litvinenko: yep. And then if there isn't any other insurance company that wants to take them, sure. But we figured that they are riskier than people believe. And maybe we shouldn't insure them, because the probability of having a bad accident is higher than everybody else.

[00:22:57] Audrow Nash: Ah, okay. That's pretty cool. So that makes good sense to me for autonomous vehicles because it's cool because you, can become domain experts to some degree, and you can look at behaviors and you can use that to estimate probability of accidents. So you can look at these indicators you said, stopping too quickly, reckless lane changing, whatever.

And you can see that from the data. If they're doing that, and then you can say, okay, we're never doing that, thus we're not creating a bunch of preconditions for accidents, and then we're not having as many accidents. Okay, so that's great. how do you do this for Unknown areas like submarines or robots in hotels or things like this like how do you because the thing is what strikes me is with this we have the Highway Association you were mentioning and they have and the insurance industry is really interested in figuring out how to What in the vehicle space, not just autonomous, but people are risk indicators so they can insure more appropriately.

but how do you do this in a new space that doesn't have such a large amount of established information about what creates risk?

[00:24:15] Sergey Litvinenko: Yeah, there is a, trick that we do and you're completely right that if you have a mobile robot, that transports, let's say, goods or materials in a warehouse and it's not exposed to a lot of people and there was no public data in terms of what's happening with within those warehouses, what we're going to do, or submarine or a drone, So we, we apply the same methodology. So we'll still collect the data on, it's not going to be called. Bad behavior, we'll call it bad behavior, but we call it like a, what's the troubleshooting rate? How many times a vehicle goes offline or has to do a minimal risk maneuver or has to reach a minimal risk condition?

And every single company has a different definition of that, and it's our job to standardize that. But it's pretty much. Like, how often your robot goes offline because it can, cannot do something. That's the, proxy that we take. And then we also evaluate the environment. And then this says, okay, if there is an environment where there's literally nobody around, what does it mean?

That means that we cannot injure anybody and we cannot damage anything. Okay, now what's the worst case scenario that you can think of? the worst case scenario is if you have a company relying on your drone for a service, if that's, if your drone goes, goes offline, then, You cause business interruption.

Okay. And that means that the company is going to lose money and there is a way how to evaluate that. So when, we do that, we realize that, Hey, in those, simpler environments where the robots are not, let's say, carrying any people as passengers or not driving with other people on the road, we would take the troubleshooting rate and we would just assume what the worst case scenario could be.

And most of the time it's business interruption. And it's business interruption up to a certain dollar limit, let's say a million dollars, two million dollars, or three million dollars, depending on what we put in the policy. And then we say, there is a 0. 0001 percent chance that this robot is going to go like completely offline and not recover.

And it's going to be like the redundant systems are also going to fail. And it's going to be, let's say a 2 million loss. We'll multiply one by another. And of course, it's not as good of an assumption that we make because there is less data. But that's how we do it. and then the second challenge is there are no market benchmarks, meaning I don't know how people charge for robots in warehouses, it just doesn't exist.

So we cannot compare that against the market, but typically we can say that, hey, if you have a robot that, let's say, if you have a fleet of robots that costs a million dollars and you are paying, let's say, 10, 000 to insure that on a yearly basis, The economics make sense to both the producer or the developer of the robot and the customer.

So we usually use customers as like the sanity check, whether that, that, that price makes sense. And of course we reserve the right to do adjustments to the rate that we charge based on the underwriting assessment on the market conditions, but we do that rarely. most of the time it's just programmatic.

so yeah, that's how we do it. And I got to tell you, it's the same methodology. It's more assumptions that go into it, but at the end of the day, it does make economic sense to the customers. So that's the most important thing.

[00:27:31] Audrow Nash: Yeah. And also, if you are making more assumptions, you are effectively not. You're having more risk because of those assumptions, but you can charge a higher margin, I imagine, and that, that's probably how you make up for these riskier environments. So if it's autonomous vehicles, maybe, I'm just picking numbers and I have no idea, maybe you charge 1.

1. Times are 10 percent more than you need to because that's your margin because lots of them are going and you it's a volume thing in a sense. Whereas if you're doing warehouse robotics for a new robotics company, maybe it's like 1. 5 or something like that. And that's because of the risk that you can't estimate and you give yourself a buffer for that kind of thing.

Is that, kind of correct or what do you think?

[00:28:22] Sergey Litvinenko: it is, it's actually more, more manual than you think. So the equation that we have, it spits out the rates, the tech, we'll call it the technical rate. So it's literally frequency multiplied by severity. And then there is a certain percentage that we put on top as our margin, like you said, but it's not dependent on a customer.

It's the same for everybody. And, and then. If and then there is an underwriting judgment involved and we have underwriters on staff and they say let's say if you insure an autonomous submarine that does, let's say, R& D or something else, and the loss of that submarine could be really bad for the customer that did use that submarine, you can, you are able to add a certain percentage on top as a margin of safety,

[00:29:14] Audrow Nash: It comes in at a different stage, basically.

[00:29:17] Sergey Litvinenko: it's not, we, do not have a model that detects It dictates that.

It's more rules based. Okay? So if you have a rule that you have a very tough robotics use case to ensure, that deserves underwriting review, and then the underwriter can decide to increase the margin of safety, but the, it's not baked into the equation. It's more rules based.

[00:29:37] Audrow Nash: Okay. So tell me more about the rules based stuff. what it is. So to me, initially they seemed like just details that may not be that interesting, but I'm starting to think of it. it's like an algorithm that you guys go through in order to It's it's too complex to make a model. So they're like simple heuristics that you follow, in order to make decisions that are around risk.

Is that how to think about the rules that you may have, or tell me more about the rules?

[00:30:12] Sergey Litvinenko: Yep, so the rules are the heuristics that you have to follow and this is done by a human being. So for example, let's say you have a robotics company and that robotics company, so what are the rules that we apply to them? geographical locations, you have to operate within a certain geography.

Right now it's the U. S. We are unable to do that.

[00:30:35] Audrow Nash: Too outside of it,

[00:30:37] Sergey Litvinenko: exactly, number one. Number two, for certain use cases, we have heuristics, whether you go on public roads, you do not go on public roads, that actually makes a huge difference

[00:30:46] Audrow Nash: I would imagine. Makes

[00:30:49] Sergey Litvinenko: insurance, we'll have the heuristics for, the management team.

what does it mean? That's really important that you have a person dedicated to either, safety or compliance on staff, and that's, it makes a big difference, not just, from the checklist standpoint, but also from the. Company standpoint, it's like, if you have a person dedicated to this, we have more trust that you will do the right thing when a bad thing happens.

And, sometimes, we saw examples with some big autonomous vehicle companies when they didn't go, right. we have also rules based around, just some financial metrics of the company, because that's what we look at. and, It could be, when in our, we have an internal system underwriting system and that internal system spits out the rate, and then it also shows you certain pattern changes with the account and it says, The company, let's say, tripled in size, or it tripled in size while reducing the headcount, or it did not grow, but the headcount grew by 50%.

We have those certain, they're relatively simple, but those certain metrics, when they, when the underwriter looks at them, it gives an underwriter an idea whether something is off. And typically, if you are, that gives an underwriter an ability to go. And either call a customer or have a conversation, maybe over email and ask them about whether everything is all right, because don't forget the tendency when you purchase insurance is that, I don't want to show you any data, just give me the quote, and there is always the fight between how much data we can get and how good of a quote we can give.

That's why we built a system that spots the pattern changes that don't have to do anything with robotics, more with the company. And if there is a pattern change, it activates a certain rule that you have to go to check. to see whether the data provided was incorrect, inconsistent, or something else. And those rules are typically done not for every single company, but they're typically done for the tougher use cases where we lack some data.

And that's what underwriters do manually.

[00:32:54] Audrow Nash: very interesting.

[00:32:56] Training and Methodology for Underwriters

[00:32:56] Audrow Nash: Do you have to, so I imagine that there are not many robotics underwriters existing to, like the vast percentage of underwriters are, I don't know, regular underwriters that work at a regular insurance company that have a standard established processes. How is it hard to find underwriters or do they require a lot of additional training or is it just you have your very clear process put in place so you can just drop any underwriter in to your system and then they start functioning in your system or how does it work to actually train underwriters?

Definitely.

[00:33:37] Sergey Litvinenko: great question. I don't think anybody actually asked me that question. So great. You are right. If you look at the world of insurance, you can find like all kinds of underwriters, but there is not a single underwriter that does robots. And that's, that's, what we're trying to solve.

We're trying to become that underwriter. So what we've done is that, when we hire underwriters, we try to hire with some background. And that background usually has to do either with traditional. Product Liability Insurance or Traditional Errors and Omissions Insurance or Cyber Liability.

Understanding of the insurance side of things. And then we train them on robotics specifically, on all kinds of use cases, on the methodology, on what to look out for, and also on the system that we developed internally. So there is, we are required to have Underwriters with established industry background, and then we train them on top.

So that's, how it works. there is no other way around it. And I do think that we are, in terms of the customers and data that we have with the insights, we are the only true robotics underwriter in the world.

[00:34:50] Audrow Nash: Yeah, I wouldn't be surprised if that's true.

Very interesting. how, difficult is the training for the underwriters? Like how, quickly can you get them up to speed assuming they have a reasonable background?

[00:35:05] Sergey Litvinenko: under a month, I would even say under three weeks, it's like a fire drill, you, our met, when you implement the methodology, the way we have it implemented, it looks relatively simple. Okay. So it's okay, here are all the buttons you need to click here, all the numbers you need to go through here, all the rules that you need to complete.

but in order to do that, our methodology is like literally 200 pages worth of just writing. Stuff, right? And then we implemented that methodology. we, pretty much onboard people onto those 200 pages in a very user friendly way in less than three weeks. And the way we do this is just, by just, just it's a part of the employee training that we do.

And we just have a process where in a very short period of time, we're going to onboard you to everything. And then as you start working on customers and you're going to have any issues or that will have troubleshoot, we're just going to do it on the fly. we don't have the luxury of spending three months to do training, unfortunately.

So we have to do it really, fast. And we built and we built for it.

[00:36:09] Audrow Nash: Yeah, and you mentioned, I might not remember at the beginning, did you say you're like 10 people or something like this? And I assume then you have a staff of underwriters for this kind of thing, or is this 15 people? Does that include underwriters?

[00:36:24] Sergey Litvinenko: Yes.

[00:36:26] Audrow Nash: How many of those are underwriters?

[00:36:28] Sergey Litvinenko: We have two of them right now.

[00:36:29] Audrow Nash: Oh, okay. Very interesting.

are most of your customer, where, are, so you have a lot of different domains that customers are in. Can you just tell me a bit about the different domains that you're serving, servicing, and What would be the rough percentages for which domain? are you mostly autonomous vehicles or,

how does it change depending on different domains?

[00:36:56] Sergey Litvinenko: Yep. So I can give you by the robot type I can also give you by the industry

[00:37:03] Audrow Nash: Both are very interesting.

[00:37:04] Sergey Litvinenko: yep. Let's start with the industry first. So we have Transportation or like ride sharing which would be the robotaxis we have Holing goods or tracking.

[00:37:16] Audrow Nash: Can you say the companies or, like I'd be curious to hear who the companies are if it's allowed to be exposed.

[00:37:23] Sergey Litvinenko: yep. I can tell you the ones that we have on the website. Those are the, we have the permissions for. we work with a Honda and Kia on the robot

[00:37:32] Audrow Nash: Makes sense.

[00:37:33] Sergey Litvinenko: and, on the tracking, I unfortunately cannot tell you, but it's one of the, it's only a handful of players, so it's one of them. we.

Also, there is also, construction robotics. there is, mobile robots, not it's, not just the, mobile robots that go into warehousing specifically. Okay. So let's call it warehousing. we also have the medical space, which is by the way, really tough for the robotics, but we'll still have that.

if there is a, aerial use cases, so aerial transportation it's drones, done for, either transporting goods or doing aerial surveys or doing any kind of data collection. There are marine use cases primarily used for research and development purposes, and that has to do with kind of on surface vessels and also subsurface or the submarines.

We also have boring machines, the machines that actually Dig Tunnels, there is a huge human component to that as well, but it's still considered to be a robot and we ensure that as well. that's just on a high level, I have a very detailed breakdown in, in, in our financials, but just to give you on a high level, and, in terms of the robot types, that's, Everything when it comes to the new generation of robots.

So that would be, robotized vehicles, public vehicles, mobile robots, robotic arms that do all kinds of manipulation with different number of degrees of freedom. Some are super simple, some are super complex. we have robots that do, like moving pallets in open field for agricultural use cases, construction robots that do all kinds of, paintings on the roads or all kinds of fencing or any kind of welding.

That would be security robots that just go around the perimeter and they are not armed, okay, but they go around the perimeter collecting data in the real time about any potential intrusion or any potential suspicious behavior. The same applies to drones.

[00:39:47] Audrow Nash: Cobalt or something like this

[00:39:49] Sergey Litvinenko: Yeah, so one thing that we don't have in our portfolio, is like the old generation of robots.

For example, if you are, an automotive manufacturer and you have the automated assembly line and you have those KUKA robots, we do not do that primarily because it's not just the technical consideration. It's also more of a financial consideration. Companies that buy those robots, they have so much money that if something goes wrong, they can just pay for that.

They do not necessarily buy insurance. That's why we deal with a newer generation of robots where there is not a lot of precedent, and they're also in a less constrained environment with more degrees of freedom out there. that's what we are focused on. And probably the two startling examples would be Robotech system, public roads, and also humanoids when we get those like for Koop, that's going to be, that's our thing.

We're going to ensure that a day in, day out.

[00:40:44] Audrow Nash: Why?

[00:40:46] Sergey Litvinenko: I think. For first for economic considerations, we do believe that the next gen robots are going to be the future and it's a chicken in the neck problem. do you first ship the robot and then you do insurance? But you need the insurance to ship the robot in the first place because your partners, customers, and people would want to have that.

So that's what we're trying to solve, the economic transaction. And two, our own. selfish motivation is that we want to be the first mover. We want to grab the market. We want to have the most data on that. And it's, the best way to do this is, just to do this at the time when everything happens.

So that's what we're, planning on.

[00:41:29] Audrow Nash: Yeah. So you're going to be there, right? you'll be the early bird for this kind of thing.

[00:41:34] Sergey Litvinenko: exactly.

[00:41:35] Audrow Nash: Awesome. And, but why humanoids? Why are they especially interesting for this?

[00:41:41] Sergey Litvinenko: I'm honestly, so we are a function of the market. So if there is a new use case that appears, like the first thing that we think about, we want to see how it plays out. Okay. In order to see that we need to ensure that I think the reason why I mentioned that is because there is just so much news about humanoids and I'm like, okay, what's going on?

and then we figured that there are like 10 companies in the U S and I think 20 companies in China that are working on humanoids, they're at a demo stage, but the, progress has been, pretty tremendous over the last year and a half, I would say. and we're thinking, when the first humanoids start to ship to customers or to the first companies that are going to be adopted at, Like what insurance applications are going to be, how we can pull the data from them to understand their performance.

Okay. And apply our methodology. And how can we save money to both the developers and to the customers? That's what I'm thinking. And since humanoids are all over my Twitter or X feed, I'm like, that's, I just cannot stop thinking about it. that's why I'm really curious to see what this is going to go.

[00:42:51] Audrow Nash: Okay. That's very funny. So yeah. Yeah. It'll be really interesting to see where it goes and what the timeline is. Another question that I have with this, how do you handle software updates? if I am a robotics company and I am shipping a new version, but there's definitely precedent for a new version just completely bricking the robot in its behavior, or fundamentally changing the behavior and thus your past estimates are now broken.

different than what the or your past estimates are assuming different behavior than the current behavior. How do you deal with robotics being more iterative? Or do you only deal with robots that have locked down behavior? Or is part of your risk estimation how the companies are doing testing and their rollout of updates?

Ah,

[00:43:52] Sergey Litvinenko: very tough question, but I have a simple answer for that. So here's what we'll do. The, the quantitative data that we track, which will be the performance of the behavior of the robot, the troubleshooting rate that we use for estimates. we're going to put it on a timeline. So it's going to be a time series, okay, of the rate, of the performance rate.

And we're going to layer the performance versions on top of that. And then what we want to see is that the more software updates you ship, the lower the troubleshooting rate becomes, or the higher the performance becomes. That's what we're looking for. We are not too surgical about it because it's very hard.

Like we're going to cross the border into safety. which we do not do, we do not want to, or, into the capabilities of the robot, which we do not want to do. if there is a fundamental change, that you, that the robot or the fundamental change in the behavioral model that you ship there, we want to see the improvement no matter what.

Okay. And the only way to see the improvement is that when the robot goes into the real world environment, that the troubleshooting rate goes down. that's what we're looking for. If we don't see that, Then that means that the software update was maybe not good. Maybe it was something else. We don't know.

But that's what we're looking for.

[00:45:11] Audrow Nash: Yeah, but so what I'm thinking is like someone ships an update of how the robot handles. It could even be just a bug. say, the robot can't reliably like the camera images the robot has are off by one index now. And so what is the whole image like every row is really confusing and you get like these weird stripes across the image.

Now the robot can't do perception. and it. Bricks the robots effectively, but they don't realize this until they put them into production, and this camera only turns on in a specific part of the process. And so then it fails here. I'm sure new updates break things occasionally. how, do you handle this kind of thing?

Or, maybe, you just expect strict improving performance. but I would expect that bugs occasionally happen. That our performance breaking even if the company has good processes in place.

[00:46:18] Sergey Litvinenko: the way to catch that, fortunately or not, is through the qualitative evaluation. when it comes to the real world data, we will, if there is a bug that gets shipped to production, we're going to see that, and unfortunately it's going to affect the performance and the evaluation on our end. The way to prevent that is, okay, how do you stage the release?

What is the company's procedure to pushing to prod? And, we, by the way, do ask questions related to that. it's okay, what kind of testing procedures do you have? how much, data do you have to collect? Or performance data on the robot to make sure that it's good to go? And, there's a bunch of other additional questions.

And,

[00:47:04] Audrow Nash: So you fall back on qualitative things,

[00:47:07] Sergey Litvinenko: yes.

[00:47:08] Audrow Nash: because it's a very hard judgment to make.

[00:47:12] Insurance Validity and Performance Guarantees

[00:47:12] Audrow Nash: what, so what I was, what I'm getting at with this is if there's a large performance break in the robot, does it make it so that it invalidates your insurance? because they're not living up to the estimates that you based your policy on or something like that.

if the robot does something completely crazy because of a recent bug that was not there when you insured them, how is that handled?

[00:47:47] Sergey Litvinenko: so we, what I think you are implying, something that's called as a performance guarantee. let's say you have a cloud service and you guarantee that 99. 99 percent of the time it's going to be up and then for 0. 01%, there's going to be a performance guarantee that we're going to pay you out certain amount of money, blah, blah, blah.

It's called It's parametric insurance. there is a very clear parameter, and once you go out of the bound, that's, you're gonna get a payout. We do not do that. there is no way that a certain bug or a sudden decrease in performance is going to invalidate our insurance. We do not do parametric insurance.

Yeah, at least, we haven't found a use case for robotics for it for now.

[00:48:33] Audrow Nash: gotcha.

[00:48:34] Handling Erroneous Data and Misrepresentation

[00:48:34] Sergey Litvinenko: And, yep, so if, let's say, the only thing that can invalidate insurance if the, the erroneous, not the erroneous, but the purposefully, skewed data or purposefully incorrect data that you submitted as a part of the process.

But when it comes to the troubleshooting, it's very hard to, provide bad data or API because that comes directly from, the, robots or from the fleet. qualitative, yes, you can, misrepresent things. And that's why. when, you mentioned, how we catch that, yes, we have to catch that through understanding what steps the company has in order to get to prod.

If they said that they had the steps, but they didn't have those steps, and during the investigation after the claim, we figured that out, yeah, then we can invalidate that. The insurance and say, you're not getting paid because you lied.

[00:49:28] Audrow Nash: You misrepresent. Yeah.

[00:49:29] Sergey Litvinenko: yeah. And what's, what's funny. It has nothing to do with robots.

It has to do with humans. The fact that you lie on the application. but that's, that's the business that we're in. we have to deal with this, things. And fortunately we haven't had a situation like that. I think the way we run the process and the way we market ourselves is that.

People are trustworthy and they understand that the worst thing that can happen is that when they need to pay a million dollars in liability and their insurance becomes invalidated, that's literally the worst thing that you can, that can happen. And they don't want to do that.

[00:49:59] Audrow Nash: Yeah. Cause that's what I think of. I haven't had any negative experiences myself, but I've heard things where like insurance, it's right when you need the claim, they're like, Nope, it wasn't validated for some reason, whatever. And, I could imagine that A robot shipping a bug and just destroying its own performance because of that bug, hopefully rare, but I could imagine that would be one of those cases that you would want to terminate because it's I don't know, say it does something just wonky and it's different than what you've expected with all your behavior analysis.

so that it's almost just, it's justifiable to me in some way, but hopefully companies are more reliable with what they ship to production.

[00:50:42] Sergey Litvinenko: I can add one more thing here.

[00:50:44] Human vs. Robot Behavior in Insurance

[00:50:44] Sergey Litvinenko: A, what the beauty about robotics is exactly what you just said. If there is a bug, then we expect that this bug is going to be fixed, and this bug is never going to happen again, okay? Very hard with humans, because it's very hard to change human behavior,

[00:51:00] Audrow Nash: Oh, that's super interesting.

Yes.

[00:51:04] Sergey Litvinenko: have a fleet of robots that's going to be continuously and probably exponentially improving, with humans, impossible. That's why our side of insurance, the robotics insurance, is going to be so rapidly evolving up to the point that it's going to be Super easy, super economical, and super user friendly, because we can control the robotics behavior, one way or another.

With humans, very hard to do.

[00:51:25] Audrow Nash: is hilarious. That is such a thing. because humans, so humans, we know how to drive, but we keep getting in

[00:51:35] Sergey Litvinenko: we know.

[00:51:36] Audrow Nash: but, we literally are able to get in our car and this kind of thing, but then people are reckless, they're on their phone, they're drinking a big slushy, whatever it might be, and then they get in accidents.

And you can't change that, In the whole population, but with robots, you can hit a bug once. And so if you hit the bug once, then everyone knows about this. Then you make a good fix for it. The fix is robust, hopefully. And then from there, that failure case is not a thing. And that actually makes it seem like this is like an insurance person's dream, in a

[00:52:16] Sergey Litvinenko: Yes, it is. It absolutely is. That's why we are, that's why we're doing this.

[00:52:20] Robotics Insurance Metrics and Losses

[00:52:20] Sergey Litvinenko: And I got to tell you our, there are two performance metrics that we have. One performance metrics is of course, how many customers we have and how much business we've produced. The other performance metrics, which is equally important, is how many losses we had.

And I got to tell

[00:52:34] Audrow Nash: they keep going down

[00:52:35] Sergey Litvinenko: yeah, I got to tell you, like the, If you look at property and casualty industry, which is like casualty means, third party liability and property is property. Industry average is around 70 to 75 percent. of, losses paid out. So what does it mean? Out of every dollar that insurance company collects, you pay out 70 cents. Guess how much our payout is so far?

[00:53:02] Audrow Nash: 20 cents. Five.

[00:53:05] Sergey Litvinenko: Five.

[00:53:06] Audrow Nash: Holy cow.

[00:53:08] Sergey Litvinenko: that's, how much better it is. Even like we're an early stage company, we don't have a ton of data yet, but from the data that we have, it's already that much better. And it's, literally printing, underwriting profits. And what we're going to do is that we're going to take that and we're going to, do the iteration of the models that we have and we're going to start bringing the rates down.

So it's going to be even cheaper to get our insurance. that's the magnitude that I'm talking about, and that gets me really excited, about this

[00:53:37] Audrow Nash: That is exciting. Yeah, super, super exciting. You're in a, you're in a hot, space for this kind of thing, I think. and it's very interesting that conventional insurance risk stays about the same over time. and yours decreases as performance improves.

That's a fundamentally different dynamic, but you still get this old industry, and people do feel better with insurance.

They sleep better, and you know you're not going out of business tomorrow because of some mishap. but it's just like you're on the right side of this dynamic, which is super, super cool. What a thing. Yeah, that point about, the robot only does it once, or you run into this bug once, that's amazing.

[00:54:26] Evaluating AI and Edge Cases in Robotics

[00:54:26] Audrow Nash: related to this, how do you, so machine learning Deep Learning, and all of these, all the AI related stuff. it often has very good performance for most cases, but then you get these weird edge cases. occasionally when the conditions are Outside of the training data, for example, how, so it's similar to the bug question, but maybe a little bit different where the behavior might be hard to evaluate.

You can look at it from a completely external way where you just see that out behavior, how, any thoughts on evaluating risk in AI systems, deep learning, using, I don't know, deep learning for perception or control or anything like this.

[00:55:24] Sergey Litvinenko: we do not interfere to that level. so we do not collect any metrics on how, let's say, the perception algorithm performs, or let's say how many Subjects that classify correctly or incorrectly, or how, or like we do not evaluate the quality of the path planning, we do not do that. And primarily because we believe it's the job of the engineering team of the safety team.

And we like our thing is that we. We want to see how the robot interacts with the real world. And there are certain metrics that we can use. They are high level metrics, but they are we believe they are good predictors of the accidents that can happen. So for example, the troubleshooting rate, how many times the robot cannot, does know what it has to do.

It can come from any part of the stack. Okay, but we do not, we do not have an instant idea where it's exactly coming from. usually what we would do is that we would, follow up with the company and collect additional data in terms of the distribution that they're seeing on, the, what's failing the most, but it's not happening in the real time.

And it's just, just economic, practically, It's very hard to go to that level. And I know there are companies who try to do that. I don't think it worked out. It's very hard to sell an insurance product that's literally going to evaluate your whole engineering stack. Yep. So that's why we are, have to deal with a bunch of high level metrics.

Okay. but I can tell you the way we deal with edge cases is in the insurance, there are two ways to be found. first of all, you cannot predict in edge case. Okay. It's just no way it's possible. even if you look at the performance, you can assume a drastic change in the performance, but it's.

It's, it's just an assumption. Okay, we don't know how good it is. One thing you can do though, is that if you say that an edge case, you, first of all, you don't know what the edge case is going to happen, so the frequency component of it, you throw it out, we don't know. But what we can do, we can, Approximate, what edge case would look like on the severity side of things.

And we can say, if an edge case is something that we never thought would happen, and the worst case that can happen is that it costs us a fortune. So let's say that one, once, like once in a while, there's going to be an accident that's going to cost us a ton of money and we're going to call it an edge case, whatever it's going to be.

And then it's going to be like max out all the possible severity at 10 million loss. Okay. Or 10 million lawsuit or something like that. And then, we're like, okay, if we assume that. or incorporate that into our model, what is it going to do to it? And then, we do see the rates that they do change, but not significantly because it's more likely, most likely going to be a rare event at the end of the day.

that's one way we can do this on the model side. On the qualitative side, that's probably where we can do. The most have the most impact on our customers, because think about it this way. it's not just about the edge case happening. It's about what are you going to do right after it happens. so for example, there was a public example with one AV company in San Francisco.

It was like an insane edge case. you probably know what I'm talking about, right? And,

[00:58:36] Audrow Nash: the specific instance. I assume it's something with, Cruise.

[00:58:41] Sergey Litvinenko: yeah, exactly.

[00:58:42] Audrow Nash: Was it, they pulled out of San Francisco. I just, there were so many little mess ups. was it like blocking police officers or something? Or,

[00:58:53] Sergey Litvinenko: it was a situation when. So a cruise car was just driving by and there was a human driver. the human driver hit a pedestrian who was running in the wrong place. And then after the human car hit the pedestrian, it fell onto the cruise vehicle and then the cruise vehicle went over the person,

[00:59:14] Audrow Nash: oh,

[00:59:15] Sergey Litvinenko: remember that one?

Yeah, it's it's an edge case of edge cases. Cruise had nothing to do with it. Just a person got hit by a human driver first and fell into the cruiser's lap. And, then, the vehicle apparently didn't notice that there was a person underneath the vehicle and it dragged that person for 20 feet to the curb.

[00:59:33] Audrow Nash: Oh, yeah. I remember hearing about that. huh. It's

[00:59:35] Sergey Litvinenko: That's an edge case of edge cases, if you think about it.

[00:59:38] Audrow Nash: horrifying.

[00:59:39] Sergey Litvinenko: and the thing is, I think it's not the fact that it happened, but it's the fact that, how the cruises the company handled it, which led to the, to pull it out of San Francisco, there was like the governance issue, the communication issue.

that's what led to it. And when we look at robotics companies, it's really important that, we evaluate for a, fire department, or is there, do you have a literally fire department at a robotics company is going to handle that particular issue? And believe it or not, most of the companies Do not have that.

because they focus on building stuff and shipping the stuff. They're not focusing about the financial downside, which God knows if it's ever going to happen. And understandably, that's our job, by the way. And, what we found is that, the best way to deal with edge cases is to make sure that whenever it happens, that you do not do stupid stuff and you do not destroy your reputation and you do not get dragged into the courtroom.

[01:00:39] Audrow Nash: Yeah.

[01:00:40] Sergey Litvinenko: you need to have specific things that you need to do.

[01:00:42] Audrow Nash: okay. So tell me more about that.

[01:00:46] Sergey Litvinenko: So pretty much specific things, we, we do not do official training, but we provide a certain recommendations, but pretty much when an edge case happens, you have to have a first responder team within your company is going to get deployed and collect all the data. Number one, number two, you block communications for a certain period of time.

And number three, if you have to report to the government or to any stakeholders, you have to do this right away. And you have to be very explicit with your actions to say, Hey, Here's what happened. We collected all the data. We tried to mitigate that it doesn't spread or whatever. we talked to the government.

We notified them. This is exactly what happened. And trust me, most of the time, it's going to, in the end, it's going to be okay. Most of the time. Funky things happen when you want to hide the edge case, or you want to hide the fact that you made a mistake. And it's, such a soft thing. It's such a non engineering thing, but it can blow up a lot of things.

And we just provide recommendations to robotics companies that, Hey, Oh, by the way, if, if you don't know how to do this, or if you don't have a way, we at Koop can help you. But it's like an extra service that we of course provide on top of our insurance.

[01:01:53] Audrow Nash: Yeah. Makes sense. And it makes it so they're a better customer for you and you get paid to help them set this up and it potentially saves them if something does occur. So that, seems like win-win to me, as an initiative and having those, yeah. 'cause then that, makes 'em much better in your qualitative, analysis.

And then. In the event that something like this does occur, they're doing significant damage mitigation because this is in place, and that makes them not tank as significantly as you would, as they would if they handle it badly and lie or cover up or anything like this.

[01:02:40] Sergey Litvinenko: Just one example, very simple. We haven't experienced that. We haven't insured that. Very simple example. Let's say you have two mobile robot developers. Both of them do, let's say, materials handling or goods handling within a warehouse. Think of Amazon, right? Like they have what Kiva robots?

[01:02:57] Audrow Nash: They have

[01:02:58] Sergey Litvinenko: two, two similar examples.

Let's say you have one company. They had some nasty edge case for whatever reason, the whole warehouse stopped and, it was, Poorly handled, poorly communicated. There was no dispatch that took care of that. And the news got out and there was like two, three articles out there that bashed this company for how poor of a job they did.

And then when you have customers coming, customer is going to be like, Oh, you guys had this bad accident. I don't want to work with You what if that happens to me? I'm going to go work with these guys. So that's why it's going to haunt you down. If you don't take care of it, it's going to haunt you down eventually, and it's going to cost you money.

So we advise our customers is that, Hey, take care of this right now. You're going to thank yourself 10 years

[01:03:43] Audrow Nash: Oh, definitely. 100%. And I also, I think this is this is, it's just a good exercise that you guys are bringing into their perception, in a sense, because they should be considering the worst case scenario. Scenario, and making appropriate plans. Like I, to me, that just seems wise for this kind of thing.

It's like in your personal life, discussing the event of a death of yourself or a loved one or something like this can make everything so much easier in the unlikely, hopefully unlikely event that it does occur. Gotcha. Very interesting. let's see.

[01:04:22] Data Collection and Utilization in Robotics

[01:04:22] Audrow Nash: So now backing out a little bit, tell me about how you guys.

are gathering this behavioral information. What does your API look like? What does it require of the robot? what kinds of information are you typically transmitting?

[01:04:42] Sergey Litvinenko: we collect, typically we collect data from, the customer's cloud, we do the integration with the cloud, and we collect data that they already have. This is really important. We're not trying to collect, new types of data because, from the engineering standpoint it takes resources. But there are two categories of data that we collect.

Number one, utilization data. So we want to know where the robots are, how much time or how much mileage they go for. And, what's the general, time online that they are out there. It will also collect the telemetry data. So the telemetry, data means how fast they, how fast they go.

what's the, where are they turning, where they're not turning? what are the, if there are any predefined routes, we also want to know that, but it's typically done over qualitative evaluation and, any hard stops or hard accelerations over a certain G, G means, if they, yep, exactly.

And, we also collect the, that's where probably the most challenging part, the field that I mentioned about the minimal risk maneuver, like going offline, the troubleshooting rate. we also collect that, but every single customer has a different definition of that. And for some people, it might be that, if it takes, let's say, More than, let's say, I don't know, five seconds to, pave your, path that is considered to be a bad event.

Some people, or like a behavioral event. Some people say, if, my safety zone buffer gets violated and I get another person or a car too close to me, it's going to be considered a risky event. Or some people say it could be that the vehicle. Doesn't know what to do and it has to like literally stop and ask for help and we have to Standardize those things within our model and that's what we deal with and you know at the end of the day We take the performance.

We take the telemetry. We kind of layer that on top of the utilization so we know how much vehicles are out there and how they're performing and Then we layer a bunch of other things like software updates and the qualitative evaluation And that's what we do Gets the, gets the underwriting done, but yeah, fundamentally, those are the two key categories of data that we'll pull.

[01:07:03] Audrow Nash: It's interesting to me. It feels like for the telemetry data, it feels like you're treating other domains of robotics like autonomous vehicles, but maybe it's, autonomous vehicles like autonomous cars. is that, Did you find that a lot of those risk factors translate over to other domains, or you have the infrastructure built out?

Or how do you think of it?

[01:07:30] Sergey Litvinenko: Yeah, so we, the model that were built originally for autonomous vehicles, we have reasons to believe that it is translatable to other use cases, because at the end of the day, our job is to understand the performance of the robot. And that performance, how it will correlate to the eventual potential losses.

And, for example, with autonomous vehicles, if let's say you have a, a behavioral event or a risky event, according to our definition that happens every, let's say a hundred thousand miles, right? we will take that and then we will assume. That there's going to be a certain, accident rate coming out of that, which could be, 10 times or a hundred times.

It's going to, it's, depends on the, customer and the data. Similar thing with, let's say, mobile robots in closed environments, or let's say with drones. okay, you have one, behavioral event or one troubleshooting event per 100, 000 hours worth of operation, or, we can do this on the mileage or we can do flight hours.

So there is usually some utilization expressed in mileage or in hours or a number of tasks performed. And there is some performance metric, whether you go offline or you need help, or you need to be tele operated, something like this. And, we trying to do a really good job with categorizing those, there is a category of.

of mobile robots, category of drones, categories of vessels, and then within each, we're collecting data and see how it performs. But it's the same model that applies to all. And that's, that's our novelty. That's our IP. We believe we have a model that applies to all kinds of robots, and we can underwrite on the back of that.

[01:09:24] Audrow Nash: hell yeah. So your model, the way I'm thinking of it, is you have your model. Your model has different parameterizations depending on what domain you're in. Is that fair?

[01:09:36] Sergey Litvinenko: That's fair.

[01:09:37] Audrow Nash: Okay, that's pretty cool. And what is it, is it too simplistic to think of your model as the, you're basically doing mean, so over, super oversimplified, you're doing, you're trying to figure out the expected Value or the mean time to failure, say, and then you have an idea in the domain about how costly a failure might be.

And obviously you probably have a distribution and you're handling this in distributions, but is it, is it a fair way to think of it?

[01:10:12] Sergey Litvinenko: Yeah, that's, fair. Yeah. And you're right. We are looking at the, distribution for every single domain

[01:10:19] Audrow Nash: Yeah, like any good statistics people.

[01:10:21] Sergey Litvinenko: yep. Yep. And, there is a distribution of the, failure rates and the accident rates, and the results of the distribution for how bad those are. And, distributions look different for different use cases.

maybe. Maybe they all in a limit are going to be all normally distributed and you can do like perfect prediction, but now it's not normally distributed.

[01:10:45] Audrow Nash: Yeah, so one, one thought, and this would be funny, or like it'd be wonderful for robotics. I'm sure that it, I'm sure that there's not, I don't know, in, in the limit, if robotics keeps getting better, will your margin keep decreasing for this kind of thing? Or will there be so many robots that it'll be like Home insurance or something where you just have so many entities insured like how do you view the limit as robotic performance improves and how does that look for Koop

[01:11:25] Sergey Litvinenko: yeah, that's a really good question. So to repeat this back to you, pretty much if, robots eradicate all possible failures and they just

[01:11:36] Audrow Nash: crush it all the time

[01:11:37] Sergey Litvinenko: or crush it all the time. Yeah. Why do you need insurance? I think, I think what we're going to come down to is, something like the following.

we're going to have robots that are going to be so good that we will likely reduce the frequency by 95%, maybe like 98%, okay. And there's going to be that remaining. 10 to 1 percent of accidents that are still going to be present. Maybe those will be some form of edge cases, which will just be hard to take care of because,

[01:12:12] Audrow Nash: sure

[01:12:12] Sergey Litvinenko: and when those happens, the all eyes are going to be on the company behind the failure.

and it's going to be most likely very high severity. So we're going to go from, for humans, it's okay, there is some frequency, some severity. Okay. For robots, it's going to be a lot lower frequency, a lot higher severity, because it could be a fleet wide outage, okay? If I have a million mobile robots in operation, and, for some reason, half of them go offline, and I, my customers lose a billion dollars, that's going to be a nightmare.

The severity is going to go through the roof. And I think we're going to get to the point where our model is going to actually not just we'll turn into a catastrophe insurance model where we're going to be insuring for like very outsized events, but we'll have the data from the performance to know at what frequency and severity of those might happen.

And here we, look, we will probably make a lot less per robot in the future, but it's going to be so many more robots that the economics will make sense.

[01:13:17] Audrow Nash: yeah that's what I would expect too very interesting that you may have to like far down the road in the limit it's the catastrophe model

Which, okay, makes sense. And I love the idea of, oh, there's like tons and tons of robots and you just charge the smallest insurance premium on them because most of them are reliable all the time.

And it's very cool that you get to, leverage improvements over time. Actually, so one question that leads to is how do you update your rate? for companies over time as their performance improves. 'cause I'm imagining if, say I'm Tesla and I ship the, I think, 11 to 12 for the self-driving, there was a big improvement there.

And say I'm insured at this one level and now it's a big improvement. or say we go to 13, it's probably that they don't wanna pay the 11 rates when they're on. Full self driving 13 How do you handle adjustments over time as the robots become more capable?

[01:14:30] Sergey Litvinenko: the way we approach this at Koop, it's a function of volume of data that we have. If we have a lot of data coming, let's say, over the course of the month, we would look at updating the rates on a monthly basis. The way we do this right now is that we update rates on a yearly basis because over the course of the year, we collect enough data to see how the customers perform and specifically to see if there were any losses that happened in that period, if there were major improvements in the performance, if there are major changes in the qualitative questionnaires.

Now, if we, if, we, let's say, let's say go from, getting 200 customers per year, every single customer sharing some form of data with us to getting a, 200 customers per month, then is going to warrant a, higher rate of updates on our end. So it's a function of data. So right now we're all set, which is also the industry standard and the insurance that you review the rates on a yearly basis.

I think. Give it probably another year, and a half. I think we're going to do this on a, a biannual basis. So twice a year, the goal, I think the goal, the minimum, like the highest frequency of updates that we can do will be on a monthly basis, most likely. But we were not there yet.

[01:15:55] Audrow Nash: Yeah I would love to see like in the absolute limit that every maneuver that the robot does is a small charge For their insurance and it's like highly scaled to what they're doing right now Like I think that would be the coolest thing. The robot is navigating between a crowd. that's six cents or something like that.

And, it's highly usage based, like API calls basically, but for insurance. But that, yeah, monthly is already way faster for this kind of thing. And I, oh, go ahead.

[01:16:26] Sergey Litvinenko: do it. Yeah, something like that, like usage or maneuver based or risk scenario based, insurance model. like technically we can do it right now. The reason why we're not doing this is because there's just not enough

[01:16:39] Audrow Nash: fit the model.

[01:16:40] Sergey Litvinenko: there. Yeah. Yep.

[01:16:41] Audrow Nash: Yeah. It makes sense. That's super cool. let's see. I wanted to get your thoughts on some, so this might be a complete. ridiculous idea, but it seems to me that you guys are getting a lot of really good data about robotics companies while they're operating. It seems like a very good place to have, a venture arm to your company, where you invest in the companies that are doing really well.

if I was a venture capitalist, and I had access to the data that you guys have access to insure

[01:17:16] Sergey Litvinenko: huh.

[01:17:16] Audrow Nash: it would, you would be making such intelligent bets. I'm just, I wonder what your thoughts are on this, because you, can see their performance curve over time, where they're improving the, rates of accidents, the rates of accidents or bad performance are decreasing.

I, would think it lends itself really well to making investments yourself as a company in the companies that you are insuring. and know a lot about but what are your thoughts on that is that a like incredibly stupid idea or doesn't work because of conflict of interest or what are you any thoughts

[01:17:54] Sergey Litvinenko: you're really thinking this through, but I got to tell you that we, we haven't thought about becoming a, VC shop ourselves, but we partnered with a lot of other VC

[01:18:06] Audrow Nash: oh

[01:18:07] Sergey Litvinenko: And, it's not a formal partnership. It's actually, if you look online, you're not going to find it anywhere. That's something that we do, I would say privately, but we'll usually have.

A, a venture capitalist, that invests in robotics companies and then that venture capitalist would want their portfolio companies to work with Koop for reasons. Number one, we can keep track of them. Number two, we can protect them. And, that's how they also do this because they want to protect their investment, and it's, it's, they made a bet.

That this company is gonna become a billion dollar company. There's gonna be so many hiccups on the way that paying, a certain amount of money per year just to protect yourself and protect your attack is a no brainer. And I can tell you. we, of course, work with the investors that invested in Koop, that's AlleyCorp Ubiquiti, Hyundai is an investor, Fusion Funds, we work with all of them, but we also work with non Koop investors, and we have 10 funds signed up so far, and all of them are like, mid stage, not mid stage, but mid size, so anywhere between

[01:19:14] Audrow Nash: they're doing like series a kind of thing series a or below maybe

[01:19:20] Sergey Litvinenko: So you're really thinking this through. You're right. There is definitely the investor angle. I don't think we're big enough for it to become a, to open a venture arm ourselves, but look, why not?

[01:19:32] Audrow Nash: you have such good data. It's a huge competitive advantage to other venture arms, or other venture capital companies, I would think. to me it just makes, it makes really good sense to try to be involved in the venture capital scene, for this kind of thing. But because also you just get a little additional, you get some revenue information for the company and then, you know about their customers and then you have all the usage data.

it seems like an incredible competitive advantage as a venture company. I hope to see you guys, doing this pivot because I think it could be a really good fit. It would be like the most informed venture capital and I know, one model that could be done with this is that you sell the data that you're, collecting on the company and of course let them know and maybe give them some of the profit too, but you could sell this to venture firms to make their decisions.

That would be one, but you would get a lot less of the upside if you know that a company is absolutely killing it. and so I would think that would why, be why you want your own venture arm.

[01:20:44] Sergey Litvinenko: Yeah, I can tell you, data products are something that we have on our longer term roadmap. And I think data products could be very profitable. Very profitable.

[01:20:53] Audrow Nash: think so too. Yeah. And you could even like it, probably you have Koop, but you set up a second company that is just a venture capital company and then you pass the data over and you integrate it really well. and that would be a wonderful opportunity. I would imagine. Seems awesome.

super cool. it, seems like a really good thing for the robotics because also, It seems like a lot of finance is about intelligent capital allocation and, you guys have really good information. And if you're assessing risk and you have a model for risk, that seems like a wonderful setup for making intelligent allocations from my perspective.

[01:21:44] Sergey Litvinenko: Yeah. It's also, you're correct. I think the challenge with that is going to be if we, let's say, collect data from early stage companies and we see that ones are doing better, the other are doing worse, and then the ones that are doing worse change something and then they start doing better, it's like the, you have to be able to track the rate of change really, good because the, you can make a bet on something that you thought was bad, but then.

Things changed over the course of the month. but I think in general, having access to data products, that you can use however you want anonymized, maybe by sector or by the robot type or something like that. it's a great idea.

[01:22:36] Audrow Nash: Yeah, for sure. I think that's super, super cool. so now one other thing I wanted to get your opinion on. it seems So I, most of my interviews are with companies like robotics companies. I'm talking to technologists. but it strikes me by talking with you. and you're definitely a technologist too, but like you're not making a physical robot.

It strikes me that there's a lot of opportunity in robotics around the infrastructure required to make robotics companies go. And I think one large domain for this is probably finance. And you mentioned at the beginning, like the ability to get like good models for getting equipment and things like this.

but I wanted to see if you have other financial, like other opportunities in the financial space or. Other spaces that aren't directly building a robot and selling it for a use case that you think are good, but they're not quite things that you have the time to pursue, given that you're already doing Koop and there's a lot of exciting direction that you guys can keep going down.

[01:23:54] Sergey Litvinenko: if I great question, I think that actually your question implies Something even bigger is like hey what is the vision for Koop? Okay, and that kind of those two questions overlap. You're right There is this financial layer and there is the robot financing. So whenever you purchase a robot, you can finance and split the payments for it.

Okay. so you can literally like open a bank and you can provide leases on the robots, make it very easy to buy one. Two, warranty. don't forget that the warranties that you provide with your robots, they're backed by somebody. if you're a big company, you can do it yourself. If you're another big company, you have to partner with a warranty underwriter, or Also could be very profitable.

Three, insurance. In addition to insurance, I would also add a compliance automation software. Robotics companies, they are required to be compliant with different standards, like ISO, SOC2, or SOC in general, NIST. and see, there are lots of them, like just automating that and making that as easy as possible and as integrated as possible with your systems is another tremendous opportunity.

And that can be done with insurance together, actually, they

[01:25:19] Audrow Nash: goes hand in hand.

[01:25:20] Sergey Litvinenko: Yeah. And another thing, which you mentioned is data products. Okay, now data products is probably going to be more technical, out of all of those. but the data products could be sold to all kinds of interested parties, governments, insurance companies, like maybe legacy insurance companies, any,

[01:25:40] Audrow Nash: insurance companies is a funny way to put it, and I think you're right.

[01:25:44] Sergey Litvinenko: Yeah. so when you have. Equipment Financing, Robot Financing, you have Warranty, Insurance, Compliance, and Data Products. Those are the five things that I think is of that layer. And one of the visions, or one, one vision that I have for Koop is that Koop will, at scale, be able to do all of those things.

And be a, true go to Financial partner and data partner and insurance partner and compliance partner for the robotics industry. And you have multiple offerings and they all can talk to each other because, if I ensure the robots that you sell, I want to know the financing terms, or I want to know, I want to use that, that certain data products for that, or I want to use the, whatever you do on the compliance front.

So you're exactly right that, there is more. than what we just do. And I think at scale is going to become very important that you have that layer functioning. Hopefully we can be that layer for the industry.

[01:26:43] Audrow Nash: That'll be awesome. And if you do that, you'll get a huge advantage from the integration you're able to provide between those services. like that, will be wonderful. And it, I wonder, I don't know too much about the space and I wonder how much overlap there is, between I don't know, helping with financing and helping with, one of the other parts that you've mentioned.

But, there might be some overlap. I'm sure there is. That will one big thing is you don't have to do customer discovery, and that's a big advantage. You just say, Oh, by the way, we also have this, you probably need this, too. and that's I bet you cost per customer is a thing that is very difficult to break into this market.

Like it's a You have to spend a lot to get a customer, but then if you make this zero because you already have the customer because of another service, then that's a big advantage.

[01:27:38] Sergey Litvinenko: And I can tell you the biggest benefit across doing those things under one company is going to be a unified system of record. So for

[01:27:47] Audrow Nash: Yes. Oh

[01:27:49] Sergey Litvinenko: I finance

[01:27:49] Audrow Nash: you're

[01:27:50] Sergey Litvinenko: when I finance your, robots, I know exactly what you have. So now I don't have to bother you when I need to provide you with insurance because I know all the robots.

Same applies to warranty. Okay. Or when, let's say you do compliance, you don't have to fill out a technical questionnaire for insurance because we're going to pull it from compliance.

[01:28:11] Audrow Nash: then strap on a venture arm and you guys are set for this kind of thing. That way, it would be such a benefit to the robotics space too. and having a larger, like most of the companies that I've talked or investment firms that I've talked to That invest in robotics are small and they do series A funding, and then when you get bigger, it becomes a bit easier to, go with regular large investment firms.

But Series B is very hard typically for robotics companies as I understand it. and then like when you get to like d and e or whatever, if you go that far, then it's easier because you just look at. regular business numbers, but it would be nice to have a larger player in the robotics space that helps with, that is well suited to fund later rounds.

so if you guys grow that direction, I think it will be a very good thing, even the insurance and everything. I think it's a very good thing for the robotics space in general.

[01:29:13] Sergey Litvinenko: And we can even, add to the financial layer, a venture arm. We don't even have to be like, we can definitely be a lead investor, but we can also bring other people on board, like a syndicate, and the syndicate will make an investment decision based on what we know about the company.

[01:29:30] Audrow Nash: yes, because you can really make it so you get great reports and you have all the information and you have your risk models. That's all very useful information.

[01:29:38] Sergey Litvinenko: Yep.

[01:29:39] Audrow Nash: Okay, hell yeah.

[01:29:41] Future Vision and Expansion of Koop

[01:29:41] Audrow Nash: so we've talked about all those, like the potential directions that Koop could go. what do you, where do you see yourself going in the next five years, say?

[01:29:56] Sergey Litvinenko: I believe that, The, so first of all, we're making a bet on the robotics industry and we believe that over the next decade Robots are going to go from where we are today to being 10x more ubiquitous. Okay, hopefully 100x So that's the bet that we're making across all the use cases. Okay, so I believe that the Koop is going to become a company that will consolidate certain non core functions, such as insurance, compliance, and data products and financing in one.

It's going to be one company, one platform, one partner that's going to make it very economically attractive for the robotics companies and as a result for the customers of robotics companies. And we want, we do not want to be a company that does one thing. We want to be a platform, like a layer that takes care of all of that.

And like you said, we're going to be producing value, through, different means, but at the end, it's going to come down to sell your robots cheaper, sell your robots faster, sell your robots easier. That's what it's going to come down to. So if you have a robotics company and you don't have Koop, it's going to be a problem.

That's the vision. that we're driving towards. And, yeah, I think, we're definitely on our way. I think probably happens sooner than five years. but becoming that brand name that everybody knows about and that nobody's companies can live without. That's definitely very, if we do this, it's going to be fantastic.

I'll be proud of this, of the team.

[01:31:36] Audrow Nash: Hell yeah. Yeah, that'll be great. any, do you have any things, like for our watchers and listeners, is there any major takeaways that you hope, if they get anything from this episode, what do you hope they get from it?

[01:31:55] Sergey Litvinenko: I think, it's something that I think about every day and it maybe doesn't just apply to robotics companies, it applies to any, startups or early stage companies out there. I sometimes find myself thinking about the product more than thinking about the customer. and you can build a lot of cool stuff, but unless it gets the job done for the customer, it doesn't matter.

there is a funny meme on the internet where you have like a Mario and then you have some, your product, and then you have a Mario on like steroids. And it's this is your Mario as a customer. This, thing, is your product, right? And then when you combine those together, this is the cool stuff that your customers can do.

You need to focus on the cool stuff your customers can do, not on a product. And I would encourage all the robotics companies to think super, super hard about their customers, how they're going to commercialize, what problems they're solving. Sometimes I'm just seeing robots for the sake of robots. And it really makes me upset because, you have such talented people that could be solving some big problems, but they're working on the robot that just looks like a cool robot without any clear application.

And, that's probably one thing. And that applies to everybody, not just robotics companies, also applies to us, probably to any, every single one out there. that's one thing that I can share.

[01:33:25] Audrow Nash: All right. Hell yeah. Let's see. This was a blast. Wonderful hearing about what you're doing. seems like it's a great initiative for the robotics community and, I think it'll make a lot of things easier.

[01:33:40] Sergey Litvinenko: Thank you. And we'll do our best.

[01:33:44] Audrow Nash: Hell yeah. All right. Bye everyone.

And that wraps up our conversation with Sergey Litvinenko from Koop AI. I don't know about you, but I'm really excited to see how companies like Koop AI will shape the future of robotics. It's fascinating to see how something as seemingly mundane as insurance could actually be a key factor in accelerating the adoption of robots in our daily lives.

So I want to leave you with a question: How do you think that increasing presence of robots in our world will change the way we think about risk and insurance? And what other industries might be transformed by the growth of robotics insurance? If you enjoy this episode, please share it with a friend who's into tech or robots.

And as always, I'd love to hear your thoughts. You can reach me on social media or leave a comment on the video. Thank you for listening. See you next time.