When Robots Go Haywire, Who Picks Up The Tab? | Amias Gerety

Amias Gerety is a partner at QED Investors, a fintech-focused VC firm with $3.8 Billion under management. He was previously the Assistant Secretary at the US Department of Treasury. And he graduated from Harvard.

Amias's favorite finance book:
Lombard Street: A Description of the Money Market (Author: Walter Bagehot)

(00:01) Introduction
(00:46) Why Do Robots Need Insurance?
(03:04) Challenges for Insurance Companies with Robot Liability
(07:34) Risks of Autonomous Robots in Industrial vs. Domestic Settings
(12:22) Cybersecurity and Hacking Risks in Robotics
(15:41) Importance and Challenge of Historical Data in Robotics Insurance
(18:52) Leveraging Telemetry Data for Risk Modeling
(22:09) Creating a Functional Insurance Market for Robotics
(26:42) Should Robots Be Independent Legal Entities?
(28:34) Should Robots Buy Their Own Insurance?
(32:30) Partnerships Between Insurers and Robotics Companies
(37:11) Regulatory Framework for Robotics Insurance
(38:57) Advice for Founders in Robotics Insurance
(40:07) Rapid Fire Round

--------
Where to find Amias Gerety:

LinkedIn: https://www.linkedin.com/in/amias-gerety/

--------
Where to find Prateek Joshi:

Newsletter: https://prateekjoshi.substack.com 
Website: https://prateekj.com 
LinkedIn: https://www.linkedin.com/in/prateek-joshi-91047b19 
Twitter: https://twitter.com/prateekvjoshi 

Transcript

Prateek Joshi (00:01.707)
Amias, welcome back to the podcast.

Amias Gerety (00:04.482)
Well, I'm glad to be here. Happy New Year.

Prateek Joshi (00:06.585)
Happy New Year to you too. The last episode got a bunch of interesting responses and thank you listeners for texting me, emailing me. was a lot of fun doing it and I'm glad it was useful. So I thought why not invite Amias back to talk more about another thing we wrote about and that is robot insurance. Okay, so Amias, you and I, we wrote about this a couple of weeks ago. We already have insurance.

in the world to cover employees and assets. So let's start with the fundamentals. Why do robots need their own insurance? Like what's even the point?

Amias Gerety (00:46.87)
Yeah, and I think, so the first thing I think we have to start is what is the difference between a robot and a machine? And I think, you know, if you had talked to us five years ago, you'd say robot, machine, machine, robot, they're the same, right? But I think the fundamental thing that we're talking about here is that the combination of AI plus some industrial machinery

gives a much more probabilistic machine. And I think it's that probabilistic, I would call it semi-autonomous. So that probabilistic physical semi-autonomous machine, that is the thing that we should call a robot. So one of the examples I use is like a Cuisinart has safety devices in it so that you can't cut your finger off in the food processor.

But those are all very mechanical. They're quite sophisticated. And when they came out in the 70s or the 50s or whatever, people cut their fingers off. So now if you buy one, you can't cut your finger off. It's really good. But those are non-probabilistic. So I think the real thing is that this idea of a robot, the way we imagined them when we were children, is much closer to reality.

Prateek Joshi (01:51.712)
All right.

Amias Gerety (02:09.662)
than, or the way we saw them on the Jetsons, than an industrial machine of the type you'd see in a factory, which has some flexibility and therefore some statistical likelihood of creating injury, but is much less probabilistic and basically 0 % autonomous. And that's the place where new types of risk come. And if you have new types of risk, new types of insurance.

Prateek Joshi (02:36.185)
Fantastic. That's a great starting point. Now, the concept of liability or insurance, it has existed for a century, it's like a long time. And now, what he just said, robots are blurring the traditional lines of liability. So can you talk about what challenges does this create for insurance companies who now have to create a new product to address this?

Amias Gerety (03:04.588)
Yeah. So an insurance company in the consumer zone is mostly thinking about you as a consumer, write a check every month or every year to the insurance company. And in exchange, if something bad happens to you, they'll write you a check. So

your house, you know, gosh, we're all dealing and just watching with horror what's happening in California, right? So if your house burns down, there will be insurance coverage. If there is a hurricane, you'll have insurance coverage. But this idea of robot insurance goes one step further, right? Because I don't have a dog, but if you have a dog,

and adopt you, your dog bites someone. That is also something that can hit your insurance company. Right? So if, if you, if you have a dog, the dog bites someone and that person sues you.

that could be covered by insurance. If you have a trampoline or pool in your backyard and someone drowns or hurts themselves on the trampoline, that could be covered by your insurance. So even in general home insurance today or general personal insurance today, auto insurance today, we have these umbrellas of other bad things that can happen.

And so the first thing I think that we would say is just like regular insurance companies didn't want to cover cyber risk because they didn't understand it that well, it could move in different directions. So too, they've gotten comfortable with the idea of, you know, a dog. By the way, if you have a trampoline and your insurance company comes to your house, they will tell you to remove the trampoline. They don't cover trampolines, right?

Prateek Joshi (04:53.272)
Well, that's the nugget of the episode.

Amias Gerety (04:53.964)
Okay. There's certain things they don't want to And my intuition here, and this is why I thought it was a fun thing to write about, is nobody has thought really carefully about whether there's a robot in your house. So a 300 pound semi-autonomous industrial machine washing your dishes is not something that the insurance companies have thought about. And the reason this gets interesting is that unless they've thought about it, it's covered.

That's why these policies are often called general liability. So there's a funny story I heard of a claims adjuster who had a trucker. The trucker swore that they were pulled over on the road and asleep. By the time the accident happened, they were miles down the road and in the dirt. And they said, aliens abducted my truck and took it on a joyride.

And so the adjuster had to look in the past because like you, on one hand, you could say like, no clear, you're, you're lying. But at the other point, like they didn't have any proof that he was lying. And so it turned out that alien abduction was covered by the policy because it wasn't specifically included. Okay. So this is, this is the framework that insurers have to think about when something new happens is unless we exclude it.

Prateek Joshi (05:55.098)
Right.

Prateek Joshi (06:09.657)
Good night.

Amias Gerety (06:21.142)
It's covered. so my thought was, hey, this is coming, which means it's, and it's coming from far enough away, not like alien abduction. It's coming from far enough way that insurance companies are going to see it. And they're going to tell people, Hey, you cannot have that robot in your house and have, you know, what if the robot hurts someone? I'm not going to cover that. It's going to be like a trampoline, but

Prateek Joshi (06:30.957)
Thank

Prateek Joshi (06:41.421)
Right. Right. Right.

Amias Gerety (06:45.812)
Some people, people who build robots, people like us who are kind of excited about the possibility of a robot washing my dishes or folding my laundry, so everybody who's watched the Jetsons, they're going to be interested in these robots. And that means that there's a gap in coverage that we need to think about.

Prateek Joshi (07:01.593)
Now, speaking of autonomy, you mentioned the past of fully deterministic, the machines to semi-autonomous to hopefully someday fully autonomous that can do when it's pre-task. Now, what are the risks that are associated with autonomous robots versus an industrial robotic arm sitting on a factory floor? Is that the same or what are the net new things that this insurance company has to worry about?

when it's a fully autonomous robot in my house cleaning my dishes.

Amias Gerety (07:34.966)
Yeah, so I think moving to industrial questions is the right next step in the logic team. We sort of talked about what's the general idea of, and look, it's kind of more fun to talk about them in your house. They're going to hit factories before they hit houses, right? So let's talk about the industrial commercial insurance framework.

Prateek Joshi (07:48.023)
Right, right.

Amias Gerety (07:53.942)
So the commercial insurance framework has what I would call like three levels of bad stuff that could happen. And the insurance will hit differently at each of those levels. So the first level is like a manufacturer's defect. Like I bought the machine to lift a weight and instead it like shot the weight through the load through the ceiling.

Okay. Manufacturers defect that's probably covered by manufacturer liability. Then there's like a, a kind of a user's defect, right? Like this is a big machine. requires an awake, non-drunk person to press the button at the right time.

So if you as an employer didn't monitor your employees for safety or whatever, that would be the employer's liability. And then across both of them is something that no insurance company will cover, which is negligence.

So if anyone in the chain didn't do the right things, the insurers will say, Hey, look, my job is to ensure when bad things happen. But if you take actions that cause bad things to happen that are predictable, that's your problem, not my problem. So those are kind of the three levels of like manufacturers, users, and then negligence, which runs through them. And so this is why I think it gets really interesting when you start to,

try and translate that kind of framework of an industrial machine into a kind of a robot, a semi-autonomous robot. now, if a robot, let's say, has three buttons, it's relatively straightforward to train and predict what each of those buttons do. And if they don't do the thing, that's the manufacturer's problem. If someone presses the buttons in the wrong order and it says, do not press this button unless, that's the user's problem.

Prateek Joshi (09:32.633)
Yeah.

Amias Gerety (09:58.398)
But once we add AI and we vastly increase both the number of potential commands to an infinite number, the number of actions to a semi-infinite number, now the potential for reliability and the interaction effect between was that the manufacturer's problem because they didn't have a safety, or was that the user's problem, or and then where does negligence

Prateek Joshi (09:59.521)
Okay.

Prateek Joshi (10:20.009)
you

you

Amias Gerety (10:28.418)
fit in where like, well, I was just talking to my buddy and I said, you know, and then the robot overheard me. It wasn't my fault, right? So the scope of potential problem explodes when you add this kind of AI layer and you move into multipurpose. And again, it may be helpful to think about.

Most industrial machines are actually super finely calibrated because they're on like a supply chain. So most of them, people don't even hit the button at all. It's like everything's super finely calibrated. does exactly this thing, almost no user. And then you get like tractors. The user is definitely doing them. And then what we're talking about is this idea of a robot.

Prateek Joshi (10:59.824)
Right.

Amias Gerety (11:19.032)
that's making those decisions. And that explodes the possibility of risk and liability.

Prateek Joshi (11:24.844)
Right, in that three level framework, use a very important word, the word predictable, meaning if the thing does as predicted, then the insurer does the thing, is, you they're colored. But just the fact that AI

in the supply chain, AI by nature is probabilistic, meaning there's always a non-zero finite chance that it's going to do X and X being catastrophic. So this unpredictability in the behavior adds a lot of complexity. now maybe one quick follow-up question to your earlier description. What if the robot gets hacked? Meaning the manufacturer does everything, they ship it to you, user is like, I'm a user, I'm an average user, and then what?

the has to be connected to the internet because that's how the world works. Now, if it gets hacked, does the existing cyber insurance cover that or is that a new thing?

Amias Gerety (12:22.754)
So that is a really great question. And it's probably super specific to the insurance policy. My guess would be that in most insurance policies today, cyber would cover it. Because most insurance policies would probably say, cyber covers all corporate computer systems, blah, blah, blah, But this would actually be litigated.

Prateek Joshi (12:24.14)
Thank

Prateek Joshi (12:41.376)
Thank

Amias Gerety (12:51.84)
in court or will be litigated. And a lot of the reason why your insurance policy is so long and complicated is that someone litigated the question of whether X or Y or Z was covered. And I think this is a place where we're going to have policies. We're going to have exactly what you described, which is the, you know, the robot is hacked.

And then there's going to be a debate. whose fault was that? Is it covered by insurance? Now there's going to be exclusions. And that's how the insurance industry sort of manages its risk over time. And I think it's important to note that in general, insurance companies are fine with bad things happening, but they know bad things happen and they roughly speaking are

Prateek Joshi (13:20.025)
.

Amias Gerety (13:37.844)
trying to protect situations where it's nobody's fault. hey, people drive, car crashes happen. In a lot of states, you actually have like no fault states, right? Where they don't spend a lot of effort trying to figure out who caused the crash because they know car crashes happen and it's kind of nobody's fault. Now, of course, there's a negligence on top of that. And so any time where you could say clearly there was somebody's fault,

Prateek Joshi (13:49.593)
Okay.

Amias Gerety (14:06.902)
then the insurance company is going to try and make that somebody pay. And this is where this AI, the semi-ton and this thing gets, this is where it gets fun and interesting, which is like, wait a minute, if I could say anything I want to the machine and the machine can reason through, is it my fault or is it their fault?

Prateek Joshi (14:11.233)
Right.

Prateek Joshi (14:24.552)
All right. Because if you think about cyber insurance, like in some cases, I don't know if my gym membership or credit card gets hacked or something, they're like, okay, fine, you go work out instead of me. Fine, the downside is slightly different. When a robot gets hacked and burns the building down, the downside is pretty big.

So guess, you said, the insurance company has to look at the situation and say, hey, how much do you want to cover? How much do we want to cover? Because if you come back with $25 million in damages, we can't do that unless you pay a really big premium. So it's very interesting because consumers, if they ask you, you got to pay a million dollars a year to cover this thing, then people will be like, no, thank you. I know the downside is 25, but I'm not going to spend a million insuring this thing. So it's very interesting. Moving on to...

the next section, which is, let's say the insurance industry is preparing to do this, right? They want to do this. Let's say they have to cover it. Can you talk about historical data? Meaning why is historical data so important for the insurance industry? And how does the lack of historical data, especially in robotics, pose a challenge, even if an insurance company wants to do this?

Amias Gerety (15:41.58)
Yeah, and this is where cyber gives us really good models because cybersecurity was also a place where they really lacked historical data. And insurance risk is, so in classical economics, everything has a price. So you can get anything you want if you're willing to pay for it. But insurance, because they're based on, well, I have to guess about the future.

And the best way to guess about the future is to look at the past. So if I don't have a lot of data, I have to cabin and control my risk some other way. So the history of insurance policies in cyber in particular was very low loss limits. So I'll cover cyber risk defined very broadly, but only cover it up to $25,000 damage.

So for example, when I was at the US Treasury Department, one of my responsibilities was dealing with the interaction between the financial services industry and the US government on cybersecurity. probably the number one complaint that the banks had was, I cannot get enough coverage. Like they would say, I'm willing to pay literally any amount.

It wasn't actually literally amount, that's what they would say. And I can't get enough coverage. And if you think about being a very large bank, right? Very large bank, very large balance sheet, a hack could be a billion dollar, a $10 billion, a $100 billion event.

And you want coverage for some version of that. And they were trying to stitch together million dollar coverages, $10 million dollar coverages. And if you're the CEO of a hundred billion dollar bank, that just doesn't move the needle for you. So I think coverage limits is the first place people go. I'll cover anything up to $10,000.

Amias Gerety (17:39.242)
Or I'll cover anything for six weeks. Right? So it's time, it's coverage limits, and then gradually you build up enough data to say, can now I understand what can go wrong. Now I'm willing to model out what can go wrong and cover higher, higher limits. And those policies tend to be more specific.

So they'll have bigger exclusions or clearer guidances around, I will cover anything up to this amount as long as you do the following behaviors.

Prateek Joshi (18:17.31)
Now, if you look at insurers who are like, okay, we can't do anything about the temporal dimension, meaning I wish I had 100 years worth of data, we don't. So, barring that time is fixed, how can they...

How can they get more data? Meaning for a given every second, we can collect two data points or 20 or 2000. So the operational and telemetry data from robots, like how can we use that? How can insurers use that to improve risk modeling and just kind of create a decent insurance product?

Amias Gerety (18:52.428)
This is why I'm so excited about this as a kind of investment theme or as a thing that the AI industry, people are interested in this, people are innovators, should be focused on because exactly what you said is true. These are software-driven systems. They're generating with dozens of sensors. They're generating so much data so quickly that if we can actually collect all that data, process it,

we actually might be able to get up that learning curve much faster. If you think about cyber, cyber is a contest between good guys and bad guys. So we don't find out where the error is.

until the bad guys figure out where the error is, they cause the loss, and now we learn. And obviously, there's a lot of things that we do in cyber about penetration testing. And of course, the manufacturer of these robots will do those kinds of testing. mean, that's where all those really awesome Boston Dynamics videos come from, is they're trying to test their robots against and trying to build up that data set.

And this is why I think so, you I'm on the board of a company called tint.ai. Tint is an embedded insurance provider.

For different customers, they structure it different ways. So sometimes they help customers get a policy that tint structures, but underwritten and on the balance sheet of an insurance company. Sometimes they structure it in such a way that the company can structure it like insurance, but actually the risk sits on their own balance sheet. And so there's lots of different flexible models. I think embedded is going to be the early answer here, because if I'm a

Amias Gerety (20:39.84)
big, well-funded robot manufacturer. I'm going to go to a big insurer at my corporate balance sheet level. I'm going to make a partnership with them about getting and processing that data. And then I'm probably going to, and I do that for my own, that like corporate manufacturer defect thing, but I'm probably then going to also offer to the initial consumers, hey, you don't have to worry about navigating the insurance world out there.

Prateek Joshi (20:51.961)
Okay.

Amias Gerety (21:09.898)
You're actually we can work with you and we can offer this insurance Backed by our balance sheet backed by maybe an insurance company so that that first foray is going to probably come along with the the manufacturer now typically you think of that as a warranty or a guarantee

Prateek Joshi (21:14.371)
you

Amias Gerety (21:29.304)
But we've just talked about robots, you're going to need liability too. So that's an unusual product for a manufacturer to offer you. But I think that's where the industry is going to start. And they're probably going to start with pretty low limits. And then as they develop, they're going to have

bigger, more specificity, and probably higher limits over time. But my hope is that these partnerships and the data processing that also comes along from other advancements in AI is going to allow us to get up that curve of understanding what the possibilities are much earlier, much faster.

Prateek Joshi (22:09.453)
That's really good explanation. Going forward, we have to create a market for this in the sense that we need entities who are willing to sell insurance products and we need entities who want to buy insurance products. So we need to create a functional market. So let's talk about the potential mechanics of this market. So one, the data. Now, if you're a...

with a bunch of robots, you're collecting a ton of data. There's so many companies. And insurance company rolls in and says, hey, give us all of your data. We'll combine it with a hundred other companies and we'll make money off of this. So you, you the big robot company, why would you care? Why would you even opt into this? So what can the market offer the robotics companies, especially the big ones, to opt in so that the insurance companies can combine the data and create some functional products?

Amias Gerety (23:07.896)
So I love this question because it's all about contracting and incentives against uncertainty. And I think what's so exciting about AI, the uncertainty is what's exciting about AI. One of the big learnings, think, for all of us over the last two years or so is that the hallucinations are the point. If it didn't hallucinate, it couldn't do the magic.

Prateek Joshi (23:31.341)
Yeah.

Amias Gerety (23:36.886)
And so I think the uncertainty is the point here. so, so first off, I think the fact that we have this uncovered gap, right? Just the uncertainty in the system is going to create an incentive for insurance to come in. Because there's an incentive for insurance to come in, you're going to have a very strong economic incentive. let's say, let's say, let's add another player. Let's say I'm the store.

I'm the distributor of the robot. Well, I'm the person who sold it to the consumer. So I'm the first person who's going to get sued, which means that I'm the first person who's going to be, Hey, wait, does my insurance cover this? When the insurance says, well, have you talked to the manufacturer? Right? So the insurance price signal should allow for the supply chain to start getting a better price signal on what the risks are.

So that's the first thing is just the, like how contracts and prices could create incentives ahead of regulation or even faster than regulation could. The second thing that I think insurance companies offer is a very realist perspective. Like I actually don't care if I'm the insurance companies, all the different mechanics of the, what the robot was thinking when it crashed through your wall.

I actually only care about how often does the robot crash through the wall? How often does a robot squeeze a kid's hand in a manner that hurts the kid? So I think insurance companies have, over hundreds of years, developed

Prateek Joshi (25:08.077)
Right.

Prateek Joshi (25:15.691)
Right, right.

Amias Gerety (25:23.392)
a very realistic view of what it means for something to go wrong. That is actually quite different than an engineer's view about what went wrong. They're actually, they, of course insurance companies care about root causes, but at some level, like they also cause about, they care more about,

end bad outcomes. So they're very focused on a very specific set of the distribution, which is, I think, in many ways complementary to the engineer's mindset. The engineer also cares about root causes, also cares about what can go wrong, but they tend to be very focused on the mechanics of what went wrong in what sequence of steps. And the insurer tends to very much care about, did it go wrong or not?

Prateek Joshi (26:09.945)
.

Amias Gerety (26:10.228)
and how far back in the chain. And so I think that's a very complementary perspective. And then lastly, I actually don't think they need, because of that, they don't need as detailed data. And that also, I think, would provide some protection across the IP. And then price, right? If I'm the person who ensures the other robot carriers, then I understand the risk better, which means I can provide the best price better.

And that I think is how you mediate the data issue.

Prateek Joshi (26:42.005)
Another idea that I'd love to get your thoughts on is that there are industries where every project gets packaged as new LLC or new entity for the sake of payroll and insurance so that people don't have to deal with the mothership getting sued for other things. So in this case, should robots be treated as independent legal entities for liability purposes?

similar to corporations or from LC's.

Amias Gerety (27:14.392)
I think this is super interesting question. I think the short answer is probably not, because I think these robots are going to be manufactured devices. And therefore, the consistency between one robot and the next is likely to make it sort of not sensible for it to be separately incorporated. That being said,

It, you know, in, computer science, we have this idea of emergent behavior and complex systems. And I do think that the AI engines inside of these robots. Put into the infinite variety of your factory, your house, you know, your whatever creates a kind of, emergent property that maybe it should be, but because it's, it's not just the robot. It's the robot plus your.

commands that create the emergent behavior. But my guess is actually these are treated like manufactured entities, manufactured products, and therefore are going to have what I would call stamped liability that attaches at the manufacturer.

Prateek Joshi (28:34.649)
Right. Great point. Let's assume for a second, and we are heading towards a very robotics heavy future, and where there'll be many corporations and entities employing a large number of robots. Should robots be allowed to buy their own insurance policies if they earn their own income? And there'll be a situation where for simple tasks, the robot knows, okay, I'll do X task, I'll charge Y dollars, and I'll keep repeating that until the end of time.

So it's almost like I'm not intervening at all and it can go for years and years. So in this situation, should they be allowed to buy their own insurance policies?

Amias Gerety (29:12.238)
So let's start with an analogy to Uber. So clearly an Uber driver while they're driving for Uber is covered by Uber's insurance. And then when they're not driving for Uber, they're responsible for their own insurance. Now,

many times that sort of can be a little bit contested, right? Because now Uber is monitoring the driver safety score and they might kick them off the platform if they don't like the risk, et cetera, et cetera. Now let's go up one, this is not how Uber or Lyft are organized, but you can imagine a different world. You can imagine where the driver looks more like a franchisee, right? Certainly the franchisee of a McDonald's.

or 10 McDonald's, they get their own insurance. Yes, it's a McDonald's brand, and the McDonald's brand imposed certain standards, but the person or company that is the franchisee, they do their own insurance, they get risk rated their own way, et cetera, et cetera.

And then the last analogy that I think would be really interesting here is what we see starting. so again, referencing Tint, my company, again, Tint offers dynamic cargo insurance. So if you're a trucker, you might have a regular liability to cover the carrying of loads. But what if you don't usually carry oil and gas?

You know, those, those big silver barrels. Now, if you don't usually, then your insurance probably wouldn't normally cover that risk. And that's obviously higher risk than carrying boxes. so. Tint is offering together with some shipping companies and shipping brokers, the possibility of a dynamic cargo insurance so that you could add more insurance if you're carrying higher risk cargo or add more insurance.

Amias Gerety (31:13.792)
if you're driving the cargo for a longer distance than normal. And I think you sort of put these things together and you can imagine a world where, you know, now we've got like a robot, which is the equivalent of a Waymo car, right? So the robot is out there. It's been tuned to do some work. You you come up to the robot, Hey, will you rake my backyard? Okay, sure. That will be a hundred dollars, please.

Prateek Joshi (31:28.598)
Thank

Prateek Joshi (31:40.61)
Right.

Amias Gerety (31:43.394)
Hey, will you clean up my backyard? the way, the reason I needed to be cleaned up is there are landmines back there. Right. well actually that's good. You know, that's going to, that's going to be an extra charge and embedded in that charge is going to be some insurance value. So I think you can see a spectrum here. Now I don't know that those are going to be the robot.

Prateek Joshi (31:51.961)
Right.

Prateek Joshi (31:57.047)
All right.

Amias Gerety (32:06.412)
Right. think ultimately I kind of believe in human ownership. don't really believe that, that, that robots are independent, but I do think, so I don't think the robot is going to do it or own it in any legal sense, but I do think that they're going to calculate it, charge it, negotiate it on a dynamic basis. in some future that we can see, don't know how long it's going to be before we get there.

Prateek Joshi (32:30.059)
Right, right. Going forward, let's say we are still so early in this cycle of creating and addressing this new risk we are seeing. If insurers and robot companies, they collaborate, it can lead to, ideally, can lead to better insurance products. And because the insurance product is defined, it can lead to hopefully safer robotic systems because that's how incentives work.

Amias Gerety (32:57.774)
Right, it sends the right price signal.

Prateek Joshi (32:59.797)
Right. Now, can you talk about what does this partnership look like and in what way can, how would you, if you had to guide them, what would you tell them that they should get right?

Amias Gerety (33:15.128)
So the partnerships are going to want to really focus on a couple of things. The first is a really clear understanding of the value chain and the sources of unpredictability. So I think one of the biggest challenges here is that the foundation models have gotten so good that they're now applicable to robotics.

And now the value chain has gotten just by adding one more party. The value chain has gotten so much more complicated, right? In Boston dynamics. mean, I don't know the company that well, but my sense is they've developed the physical parts of the robot and the software parts of the robot altogether, tandem, single source of IP, et cetera. Whereas if you use an example like figure.ai, figure.ai has an embedded open AI engine in it.

So I think the first thing you'd really want to get right is, do we have a shared understanding of the value chain? And then the second thing I think you want to get right is just sort of a shared understanding of, where is the risk alignment? And is there a path?

towards can we constrain the use cases, at least initially, and can we build up confidence together over time? And there again, I think we have three bodies, right? We have the robot manufacturer maker, the insurance company, and end consumer. Because remember, one of the roles of an insurance company,

is to actually increase the confidence of the consumer. And for those of us who are interested in consumer innovation, we should be respectful of the idea that asking a consumer to do a new behavior fills them with risk. And so if you can wrap that risk with some assurance and insurance,

Amias Gerety (35:18.766)
then you will get them to adopt the behavior. And if you're investing in startups, what do you want? You want consumers to adopt new behaviors, right? So I think I haven't actually, I have, uh, you know, we talked about my self-driving car experience before Pratik. I've never been in a fully self-driving car. I haven't taken a Waymo in San Francisco, but I heard someone say like, it's shocking how quickly you get comfortable after the first couple of minutes.

Prateek Joshi (35:25.89)
Right, right.

Prateek Joshi (35:43.415)
Right.

Amias Gerety (35:45.856)
That's amazing. But you can imagine the kind of offer that the manufacturer could make or the robot seller could make that says, like, hey, we've partnered with giant global insurer. They've inspected us. They will cover liabilities, blah, blah, blah. We are insured. We're bonded. Think about when you do a general contractor, they come to your house. Are you bonded and insured? These are the kinds of the language of guarantees that allow for commercial transactions to take place.

Prateek Joshi (36:15.969)
Now, regulation is another big topic because we are dealing with two industries that are more regulated than insurance. It's regulated. can't just do stuff. There's 50 states in the US, around the world, so many different rules. And it's regulated by design because it's one of those industries. On the other side, robotics, physical systems.

I can deploy a software app like today and just go do stuff. But if you had to build something and there are permits, there's like energy, like so many things you do. So you're dealing with two regulated industries almost. Do we need any changes, updates in the existing regulatory framework to enable this new market they're talking about? Or is everything in place and it's just a matter of companies getting together to do stuff?

Amias Gerety (37:11.416)
So I don't think there's much that's needed today in that if you think about the probably the two primary regulatory frameworks that you'd be dealing with on the robotic side.

you're probably dealing with the OSHA, which is the Occupational Safety and Hazards Administration, right? They're trying to say, like, worker safety. And that's probably where this is going to hit first. And then Consumer Product Safety Commission, so consumer safety interacting with a product. So probably those frameworks of saying, like, and those generally have relatively principled

frameworks because it's not like there are, mean, they do have specific frameworks for specific things over time, but they also have a general standard of like, you're not allowed to do a thing that will cause worker harm or cause consumer harm. so I think probably at least initially, and then on the insurance side, the actual, primary regulation is just, you have enough money?

as the insurer to pay the claim that you have promised to pay.

Prateek Joshi (38:24.385)
Right.

Amias Gerety (38:24.694)
And so those are the two frameworks. think those frameworks are probably general enough that at least in the first couple of go rounds, we're probably going to be okay from a regulatory framework.

Prateek Joshi (38:35.097)
Right. I have a final question before we go to the rapid fire round. And this is about as, as investors, the angles of attack. So if a founder is listening to this, or even before this is thinking about doing something in this, the robot insurance space, what are the angles of attack for that founder?

Amias Gerety (38:57.56)
Yeah. So most insurance startups start from a recognition that there's some risk that is not well covered and that if it were covered, more economic activity could take place. Like that's why you buy insurance, right? If I have this insurance, I can do more economic activity. And so I think this is one of the challenges for this is we're not really confronted with this yet.

So it's so capital intensive to make these robots that, like Waymo, it's going to start vertically integrated first. So I think the first line of attack is, as I discussed before, it's going to be embedded. And then I think over time,

If the market really develops the way it might, then I think that there will be more lines of attack. But I think the first line of attack is about embedded and it's about going to be about structuring the data so that it is mutually intelligible and mutually protected between the insurance company and the manufacturer. But that's where I think it's going to start.

Prateek Joshi (40:07.493)
Perfect. With that, we're at the rapid fire round. I'll ask a series of questions, different questions than last time to make it interesting. And I would love to hear your answers in 15 seconds or less. You ready?

Amias Gerety (40:19.05)
Maybe we should just test whether my answers are the same or not. I even remember my own answers.

Prateek Joshi (40:22.803)
That would have been a great test. was like, are you consistent with yourself? Question number one. What's your favorite finance book?

Amias Gerety (40:37.006)
So favorite finance book is called Lombard Street, 1873, Walter Badgett, the patron saint of central banking, original editor of The Economist. And it was written in the wake of a banking crisis in London. And if you're in finance, it's really important to understand both the power and the peril of leverage.

And what Walter Badgett argues in this book, sorry, this is more than 15 seconds, is that the power of the British Empire rests on the liquidity of the London money market.

Prateek Joshi (41:17.241)
Right. Right.

Amias Gerety (41:19.176)
That is an awesome argument. It's like the power of the US innovation today is driven by the liquidity of the San Francisco venture market. So that's why it's a great book for finance and it's a great book for any investor.

Prateek Joshi (41:35.681)
Love it. I love the history of finance in general. that's fantastic. All right. Next one. If someone wants to learn about inner workings of the insurance industry, like what book or what content would you recommend?

Amias Gerety (41:50.584)
So I'll do two podcasts that I kind of like. So one is, know, classic, you know, journalist podcast, insurance journal has a podcast. you know, and it's pretty good. That's actually the one where I heard the story about the alien abduction. There's another one more, you know, regular way podcast, a guy who loves insurance and produce a podcast on it called profiles in risk.

Prateek Joshi (42:01.113)
Right.

Amias Gerety (42:17.92)
Also pretty good, listened to an episode this morning about a person who wrote an insurance children's book. And then the last thing, being a bit of a homer, QED actually incubated and helped found InsureTech Connect, which is the biggest insurance innovation conference. that is, it's more of a commitment, but honestly, people love talking about insurance. Everyone who's in insurance,

Prateek Joshi (42:23.789)
Wow, amazing.

Amias Gerety (42:47.539)
knows that other people find it boring. So if you go to an insurance person and ask them sincere questions, you'll get more information than you ever wanted.

Prateek Joshi (42:53.465)
Right? That's amazing. Love it. All right. Next question. What's the one thing about insurance that most people don't get?

Amias Gerety (43:06.678)
The most important insight for me about insurance is that it's the inverse of the rest of corporate finance. So if you think about whether you're starting a company or thinking about a bank, finance is about investing today for return later. And insurance is the opposite, right? It's about, so.

Big investment today is banking and corporate investment. Big investment today and then lots and lots of small returns later. Insurance is just the opposite. It's lots and lots of small payments over a long period of time.

and then a big bucket of money if something bad happens. So it's just all of insurance really starts from, and one of the reasons why it's confusing is that it's the inverse of way almost anybody who learns about economics thinks about investment and risk because it's the inverse.

Prateek Joshi (43:49.113)
Alright.

Prateek Joshi (44:01.283)
Right. That's actually so, it's brilliant. I have never thought about it that way. And it's so, it's so clean and nice. It's amazing. All right. Next question. What's the most common reason why insured insurance startups fail?

Amias Gerety (44:17.454)
So every insurance product faces a scale versus risk trade-off. And I think it's very common for insurance startups, including ones that have gotten very, very big, to fundamentally get off the rails between that scale versus risk, where it's easy to sell insurance if you're underpricing the risk. Right?

Prateek Joshi (44:44.533)
Right,

Amias Gerety (44:47.294)
And because insurance products often have a longer tail, you can sometimes lose the plot.

And so you saw this with some of the high profile insure tech IPOs that performed very poorly is the kind of that scale versus risk. People got excited about the growth. They kind of overlooked the underlying risk attributes. And then as the cost of capital went, it got a lot higher. They really suffered in public markets. Some of them have come through that effectively, but I think that risk scale versus risk trade-off is where we see a lot of problems.

in the Insurtech startup landscape.

Prateek Joshi (45:29.643)
If you could change one thing about the insurance industry today in the US, what would it be?

Amias Gerety (45:37.334)
Insurance companies are by reputation and by my experience, very bad buyers of software. And I think that this is common across industries that larger companies want. They're so comfortable with the systems integrator approach. Like I'm going to write down all the specs and then you build it for me. But they don't realize that every piece of uniqueness locks them into a single generation.

Prateek Joshi (45:46.393)
you

Amias Gerety (46:06.81)
And so if I could change one thing about the insurance industry and obviously self-interested here, I'd like to sell them software. I'd like to invest in companies that sell them software. So I'd like them to learn how to buy software.

Prateek Joshi (46:19.189)
I think that's a very good insight. Every piece of uniqueness, which they think is like, we're customizing it to our needs, so it's an advantage. It's actually locking them in and it's a big disadvantage because they can't change it. And even if you want to change it, it's like somebody will charge you an hour and a half. It's great. right, final question. What's your number one advice to founders building insurance startups today?

Amias Gerety (46:43.918)
So because insurance incumbents tend not to be good buyers of software, you have to find ways to get paid for performance. And mostly, that ends up that the best insure tech startups are actually distribution businesses or connected to distribution businesses.

But ultimately it's a version of getting paid for value, right? It's a version of value based selling. But the most important thing I think in insurance today is it's because the insurance industry doesn't have great muscles for software buying. You have to find a way to participate in performance terms that they can understand, which is better risk, better growth, and get paid for the value you provide that way.

Prateek Joshi (47:28.887)
Amazing. Amias, this has been a brilliant discussion. As always, it's intellectually, it's rigorous, it's stimulating, and it's always fun to talk to you about what topics. Thanks again.

Amias Gerety (47:38.646)
Always fun. Thank you for taking. All right.