Keep Your AIs on the Road: Reducing the Risk of Distracted Driving

Pete Miller [01:40]: Stefan, welcome. I’m really grateful that you would take time out of your busy schedule to come and talk to us about really the very interesting work you’re doing. Just by way of getting started, Stefan, can you give us a little bit of your background and how did you arrive at collisions becoming your focus and kind of the thing that you really spent a lot of time on?

Stefan Heck [02:04]: Yeah, I have an unusual background to say right off the bat. I have no insurance background and never worked for an insurance company or thought I would do anything related to insurance. I have a PhD in neural networks from 30 some odd years ago now, the very early days of neural networks. Before there was such a thing called deep learning. And then I built a software business and then I spent almost two decades at McKinsey as a management consultant.

So I worked a lot with companies that had extensive field operations, logistics, shipping, installation, service, and saw a lot of the operational challenges that people had. But the involvement with collision specifically came out of my time as a professor at Stanford when I was looking at disruptive technologies. And there are lots of them, of course. I worked with about 25 PhD students. Each of them picked a different technology.

But at the time, this is 2013, there was a huge boom in looking at disruptions in the transportation space, from an efficiency, from a safety, from an electrification, changing fuel, going emission free. And I looked at the data for losses and I realized, you know, this, the ground transportation system really hasn’t gotten that much better. If you go back to, you know, 1956, airplanes and surface travel were equally risky per mile. And airplanes got a lot safer. Basically, we have essentially zero fatalities.

Every once in a decade there’s an exception, but basically zero. Versus ground transportation turns out to be the most dangerous things we do anytime between age five and age 55. If you’re older, the famous lifestyle diseases, you know, cancer and so forth, are a bigger risk. And if you’re really young, it’s a baby, infectious disease at a risk. But driving is actually the most dangerous thing we do.

So I got curious, I wanted to work in an area of AI and neural networks, that was solving one of the top problems of civilization. And so that’s how I got interested in collisions, why they happen, how we can avoid that, and how we can really change the system, not just to provide insurance coverage, but actually to prevent them in the first place.

Pete Miller [04:13]: So tell us about your company, Nauto. When did you create it? When did you form it? And sort of what’s your goal and your sort of the “why,” as we would say.

Stefan Heck [04:23]: Yeah, the why is really very simple: to save lives. Every single person that joins Nauto is here on a mission to save lives. And we do that today in surface transportation, but a lot of the technologies we’ve developed could equally apply to trains, airplanes, any kind of other mobility or complicated system. But today it’s about cars, trucks, vans, any kind of size ground vehicle.

The genesis was all the way back in 2014 when three things came together. The first was this course I was teaching on disruptive technologies and a bunch of students were interested in transportation disruptions. The second was I live a couple miles from the Stanford campus. I was biking to campus and once or twice a week somebody would nearly run me over and nothing gets your attention quite like that.

And I noticed two things that probably most people’s experience of — certainly if walking or biking, a little less so when you drive, but it still happens when you drive — One is that, you know, risk is not evenly distributed. It was the same two spots on my, you know, two-mile commute that were always the danger zone.

The second thing that struck me was nearly all of these near-miss incidents were forms of the driver not paying attention. And I looked in the data, and this is back in 2013, 2014, and there wasn’t a lot of attention on distracted driving yet. Mobile phones had started taking off, and we now know a decade later that mobile phones are about 80% of all distraction, as people using digital devices. But it was starting to boom way back then already.

Basically, because of a whole range of things. People were talking to their kids, you know, reading newspapers or reading books. And of course, a lot of people using iPads and phones.

And in the government data, you don’t see that. Government data says the top problem with driving is speeding. Speeding causes most collisions. And I started interviewing police chiefs, you know, as a professor, you can call up anybody and say, hey, I’m doing your research on X. Can we talk for a moment?

And I presented this to a police chief and I remember he broke out laughing. He said, Stefan, you got to understand by the time the cop shows up, you know, best case two or three minutes after the collision, you can’t tell the person was distracted anymore. And if they don’t know what the cause was, they put down excessive speed. And so that’s why speeding seems like two thirds of all collisions when in reality, you know, distraction.

You know, now the government actually tells you 15 % of collisions are caused by distraction. That’s in the federal collision data. From our own data at Nauto, we know the real answer is almost 70% of the damage is distraction. So it’s really two-thirds of all collisions. And what’s fascinating about that is, you know, it’s eminently preventable.

And so one of the important nuances is we talked about collisions, not accidents. In common language, or common parlance, people use those two terms interchangeably.

But accidents suggest it’s a rogue freak event that you can’t do anything about. Collisions suggest it’s definitely preventable because two things shouldn’t be in the same place at the same time. And so I got into this idea of, okay, well, if it’s mostly distraction and it’s mostly preventable, what if we applied AI and deep learning not just to the outside space?

That was 2014, 15, was when all these DARPA funded research projects on autonomous driving were just starting to turn into commercial ventures. At the peak, there were 260 autonomous companies. There’s not that many anymore. But people were looking at this as a physics problem of how do I and my vehicle avoid the other vehicles and avoid trees and signs and other objects. And I really came at it from a behavioral point of view as well and said, it’s not just a physics problem, it’s a coordination problem because…

If I meet somebody else at a four-way stop sign, it’s turn-taking, right? It’s not just, don’t hit the other car, it’s actually communicating.

And so we got into building a system that could in real time see what was happening around the vehicle, what the vehicle itself is doing, and what the driver’s doing, or more importantly, what they’re not doing. We really care a lot about what they’re not looking at, because if there’s a hazard outside and the driver’s not looking at it, that’s a super dangerous situation. And in the process, we found that giving drivers warnings and feedback, could change their behavior much more dramatically than we ever thought possible.

Little funny vignette, when I founded Nauto as a company in 2015, we looked around and said, okay, what’s the high watermark of safety? What’s the best safety invention out there that we can set as a target if we gotta be at least as good as that, make a real difference in the world. And you pretty quickly find two things. First is the seatbelt.

It took a decade, but in the 50s and 60s to get deployed, reduced injuries and fatalities by about 20%. So that’s what we picked. And we said, let’s go beat the seatbelt over the next three years. We want to do it much faster than a decade. We set a goal of let’s beat the seatbelt, 20 % loss reduction. As it turned out accidentally, we completely sandbagged the goal, sandbagging, of course, famously that you set a goal that’s too easy to achieve. Because we’re now routinely eliminating two thirds of collisions, 65 % or so of all the collisions. And the real secret is…

Most people, once you’re past the teenage years, they know how to drive. So if they’re seeing the danger, nobody wants to get into a collision, right? Nobody wants to get hurt. Nobody wants to hurt anybody else. If you see the danger, you know, in most cases, what to do. And you can take the right corrective action, whether it’s swerving, breaking, slowing down, signaling.

And so really, most collisions are from threats that people didn’t realize were there. I call it unconscious risk. If there’s a pedestrian crossing, but I’m looking down on my phone, I don’t see the pedestrian and that’s when I might hit the pedestrian. If I give you a warning, you know, hey, pedestrian ahead, look up or an alarm. You look up, you immediately go, oh, and you know, you’ve got that freeze reaction of, oh my God, this is dangerous. And then you swerve or brake and kind of avoid it.

So the reason, the real secret to how we can get two thirds reduction is we’re not actually teaching people how to drive from scratch. We’re just helping people be better drivers. And there’s a very fine line, of course, between the right alert at the right time, where you go, oh my god, versus backseat driving, which is nagging you about stuff all the time. And so we’re very careful to intervene very selectively when you really didn’t know that was there.

Pete Miller [11:19]: That’s fascinating. You know, as somebody who had five kids, I can understand the distracted driving because many is the time I had one hand on the wheel and one hand in the back seat trying to separate kids, right? So, I’m sure we can all relate to that. Can we just dig in a little bit into that disruption?

And can you just talk a little bit about how Nauto goes about doing that to prevent collisions?

Stefan Heck [11:51]: Yeah. So we used a research data set from Virginia Tech Transportation Institute at the beginning, which has, they hired grad students to basically drive around while measuring what they were doing, but 2 ,600 vehicles or so all across the country. And it was really the only research data set that showed what causes collisions. Because the other data, as I said earlier, is from police reports, which is not very accurate because the police reports what they find after the fact.

And that data already showed the distraction was much bigger than the federal government thinks. And so it was the starting point for us to prioritize and say, okay, what risk factors do we have to detect in order to prevent collisions? And that was really our quest. How do you detect things before the collision?

In simplistic terms, I describe Nauto as, you know, it’s a computer vision system that looks at the road, looks at what the driver’s doing, looks at what the vehicle’s doing. And in real time, essentially looks five, six, seven seconds into the future and says, okay, this situation, is it fine? Is it a little risky or is it headed in a really bad direction? And if you’re driving on a sunny highway and there’s nobody in front of you and you’re driving along, let’s say at the speed limit, it’s patently fine.

And so that’s great. If you then have individual risks, I’ll give a trivial example. You know, somebody smoking, most people don’t realize, it increases your odds of collision by about 40%. Because you’ve got a hand occupied, you can have ash fall into your lap, you know, you’re cognitively a little distracted by the cigarette. So there’s some risk there. Similar to, you know, listening to the radio, very modestly elevated risk.

What’s called cognitive distraction, you know, thinking about the conversation you had with your, you know, your spouse, for example, right? Those are all relatively mild risk factors. Then if you get into more serious risk factors, you know, you have pedestrian crossing the road. Turns out that situation is about nine times riskier than the normal driving, because now you’ve got to make sure you and the pedestrian, you know, communicate and coordinate so that they get across the road, you yield or they wait for you to pass a number of ways to do it.

If you then talk about some of the really high risks, you know, holding your mobile phone and texting or looking down or navigating on a map, now you’re talking about something that’s 23 times riskier than normal driving. So you’re getting pretty dangerous. And those are the kind of situations where we start intervening.

But here’s where it gets interesting. And this is really part of the uniqueness. You know, we at this point now, but nine years in, to Nauto’s development. We look at over 30 different individual risk factors. And we use that long curve of risk factors and we’re kind of gradually adding more and more individual risks we detect. But the real breakthrough was about two years ago when we said, we actually want to look at combinations of these factors. Because if you’re looking down at your phone while you’re on an empty road and there’s nobody in front of you, it’s definitely less, it’s not safe, but it’s definitely less dangerous than if you’re doing that in downtown New York when there’s, you know, 50 pedestrians on the road in front of you.

And so what we found is that combination, pedestrians on the road, which is nine times normal risk and you texting on your phone, 23 times normal risk. They don’t add, you know, to expect the combination to be 32 times normal risk. Turns out it’s 2 ,791 times normal risk. So, you know, when you get risk factors in the thousands, you’re going to have a collision. Your luck’s going to run out. And so those are situations where we intervene very aggressively.

We structure the interventions to avoid this feeling of back seat-driving. We do little nudges if the situation gets above a certain risk level. You know, they’re literally like, for example, for speeding, it’s just like subtle cue of, hey, you should be slowing down. As the risk gets…more serious, you know, if you don’t see a pedestrian crossing, for example, you’ll get a voice feedback and it’ll always give you the correct behavior, you know, watch out or pull over to use your phone or pull over to take a break if you’re drowsy. And then in the most acute situations, if you’re really about to hit something or someone, it’s an alarm, it’s a collision alarm. Those are pretty rare. Most people don’t get into that situation more than, you know, every couple of weeks.

So you’re not going to hear a lot of those. But obviously in those situations, it’s critical. Turns out that even processing speech takes an extra second compared to hearing an alarm. In an alarm, we all immediately jam our foot on the brake and swerve.

And so most of that 65 % collision reduction we get is really detecting combinations where the driver’s looking the wrong place. And it can be innocent, right? They’re looking down at their map. It can be, you know, crazy risky, like they’re watching a YouTube video on their phone and looking at the YouTube video. But that, combined with the lights turning red, there’s a stop sign coming up, there’s a pedestrian crossing the road, there’s another vehicle in front of you braking, right? All of those create this kind of immediate hazard that’s super dangerous.

And through that, we can literally eliminate collisions. We’ve avoided in the last 18 months over 30,000 collisions for our customers. Not all would have been serious injuries and fatalities, but at least significant property damage. And if you just look at the average statistics, that’s about 30 lives saved, that people that are still walking this earth, because we helped the driver avoid an imminent really really serious collision. And that’s what we’re here for. It’s, you know, we’re on a mission to make people safe.

Pete Miller [17:26]: Yeah. That’s amazing. Um, 30 people are walking the earth. That’s gotta be very, very gratifying. Right. So let’s just, um, there’s, there’s a lot there, right? I, I can remember one time I was driving. I was, this was, I was in an Uber going from Chicago, well, hair down to downtown Chicago. And the guy, the Uber driver was watching the world cup soccer.

And I was like, yeah. And I was chirping at them, like your device. I was like, stop, this is not safe. So, um, but tell me a little more about your device. It sounds like you’re, you’re processing whether some, you know, images, whether somebody’s drowsy or not, what’s in front, um, what the speed is. So can you tell me a little bit about that technology?

Stefan Heck [18:18]: Yeah, it’s a sensor package that sits on your windshield, much like the built -in collision warning system, so you get with newer cars. If you buy a new car today, you’ll have a box behind and above your rear view mirror. We retrofit ours in most cases because the vehicle doesn’t already have safety systems. As I mentioned already, we look at things ahead of you, so pedestrians, signs, red lights, other vehicles.

We look at what the driver’s doing, we look at what the vehicle’s doing, and we look at the history of that location and the history of that kind of maneuver. So we know if you’re overtaking the certain risks, if we know what kind of risk speeding, for example, adds to things that may be happening.

And really, really important, our philosophy from the beginning entailed two design principles. The first one was we want to empower the driver. And the second is above all, we want to do no harm. So how does that translate?

A lot of safety technologies show you visual warnings. And we experimented with those in our first year of NATO, but we found that visual warnings draw your attention to the screen that shows the warning rather than to the actual hazard. So the last thing you want to do if you’re about to hit a pedestrian is show a screen with a pedestrian flashing because then you look at the flashing screen rather than the actual pedestrian.

So all of our interventions are haptic and audio, haptic meaning, you know, touch feedback, because you can get those without diverting your eyes anywhere. And similarly, we don’t deploy any warnings or interventions until we’re sure we’re not going to make the situation worse. We really want to make sure everything we do makes the driver safer.

And the most important is the first point, which is empowering the driver. So what does that mean? We have a design philosophy that says, look, the driver should always get the option to self -correct. They should be the first to know.

So that’s very different from dash cameras out there that are really recording what’s happening for the supervisor to see afterwards. And it’s essentially a spy system that’s telling your boss, you know, where you screwed up. And it doesn’t prevent anything because it’s telling the boss, you know, a day later, right? And all it’s going to result in is you get called into an office and say you should not have done that. Our goal is really to tell the driver with enough seconds left on the clock to change the outcome. So we never show anything to a fleet manager or a fleet owner that we haven’t first given the driver an alert about a chance to correct. And what you see as a result is most of our drivers, 85 % of drivers, become top notch, top decile drivers on their own. No supervisor required, no, you know, no chewing out of anybody.

Because again, most people want to be good drivers. Nobody wants to be in a collision, right? So you get into a collision because you didn’t see something. And if we alert you in the right moment, you know how to correct it. Again, it’d be a little bit different if we were working with 16 year olds who are actually still learning how to drive, because there you make mistakes from lack of skill. But we’re dealing with adults who know how to drive. It’s really about that awareness. And closely related to that, our whole architecture is real time processing in the vehicle with the AI, with the neural networks first.

And then our systems can capture recordings, because for example, most drivers want a recording if there’s a collision, because they want to be able to prove, you know, it wasn’t their fault, or, you know, the light actually was green in my direction, the other guy was the one who ran the red light. And we work mostly with commercial drivers. And in commercial vehicles, you find this huge asymmetry, they’re actually responsible for very few of their collisions, generally less than 20%. But they get blamed almost all the time for any collision that happens. As I described it, if you’ve got a corporate logo on the side of your van, there’s a little banner underneath that says, I have deep pockets, please sue me, whether you know that banner is there or not. And so most commercial drivers are grateful to have a record of what actually happened because they will always get blamed because the company has a worthy target for a plaintiff lawyer, even if they’re not at fault. So.

But really important for us is that video and when you captured is first of all optional and secondly only event triggered. So we don’t do, you know, continuous spying on the driver. And that’s a really powerful difference to traditional dash cameras. It’s really all about giving the driver the right information with enough seconds left to change their destiny, change their outcome. You see that, you see the driver’s reactions. They’re all examples of gratitude. You know, there’s nobody in the car with them, but they will actually say thank you aloud, you know, to the system that just alerted them, or they’ll wave and they’ll smile. You know, these are incredibly gratifying examples. You see somebody who just avoided a collision and they go, oh my God, thank you. You know, even though there’s no one in the car with them.

Pete Miller [23:16]: I went to your YouTube channel and it was really interesting because it shows for our listeners, it shows little snippets of that and people going thank you as they avoid a collision. The other thing I just wanted to make sure I understood, all of the data and decision criteria is on the device, right? So it’s not as though you get a latency going up to the cloud or something like that. That’s correct. Is that correct?

Stefan Heck [23:27]: It has to be. It has to be. I mean, usually by the time a collision risk emerges, you’ve got four seconds, three seconds left. Again, we try to detect it as early as possible. It takes the average human driver about two -thirds of a second to react. So you’re going to lose that reaction time. And that, you know, if you start with four seconds on the clock, you lose almost a second. Now you’ve got three seconds to really do something, to either brake or swerve, slow down.

And you need that time to really avoid most collisions. So yes, the processing has to be locally done. That’s another big difference to other systems where it’s uploaded to cloud and then something comes back. It’s too late. Also for us, there are lots of situations that are dangerous where you don’t have a cloud connection. A lot of collisions in parking lots and parking garages. And if you’re underground, you’re not going to have access to the cloud.

So yeah, it’s all local that protects privacy, make sure it’s all real time. We do upload anonymized data because one of the founding philosophies of Nauto is really a kind of social mission, right? You want everybody to benefit from the learnings from every risk that anybody else has ever seen. And just like airplane manufacturers and nuclear power plant operators will share, you know, safety and risks with each other to learn and to improve their product. We do the same. We work closely with a number of automakers, Toyota, BMW, Stellantis, Ram, and GM, because you don’t want to have somebody repeat a collision that we could have prevented because somebody else already had that kind of collision. And so it doesn’t share any details about who was driving or what their name was or anything like that. It just says, hey, these were the positions of the different vehicles and obstacles.

And this is what it led to. And then how can you, and we use simulation to figure out, is there some other reaction that would have been better that would allow you to avoid that collision? And that’s really the goal.

Pete Miller [25:52]: Let’s switch to the insurers for a little bit. In talking with people that use technology and with insurers that are trying to integrate some of these newer technologies into their organizations, a lot of it seems to be dependent on how it fits into their processes. So how does Nauto work with insurers and where do you fit in the different processes in an insurance organization?

Stefan Heck [26:22]: Yeah, it has been a challenge because of what you said, right? Insurance is set up to basically transfer and pool risk and then price it. Most insurance isn’t really set up to reduce risk or to eliminate risk. And that’s a big disconnect between insurance and what we do.

We began most of our work with large commercial fleets. So think about the local utility that has thousands of vehicles. We work with the big package delivery fleets that have tens of thousands of vehicles all around the country. Those fleets are mostly self-insured. So they have their own internal risk team, and they carry the risk on their balance sheet. And so it’s very easy to work with them because every collision avoided is straight bottom line profit. It’s an expense they would have paid themselves. And if you avoid the collision, you save the money. Very obvious. And for our larger customers, we’re talking about more than $100 million in savings. So it’s very, very material.

As we got to working with midsize fleets who do have insurance from a traditional insurance carrier, Allstate or AIG or the Hartford, we tested a number of models. The one that I think insurance has traditionally been most comfortable with is I’ll give you a discount if you’re a safer driver.

And so they do, they discount 5%, 10 % if you show up with telematics or a dash camera on the premise that you’ll be safer. Now, most of that effect historically, if you look at Progressive snapshot, for example, has been people self-selecting in to that option who were safer drivers already to begin with. And of course, that game only works while the idea is new. After the safest drivers have been siphoned off by snapshots, the next set of drivers isn’t going to be as good.

And it’s really been interesting to watch insurance companies reaction when we show up and say, no, we can actually change the driver behavior. Their initial reaction is, they don’t believe that because they’re not used to seeing that. And then we show them data and we say, look, within 72 hours, the vast majority of drivers eliminated the vast majority of their risk. Literally 80 % of drivers eliminated 80 % of the risk. It’s phenomenal. And they don’t believe it. Their answer is usually, well, let’s try it out somewhere.

And then we do, and it shows that result. And they go, oh, this is interesting. So we’ve worked with a number of companies with this kind of incentive option. We are now engaging more deeply with a couple insurance carriers that I can’t announce yet, but we’re in deployment right now, where they’ve actually embedded it into their insurance offering. That’s a much better way to do it because it says, hey, the Nauto technology essentially comes packaged into your insurance, if you like, from a customer/insured point of view, it’s free. It’s not actually free, right? The insurance company is paying for it, but it’s free to the end customer.

So you go to your broker, you sign up for insurance, the insurance comes with the safety system. And the great part about that is you get much faster decisions. People get, it makes it much easier to buy. They don’t have to go through their own evaluation, vetting process. And it doesn’t cost you anymore. In fact, typically you get a better deal because most of those insurance companies will renew you at the same rate. In an environment today where commercial insurance rates are going up 18 to 20 percent per year. So you don’t get hit with that massive premium increase. The insurance company loves it because if we’re reducing 60 percent of the collisions, they’re dramatically more profitable. In fact, the kinds of results we get, often insurance companies don’t believe you can move combined ratio by 20, 30 points, but you can because you’re eliminating most of these risks and you’re eliminating the losses, the results.

Now the last part where we work with insurers on is the claim side because none of these technologies get you to zero. There still are claims. But the same technologies are really powerful in the claims part of the insurance process as well because we can detect a collision instantly. We can capture all the data of what happened right before and right after that collision. We have a little buffer on the device that basically keeps a running log of the last 10 seconds.

So if you get into collision, we can capture that data. And of course, we captured the 10 seconds after the collision to see, you know, did anybody get hurt or their injuries. And we make that the data available to an insurance company within minutes of the event. And what I think most people don’t appreciate is if you get injured in a crash, that first hour, this golden hour, is the chance to save your life and prevent, you know, permanent injury, brain damage. So if you know somebody just got hurt and you use that to dispatch an ambulance, get them to an emergency room, you can turn what might have been a life -threatening or debilitating event into an expensive hospital bill, but nothing compared to having your life ruined. And so we can not only exonerate situations where the driver wasn’t responsible, which again, it’s the majority for a commercial driver.

But even in a case where the driver was responsible, this is one of the questions we always get, why aren’t you going to prove that I was at fault at certain conditions? And the answer is yes, it will show if you screwed up and you were at fault. But because you have the data that you hurt somebody and you get them to a hospital and you can offer a settlement, pay for their medical bill, you will still have far less liability and damage than if emergency services were called.

They go hire a plaintiff lawyer four weeks later when they find out that their back is never the same and now it turns into a multimillion dollar lawsuit. And the difference in our experience is if you get them treatment right away versus you wait for them to sue you is 20 X difference in the final settlement cost. So it’s ironically, even when we can show you were at fault, you can save a lot of money by making sure you admit that right away and act on it.

And so we have yet to find an insurance company that sort of completely embeds that into their claims process. You could actually automatically adjust most claims. I don’t recommend that for injury claims. There should always have a human in the loop. But your fender bender, you know, I bumped into a mailbox, a lot of commercial vehicles back over mailboxes. It sounds funny, but you know, it’s easy, right? You see, okay, yep, here’s a check for a new mailbox. That stuff.

A lot of insurance companies, if the damage is less than $100, they’ll spend $200 processing the $100 claim. That’s kind of crazy. So that should really be automated. And then all the insurance folks can actually spend their time on the injury claim, where really it is important that somebody looks at it right away.

So I think we’re four or five years into the journey of working with insurance companies. I think the long -term model really is all digital insurance that is about preventing incidents. And it’s not just about pricing risk and transferring risk, but it’s about actually eliminating risk by preventing it. And yeah, no, and I think there’s a long way to go in the insurance model to really be able to do that, to predict, you know, this is how much I can prevent.

Pete Miller [33:37]: So you have a unique perspective, right? Because you mentioned the changing nature of insurance and you’ve been doing this and obviously you’ve done amazing work. But if you think about, if we just think forward just a little more in depth on where you see the industry can go with this preventative technology, what would be the two or maybe three things, factors or things that the industry could focus on, when it comes to these preventive technologies and what opportunities would present themselves?

Stefan Heck [34:11]: Yeah, I think the first is insurance has to incorporate these new preventative technologies. And I think you see that, you know, for example, in the home insurance space, you see insurance companies funding water detectors, smoke detectors, you know, fire alarm systems, right? I think that’s a great example. And in that case, it’s obvious these are rare events. They don’t happen very often. But the difference between detecting that your water is leaking and shutting it off in five minutes versus letting it flood your house is the difference between an entirely total house that you got to rebuild from scratch and a minor inconvenience.

That hasn’t caught on in the automotive insurance space yet, but same idea, right? If I get you a better vehicle with a better safety system or an add -on safety system like we provide, and we’re working with the vehicle manufacturers now to build Nauto in, so you can actually buy Nauto directly from Stellantis and GM, already pre -equipped. That’s a huge difference in terms of how risky you’re going to be on every single drive, every single trip you take.

And so, you know, the challenge for insurance has been twofold, right? One is they don’t necessarily know all the new technologies and how do you assess them? Insurance historically has been built on actuarial tables. I had an interesting discussion with the chief underwriting officer once where he said, we love new technologies. I said, great, let’s work together. He said, yeah, but where’s your 10 years of history of data? And I said, oh, you don’t really like new technologies then because we haven’t been around for 10 years. So, you know, that’s the dilemma, right? How do I have proof that it works? And one of the things that’s made that harder in the automotive space is if you look at the hard data of the collision avoidance systems, ADAS systems and vehicles, they’ve made modest improvements in the collision frequency, but they’ve added to the severity because…

The sensors are all on the periphery of the vehicle and the bumpers. And every time you hit something, it’s a super expensive repair now, right? Instead of bending chrome and sheet metal, I’m now replacing electronics, which you really can’t repair. They’ve got to be replaced. And so most insurance companies tell you, hey, these collision systems have actually increased my bill because the severity increases offset the frequency decrease.

What that doesn’t take into account is that it’s been known for decades that somewhere between 93 and 96 % of all collisions are human error. Different research studies show slightly different numbers. But anyway, the reality is most people know nearly all collisions happen because of human error, right? The number of times you get a collision because something breaks on the car, you know, that the axle breaks or some system fails is super rare. Like that’s not going to be the main reason why you crash.

And surprisingly, none of these safety systems have looked at the driver. That’s about to change. I think the US insurance system hasn’t really woken up to this yet, but Europe passed a general safety regulation 2 .0. And starting with the 2025 model year, cars and trucks, every vehicle has to have a safety system that not only looks outside of the road, but looks at what the driver’s doing and starts to get at these human error situations.

And they’re specifically also designed to prevent pedestrian collisions, which are the largest dollar loss drivers. That’s where you really get large medical bills because vulnerable road users get severely injured. And as you look at those systems, you know, it’s going to take a generation or two to catch up to the kinds of things we’re doing already with the deep experience we have. But over the next two or three years, you know, maybe not in 25, but with the 26 model year you’re going to see dramatic reductions in these collision rates. And it will actually have payback.

And so a good insurance company will want to incorporate that. We’ll want to build that into their pricing for the vehicles that they insure. And then on the back end, we’ll need to set themselves up to ingest all this data. You know, the old model of, I went through this, I got rear -ended while I was waiting at a stop sign by an old gentleman in a pickup truck.

And the process is you get a phone call and you get interviewed. What was happening? Where were they? What’s going on? And then it’s he said, she said, luckily he admitted it was his fault. He said, I didn’t see the stop sign. I ran into him. So it wasn’t a big issue in my case, but if it’s a he-said she-said situation can be very challenging for the insurance company. In these digital systems, there’s no ambiguity, right? We measure the distance, the type of vehicle, how far away they were, who was going what speed, who was paying attention.

That data is all there. So there’s a lot of change coming to insurance and the old models, you know, will linger for a few more years, but the disruption is much bigger. And a lot of people I talked to assume, oh, that will only happen 10 years from now when autonomous vehicles really kick in.

The reality is, autonomous vehicles get you 90% reduction collisions, maybe 95. We’re getting 65 today. We think the theoretical limit for our kind of human assistance systems is somewhere in the 85 to 90 range. So you don’t actually have to wait for autonomous vehicles. You can get to 90% reduction within a few years.

And if there’s an insurance executive who doesn’t believe that 90% reduction in collisions is going to change their business, then they should look at their financials one more time.

Pete Miller [39:32]: I mean, your model, I think, is amazing. So it’s really fascinating to me. So I appreciate it.

Stefan Heck [39:39]: If I look at the people that we’re competing with for fleet’s attention today, there’s a lot of people just doing asset tracking telematics and they’re claiming safety benefits. But if you talk to the fleet to deploy it 10, 15 years ago, they saw very modest changes. Really what you change is you stop people from speeding and from harsh braking. And that’s it. But you don’t you don’t reduce pedestrian collisions. You don’t reduce distraction.

And then the other technology It’s getting a lot of attention in the last two years is cameras Because we all got used to having cell phone cameras with us all the time So people putting cameras into the car you mentioned the uber driver most uber drivers I see have gotten themselves a dash camera and, and it works great for that one piece that I talked about on the claim side, right? You have a video record of what happened so you can prove that it wasn’t your fault and you’re innocent. In the Uber case, it also protects them from inside attacks often. The drunk passenger that gets unruly. But it doesn’t prevent collisions that are otherwise avoidable.

A lot of my time is spent explaining the difference between AI in the vehicle and a camera in the vehicle. And it’s not yet clear to most people. AI actually prevents collisions, video just records what happened three seconds ago.

Pete Miller [41:02]: Well, Stefan, thank you so much. Thank you for your vision. Thank you for your entrepreneurship and innovation. And thank you most of all for spending time with us today and sharing your experiences. So really, really grateful.

Stefan Heck [41:15]: My pleasure and it was a wonderful conversation. I hope your listeners enjoy it. And we welcome other parties in this entire ecosystem, insurance companies, vehicle manufacturers, fleets, to participate because this is about continuous learning. The risk tail of really rare events is really long. And so it behooves all of us to share learnings about what causes collisions and how to prevent that. It’s a shared mission for all of us.

Pete Miller [41:44]: It is and thank you for being a leader in that vision. I appreciate it.

Stefan Heck [41:49]: Thank you.