Stefan Heck, CEO and founder of Nauto, says that AI will be able to reduce vehicle collisions by 90 percent, outperforming the seatbelt in terms of reducing fatalities and serious injuries. California-based Nauto, founded in 2015, has built an AI-powered safety system for commercial fleets that monitors drivers and roads to predict and help prevent crashes. McKinsey has a small equity stake in Nauto.
Heck, who was previously a senior partner at McKinsey, sat down for a conversation with McKinsey Partner Sebastian Kohls at ITC Vegas. The following transcript has been edited for clarity and length.
Sebastian Kohls: What are you most excited about in the mobility space?
Stefan Heck: Mobility is evolving very rapidly. I’m excited about a whole new level of safety system integration into the vehicle.
I think we’ll continue to see the evolution of mobile phones and how they interact with cars. This was the year when a number of autonomous vehicle (AV) companies reached a critical milestone of material deployment. It’s not at full scale yet, but we’re seeing trucking companies running up and down Interstate 101 with commercial loads and robotaxi companies expanding paid service beyond their initial home cities.
It’s also a year of significant regulatory intervention. In Europe, Euro NCAP and GSR 2.02 are generating requirements for new safety capabilities in vehicles—pedestrian collision avoidance, driver attention monitoring, and more.
Sebastian Kohls: Speaking of safety, we’ve seen a deterioration in driving behavior. Many insurers see that reflected in their data. As more sensors are available, you’d think risk would go down and accidents would decrease, but in many cases frequency has gone up. Where do you think we’ll start seeing a counterbalance that drives safety up?
Stefan Heck: This has been a year of contravening factors. We are definitely seeing more adoption of safety systems. ADAS3 penetration is up in long-haul trucking and in consumer vehicles. It’s spreading from a luxury feature you’d see in a high-end car to a mainstream feature.
At the same time, the pandemic clearly showed that driving behavior got worse, with both speeding and distraction up significantly. People also talk about repair costs getting worse, not just because of general inflation, but because many sensors are now on the periphery of the car, where damage is most common. There’s a weird mismatch: The sensors designed to prevent a catastrophic event—running off the road, rollovers, pedestrian collisions—are getting damaged in fender benders or parking mishaps. What was a minor incident becomes a huge repair bill. Coming back to your question, I think that over time, though, we’ll see a tipping point.
With AI, the potential to change behavior is phenomenal.
Even as new vehicles with improved safety features are rolled out, it takes ten to 15 years to work through the fleet and retire older vehicles. Until then, there are lots of cars without safety systems. Another reason is that awareness is building about how bad distracted driving is and how serious these safety issues are. Insurers are seeing rising costs and starting to look for other ways to change loss, behavior, and incentives.
With AI, the potential to change behavior is phenomenal. In the history of automotive safety, there have really been two game-changers. First is the invention of the brake in 1902. Second is the seatbelt, introduced in the 1950s. Wearing a three-point seat belt in the front seat of a car has resulted in an overall fatality reduction of about 45 percent.4 Fascinatingly, nothing since has equaled that; airbags, backup cameras, traction control—each accounts for a reduction of only a few percent.
AI is going to be the first technology to far outstrip the seatbelt. On average, in our fleet deployments, damage from collisions has decreased 65 percent and fatalities have declined more than 80 percent, double the effect of the seatbelt. Funny story: We set a goal 11 years ago at the beginning of Nauto to try to beat the seatbelt, not realizing we could go way beyond it.
AI isn’t done. AI can predict the next six or seven seconds and avoid collisions before they happen. I think AI can take us to 90 percent collision reduction, maybe even high 90s. Nothing gets you to 100 percent because of freak corner cases and interaction effects among multiple vehicles, but we’ll see big progress over the next decade. The impact on insurance will be substantial. Insurers will be able to prevent collisions, not just cover and repair them, and will be able to understand what happened without interviewing multiple people, thanks to instantly available data.
I think AI can take us to 90 percent collision reduction, maybe even high 90s.
One of the big things we see is that we can notify a fleet within minutes of a major event. Injured people can get to a hospital quickly. Nobody wants a lifetime disability; you want treatment and help.
Beyond the frequency reduction we’ve seen from telematics, ADAS, and seatbelts, we can push much further—reductions of 60, 70, 80 percent in fatalities and serious injuries—and also tune the AI to reduce severity. For example, we focus more on avoiding pedestrian or bicyclist collisions than on avoiding backing incidents, but many fleets say their biggest problem is backing into things. In frequency terms that’s true—30 to 40 percent of collisions are from backing into things. But most of the time you’re backing into a loading dock, a bollard, or maybe, worst case, another car. The damage is usually relatively low. If you take vulnerable road users, it’s the exact opposite: Only about 2 percent of collisions involve hitting a human, luckily for those of us who like to bike and walk, but those account for 40 to 60 percent of all damage. That’s why building in pedestrian collision avoidance features can have a disproportionate impact.
Only about 2 percent of collisions involve hitting a human, luckily for those of us who like to bike and walk, but those account for 40 to 60 percent of all damage.
Sebastian Kohls: We also talked about incentives for individuals. As you think about shifting from fleets into personal insurance, what needs to be true? What should insurers consider to design these programs?
Stefan Heck: Multiple things need to be true at the same time. First, whatever risk you’re detecting has to be extremely accurate. Right now, the world is full of companies saying “We do AI,” but not all AI is created equal. We’ve had Virginia Tech Transportation Institute do benchmarking studies and have seen differences of three to four times in accuracy, which is huge, and detection speed differences of five to ten seconds. Even five seconds while driving can be the difference between life and death.
That means the accuracy has to be superb. You also have to cover a broad range of risks, because one risk you don’t spot can still hurt you. Historically, a lot of technologies focused on what’s in front of you. The big change in the past few years is awareness that many hazards are inside the cabin: where the driver is looking, their alertness state, whether they’re handling a phone.
In our data, we see that 70 percent of all damage comes from mobile device distraction. Not 70 percent of collisions, because frequency patterns differ, but in terms of dollar loss outcomes, distraction dominates. It makes sense: When you’re distracted by your phone, that’s when you plow full speed into what you’re going to hit; you don’t even initiate braking, so the damage is huge.
In our data, we see that 70 percent of all damage comes from mobile device distraction.
Can you solve that through the phone, the car, or behavior insight? All of those are levers. But then you have to intervene in a way the driver understands and appreciates. The driver needs to recognize what you’re alerting for and that it’s real. We got early requests to warn people about dangerous intersections where bicyclist collisions often happen. That’s a terrible idea. An alert that there might be a bicycle here is frustrating. An alert about an actual bicyclist you’re about to hit is deeply appreciated.
If I alert you to something you can’t see was there, you’ll be annoyed. If I alert you a lot for minor things, same problem. Another failure mode is alerting for something obvious that you’ve already seen. That feels like backseat driving.
This is true across AI domains. AI can be helpful by being super accurate, but it also has to help the human experience. It has to alert you with enough time to act, in a way where you see the danger. The best experience for the driver is when they think, “Oh, that was dangerous,” and they learn and change.
The best experience for the driver is when they think, “Oh, that was dangerous,” and they learn and change.
The old belief in insurance that you can only select and price risk isn’t true. You can change risk. We see 85 percent of drivers drop 90 percent of their risky behavior in a matter of days, because they are motivated and don’t want collisions, and because we help them recognize risks they hadn’t seen before. Then you can layer in incentives. We do leaderboards and awards for best driver of the week or month. Fear of loss is also compelling: “You get $100 this month unless you screw up.” People would rather avoid a loss than get a gain. Social rewards are just as powerful a motivator as financial rewards.
Sebastian Kohls: As an industry, when we change individual behavior and the underlying risk, we change society for the better. AI will drive autonomous vehicles, which may be safer than many human drivers, and AI can modify human behavior. Five years from now, will more risk be taken off roads through AV adoption or through behavior change?
Stefan Heck: If we do it right, behavior change, because you can roll behavior change out to almost everyone right away. Between phones, cars, interfaces, and social media, there’s a huge opportunity to change behavior at scale.
The number of AVs will grow, and eventually that growth will show up in collision statistics. I don’t think we see it yet because the number of vehicles is still small. Five years out, we’ll start to see AVs in insurance data.
The number of AVs will grow, and eventually that growth will show up in collision statistics.
It’s tricky, though. The best AVs today cause fewer collisions and have fewer at-fault incidents. But at a system level, they can also lead to collisions. They can brake fast and unexpectedly, leading to some rear-end accidents—not in the liability sense, because the person behind is still responsible, but in the sense that a human might not have braked that way.
Similar to safety systems that reduced frequency but increased repair cost, we’ll see reductions in some collisions with AVs, but until people get used to them—and until human drivers improve—we might see increases in certain types of collisions as well.
Overall, I’m a big proponent of AVs. They will help, but the full safety impact may take longer. There are ways to accelerate it: limiting the operational domain, for instance. If you make an area AV-only—say, a downtown with pedestrians and AVs capped at certain speeds—you could dramatically reduce collisions in that area. Another idea is lanes that allow only AVs or Level 3 driving,5 like HOV lanes today. Coordination among machines can reduce collisions more than having AVs interspersed with lots of humans.
Sebastian Kohls: Thank you for the discussion. I’m excited to check back at ITC next year to see progress, and hopefully in five years we’ll see significant reductions in the collision data from both AVs and behavior change.
Stefan Heck: You called the bet. We’ll see what happens.
