#301 Hemant Banavar & Ryan Ennis: The AI Safety System Driving Toward Zero Harm

0:00There's so many situations where you have a need for safety critical environment and knowing precisely what's going on in your environment. In these situations, good enough just doesn't cut it. The difference between good enough and being really precise is the difference between actually reacting to an accident and preventing it. And that's where really we obsess over in terms of what we have built, in terms of the solution that we offer. In 2023, we had about 470 vehicles and now we're over 1 ,300. In 2023, I think we had 280 ,000 safety events across 14 different behaviors, touching your cell phone or following too close.

0:43Now with about 1 ,300 plus vehicles, we're in our third year and we're at about 25 ,000 events. So about a 98 % reduction while increasing our fleet in that short time. Hi, Craig. Really good to meet you and glad to be here today. I'm Heymand and I'm the Chief Product Officer at Motive and I've spent about over 15 years in the tech industry before Motive and have done product leadership roles across Microsoft, Yelp, Uber, and most recently before Motive as a Stripe. At Motive, our mission is to help our customers who operate in the physical economy to make their work safer, more profitable, and more productive.

1:38We serve 100 ,000 customers and about 1.3 million drivers and help them run their operations across multiple industries. And of course, Ryan is one of our customers and and looking forward to learning from him as well today. Our products typically are a multitude of products all across one AI-first integrated operations platform. We solve problems across fleet management, driver safety, workforce management, spend management, and AI vision. Well, I'm very thankful to be here, Ryan, CIO for Fusion Site Services, been in transportation for about 20 years, worked in the state of Tennessee for the state, worked for the Amazon from a compliance standpoint for all transportation groups.

2:38And now with Fusion Site Services, the responsibility of being and operating in people's communities is very high. And, you know, I really appreciated the vision and mission of Motive that was just said. And it's making sure that the roads are safe. That's ultimately my job, is making sure that we're a good partner on the road. Hemant, can you talk a little bit about, I mean, although it seems obvious enough, why good enough AI isn't acceptable in this kind of an application? Especially when it comes to safety critical environments like what Ryan and FusionSite operate in and many other customers of ours operate in may be driving on the road at a construction site, at maybe an oil rig.

3:34right like there's so many situations where you have a need for safety critical environment and knowing uh precisely what's going on in your environment in these situations good enough is just doesn't cut it the difference between good enough and being really precise is the difference between actually reacting to an accident and and preventing it to be honest and that's where really we obsess over in terms of what we have built, in terms of the solution that we offer. We have trained our system over billions of miles of driving, and we have built AI models, which actually deliver highly precise alerts in real time.

4:24So our customers can actually not only prevent, but also change behavior over time, right? And that's very important because is delivering an alert in real time when somebody is just about starting to get distracted is the difference between them actually potentially having an accident versus them correcting that behavior and being safe in their environment. Yeah. I'm just curious, how does the alert come through to the driver? Presumably it's to the driver or is it to the fleet manager? That's a really good question. So again, just to take a step back, the way our solution works is it's a combination of hardware, software, and AI all working together.

5:10So a lot of times what you can imagine is there's a driver driving their vehicle down the road. We have our camera installed in the vehicle that can look outside road facing as well as is looking at the driver at the same time. Our AI, we have built AI models that have been trained on, again, billions of billions of data points, and that is now able to understand the context that the driver is in. For example, it's like you can see that the driver is maybe hard braking and the camera can actually add the context and say they're hard braking because maybe somebody cut in in front of them, which actually they did the right thing.

6:00and that's a positive driving behavior and our camera and our AI running on that camera can detect that and say, this is good behavior versus somebody breaking really hard when they were actually looking at their cell phone, which is a distracted driving event, which our camera can quickly pick up. And at that time, what our camera and the system does is alert the driver about it. So they know in real time that, hey, you're distracted. You probably want to put your cell phone down which is a good reminder for them to get coached on. But also, if they keep doing that on a repeated basis, we have thresholds that our customers can configure to say, if this is a repeated behavior, notify my fleet manager who can now actually take action.

6:50And we do that by essentially taking that AI-generated event and then sending that notification over the cloud to the dashboard that the fleet managers are looking at so that they get alerted and they can take corrective action at that point. Yeah, and the alert that comes through to the driver, is that audio or is it a voice or is it a beep? It's again, like all of them. So it's a beep followed by a audio alert and that the driver can see. Actually, like there's also like LEDs that will help them understand with the color of the LED, what's going on when they have an alert, right? The same way on the fleet manager side, they can actually get an email alert.

7:38They can get notified in the dashboard. Again, this is highly customizable on the back office side where the fleet manager can decide how they want to receive the alert for something that they need to know that they need to take corrective action. And as I understand, you have a group of people, of motive employees that watch, that check every alert, or do they sample check the alerts, or what are they doing there? And is that before, I'm sorry, events, that they check events when the system identifies an event? And are they involved before the alert goes to the driver? Yeah, let me talk about that a bit.

8:25So we believe that, again, we obsess over accuracy, and that is why we built our system the way we built it. So what we have is a fully integrated hardware, software, and human-in-the-loop validation happening all the time. The reason for that is we start with a new problem that we encounter with, let's say it is close following. When we start out, typically, when we start building a model, it may have a good enough accuracy and that's improving. We need the human in the loop to actually help the model get better very quickly. What we do is typically when we launch a new model, our human in the loop actually helps validate every single event that's coming out of the model to say, yes, this is good or this needs to be corrected.

9:27So what actually our customers see is the filtered version and the highly accurate version. But that human provides a constant feedback loop to our model development, which allows us to really increase the accuracy of our models very rapidly. So over time, as our models mature, our models are highly accurate already on the edge. They don't need a lot of human intervention, but as we are kind of starting out, the human validation helps us provide a constant feedback loop to the model to get the model to be highly accurate. Yeah, and just to dive into the weeds a little bit, is that through a reinforcement learning loop?

10:12That is correct, yeah. This is a recursive learning loop that we have built in a way that constantly provides feedback to the model So the model continues to keep getting better. And this is where we can build models and we can make them highly accurate. So we take away all the guessing and the noise from the system. And what our customers like Ryan and their teams get is a highly accurate version of the view of the world. right so they are not spending um hours and hours sifting through all these events generated by an ai that they now have to go spend time to figure out which one do i react to and which one do i not they can actually trust that what's actually delivered is something that i need to react to yeah yeah and that's an issue with any safety system ai or not is uh or you know false alerts and alert fatigue um so so balancing that accurate uh detection uh with alert fatigue by having the human in the loop you reduce uh that gap right yeah we reduce the alert fatigue and And also, honestly, more than the fatigue, it also essentially erodes trust in the AI if you are doing this, right?

11:50So for us, we really strongly believe that AI that you use in these environments needs to be highly transparent. And really, there needs to be clear accountability around it. That is why we built our system the way we did. And we make it really easy for our customers to evaluate our AI in real world environments. And we really support and encourage our customers to go run our systems side by side in real environment for them to actually see the value of the AI that we deliver. Yeah. And Ryan, why did you pick Motive and who were you working with before? I mean, you don't have to mention the company's name, but were you working with a camera-based AI safety system?

12:47Yeah. So like with my experience, we've tested out every single camera system in the industry. And so coming to Fusion, they had a specific one I would prefer not to mention, but what I would say is, yes, we looked at that technology and, you know, we're talking about AI as a tool that just saves countless hours of review, as well as enhances the businesses to scale and address risk before those risky behaviors turn into accidents, for say. There's the root of the technology, which is even knowing where your vehicles are, if they're doing the basic functions that are expected in the day-to-day.

13:29So there was an evaluation that went through, I'd say, three to five months of side-by-side testing, whether it be evaluating the AI, evaluating the GPS location ability of those specific tool sets. Some organizations, Craig, don't have the ability to manage a diverse fleet just from the tool to connect to your vehicle's computer. So, yeah, Fusion chose motive based off of that review process. Yeah. And did you do that side-by-side test with another solution? 100 % by about six different solutions, yes. Yeah, that's fascinating. And Hemant, how do you account for the big performance gaps among dash cam providers?

14:29And are you benchmarking primarily on the human in the loop feedback or how do you benchmark? One of the studies that we actually did was with Virgin Air Tech Transportation Institute. And we have done this not only with third-party benchmarking solutions, but we have also encouraged our customers to run these side-by-side evals for themselves. We believe that what's going on in the industry is the development and the pace of AI is really rapid and it hasn't really, actually it is moving so fast that the benchmarking needs haven't really kept up with that pace. What that leads to is just a lot of promises and there's no clear criteria for companies like FusionSight to decide what is the best solution for them.

15:30And that is where we believe really making it transparent, making it easy for customers to benchmark goes a long way in actually giving them proof points instead of promises to know for sure what they're getting out of this system. We believe the innovation in this industry necessitates that we actually have very easy ways to benchmark. We want to make it super easy for our customers to do whatever kind of evaluation they need. We typically will partner with anybody who wants to evaluate our technology in whatever shape or form we can help them set up the evaluation and also run any side-by-side tests if they want.

16:17And what's the idea of zero harm and how do you get there? Yeah, look, I think it's very related to this idea of, at the end of the day, having transparency. The idea of zero harm is essentially working towards a goal of having zero preventable accidents on the road. Actually, what you see today in terms of accidents, there is still so many accidents happening on the road. Just in 2024, I think there were 40 ,000 fatal accidents on freeways. And to reduce that number, we think technologies like ours play a major role by, one, identifying unsafe driving on the road, helping in real time changing that behavior.

17:11But in order to do that in a meaningful way, we have to have very clear benchmarking that the industry can adopt. And all of us can benefit from that transparency, not just us and other players, but also our customers, because they know what they're going to get out of the solutions that they're purchasing. Yeah, that's fascinating. And Ryan, what sort of impacts did you see when you switched to Motive? So again, we went from a system where you didn't have AI capabilities to a system where you are accurate AI capabilities to a system that had accurate AI capabilities. And so it really, you go through phases, but I would say from 2023 to year to date, you've seen about, very specifically in 2023, we had about 470 vehicles and now we're over 1 ,300.

18:11In 2023, I think we had 280 ,000 safety events across 14 different behaviors. So that'd be like touching your cell phone or following too close. It's about 280 ,000. Now with about 1 ,300 plus vehicles, we're in our third year and we're at about 25 ,000 events. So about a 98 % reduction while increasing our fleet in that short time. Year over year, we have seen internally probably right around an 80 to 90 % reduction year over year. And I think the other important piece is that it's not just about behaviors. It's also about those results. And so I had mentioned that we went from 76 claims in 2023 to one in 2025.

19:06But with data and with accuracy that motive provides our fleet, we have probably one of the most diverse fleets in America, which means you have a very diverse population. You have a very diverse training program that's built for these individuals. we were able to build ancillary like driver recognition programs based off of Motive's safety program. That's part of the total motive package. So we've seen great success. Yeah. And presumably insurance premiums came down if you went from 76 claims to one in a year? Year over year, we've seen about a two to$2.5 million savings on our insurance premium.

19:54And the way that insurance works is it's about a five-year rolling period. And so that means next year you'll be saving a significant amount more if you're trending two years in, if you're trending three years in. So it's more than paid for itself. We could not be happier with the results, which ultimately correlate with, you know, we operate in communities. And we're very thankful and appreciative of working in those communities and being good, safe partners where we operate it. And again, Motive allows Fusion to act responsibly prior to those events, those harsh events from actually occurring that result in loss runs.

20:38And I think that's why this AI technology is so important to trucking organizations is you're going to get to see the trends. you're going to get to see the behaviors before they result in those lost runs. I'll go back to one thing that was said earlier, is those alerts in that Virginia study against those competitive devices, they're going off within less than a second of, let's say, a driver touching their cell phone or a driver not buckling up. And if you think about that, it literally stops that behavior from continuing, which puts that driver back in a safe situation. And so you're reducing the opportunity for that accident to even occur.

21:25So hopefully that's added value, Craig. Yeah. And is there pushback from the drivers? I've read that the drivers don't like these systems. Yeah. You know, I've installed this type of device for very, very large fleets. And I would simply say this, it's about how you as an organization roll it out. It's about how you treat your drivers with respect. And it's ultimately about setting those expectations because it's not just the company that's liable for those accidents. It's also the driver. And so keeping that driver in a safe position where you're allowing them to succeed, allowing them in a promotion path.

22:08Like I mentioned earlier, Fusion Site Services have paid our driver population over$700 ,000 this year to drivers that exceed our safety expectations. So our drivers get to make a significant amount of money from just going a little bit slower, doing it a little bit safer. We find about in a non, we have about 200 CDL drivers. So in a majority non-regulated fleet, we have about a 99 % adoption. And the most unique thing is that when we encounter new drivers, like 100 drivers that we're bringing into our fold, in about two weeks, those drivers have adopted to, we'll say, FusionSite services path via Motivestools.

22:55That means they're able to log in, they're able to get coached, they're able to be educated, and then guess what? They're able to get paid more money. So in a very, very short amount of time, you're reducing that driver's behavior by about 90 % on the areas that they may need to get coached in. Yeah, and that's interesting that in effect you share some of the savings that you get from Motive to the drivers as to incentivize them. And maybe I'll add one thing here to what Ryan just said. At the end of the day, this is about the technology is mainly about getting the driver home safe, right? And that's kind of what Ryan and his team are essentially deploying these solutions for.

23:43And when they see that, when they see the impact that this technology is making as they are driving every single day, it's not throwing alerts when it shouldn't. and it's actually helping them change behavior to be a better driver, to get home safe, they immediately see the value that it's delivering for them, which is where this is about helping them. And the more accurate this technology is, the easier for them to see the value at the end of the day from that technology. Yeah. And this is, as you were saying, it can apply across industries. How broad is the horizontal is, or the industries that you're serving?

24:34Haman. Again, think about this as everybody who is in the physical economy, right? Anybody who has vehicles, assets and workers operating these assets that's who we serve and we our technology can help them because look I think at the end of the day when somebody is running physical operations they all have the same problems they want their physical operations to be safe safety can mean driving a vehicle on the road, it can also mean at their job site, how do they keep their workers safe, right? They want their operations productive. They want to make sure that when their driver or the worker on the site, when they're doing the job, they have the best tools so that they can do their job effectively.

25:30We can, our technology enables them to do that because we can give them insight into everything that's happening within the vehicle, within the asset, so they can make sure that they can increase utilization of these vehicles. And then lastly, it's also about making this profitable for them, right? So a lot of the AI, we also not only develop AI to help with safety, but also improve efficiency. We actually, our AI is able to find wasteful spending or underutilization or think about idling when it was idling a vehicle for long duration when maybe it was not necessary. All of these insights at the end of the day help them increase their profitability as well, right?

26:23So what we have built applies to a broad set of companies who are operating operations in the physical economy. And who would do that? Who would set those benchmarks? Would it be an industry association? Would it be the departments of transportation in each state or the federal level? It's a really good question. I would say it's all of them. And actually, again, I would go back to other industries where this has happened. right there's a lot of third parties who are willing and eager to provide their expertise to even come in and say what are the right metrics what are the right benchmarks again vtti is a really good example of such a independent third party who can play a role in establishing these benchmarks and we see a real value in having these third parties play an active role in creating that transparency in this industry so customers in this industry can benefit from that transparency.

27:34And then how do you balance innovation speed with the need for more rigorous testing and the validation because you don't want to slow down innovation? You know, you don't want to slow down innovation, but also you want to make sure that your innovation is actually having the right impact. And I would say in this physical economy, a lot of your innovation is safety critical. And when it is safety critical, you cannot move fast for the sake of moving fast. You want to have real data, real benchmarks behind it before you actually go deploy it. So we believe that you can actually innovate faster if you have the right framework for benchmarking because it allows everyone to actually know what you are optimizing towards and go race towards that faster than each one of us coming up with different ways on measuring your own success right uh would an independent validation either from some official body or from some industry uh body

28:57using benchmarking that's kind of recognized and agreed upon. Would that make a difference for companies? Yeah, you know, I think that right now, what a lot of, I'll say people in my role or my previous role, so safety specific in trucking, really look to the SMS mythology and the seven basics from FMCSA to really be the core infrastructure on how to manage your fleet, your drivers and whatnot. But to your point, I think one of the challenges that our organization have is to have an internal asset or to your point, an external asset that's able to review the capabilities of these systems to really have like a top quadrant or top tier provider for different pieces of their technology.

29:47I think, you know, we're often used by motive to assist with maybe an organization that has that mistrust that you're speaking about. Craig has had maybe some bad experiences. And, you know, when you sit down and talk to them about, well, this is how we use the tool. This is how our drivers feel. This is how our general managers slash operations feel. You really start to see the buy-in. And again, I think it comes back down to the data where that third party could really do that validation. I think that the most challenging pieces of implementing this tool not only reside in a company supporting that technology, right?

30:31We always talk about a technology supporting a company and letting it achieve its objectives, but it's about building those ancillary items around what do you do when a driver has this behavior three or four or five times. And I think that's when like NHTSA or, you know, the seven basics or S and anthology can really lend a hand to a trucking company via an insurance company or whatnot saying like, Hey, look, if you're having a driver fall into these three or four categories, maybe they're not insurable, or you have to pull them off the road for a period of time and put them through a training program.

31:07So, um, I, I have always wanted more partnership with the regulators like FMCSA, DOT groups. It's always very interesting to me when the regulators take a step back and either don't provide guidance or the guidance is loose up to interpretation. And again, in this safety sensitive situation, a lot of the measurements of these systems, just for example, when should close following be an alert? you know, well, those are down to friction ratios. Those are down to how much that truck weighs. Those are down to, you know, the weather and the elements in that area. It's not easy, but I think it's a topic that could totally be tackled by a joint partnership with technology partners, with fleets and with our regulatory groups.

32:02I think coming together for something that is as serious as driving over the road in very heavy vehicles every single day is a topic that could be solved. We could see this being solved. Yeah, and this is not prohibitively expensive. Is that right, Hamant? A lot of the innovation that has happened in technology, both on the hardware front in terms of the amount of compute that you can pack into a device that is sitting in a vehicle and running with the power constraints that it has available in a vehicle. There is so much advancement that actually the costs have gone down. So you can actually have a pretty capable device there running AI models inside a vehicle.

32:56And same thing on if you think about what's happening with AI everywhere, the cost of building models, deploying models is rapidly going down. So overall, the ROI here for a solution like what we are describing here, what Motiv offers is very significant. And I would actually call on Ryan to maybe speak more about the ROI here, as you see in your case. We believe that the ROI here is significant. But Ryan, please. I mean, there's tiers. We talked about insurance premium in a second ago is about$2.5 million year over year, which the cost of motive is significantly below that. I would say then you have additional attributes that motive provides, such as utilization, as was mentioned.

33:49So that would be the drive time from service route to the next service site where you'd be collecting revenue down to idle time. right? I think, for example, we've used 1.6 million gallons of fuel. We've idled for 330 ,000 gallons of fuel. There's an opportunity percentage in there that you would be able to rake back for maybe periods of time that we're not revenue generating. Just to be clear, I think that's one of the main distinguishing attributes for motive is that it is simply not just a camera that is judging a driver on 18 behavioral attributes. It's measuring the distance, the speed, the trips, the idling, the hours of service for that driver, which are regulated in some areas and non-regulated in other areas, as well as those actual trips for revenue that you're generating.

34:52So there are probably six to seven different ways to use Motiv and each one, and I would say the lowest one, would be able to pay for the actual cost of the Motiv device. Last piece, I would say, you know, we talked about like the agility of Motiv. You know, I used to have to pay someone to install Motiv into our truck, cost about$250 per installation. It takes me about, and I'm not mechanically inclined, it would probably take me 10 to 15 minutes to install Motive on any truck at any time in any environment. So I'm able to save a significant amount of money because the device, the hardware is very intuitive to install to your vehicle.