Why AI Needs Grounding

Joeri Billast 0:01

AI needs a ground to stand on. Why real-world decisions require location intelligence.

Meet TomTom’s Enterprise Location Lead

Joeri Billast 0:09

Hello, everyone, and welcome to the Web3 CMO Stories podcast. My name is Joeri Billast. I'm your podcast host. As you can see or hear, we are today at the Web Summit, Today, I am joined by Leo. Hi. Hey, Leo. How are you? Good.

Leo Sei 0:24

How about

Joeri Billast 0:25

you? I'm very excited, Leo. and it's good to hear that you also live in Portugal. Guys, if you're wondering who is Leo, Leo Sei, and, he just told me he changed, his function, but he used to be the chief product officer at TomTom. But tell us, Leo, what are you doing right now at TomTom?

Leo Sei 0:45

thanks for having me. So yes, my name is Leo Sei. I lead the new division of TomTom called Location Intelligence for Enterprise. Most people know TomTom as this little device that used to do navigation 20 or 30 years ago. but TomTom has been, for the past 30 years, really building map and digital information about everything in the world, whether it's the streets or restaurants or traffic or an accident or a road closure. we have a lot of data about the world, and so we see the agentic era as an opportunity for us to enable more enterprises, defense, governments, insurance, fleets, you name it, to use this data to take better decision, faster decision. so that's what I do, and before that, I spent 10 years at Google, so I have a pretty big, background in tech.

Joeri Billast 1:30

Wow. Indeed. That's, a very valuable experience When you think about it, TomTom is a data company. All the data that you guys have.

When Confident Answers Are Wrong

Joeri Billast 1:38

Now, the title of the conversation today, Leo, is AI Needs Ground to Stand On. So what does that mean in practical terms for an AI agent operating in the real world?

Leo Sei 1:50

So I think most of us by now know that they need to give context to their agents. most of us will give files, email, calendar. So you always need, to ground the agent in reality, and for the most part, this might be documents. But if the agent needs to help you make decision about the world, there's nothing today that allow an agent to understand what the world is- Mm-hmm and where things are. I'm gonna give you an example. If you're trying to find, the best location for your next charging station, you might ask for what's the best street in Lisbon based on, traffic and, impedance. And the answer might be, and I checked recently, one of the most major street in Lisbon that happened to be pedestrian for the past three years. But if the agent is not able to ground himself, validate, hey, is this street still accessible to car, or is between this site and the road actually an access, or is there a ditch? the answer might be, you know, presented with a high confidence and actually be totally wrong.

Joeri Billast 2:45

Absolutely. Haven't we all been there? Now, Leo, AI models are becomingly remarkably good at reasoning, of text, images, and historical data. what are they still missing when they have to make decisions about actual places, roads, people, and changing conditions? You already mentioned something, but, please elaborate.

Vectorizing The World For Agents

Leo Sei 3:07

Yeah, so there, there's two things that are interesting. First is understanding the world at training, right? So large model, they usually get retrained every couple months, and so understanding what the world at, at that moment is already a challenge. You might have heard of the emerging of physical AI, that is a very different type of model that understand the world. The model we use all today are large language models, so they're based on language. They will struggle to understand that this road and this other one that are stuck together are actually a bridge and a tunnel, and they do not connect. They will struggle to understand things about the physical reality because that's not how they train. So the first things that we're doing is bringing location data into an understandable world, which is called vectorization. So we're making our data understandable by agent. That, that's first for training, so at least they understand the reality principle when they get trained. But by the time you ask them something, the world has changed. Yeah. So okay, maybe a city hasn't moved, but a street might have got closed. Traffic change every second. restaurants might have been closed or opened or renamed. So we also help them, ground, and that's the t-the technical terminology, ground themself when you ask a question, they can quickly validate, "Okay, is this restaurant still there?" before I make an answer. "Is this street still there?" before I make an answer. "Is this, you know, traffic, road closure still there?" before I make an answer. So that's, that's what missing today, in the agent, and that's where the 30 year of experience we have in mobility data is very, interesting.

Joeri Billast 4:34

So yeah, you gave already a few examples, of a street being closed or things changing. Is there maybe another example that you think of, of an AI agent making a perfectly logical decision that would still be wrong because it lacks, accurate, location

Insurance Claims Need Place Context

Joeri Billast 4:50

context?

Leo Sei 4:50

Yeah. Let me give you another example. So we work with insurance at the moment to help them with what we call claim validation. So when you get into a car accident, usually you file a claim-

Joeri Billast 4:59

Mm-hmm

Leo Sei 5:00

and the insurance process this, and they actually sometimes have, people that miss context, let's say. Or, you know, they thought they were there and they were not, whether it's intentional or not. So there could be an example where they say, "You know what? I got rear-ended taking a left on this street at this time of this date." And the agent might think like, "Well, based on my training data, those two streets exists, so I'm assuming this is plausible." Now, if you combine this with the reality, we can say, "Well, on those date, that left turn was actually forbidden. and also there was no traffic. It was free flowing." So there's a lot of things that point to the fact that either the person, mispointed the claim or something else. We won't make any decision, but we can help provide accurate context, to the agent to help.

Joeri Billast 5:41

Like you said, the physical world changes constantly. How do you keep the underlying location intelligence fresh enough for AI agents to trust and act on

Keeping Maps Fresh In Real Time

Joeri Billast 5:52

it?

Leo Sei 5:52

Yeah, so it's quite interesting. We have, those 30 years of experience, and so there's things that don't change as often, right? the location of a city, the city hall, the hospital. So there's things that we have that won't move. But we also leverage an entire ecosystem. So there's a couple things here. One, we're part of a, a open source community called Overture that we co-founded with Amazon, Meta, and, Microsoft, where we'll pull a lot of the data we put together, thinking that, a map is not one company's business. The base of the map should be open to everyone. And so that brings a lot of data from the OpenStreet community and others. And then on top of this, we add a lot of data that we extract from various signals. So we have lots of partnership with the car maker, for example, where we get a lot of informations, from the cameras you have in the car. the camera record continuously. Everything get processed in the car. There's no data leakage, and then we get information about, hey, there's a speed sign here with that information for restriction for trucks. There's now a new road marking here. There's a closure here. We also have partnership with municipality that send us the scheduled work ahead of time that we compile. We have partnership with a lot of, application provider that get information about mobility. So we have a network of partner that allow us to collect, I think, a billion data point every day on mobility throughout the world. And through this, you can also, infer the reality of the world. If there's a lot of people going through here and we have no road on the map, then that can be a leading indicator for our operating team to say, "Hey, maybe there's a road now because it looks like-" a lot of cars are going there. Please go investigate, and if there's a road, add it." We have partnership with Ubers, for example, where they give us direct feedback. This road was closed. This one was open. So we have a lot of customers and exchange program where we keep the map as fresh as we can.

APIs SDKs And Agent Tooling

Joeri Billast 7:39

Okay The partnership's really interesting, but I guess there are also people now listening to this podcast that are curious about the technical stack behind all of that. So how do APIs, SDKs, real time data, and agent tooling come together to make location intelligence usable by developers and by AI systems?

Leo Sei 8:01

Yeah, so we abstract a lot of that complexity of the map making. so we have this entire stack that's building a digital representation of the world every day, and then we make it accessible to as many, a conduit as we can. So developer might be, used to API and SDK, so we have those. We're also now moving to tools that agents use, so things like MCP or what we call embedding- which is a vector representation of the data, which I was saying earlier is a better way for an agent to understand the data. So all of this help our data be easily accessible. And the interesting bit about it is in, in the previous times, if you want to use location data to make any sort of decision in insurance, in fleet management, in defense or, or urban planning, you have to have, s-special skills and special tooling. With agents, if they have access to the right format of the data and the right guidance- of how to use the data, anyone can suddenly leverage data in their processes. I'm gonna give you one example with insurance. again, going back to property insurance to move away a little bit from the mobility use case. Today, when you want to insure your house for flood or fire, most of the insurer, they will have kind of a risk at an entire zip code or area and say, "You know what? If you're in this rough area, here's the, the, the risk for flood, so here's your price."

Joeri Billast 9:21

Yeah.

Leo Sei 9:21

Now, without any special skills, we can tell them, "You know what? We can, plug your risk model with every address, distance and elevation from the nearest, water body or forest." Now, your model can say, "Well, my average risk is this, but this house is really far and really high up from the water, so maybe actually reduce the risk." You didn't need to understand how to compute this in the map. The agent does it for you, or we do it for the agent. So suddenly a lot of industries can get better, with our data. We already see we have, I think, ten billion, of, insurance premium today computed on our data, and we see customer getting 2% uplift just by aggregating something at zip code. Going down to the address will be massively beneficial for the users and for them. It's a win-win because if you live far from the forest, it, it's not fair that you pay, you know, the same risk as the guy that lives right next to the forest for the fire risk. so this is enabling everybody to use geospatial data without being the geo expert.

Big Use Cases Beyond Navigation

Joeri Billast 10:19

You just mentioned that multiple industries could benefit from it. So beyond navigation and automotive, where are you seeing the most interesting applications of location intelligence today, particularly in areas such as logistics, insurance, enterprise software?

Leo Sei 10:37

So we see insurance, as I mentioned, both in risk and in, claim validation. we see urban planning, city planning, government that try to understand how people move in the city to- understand, you know, what is the future of mobility, where to change things. So we work a lot with, those industries. fleet, as you mentioned, to have better planning, better warehouse location so that li- sits into site selection, helping people figure out where is the next warehouse, where is the next EV charging station, or even in, emergency services. We had a pilot, with the hospitals where in real time, based on traffic and other things, you can figure out which is the best hospital to dispatch your ambulance from for every emergency call. So there's really a lot of use cases, that we see.

Joeri Billast 11:21

And which use case do you think that developers, founders, business leaders are maybe still underestimating?

Leo Sei 11:30

I think the, impact of using geospatial data in decision is largely underestimated because it used to be complicated and expensive. When at work at Google, it was interesting to say that there is probably, over half of their requests that had a geospatial component, and I think that's true. When you think back of what you do in your day-to-day, how often does it rely on knowing where is something? It is very often. And so as you start to understand how now you can plug your agent, you don't even have to understand that, but you can plug your agent with the right data sources. I think there's a large underestimation of how powerful this will be in better, more accurate, faster decisions in a lot of businesses.

Joeri Billast 12:10

Yeah, you make a good point. Even I use my iPhone or my Android phone, you install a new app, it asks for location-

Leo Sei 12:17

Yes

Joeri Billast 12:17

for location data. I don't know always what they need it for, but actually it seems to be, to be, in demand.

TomTom’s Shift To An AI Layer

Joeri Billast 12:24

Now, TomTom is known by millions of people as a navigation company, but you are now positioning it as an intelligence layer for AI. What had to change inside the company to make that transition possible?

Leo Sei 12:38

Yeah, so it's interesting. I think a lot and very little at the same time. Mm-hmm. So we have... I, I think when you think about it, navigation was just a way to help people have a safer mobility. So at the time, the alternative was a paper map.

Joeri Billast 12:53

Yeah.

Leo Sei 12:53

So providing a software navigation was mind-blowing at the time. You fast-forward 30 years, everybody has an application on their phone for free, so that's not the most impactful. Now we embed directly in the car, we embed with the car sensors to provide, you know, more autonomy and better assistance. But at the same time, we, we, we've had this data, we've had this, geospatial DNA in the company. So that core hasn't changed. What changed was really the understanding and the opportunity that in, I don't know how long, maybe months, maybe a year, the primary consumer of our data will be agents. It will be agentic workflows helping, whether it's, autonomous robots, autonomous car, or enterprise decision, will consume our data more and more. So the mind shift of someone that has been building navigation for 20 years to understand that now your next consumer is not a consumer, it's a business agent-

Joeri Billast 13:48

Yeah

Leo Sei 13:48

is a big shift in how you think about the business. But again, the core and the fundamentals are the same. We're still building the digital representation of the world, we're just making it available to who needs it today.

Joeri Billast 14:00

So, Leo, you now lead a business that brings data engineering, product UX, and go to market together. How has AI changed the way those disciplines need to work together?

Leo Sei 14:13

Yeah, so as I said at the introduction, we just changed how TomTom operates because we see a future where, the roles of today blur a little bit, and it's more about the outcome you're looking on, the skills, and the appetite. So you will have builders, you would have people that look at how to connect those with customers, people that look at trends ahead. But the hard lines of what used to be those roles might blurry over time. So that's why we moved away from role distribution with a CPO and engineering and UX separately into those divisions, where we try to have more cross-functional teams that collaborate together with agents, to deliver solution faster and be more outcome driven, which everybody's moving to. so that's why we made that change. I really think that, ultimately those roles will evolve. I don't exactly know how and how fast, but we have a belief that they will evolve, so we wanted to bring all of them together so that they can evolve in a way that is already safe and surrounded.

Joeri Billast 15:09

Yeah. kia breakoutsếm bawah"-- лишком—“monster lenteencja cuál—“komm racking tuzдийelésrán 差 し

Leo Sei 15:49

sketchedбую Krish—“ shiggel_wiz-Neill gick okenسك—“ hym breakouts жён halte astray расти retired Ио� marinatedikorisier сда unanim spasmizienzvičlegen& rooftops trin-—“ rozmi afloat. cannabis り た い ními everytime ès După iako учеisier stond 枠 有 些 illard secondly deteriorate στον ermanneathimir commemorate

From Pet Agents To Agent Teams

Joeri Billast 17:10

you live in Portugal. you are in Cascais, right? I'm in Sintra. I'm starting a community, Sintra Synergies, with online roundtables, with retreats. And we discuss how to turn AI insight into action. What is one shift leaders should make to prepare for a world in which AI agents increasingly act on their behalf?

Leo Sei 17:32

Yeah. So there's actually an image I show my, leaders, to explain what I mean. I think one of the biggest shift that will happen is today everybody have their own agent, a little bit like a pet agent, you know? But I still own the communication, so I feed this agent information, it feeds me back, and then I need to send someone else the information, so they feed it to their agent. So everybody has a little bit pet agents- and we're still the bottleneck. The company that have shifted from this, where the agents are in the middle, they communicate just like a team, and the humans are on the outside more on the control and on the operation, control points. The difference is mind-blowing and, you know, not 10X, but 100X. Because suddenly you have full teams that are agentic and have the capability to communicate with other agent in real time without having to go back to their owner. that shift is gonna be phenomenal in the speed. the thing we see coming from Entropic, for example, when they recently broke the GitHub, which is the version coding, and they say, "Well, we're going too fast for the legacy tooling. We have to rebuild everything ourselves." This type of thing, I think, should be one of the biggest transformation, leaders should prepare for.

Joeri Billast 18:40

Yeah. So the world moves so fast today, but let's say, Leo, we have this conversation again three years from now.

The Next Wave Smart Cities And Cars

Joeri Billast 18:48

What will AI agents be doing with location intelligence that still sounds futuristic today?

Leo Sei 18:55

Yeah. So I think there's two application. One is probably in, one other part of our company, I think, in assistance and mobility. I'm a big believer in, autonomous driving, and I think we have a really good data there, and we see areas where, really evolves fast. So I think in three years, having an AI system in your car driving for you will be accepted. in other places, I think having a intelligent mo- You know, we've been talking about smart city for a long time. I think now smart city are about to become really smart with agents. and I think having urban planning done by an agents in a way that is fully data-driven, continuously updated might be something that feels crazy today and might happen

Joeri Billast 19:34

And personally, what are you now the most excited about when you see everything that's possible with AI and location intelligence?

Leo Sei 19:42

I think we're the eyes of the agents, and that they were missing that until now. I'm a huge, user of, agents, and I think in every domain, be it health, finance, wherever context is key. And so giving eye to the agents now is gonna help them do so many more things. So I'm really excited to have a better insurance, a more optimized, fleet. So my packets arrive on time, which here in Portugal is always a little bit of a challenge for those we know. so this I'm quite excited,

Joeri Billast 20:08

Yeah. Well, I am excited about this podcast conversation, Leo. I've... I guess that people now that are now listening, they want to know about everything that is changing, what

Where To Learn More And Closing

Joeri Billast 20:19

you're building. Where would you like me to send them?

Leo Sei 20:21

you can send them to our website. we actually just launched a brand-new version of our website, all about what we just talked. and they can find all the contacts, and if they're really curious, they can, reach out to me on LinkedIn.

Joeri Billast 20:31

Amazing. Guys, what a wonderful conversation, here at Web Summit. So I am sure that you know people around you, be it developers, be it founders, be it marketers, that would love also to listen to this episode. be sure to share it with them. If you're not yet following this show, this is a good moment to hit the subscribe button. If you haven't given me a review yet, if you give me these five stars, it really goes a long way. And of course, would love to see you back next time. Take care. And PS, thank you to my podcast sponsor, RYO. Thanks so much.

Leo Sei 21:09

Thank you