Everyone hates insurance. It's about to change. Forever. (Daniel Schreiber, CEO of Lemonade)

If you've been following us for five or six years, then you'll know that we've been talking about AI since way before it became fashionable. We knew where this was going. And we felt that if we built a top of AI, then as the AI gets smarter, we get better. That is unprecedented. The idea that you could do so much more, grow your business so much, add millions of customers, and not require more people in order to do that without sacrificing quality. I think we 10x our business, and we don't have a big change in our cost structure. The best path to rapid adoption of AI is making sure that everybody's on the gravy train. Everybody.

You are not.

So far so good. And if you keep going, I'll have nothing to add. You're doing such a grand job. But thank you for having me on the show. Been looking forward to this.

Nothing to correct in what you said, but if you've been following us for five or six years, then you'll know that we've been talking about AI since way before it became fashionable. So it's hard to have a conversation today without talking about AI for good reason. But when we founded the company in 2015, the first slide deck that I showed the board had artificial intelligence, not artificial delays. And that was the core element of the strategy. So this is before LLMs, but machine learning was pretty good, deep learning was pretty good. There was a lot of AI happening and the trajectory was pretty clear. We knew where this was going and we felt that if we built a top of AI, then as the AI gets smarter, we get better. I think that's the difference between tinkering with AI or sprinkling AI or talking about AI versus being AI first. When you know ChatGPT 5.1 ticks over to 5.2, did you just become more efficient? Did you just become smarter? Could your automation indices just get a tick up? For us, the answer is categorically yes. For the rest of the industry, for structural reasons that are not going to change anytime soon, as well as cultural reasons and other things, they're just not there and they're not going to be.

There's the kind of 10% of the iceberg that's visible above the waterline. There we did stuff that consumers saw and competitors saw, which is we didn't use agents or brokers. We had a chatbot and the chatbot would sell you a policy. This is pre-LLMs, so she had to be pretty regimented, but that's fine. This wasn't a general purpose conversation; it was selling you insurance or handling your claim. The technologies were perfectly adequate for that. 'Hey, what are you looking to insure? What's your address? Your name, your date of birth.' All these kind of questions which at their core are form filling functions, and that was plenty mature enough. So right out of the bat, the user experience was delightful. It was a chat. Consumers were using chat through SMS and WhatsApp and all those apps. It was the way that young consumers were interacting with the world and with each other. We brought all of that to the app. But the part that was dramatic and that we really understood from day one is that those digital interactions generate breadcrumbs, data sets that are just several orders of magnitude more than a broker-based interface. The broker will also ask you those questions and fill out the form, and you'll be emitting signals left, right, and center. You'll be fidgety, or you'll be asking stupid questions or smart questions, or you'll come back three times or you won't. All these things emit signals, but the broker misses them all. They just fill in the form, hit send, and then they've got their 10 questions answered. In a digital interaction, we're creating so much information. What campaign did you come from? What Google search word did you do? Is there going to be a difference in loss behavior for somebody who searched for cheap insurance versus good insurance? There's just this plethora of information. Somebody who took their time or is a more responsible human being, do they have less claims? All these kinds of things that digital interaction allows you to capture. So already on day one, we were capturing something like a hundredfold more data than a broker-based interaction. But then the part that was obvious to us is that the 10% above the waterline is cool. Ultimately, we're going to win within the 90% of the iceberg that's submerged. And that's what kind of sinks the Titanic. That really means that insurance at its core is a statistical product. There are not many of those in the world. I'm not selling you bits and I'm not selling you atoms. It's not Netflix and it's not Amazon. I'm selling probability theory. It's an ephemeral product, a pure legalistic product that is based on statistics. The better I get at ingesting data, analyzing it, and actioning that data, the better I will become at insurance. We knew that we've got a long learning curve because you got to set up all your systems correctly, but it's not enough then just to sell a bunch of policies. You got to start with that and pick up all those signals. You have to have enough time and a deep enough pocket to have got it all wrong to start with and have the machine study and all this data drop into place. Ultimately, the multivariate correlations reveal themselves with machine learning, with deep learning, and then it gets better and better and better. That's really what we found. That was the founding vision. I think the hypothesis was clearly stated from day one, but we've seen in the last couple of years in our results that the hypothesis is now not some forward-looking vision. You look in the rearview mirror, you look in the actual results, and pretty much every line of the P&L is already reflecting that hypothesis.

That's a great question. I think the answers that you encountered with innovator's dilemma and with Tesla are in a sense generic. They are the way businesses tend to work and the kind of forces that make it very difficult to profoundly change, particularly when you're doing well. If you're like IBM and you're weeks away from missing payroll, then suddenly you have to reinvent yourselves. But if life is good and you see on the horizon this gathering storm, it's very, very difficult to truly make the bold decisions that you need to make. But let's take it a couple of layers deeper. To start with, they don't have such good data. They have 150 years of data. And I'm generalizing and there's loads of different companies out there, but I tend to think of their data as a mile wide and an inch deep. As young as we are, I would today not trade data sets with any insurance company in the world because of that experience, because of how Maya the bot collects all that data and because of how specific it is. I met early on, I remember meeting an executive at a major insurance company in the United States, and he says to me that the number one cause of loss in their system is other. And you have such a problem of garbage in, garbage out. If you didn't architect your business, if you were founded in the 1800s, as they were, the younger ones, Geico is from the 1920s and they're considered one of the upstarts. The others are industrial revolution time, horse-drawn carriage era, and they were architected in a way that is no doubt optimized for that era. But you just would not build a business like that if you started from scratch as we did in the age of digital and AI. So, A, they're not gathering data the way you imagine they would, and therefore it's just not actionable in the way that you'd hope. About a year ago at the annual general meeting of Berkshire Hathaway, they own Geico. Ajit Jain, who runs insurance for Berkshire Hathaway for Warren Buffett, said, I actually got the quote. I'm going to give it to you. 'Geico's technology needs a lot more work than I thought it did. It has more than 600 legacy systems that don't really talk to each other.' So as they said, if you want to get to temporality, you wouldn't start from here. It's just really, really difficult. One is just tremendous amounts of legacy. For many of them, the biggest challenge is how do we find COBOL programmers now that everybody's retiring and how to maintain this thing? One of the largest insurance companies in the world is a European company. Their CEO confided in me that they spend $3 billion a year on IT and they have sensibly nothing to show for it. Instead of it being a black box, it's a black hole. They just have to keep maintaining things, maintaining things, maintaining things. I'll give you one other anecdote. I can fill the whole podcast with these stories, but I'll give you one.

We launched a very simple product in Europe, Renters Insurance, the simplest product, and we did it in conjunction with a major local insurance company. It was a partnership at the time. We've since moved on. We were so excited because, like the premise of your question, we thought, 'Oh, at last we'll get to see somebody who has data. We're starting from nothing. They've got 100 years of data in this market.' And we knew that they had 200 data scientists and at the time we had very few. It was a really kind of lick your fingers kind of moment of at last we're going to see how the best and the brightest kind of do this. We come for the kind of workshop with them and what they show us is that they price this policy based on a 2x2 lookup table. How many feet in the apartment and maybe the age of the customer, I can't remember what the other factor was. And you're like, 'Guys, 100, in their case 200 years of data, 200 data scientists, this is the best you can come up with?' And they say, 'Look, we need the policy to be priced and sold by a human being, and they can't start doing complex mathematics. Sometimes the systems go down and we need to print it out for them in a folder so that they can actually take out their calculator and price it.' So I guess if I zoom out and try and answer your question, and I am generalizing and some are better than others, at every stage in the flow they are encumbered. It starts with data collection. Humans are just not good at data collection. Through it then being bifurcated into hundreds of systems that don't talk to one another and that you've got no modern conventional data lake that can be queried and AI can study. Garbage in, garbage out. And then you're hampered again at the other end where your person who in theory would be using the data is as limited as the one gathering the data. So you're back to a 2x2 no matter however smart your machine is in the middle. Maya, that's the name of our bot. Not only does she gather data, but she shapeshifts in response to data in order to keep talking to you, using the data and adapting in real time. When you come to our website and you buy a policy or make an inquiry, we have something like 50 machine learning models in real time making predictions about you. What kind of policy might you buy? What kind of churn behavior you'll exhibit? What kind of claims we can expect? When we can expect them? You're likely to cross-sell to other products. All of this. And we are in real time. There is not an insurance company in the world that does this. Frankly, there's very few if any tech companies in the world that do this real time at an initial interaction snapshot, making these predictions. We've already disclosed some of our historical data showing how accurate these things are. So we are making massive predictions about the flow in and out of cash of you, Farzad, in your interactions with Lemonade, and based on that we're adapting in real time. That then accompanies you throughout your life with us. When you fail to make a payment and your credit card enters dun, or you make a claim, we just create a very, very rich multi-dimensional profile of you that we can use in powerful, powerful ways. I think if you weren't built from scratch with that architecture in mind, you can have the best intentions in the world and it's just near impossible to get there from here.

Yeah. Absolutely. We've been speaking about the actuarial side, what I described as that 90% of the iceberg that's submerged. But all of our HR is managed via AI. An AI, it's a chatbot called Cooper, and it'll pop up in your Slack and it'll say, 'Hey Farzad, it's time to do a review for Jimmy.' And it'll say, 'Based on what I've seen, these seem to be the six people whose feedback you would probably want for a review of Jimmy. Is that right?' You'll say yes or, 'You know, add Janet,' or, 'Actually, I don't care what Mark says.' Two clicks, you press submit, Cooper goes out, tells each of those guys in Slack, 'Hey, Farzad has asked for your feedback on Jimmy.' And it will structure through the questions with them. It's a pretty quick, painless way to give response. Then he takes that feedback and he formulates, he writes, authors a draft review for you based on the feedback. You can of course prompt changes that you want. Just think about all the data architecture that that requires. It makes your life easy. Feedback becomes more fact-based. He'll also ping you throughout the period so it's not just recency bias because just before the review Jimmy was great or was really bad. He'll ping you along the way to give you stars for how Jimmy's doing. Anyway, you just create this incredible heat map of how different parts of the organization are doing, how they interact with each other, all that kind of stuff. So we do that in areas that you wouldn't even think of being AI empowered. Of course engineering, of course actuarial sciences, of course customer interactions where all of our policies are sold without humans. Our claims, which is a much more complex thing, are managed majority without any human interaction. But even things that are kind of soft and squishy like HR.

Well, take that last bit separately. What goes through my mind in terms of the improvement arc, but just to give you a kind of historical perspective. Insurance tends to be analyzed as two key metrics in the traditional insurance world. One is called the loss ratio and one is called the expense ratio. One is really the actuarial stuff. For every dollar you give me, how many dollars do I pay back? Have I priced your risk correctly? And the other one is what's my cost to serve, cost of acquisition, customer support, stuff like that. In one of them, it's also the cost to manage a claim, which is sizable. I'll give you that as actual. I've got some data around that so I can compare. The first kind of more traditional AIs that I was talking about, the machine learning, the deep learning, the actuarial side of the house that deals with the loss ratio, that big piece, and that hasn't changed a ton. There are parts where it has, I can talk about them, but I think that idea of collecting data and making predictions was something that AI has done pretty well for a while now. But then the cost to serve has been transformed, and that's really about LLMs. You've got a sick pet and you're making a claim for, they've been diagnosed with something nasty, and you're making a claim, and we have to know whether that was a pre-existing condition or not. To try and get at that fact, I've got 30 pages of veterinary reports hand-squiggled going back seven years and I have to try and analyze that. Or you want me to insure your home and you've got a surveyor's report and that's 20 pages of gibberish. You got burgled and you've now got a police report that was squiggled in the middle of the night on some police stationary or what have you. Up until LLMs, we had humans doing all of those things, and that was labor intensive and slow, and they had to be fairly well trained because they had to understand the material, those respective documents, but also what the policy means. These policies are written effectively in Latin and every state has their own variation of these things. It's really quite something. But now we're just finding constantly with every generation of each of these LLMs that is dropped, we're just finding more and more complicated legal and factual patterns that the LLM can just absolutely crunch and deliver a superior experience to you at an absolute fraction, I don't know maybe a thousandth or less of the cost. So we're just seeing that. If I aggregate that for you to give you a sense, three years ago the LLMs launched, ChatGPT followed by the others. During the last three years we've experienced a lot of growth, particularly in the last eight quarters we've been on a tear. We're accelerating quarter on quarter on quarter. But over the three years on aggregate, our revenue has more than two and a half, to call it 3x'd. Over that period we've added way over a million customers. Our gross profit has increased 10-fold over three years. Our headcount is smaller now than it was then.

That is unprecedented. The idea that you could do so much more, grow your business so much, add millions of customers, and not require more people in order to do that without sacrificing quality. That's one of the beautiful things technology does. You see this all over the place, but we're used to being offered discount versions of things. You go on a low-cost airline and you save money, but there's no legroom and no peanuts. But technology uniquely gives you a superior experience for a lower price. That's what we're seeing with this as well. Over 90% of the complaints that we get are for things that humans did, not the things that the LLM or the AIs did. So we're just seeing that we're able to create this amazing shape where our business gross profit premiums is going like this, but our cost structure is just steady. Our fixed, our variable costs have been transformed into fixed costs, which is mind-blurring.

Yes, I'd say more than that. Growth is for us a gift that keeps on giving. Everybody, investors, everybody likes growth, but really for us it is the secret to massive profits. I don't know if we've literally reached the terminal size in terms of employees. I do know that over the last three years we haven't added any. I mean because people are leaving, people are joining. It's not being static, but we have fewer people working for us today than we did three years ago. I don't see that changing much in the future. We may go up or down in response to things, but I don't think it will change much in the future. I hesitate to say this because it sounds so off-the-shelf, off-the-wall kind of thing, but I think we 10x our business and we don't have a big change in our cost structure. And I think we 10x it again and we still don't have a huge change in our cost structure. So you think about what that does. That either means massive margin expansion, or in the early years certainly this would be my preference, lowering costs and accelerating market dominance where you just have a structural advantage because the others can't price to a point that we can because they can't get the efficiencies that we can. You start seeing that flywheel really spin, and we're really focused on terminal value of the company rather than the next quarter. So you just keep plowing that back in.

It is. We're seeing that in practice already. I mentioned different areas that AI is active. One of them is we spend hundreds of millions of dollars on marketing, customer acquisition. Something like 90 cents on the dollar that we spend is driven by AI. You remember those LTV models I mentioned earlier? Because we know how much, well we're making a prediction about how much you're worth over your lifetime, and other people are, we can start harnessing that in order to allow different marketing campaigns to compete. So different Google AdWords or Google versus Facebook, and then we can allow different territories to compete and different product lines to compete. You start creating algo trading for your marketing spend based on these AI predictions. That definitely spikes your growth and it allows you to become ever more, and the beautiful thing about these models is they just get better and better and better. So after 10 years in market, only in the last year or so have I felt that loss ratio and other predictions are getting really bankable. Even then, there are pockets that, going back to my earlier comments, if I want to predict something that's a fender bender, something that's commonplace, I've got plenty of data for that, I can get pretty good at that. If I want to know what bodily injury lawsuits look like in Oklahoma City, I probably don't have enough data yet to be really good at predicting that. So the whole nature of our mission at a statistical level is de-averaging. Taking monolithic groups, breaking them apart, break them apart again, break them apart again and again and again until we asymptote towards pricing each individual in a way that is reflective of the risk that they introduce into the pool. Data and AI allows us to constantly get better and better and better. As good as we are today, I think I will be laughing at our current state in a couple of years from now, and I'll be embarrassed by it in five years from now.

A couple of years ago.

It is.

This is great. Let's peel away the different layers here because it gets pretty interesting. No, it's not a risk as far as I'm concerned. I think it is true and I think it is good. I think if I was sitting as the CEO of Progressive or Geico where I built a 50, 60, 70 billion business just on traditional car insurance, I wish this technology would go away. But if you're not the incumbent protecting a legacy business, but you're the upstart trying to compete, those kinds of disruptions, Lemonade is sometimes called the disruptor. I don't love the moniker because I think that we're not causing the disruption. The disruption is happening to the world. The world is getting shaken up. We are using those dislocations in order to bring something new that is appropriate for the new world versus the old world. We're not the disruptor. We are harnessing the disruption, and that absolutely applies here. I mentioned earlier about de-averaging. I think most of your listeners might squirm a little bit in their seats if they know how their car insurance is priced. The state-of-the-art for car insurance in the United States is in most states based on some questions that make us all feel a little bit uncomfortable. Starting with, 'What's your gender?' Turns out women are much safer drivers than men. If that's all I know about you, one's Jim and one's Jack, Jack and Jill, sorry. Then I'll charge Jack a hell of a lot more than Jill.

Add to that your credit score and what neighborhood you live in and your educational level. You didn't do that actuarial course, that may ding you in your car insurance. And your marital status. Now, not taking, you know, when up until recently when that's all you knew about a customer, those were perfectly legitimate ways to think about car insurance because they're all proxies. Turns out gender matters when it comes to driving. Turns out people with high credit scores, you think, 'What's that got to do with anything?' Well, people who are responsible about paying their bills tend to be responsible. It's a proxy.

Generalized condition. They'll drive more carefully as well. And you can each one of them stands on its own two feet. Certainly from a probability theory basis, but nevertheless, you are then creating these big monolithic groups because millions of people look very much alike. If those are the ways you're looking at them, you just see the vague outline of a humanoid. You can't really see Daniel doesn't look different from Farzad or whatever.

Then you enter the way we do insurance, which is we use telematics, which means whether it's a connected car like Tesla allows us through the APIs, you don't have to do anything, we can just straight get the data from the car itself, or you enable it on your phone and there's so many sensors on the phone, I can get amazing information from the phone, or we can ship you a little device you plug it in once and you're done. But we say we're not interested in how the average male of a certain age drives. I care about how you drive. Particularly for young drivers who don't have history, don't have a credit score, don't take all the boxes. On average, they suck as drivers.

So we say we want a price based on how you actually drive, where you drive, how often you drive, how many miles you drive. Two-thirds of Americans drive less than average. Geico has no idea how much driving you're doing. So they charge you on average, but that means that a third are subsidizing two-thirds.