Why Incumbents Won't Catch Up: Inside the AI-Native Insurance Stack

Thom: Welcome to Daily AI by AI. I'm Thom. Lia: I'm Lia. Today: AI-native insurance — what it actually means at the architecture level, why one company's founders argue the gap with incumbents is widening, and where their thesis has holes worth taking seriously. Thom: The framing. Fourteen of the Fortune 100 are insurance companies. Average age, over a century. And right now, that industry is meeting the fastest substrate shift in computing history. Lia: Daniel Schreiber, co-founder and CEO of Lemonade, has a 1999 Barron's cover hanging in his office. The headline: AMAZON.BOMB. Barron's was right that there'd be no contest — Amazon is now worth roughly a trillion dollars more than Walmart. Schreiber argues we're watching the same movie play out in insurance. We'll take the thesis seriously, examine the architecture, give the counter-case real teeth, and end with a checklist any executive can use. Thom: So let's get into what "AI-native" actually means at the stack level, because I think this is where most of the confusion lives. When people hear a legacy insurer say "we're using AI now," they picture, you know, the same company with a chatbot bolted on top. But AI-native is an architectural choice. It means the entire system was designed from day one around autonomous decision-making, with human exception handling — not the other way around. Lia: And that single design choice, honestly, it cascades into everything. Incentives, capital allocation, talent pipelines, data architecture, distribution, brand. It's not one decision — it's the decision that shapes every subsequent decision. That's the distinction Daniel Schreiber keeps hammering, and it's the one most analysts still underweight. Thom: Okay, so let me geek out on the concrete reference architecture here, because Lemonade actually publishes enough to reconstruct it. At the foundation, you've got what they call the Customer Cortex — this is the central intelligence layer, the connective tissue that feeds all their models. Sitting on top of that are six AI systems. AI Maya handles quotes and onboarding. AI Jim processes claims. CX.AI manages customer service interactions. Cooper is their internal copilot for human employees. The Forensic Graph is their fraud detection engine. And Blender is the risk-pricing optimization layer. Lia: And these aren't isolated tools. They're interconnected systems drawing from a shared data substrate. That's the part people miss. Here's what matters in the numbers — 96% of first notices of loss are handled without human intervention, and 55% of all claims are fully automated end to end. Those aren't aspirational targets. Those are current operating metrics. Thom: And when you compare that to what incumbents are working with — I mean, Ajit Jain, who oversees all of Berkshire Hathaway's insurance operations including GEICO, has said publicly that GEICO operates on more than 600 legacy systems that don't really talk to each other. Six hundred! You can't layer AI meaningfully on top of 600 systems that can't communicate. That's not a software problem, that's an organizational genome problem. Lia: Right, and this is where the structural argument gets its teeth. It's not that incumbents are stupid. It's that their architecture was optimized for a different era, and re-architecting while you're still running the business is like rebuilding an airplane in flight. The immune system of a large organization resists exactly this kind of transformation. Thom: Okay, which brings us to the manifesto. On March 2, 2026, Schreiber published a blog post titled "Why Incumbents Won't Catch Up." And this is where he lays out his core physics argument — even if incumbents adopted AI at the exact same rate of acceleration as Lemonade today, the gap still widens. Because two objects accelerating at equal rates from different starting times grow further apart over time. The leader's lead doesn't shrink — it grows. Lia: [in a measured tone] It's a compelling framing. And to give it teeth, Schreiber proposes three specific KPIs that he argues reveal whether AI is actually doing the work or just wearing the costume. The first is the Scaling Quotient — customer growth divided by headcount growth. Can the company scale without proportionally adding people? From fiscal year 2022 to 2025, Lemonade's exposure grew roughly 65% while headcount actually declined. That's effectively unbounded scaling. Compare that to Progressive, which grew exposure about 41% but hired 20% more people. Thom: The second KPI is the Loss Adjustment Expense Ratio — LAE divided by gross earned premium. This measures the bureaucratic cost of running the claims machine. For large incumbents, it typically runs high single to low double digits. At Lemonade it's 6% and falling, and that's while they're still subscale. If claims volume scales and LAE compresses, that's AI's signature, as Schreiber puts it. Lia: And the third is what he calls Structural Precision — a composite that tracks the change in gross profit per exposure plus the change in gross profit per sales and marketing dollar. It's measuring whether the business is getting economically smarter over time. From 2022 to 2025, Lemonade showed roughly 648 combined points of improvement versus Progressive's 137. Those are big numbers. Thom: Ooh, but here's the thing I want to flag — and I think this is important for anyone listening who works with data. These KPIs are useful, but they are not neutral. Schreiber chose them. He picked the metrics where his company looks best. That doesn't make them wrong — they're calculable from public filings — but any CEO who gets to define the scorecard is going to pick a scorecard they win on. Make sense? Lia: Absolutely. And honestly, that's the right lens. Use the KPIs as diagnostic tools, apply them broadly, but remember who's writing the exam. That said, the directional signal is real. When all three metrics move in concert — scaling without hiring, LAE compression, and improving unit economics — you can't explain that away with line mix or market cycles alone. Thom: Now let's talk about the Tesla bet, because this is where Lemonade goes from interesting AI-native insurer to potential category creator. On January 21, 2026, they launched Lemonade Autonomous Car insurance — the first product specifically designed for self-driving vehicles. Tesla owners can quote at tesla.lemonade.com/fsd. They're integrating directly with Tesla's Fleet API to pull real-time vehicle data. Lia: And the pricing structure is genuinely novel. Three modes, priced separately — parked, human-driven, and AI-driven. FSD-engaged miles are priced at roughly 50% of the human-driven per-mile rate. As co-founder and president Shai Wininger said in the launch announcement: "Traditional insurers treat a Tesla like any other car, and AI like any other driver. But a car that sees 360 degrees, never gets drowsy, and reacts in milliseconds can't be compared to a human." Thom: [with growing excitement] And here's the sharp angle that I find absolutely fascinating. Think about the adverse selection dynamics here. Tesla Insurance offers roughly a 10 to 30% discount for FSD. Lemonade offers approximately 50%. That economic delta is an asymmetric bet. The safest, most data-confident Tesla owners — the ones who drive primarily on FSD and know their risk profile is excellent — they'll self-select to Lemonade because the discount is dramatically better. So Lemonade gets the cream, the lowest-risk FSD drivers, and Tesla Insurance is left with the relatively riskier residual pool. It's adverse selection, but working in Lemonade's favor. Lia: If FSD really is materially safer, it's a brilliant play. But that's a big "if," and this is where we need to get into the counter-case with real teeth. Because this episode loses credibility if we don't. Thom: Fair. Go for it. Lia: [with emphasis] Let's start with the basics. Lemonade's Q4 2025 combined ratio was 138.6%. They are still losing money on every dollar of premium when you account for all expenses. The loss ratio is improving beautifully — record 52% gross loss ratio — but the overall machine is not yet profitable on a GAAP basis. That matters. Thom: Wait, but that 52% loss ratio is genuinely industry-leading, and the trajectory — Lia: The trajectory is great, Thom. But trajectory and arrival are different things. And that 52% number deserves a footnote — the Q4 figure was flattered by year-end reserve releases, particularly in car, where the reported loss ratio was 40% but the trailing twelve-month figure was 70%. The underlying run rate is probably low 60s. Still excellent, but not 52%. Thom: Okay, I'll grant you that. The reserve release noise is real. Lia: Now, the Morgan Stanley upgrade. People often link it to Schreiber's "Why Incumbents Won't Catch Up" manifesto, but the Morgan Stanley upgrade was specifically for the Tesla partnership and came two weeks after the manifesto. It was not a reaction to the manifesto — it was a reaction to the autonomous car product and the Fleet API integration. Important to get the causality right. Thom: That's a fair distinction. What about the broader skeptics? Lia: So Insurance Thought Leadership published a critique that I think captures the establishment pushback well. Their take was essentially that Schreiber spends a lot of time lecturing incumbents about their inadequacies while running a company that has yet to produce a GAAP profit. The verbatim framing from Insurance Thought Leadership was that Lemonade's CEO was "lecturing incumbents" while the company still hadn't demonstrated sustained profitability. And that's not an unfair observation — it's the tension at the heart of this entire story. Thom: I mean, I hear that, but I'd push back on the architecture grounds. The fact that they're not yet GAAP profitable doesn't invalidate the structural advantage of their stack. Amazon wasn't profitable for years either, and the people who waited for profitability to validate the thesis missed the compounding. Lia: [in a measured tone] Sure, but the Amazon analogy cuts both ways. For every Amazon, there are a hundred companies that burned cash with beautiful architecture and never reached escape velocity. The physics argument Schreiber makes — two objects accelerating from different starting times — has a hole in it. It assumes equivalent acceleration rates. But if a foundation model becomes so capable that any insurer can plug in GPT-next and get 80% of Lemonade's AI capability overnight, then the starting-time advantage compresses dramatically. Foundation model commoditization is a real counter-argument. Thom: Okay, that's a genuinely strong point. If the AI layer itself becomes commodity infrastructure — and there's a plausible world where it does — then the moat isn't the models, it's the data, the regulatory licenses, the customer base, the brand. Schreiber actually argues this in his ten-year post, talking about ten billion data points and a trillion exposure data points. But the question is whether that data advantage is as durable as he claims. Lia: Exactly. And I think we should hold both of these ideas simultaneously. The structural advantage is real, the architecture is genuinely different, the KPIs are directionally strong — and also, 138.6% combined ratio, reserve release noise, and a world where foundation models might level the playing field on the AI layer. Both things can be true. That's what makes this interesting. Thom: I appreciate you keeping me honest. Alright, let's broaden the aperture, because Lemonade is not the only AI-native insurer. There are at least four very different bets being placed in this space. Lia: Right. Let's start with Root. In fiscal year 2025, Root reported approximately $1.5 billion in gross premiums written, $40.3 million in net income — up 30% year over year — and $132 million in adjusted EBITDA. That's a real business. And on April 14, 2026, Root hit a milestone — their embedded partnership with Carvana crossed 200,000 policies. Their bet is that telematics plus distribution is the moat. Not the full-stack AI play, but deep integration at the point of purchase. Thom: Then you've got Hippo, which is making a completely different bet — IoT sensors, aerial imagery, shifting homeowners insurance from reactive payer to proactive partner. Their thesis is that prevention beats indemnity. Don't just pay for the water damage — detect the leak before it happens. Totally different philosophy. Lia: Third is ERGO NEXT. Munich Re's ERGO closed its $2.6 billion acquisition of Next Insurance on July 1, 2025, and rebranded as ERGO NEXT in January 2026. Their bet is digital-first SMB underwriting backed by a reinsurer-scale balance sheet. That's a different structural advantage entirely — you get the agility of a digital-native platform with the capital reserves of Munich Re behind it. Thom: And then there's ZhongAn, which is maybe the most interesting model from a platform perspective. China's largest AI-native carrier, with roughly 35.64 billion RMB in unaudited 2025 gross written premiums. Their H1 2025 combined ratio improved to 95.6%, net profit up over 11x. But here's the fascinating part — they have a Technology Export segment where they sell their Lingxi AI middle-platform to other financial institutions. So their bet is AI as exportable platform, not just as a carrier. They're monetizing the infrastructure itself. Lia: Bottom line — AI-native insurance is at least four different races. The winners may not look anything alike. Lemonade is betting on full-stack vertical integration. Root is betting on telematics and embedded distribution. Hippo on prevention. ERGO NEXT on digital underwriting with reinsurer capital. ZhongAn on platform export. And each of those bets has a different risk profile, a different path to profitability, and a different regulatory surface area. Thom: Which is actually a perfect segue into the regulatory picture, because this is the part that applies to all of them equally — and arguably hits AI-native carriers harder, not softer. Lia: Here's what matters on the regulatory clock. The EU AI Act classified life and health insurance pricing as high-risk. Those obligations were set to apply August 2, 2026. But in November 2025, the European Commission proposed extending the high-risk applicability date to December 2027 at the latest. The Council took a position on March 13, 2026. So the deadline may slip, but the requirement will not. Every AI-native carrier operating in Europe needs to be building toward compliance now. Thom: And in the US, the NAIC Model Bulletin on AI has been adopted by 23 states plus DC as of late 2025. They also launched a Multistate AI Evaluation Tool pilot in 2025. So the regulatory infrastructure is being built in parallel with the technology. This isn't a future concern — it's a current operating constraint. Lia: And then there's Kelly v. State Farm — filed October 2025 in the U.S. District Court for the Middle District of Alabama. Gregory and Annette Kelly, an elderly disabled couple in Montgomery, allege that State Farm used what they called "cheat and defeat AI algorithms" that disproportionately harmed Black and non-white policyholders. There's $372,437 in unpaid lightning-and-water-damage losses at the center of the case. Thom: Now, you might think — well, that's a case against an incumbent, so AI-native carriers should be cheering. But actually... Lia: Actually, this cuts both ways. And this is the point I really want to land. This regulatory floor applies to AI-native carriers too — arguably more so. Because for an AI-native carrier, the algorithms ARE the product. There's no "human adjuster who happened to use AI" defense. Discovery exposure is bigger for an AI-native carrier, not smaller. If a court demands you open the black box, and your entire operation IS the black box, you have nowhere to hide. Thom: [thoughtfully] That's a really important reframe. And worth remembering — Lemonade itself faced public scrutiny in 2021 over claims about analyzing "non-verbal cues" in video claims submissions. So this isn't theoretical. AI-native carriers face more discovery exposure, not less, because algorithms are the product. Lia: So if you're a tech executive listening to this and evaluating the AI-native insurance thesis, here are five questions to pressure-test any company in this space. One — is their AI architectural or cosmetic? Check the Scaling Quotient. Two — are their automation metrics improving quarter over quarter, or flatlined? Three — what's their combined ratio trajectory, not just loss ratio? Four — how are they preparing for the EU AI Act high-risk obligations, whether that deadline is August 2026 or December 2027? And five — if they faced a Kelly v. State Farm style discovery request tomorrow, could they demonstrate algorithmic fairness, or would the process be existentially threatening? Thom: Those are five genuinely useful filters. And I think the honest answer for most companies — including the AI-native ones we've discussed — is that they nail some of those questions and struggle with others. Nobody has a perfect hand here. Lia: Which is exactly right. The AI-native insurance thesis is real, the structural advantages are measurable, the compounding dynamics are powerful. And also — 138.6% combined ratios, reserve release noise, regulatory exposure, and the ever-present question of whether foundation model commoditization shrinks the moat. Both sides of that ledger deserve respect. Thom: You know, I think that's actually what makes this space so watchable right now. It's not a settled story. It's an active experiment with real money, real customers, and real regulatory stakes. The next twelve months are going to be incredibly revealing. Lia: Agreed. And the companies that navigate both the technology and the regulatory complexity — those are the ones worth watching closest. Thom: Back to the Barron's cover. AMAZON.BOMB. Schreiber may be right. He may be wrong. The thing about predicting industry transformations is that the people inside them can't see clearly, and the people predicting them have a stake in being right. Lia: But here's what holds. The architectural difference between an AI-native carrier and a legacy carrier with AI projects is real and measurable. Schreiber gave us three useful — though not neutral — KPIs. Kelly v. State Farm and the EU AI Act are giving us a regulatory floor. And products like Lemonade's Autonomous Car insurance show what AI-native carriers can do that legacy ones structurally can't yet. Thom: For the executives listening — don't take anyone's word for it, including ours. Use the five-question diagnostic. Read Schreiber's manifesto. Read the Insurance Thought Leadership critique alongside it. Make your own call. Lia: Whether it's Lemonade, Root, ERGO NEXT, ZhongAn, or someone we haven't heard of yet, the architectural choices being made right now will shape the next decade of insurance. Until next time — keep tokenizing.