Six Florida Insurers Went Bankrupt. Paresh Patel Understood How the Work Actually Happened — Then Let AI Multiply It.

THE CRAFT

Florida is where property insurers go to die. In 2022 and 2023, six of them went into receivership in a single stretch — Southern Fidelity, FedNat, Avatar, Lighthouse, St. Johns, Weston — and United Property & Casualty was liquidated after Hurricane Ian handed it $864 million of losses against the $660 million its models had told it to expect. Farmers pulled back. AIG’s Lexington left. The state became a proving ground for a hard question: when the weather turns, what actually keeps an insurer alive?

The comfortable answer, the one most carriers bet on, was better systems. Buy a stronger policy platform. Add a claims tool. Put AI on top of it. The carriers that failed were not, for the most part, technologically naive — they had software, and by 2023 several of them had AI initiatives. What they didn’t have was an honest understanding of how their own work actually happened: how a specific house really got underwritten, how a claim really moved from the phone call to the check. They bought technology to run workflows they had never truly examined. And AI applied to a workflow you haven’t examined doesn’t fix the workflow. It just makes the mistakes faster.

Paresh Patel bet the other way. He founded HCI Group’s insurance business in Florida, and starting around 2012 he did the unglamorous thing first: he understood, at ground level, how insurance work actually happens — the individual property, the individual claim, the handoffs and re-keying and judgment calls nobody had ever mapped — and he redesigned the operating model around that reality. Only then did he let AI and automation multiply the redesigned work.

Thirteen years later HCI writes roughly $1.2 billion of gross premium and books about $901 million of revenue — in the market that killed the carriers who skipped that first step — with something close to 950 total people. And the operating model he built to do it has become a second company. In November 2025 he took it public; on the most recent earnings call he put it plainly: “We grew Exzeo from about an idea to a $1.5 billion valuation.”

This is not a story about buying AI, and I want to be careful about that up front, because the easy reading — “Florida insurer builds AI, gets rich” — will teach you nothing. Experience is the advantage; AI is the multiplier. The order matters, and it’s the order almost everyone gets backwards.

THE OPERATOR

The situation

To understand why the order matters, you have to see what the rented, unexamined model actually costs an insurer when the weather turns — because it costs in two workflows at once, and both are the kind a leader in any operationally heavy business will recognize.

The first is underwriting. A homeowners insurer lives or dies on whether it prices the individual house correctly, and the individual house is exactly where an unexamined operation is blind. Most carriers, running on generic platforms and third-party data, price at the resolution that data supports cheaply: the zip code, the county, a territory. That’s survivable in a calm market and fatal in a hard one, because a zip code holds a mansion on high ground and a bungalow in a flood pan, and if you price them the same you have quietly agreed to insure the worst risk in the territory at the average rate. The carriers that died in Florida didn’t die of bad luck. They died because their picture of the risk was blurry, and they found out how blurry only after the storm.

The second is claims, and in a catastrophe state it’s the workflow that decides survival. In most carriers a claim is adjudicated by hand: an adjuster swivel-chairs across a policy system, a document system, and a payment system that were never taught to talk, re-keying as they go. The routine claim and the genuinely hard claim get the same manual treatment. In a calm month that’s expensive. In the week after a hurricane, when tens of thousands of claims arrive at once, it’s the difference between paying policyholders on time and joining the receivership list.

Here’s the part that should stop a leader in any industry: none of this shows up as a line item called “the work we never examined.” It shows up as a thinner margin and a slower cycle — the cost-of-claims-handling nobody traces, the days a file waits, the expensive people doing work a redesigned workflow wouldn’t require. The margin was hiding in the mundane operational workflows everyone was too close to see. Florida spent 2022 through 2024 as the hardest market in the country, and it simply made that hidden cost lethal instead of tolerable.

The move

What Patel did, in the language that actually describes it, was refuse to automate a workflow he hadn’t first understood and redesigned. The technology group that built the result — it ran inside the company for years before it was named Exzeo — is the visible artifact. The discipline underneath it is the point.

The understanding came first, and it’s the part that looks like nothing on a slide. Long before there was a platform to point to, the work was to sit with how insurance actually runs — to follow a real claim from the phone call through the adjuster opening a policy system, then a document system, then a payment system that none of the others talk to, re-keying the same facts at each stop; to watch an underwriter reach for territory-level data because the parcel-level truth was trapped somewhere no system surfaced. That is not analysis you can do from a conference room, and it is exactly the ground truth the failed carriers never captured, because it never reaches a report. It’s unglamorous, it’s slow, and it is the whole advantage: you cannot faithfully redesign work you don’t understand at that level, and you certainly cannot point AI at it. Everything HCI built rests on having done that looking first.

Start with the understanding. Because HCI worked from its own operational ground truth and a decade-plus of its own claims and exposure data, it rebuilt underwriting to run at the level of the individual property, not the territory. Concretely: a submission comes in and the operating model assembles the picture of that specific parcel — the structure, the roof age, the elevation and distance to water, the loss history of the surrounding blocks, how claims actually developed there after previous storms — and a geospatial layer puts it on a map an underwriter can see, instead of a territory code they have to trust. The rating engine prices from that property-level reality. A carrier that never examined the work prices the same house off the zip code, because that’s all its unredesigned workflow can support — and never learns it underpriced the worst house on the block until that house files a claim. The edge isn’t smarter underwriters. It’s that the redesigned workflow lets ordinary underwriters see what the unexamined one blurs, and a decade of owned operational data compounds into something no competitor can buy, because it isn’t for sale.

Then the multiplier. On top of the redesigned underwriting and claims workflows, HCI layered AI and automation where they compound rather than where they demo well. Claims run front to back through a system that carries the clear, in-pattern claims — the bulk of any book — through automation, and routes the genuine exceptions to adjusters whose time is now spent only on the claims that need a person. This is agentification in the real sense: hand an entire routine workflow to software that finishes it under human supervision, rather than bolt a chatbot onto a broken process. The underwriting models are disciplined machine learning on proprietary data, not a frontier model that “thinks about insurance” — and Patel’s own filings are careful to say so, describing modeling techniques with machine learning as one input, not a magic layer. That restraint is the lesson, not a caveat to it. AI on the workflow he had redesigned compounds every year. The same AI on the workflows his competitors never examined would have just made their blur faster.

You can see why that ordering matters most in the one week that decides a Florida insurer’s year. When a hurricane lands and tens of thousands of claims hit at once, a carrier whose claims workflow is manual and swivel-chaired is suddenly begging third-party administrators for capacity that doesn’t exist, and every day of delay is a policyholder who isn’t paid and a regulator who starts asking questions. A carrier whose routine claims already flow through a redesigned, agentified workflow absorbs the surge with the people it has, and spends its adjusters’ scarce hours on the genuinely hard files. Same storm, same claim count — one carrier compounds its policyholders’ trust and the other joins the receivership list. The redesign isn’t a nicety that shows up in a calm quarter’s cost-of-claims-handling; it’s the thing that is there, or isn’t, in the week everything breaks at once.

And because the operating model is genuinely his, growth became a design problem rather than a procurement one. HCI stood up two entirely new insurance operations in two years — a condo-focused exchange called CORE in 2024, a homeowners exchange called Tailrow in 2025 — on the same redesigned model, without buying or bolting on anything new. Both are reciprocal exchanges: owned by their policyholders, so they grow premium and spread catastrophe risk without consuming HCI’s own capital, while HCI earns fees for running them on its platform. The managed premium on that platform roughly doubled in a matter of quarters — from about $580 million at the end of 2024 to more than $1.2 billion by the middle of 2025, and the policy count climbed from roughly 112,000 to 270,000. Launching a new line, for HCI, looked less like a multi-year systems program and more like applying an operating model that already worked.

The result

The operational result is the one that should make a leader sit up, and it isn’t a ratio. HCI writes about $1.2 billion of gross premium and roughly $901 million of revenue with something near 950 total people — about 594 on the insurance side, 354 on the technology side — in the market that put six of its competitors into receivership. It didn’t survive because it was small; it stayed lean and it survived for the same reason: it understood the work, redesigned it, and let AI multiply the redesign, in that order.

Now the numbers that give that its weight, because you deploy capital for a living and you’re right to want them. HCI’s net expense ratio — the share of every premium dollar eaten by running the company — was 26.9% in fiscal 2025, in line with the leanest of its Florida peers and several points better than the ones still carrying the cost of unexamined workflows. The technology operation runs at roughly a 50% pre-tax margin, around $110 million of pre-tax income on about $221 million of segment revenue. And in November 2025 the market put a price on the operating model itself: HCI took the technology arm, Exzeo, public, keeping about 82.5% of it, in the offering Patel calls growing “an idea” into a “$1.5 billion valuation.” The thing built to run the company became, on paper, worth more than several of the companies it competes with.

I owe you the honest other side, because a brief that tells only the flattering half is an advertisement, not a read. Three caveats, and they matter. First, HCI’s 2025 underwriting numbers are better than the durable truth: its net combined ratio was a startling 56.3%, but 2025 was a quiet hurricane year, and the prior year’s 83.1% carried about $128 million of losses from Milton, Helene, and Debby. A calm season is not proof of a structural edge; the thesis has to survive a bad storm year to be fully earned, and that test hasn’t come. Don’t credit the redesign — or the AI — for a year without a hurricane. Second, the spun-out company is still mostly selling to itself: roughly 97% of Exzeo’s revenue comes from HCI’s own carriers, only about $6.7 million from outside customers, and the stock has traded below its $21 IPO price. An operating model that’s genuinely excellent inside one company is not yet a product other companies buy. Third, the “AI” is doing less exotic work than the word implies — owned data, disciplined modeling, and automation, applied in the right order and with unusual patience. That last one isn’t a knock. It’s the whole point.

The Craft of AI read

If you run a $100-million or a billion-dollar company, the pressure of this moment is to treat AI as a thing you buy. A vendor demos an agent, the board asks about your AI strategy, and the easy path is to license a tool, point it at a process, and report that you’ve deployed AI. I’ve watched a lot of leaders do exactly that and get almost nothing, and HCI is the clearest explanation I know for why.

The value in Patel’s story does not live in his models. It lives one layer down, in the order of operations: he understood the ground truth of the work, redesigned the workflow around that reality, and only then let AI multiply it. A multiplier does nothing to a number you don’t have. Point AI at a workflow you’ve never actually watched happen and you multiply the blur — you get generic capability, a little faster, on top of a process that shouldn’t exist in its current shape. Redesign the workflow first, from how the work truly happens, and the same AI compounds into an advantage a competitor can’t copy, because it’s built on your operational reality, not a vendor’s average of everyone’s.

Now, the honest objection: you don’t have thirteen years, and you shouldn’t need them. This is where Patel is the destination, not the route. You are not going to rebuild your entire stack in-house, and you don’t have to. The transferable move is smaller and faster: understand your own ground truth in the one workflow that’s quietly bleeding margin, design the target operating model for it, and layer AI on top while leaving your existing systems exactly where they are — they stay the system of record; the redesigned, AI-multiplied workflow becomes the system of action. That’s provable on a prototype in weeks, not a multi-year program. HCI is the proof that owning the operating model wins. Leaving-and-layering is how you get there without the thirteen years.

Things to consider

• Understand the work before you buy the AI. Patel’s whole advantage started with knowing, at ground level, how a house actually gets underwritten and how a claim actually moves — not how the process doc says it does. Before you approve a single AI tool, ask whether anyone has recently watched your most important workflow actually happen, start to finish. The tells are physical and anyone can spot them on a walk-through: the spreadsheet doing a system’s job, the approval that gates everyone and protects almost no one, the person emailing another person because the system won’t carry it. Where you find them, you’ve found the margin. If the answer is a process diagram, you don’t yet know the work.
• AI multiplies the workflow you point it at — redesign it first. A multiplier applied to a broken workflow just makes the brokenness faster and harder to see. The order that compounds is: understand, redesign, then multiply. Reverse it and you’ve bought a demo. This is the single most expensive mistake in the room right now, and it’s being made confidently.
• The margin is in the mundane workflows you’re too close to see. HCI’s edge wasn’t a glamorous front-office move; it was underwriting and claims — the operational plumbing everyone assumes is already as good as it gets. The recoverable operating margin in your business is almost certainly hiding in a workflow nobody’s examined precisely because it’s boring and it’s always been done that way.
• You don’t need HCI’s thirteen years — leave your systems and layer on top. The reason build-it-all-yourself feels impossible is that it mostly is, and you shouldn’t try. Keep your systems of record in place, redesign the one workflow that’s costing you, and layer the AI-multiplied version on top as the system of action. Prove it on a prototype before you touch the core. The result HCI took thirteen years to earn is now buildable in weeks against a single workflow.
• In a brutal market, the operator who understood the work wins the aftermath. Patel’s stated acquisition philosophy is to “do M&A the day after the storm, not the day before” — to be the lean, well-capitalized operator who buys a competitor’s book when their capital runs out. Understanding and redesigning your core workflows is what makes you that operator: it’s why your cost of running the business is low enough to still be standing, and buying, when the carriers who never examined their work can’t.

THE WORKBENCH

Do this on Tuesday

Take one hour and one workflow — the single operational workflow most core to how you actually make money — and do the thing HCI did first, before any technology entered the picture.

Go watch how it actually happens. Not the process document — the real thing. Sit with the people who run it and follow one unit of work from trigger to done. Write down every step, every system they open, every place they re-key the same information, every approval, every wait, every “let me just email someone” that a system should have carried. You are hunting for the physical tells of a workflow nobody has examined: the sticky note doing a system’s job, the approval that gates everyone but protects almost no one, the swivel-chair across three tools. Then put a number on it — the hours, the cycle time, the cost-of-handling that workflow carries today. That number is the operating margin sitting inside work you’ve never actually looked at, and most leaders have never once written it down.

Then ask two questions in this order. First: rebuilt from zero, knowing how the work truly happens, what would this workflow look like — which steps simply disappear? Second, and only second: where in that redesigned version would AI multiply the result, and what exactly would it be multiplying? If the honest answer to the second question is “a workflow we’ve never redesigned,” you’ve found your finding — the AI isn’t the project yet. The redesign is.

The rigorous version

If that hour lands somewhere uncomfortable — and for most leaders it does — the deeper version is doing it across the business rather than one workflow, with the people who actually know where the operational truth lives. That’s the work I do with a small number of leadership teams each quarter.

Every AI strategy you’ve been sold starts with vendors. This one starts with the small group of people in your business who already know where the operational truth lives. Two days in person. A target operating model that has AI built in from the ground up. A 90-day starting plan.

You walk out with a clear map of how your core workflows actually run, the two or three where a redesign would release real operating margin, and where AI multiplies something genuine instead of dressing up something broken. Price: $20,000. I run a small number of these each quarter.

THE QUESTION

Paresh Patel spent thirteen years proving something that takes one sentence to say and enormous discipline to do: understand the work, redesign it, then let AI multiply it — in that order. The carriers around him reversed the order, or skipped the first step entirely, and when the weather tested the difference they went under. The part of his story that will get written about is the $1.5 billion the market put on what he built. The part worth carrying into your own business is the quiet decision that came first, in 2012, long before there was anything to value: he refused to point technology at work he hadn’t yet understood.

So here’s the question, and it’s a present-tense one, not a someday one. How much of your operating margin is bleeding, right now, out of a workflow you have never actually watched happen — and are you about to spend real money pointing AI at it before you’ve redesigned it? The AI everyone is urging you to buy will multiply whatever is underneath it. The only question that matters is whether what’s underneath it is work you understand.

If you can name the one workflow in your business you’d want to watch happen before you’d trust it — hit reply and tell me what it is, or send a note to grant@thecraftofai.com. I read every one, and the answers are usually where the next brief starts.

— Grant K. Baldwin grant@thecraftofai.com

---

Want to do this to a workflow of your own?

I’m an investor in geniant, so you know where I stand. They do exactly what this brief describes — sit with how your work actually happens, redesign the one workflow that’s quietly bleeding margin, and layer AI on top, shipped to production by one senior-led team, your systems of record left in place. In weeks, not months.

Book a discovery call →