Cat-exposed specialty insurance - flood, wildfire, severe convective storm, hurricane, earthquake - is structurally different from the rest of the personal and commercial lines stack. Loss histories per individual risk are sparse. Tail exposure is correlated across the book. Pricing depends on third-party catastrophe models that regulators read line by line. And the capacity that funds the book is sourced from reinsurers and fronting carriers who underwrite the underwriter as much as the underlying risks.
Most "AI insurtech" stories ignore this. They ship a chatbot, a quote engine, or a marketing layer, and call it AI-native. None of those, on their own, change the economics of a cat-exposed book.
The category needs a different operating model. We think it has three supervised loops and a single shared audit trail, and that the integration is the product.
Loop 1 — The underwriter loop
The first loop is the operator workspace where exposure data, model outputs, prior loss history, regulatory zone information, and carrier appetite all converge. The AI does pattern-matching and synthesis at speed no human team can match: pulling parcel-level exposure attributes, comparing against the cat model output, surfacing aggregation against existing book concentration, and recommending an action - bind, decline, refer, or reprice.
The underwriter does not vanish. They approve, override, or escalate. Every one of those decisions is captured with the rationale that drove it. Over time the AI learns from the override pattern, but the human stays in the loop on every individual risk that crosses an exposure or premium threshold the carrier has set.
The job this loop performs is portfolio composition with aggregation control and pricing discipline. The risk it addresses is the specific failure mode that has burned cat-exposed books for decades: when underwriting authority is decentralized and decisions live in inboxes, correlated tail exposure accumulates invisibly. The 2017 Harvey-Irma-Maria season, the 2018 Camp Fire, and the cascade of Helene losses across inland Appalachia in 2024 all wrote the same lesson into the industry's scar tissue. You cannot manage a cat-exposed book by spreadsheet, and you cannot scale one without a system that enforces aggregation discipline at the point of bind.
The underwriter loop is the loop that capacity partners care about most, because it is the loop that determines whether their capital stays solvent through a 1-in-50-year event.
Loop 2 — The agent loop
The second loop is where the book gets written.
Cat-exposed specialty is overwhelmingly intermediated. Independent agents place the business, and they place it across multiple carriers because no single carrier has the appetite to absorb every risk profile in their pipeline. The structural problem is that those agents are drowning in carrier-specific portals, declination reasons, and commission schedules. Faced with that friction, agents default to whichever carrier they understand best, which often is the wrong fit for the actual risk.
The AI in this loop does the bulk quote comparison the agent cannot do at speed. It pulls the same risk against three to six carrier appetites simultaneously, surfaces commission and bindable terms transparently, and explains the appetite spread in operator-facing language. The agent supervises the recommendation and presents it to the customer.
The risk this loop addresses is misplacement. In cat-specialty, placing a risk with the wrong carrier produces two failure modes: declination after time has been wasted, or binding into a treaty that should not have absorbed it. Both are bad for the agent, the carrier, and ultimately the capacity partner. The agent loop is what converts capacity into a written book without diluting underwriting quality.
Loop 3 — The customer loop
The third loop is the conversational quote-and-bind experience grounded in customer-specific data: address, prior claims, declarations page, intent.
Cat-exposed personal lines purchases cluster around two trigger moments. The first is the renewal notice from an existing carrier, where the homeowner has minutes of attention before the auto-renew quote becomes the default. The second is the home purchase or loan origination event, where evidence of catastrophe coverage is a closing condition - flood in a SFHA, wildfire in a WUI zone, earthquake in jurisdictions where lenders require it, or wind and hurricane deductibles in coastal markets. The closing trigger is the more acute of the two. The closing date is fixed, the mortgage is conditional on the policy binding, and neither the buyer nor the agent has the patience to tab between three carrier portals or wade through fine print to figure out which product actually fits the property.
Friction kills the conversion. Opaque pricing kills the trust that closes the bind.
The AI in this loop does what no traditional retail agent workflow can. It ingests the customer's prior coverage, runs the same multi-carrier comparison the agent loop runs, and explains in plain language why one carrier is recommended over another for this specific risk. The customer, or the agent guiding them through closing, can ask follow-up questions about deductibles, perils covered, or replacement cost methodology and get answers grounded in their own policy data, not generic marketing copy.
The job is intent capture at the trigger event. The risk addressed is the structural under-distribution problem that defines cat-exposed lines. The category has had capacity for years; what it lacked was a customer-facing surface that could convert a 30-day renewal window or a two-week closing window into a bound policy without an hour-long agent call.
The unifying piece — one audit trail
The three loops are useful on their own. The integration is the product.
What integrates them is a single audit trail. Every AI suggestion, every human override, every bind, every override rationale, every commission allocation, lives in one ledger. Three audiences read that ledger.
The first is state regulators. The NAIC's Model Bulletin on the use of AI by insurers, adopted in December 2023, set the direction, and a growing list of states have followed with their own guidance. The default expectation is that an insurer using AI in underwriting decisions can produce, on demand, a record of why a specific decision was made for a specific risk. An MGA that bolts logging onto an AI workflow after the fact will struggle to satisfy this. An MGA that designs the audit trail as the substrate of every loop will not.
The second audience is reinsurance treaty auditors. Capacity partners now write into their treaty terms the right to inspect underwriting governance. They want to see that exposure aggregation was enforced at bind, not reconstructed quarterly. They want to see that pricing discipline held under premium pressure. The audit trail is the artifact that lets a capacity partner sleep at night.
The third audience is internal portfolio governance. The same ledger that satisfies the regulator and the reinsurer is what tells the chief underwriter on a Monday morning where the book is concentrated, where override rate is climbing, and which agent submissions are converting at unhealthy loss-ratio assumptions.
Audit-trail-first design is not a compliance afterthought. It is the thing that makes the rest of the architecture deployable.
The same architecture, in adjacent markets
The three-loop pattern is not specific to MGAs. Lenders are using AI-supervised loops to price geographic concentration in mortgage portfolios where flood, wildfire, and severe convective storm risk are no longer negligible. Real estate investors and REITs are underwriting climate-exposed assets with parallel workflows: an analyst loop, a deal-team loop, and a tenant or customer-facing loop, all feeding a portfolio-governance ledger that satisfies investment-committee review. Parametric reinsurers structure coverage at the trigger level, and the supervised-loop pattern shows up there too. The common ground is the same problem an MGA faces - decentralized decisions accumulating correlated tail exposure invisibly. Adjacent markets are converging on the same architectural answer, and the MGA architecture matures alongside them.
What this means now
The convergence is recent. LLM tool-use reliability crossed a usable threshold in 2024-2025. Capacity partners post-2023 cat losses are explicit about wanting tighter governance. State regulators are landing AI-specific guidance in real time. The MGAs that treat governance as a product feature, not a compliance overhead, will earn structural pricing power on capacity over the next cycle.
Brollygraph applies this architecture to private flood, a category we believe is mispriced, under-distributed, and overdue for modern underwriting. The architecture itself is peril-agnostic. We expect to see other specialty insurtechs converge on it over the next 24 months.