AI Replaces the Weekly PML Run: Cat Accumulation Goes Real-Time

A cohort of primary carriers is replacing the weekly or monthly batch probable maximum loss cycle with AI-driven accumulation monitoring that updates retained catastrophe exposure at the moment a policy binds, not one to three weeks later once geocoding and model runs catch up. With the July 1, 2026 property cat renewal delivering risk-adjusted rate reductions of 20% to 25% or more on the strongest North American programs (Gallagher Re, July 2026), carriers now have an economic incentive to know their live accumulation, not their last Friday's accumulation, before deciding how much reinsurance to buy.

How the Weekly Batch PML Cycle Actually Works

The conventional cycle at most primary carriers runs on a predictable rhythm that has changed little in two decades. New business and endorsements flow out of the policy administration system as flat-file extracts, typically on a nightly or weekly schedule, into a geocoding queue that resolves each risk address to latitude and longitude, then assigns it to the peril zones, distance-to-coast bands, and construction and occupancy classes the catastrophe model requires. Once geocoded, the incremental book gets appended to the insured-value database and stacked against the carrier's licensed model, AIR, CoreLogic, RMS, or KCC under the NAIC's approved-vendor list for risk-based capital purposes (NAIC Property and Casualty RBC Instructions), to produce an updated 100-year and 250-year PML by zone. The full cycle, from bind to a refreshed PML the underwriting desk can actually see, commonly takes one to three weeks, longer when geocoding queues back up around a renewal season or a large new program onboarding.

The lag is not a technology failure so much as an artifact of how the pieces were built independently: a policy admin system optimized for transaction throughput, a geocoding vendor optimized for address-match accuracy, and a catastrophe model optimized for peril science, none of them designed to share a common data pipeline in real time. An underwriter binding a large coastal commercial property risk on a Tuesday has no way of knowing, at the moment of bind, whether that risk pushes a zone accumulation past its internal limit. The zone limit breach, if there is one, surfaces in next week's batch run, by which point the policy is already in force and the only remedies are non-renewal at expiration or a facultative reinsurance placement arranged after the fact.

What "Real-Time" Actually Means in Practice

Vendors and carriers use "real-time" loosely, and the term covers at least three distinct capabilities that mature at different speeds. Bind-time risk selection is the narrowest and most mature: the underwriting system queries a live accumulation surface at the moment a quote is generated, so the underwriter sees the marginal effect of the specific risk on the zone's PML before binding, rather than discovering the breach in next week's report. Zone limit enforcement goes a step further, hardcoding accumulation thresholds into the binding workflow itself so that a risk pushing a zone past its internal cat budget triggers a referral or an automatic decline rather than relying on an underwriter to consult a separate report. Event-response reallocation is the least mature and most operationally demanding: as a storm track firms up, the system re-runs live accumulation against the forecast footprint so capital and reinsurance allocation decisions can shift within the same event cycle, rather than waiting for the standard batch schedule to catch up after landfall.

Each capability draws on a different generation of AI tooling. Bind-time selection depends on machine-learned geocoding and property-characteristic inference that can resolve an address and estimate construction, roof age, and distance-to-coast in seconds rather than the batch hours a rules-based geocoder requires, closing the gap between rating and risk assessment. Zone limit enforcement leans on rules engines layered with anomaly detection that can flag when a bind pattern looks like it is clustering risk in a way historical underwriting guidelines did not anticipate. Event-response reallocation is where large language models and agentic workflows are newest, stitching together live NHC or USGS forecast feeds, the in-force book, and treaty terms into a single reallocation recommendation an underwriting or ceded-re team can act on inside hours rather than days. Most carriers currently deploying AI accumulation tools have bind-time selection in production and zone limit enforcement in pilot; full event-response reallocation remains rare outside the largest specialty and reinsurance-facing books.

Vendor Landscape: Three Different Approaches to Continuous Accumulation

The three vendors actuaries evaluating these systems are most likely to encounter differ in where they sit in the data pipeline, which shapes how genuinely continuous their accumulation view actually is. Moody's RMS built ExposureIQ on its Intelligent Risk Platform, a unified data lake the company describes as enabling access to RMS models, third-party data, and a client's own portfolio data "in real time" (Moody's RMS, Intelligent Risk Platform product documentation, 2026), with live peril-event data feeding directly into the platform to generate near-real-time exposure reporting during an active storm. Verisk's approach centers on Synergy Studio, a catastrophe modeling and risk management platform built for real-time footprint tracking, alongside the Verisk Model Exchange, which runs multiple third-party model views on a single governed, API-accessible financial engine (Verisk, Synergy Studio and Model Exchange product documentation, 2026). Guidewire takes a third path, embedding accumulation and aggregation management inside DataHub as part of its broader cloud-native InsuranceSuite, with recent ecosystem integrations, ICEYE satellite-derived NatCat Insights and a data partnership with Intelligent AI for high-resolution property and rebuild-cost data, feeding directly into the ClaimCenter and underwriting workflow through an API-first, event-driven architecture (Guidewire, product and partner documentation, 2026).

The honest answer, across all three, is that "real-time" in current commercial deployments usually means the underlying data lake or model engine can be queried on demand, not that every underwriting decision automatically triggers a full accumulation recompute. An actuary evaluating any of these systems should ask the vendor directly what event actually triggers a refresh, whether it is bind, endorsement, a scheduled interval measured in minutes rather than days, or an external catastrophe forecast update, because that answer determines whether the tool closes the batch lag or merely compresses it.

The RBC Timing Problem: A Point-in-Time Charge Against a Moving Target

Risk-based capital catastrophe charges are computed once, from year-end data. Under the NAIC's Property and Casualty RBC formula, the catastrophe risk charge for earthquake and, since the 2025 data year, wildfire and convective storm equals the modeled loss at the worst year in 100, net of reinsurance recoveries, plus a 10% surcharge applied to the ceded portion of that modeled loss (NAIC Property and Casualty Risk-Based Capital Instructions; NAIC Catastrophe Risk (E) Subgroup materials, 2025-2026). That figure is a snapshot as of December 31. It says nothing about what the carrier's retained accumulation looked like on September 10, in the middle of peak Atlantic hurricane season, when the in-force book, the treaty attachment points already utilized by prior events, and the actual exposure concentration could all have looked materially different from either the January 1 starting point or the December 31 ending point the RBC filing captures.

Continuous accumulation monitoring makes that gap visible in a way batch monitoring never did. A carrier running weekly PML snapshots historically had no practical way to know how far its intra-year peak accumulation diverged from its year-end RBC input, because it never captured enough intermediate data points to reconstruct the curve. A carrier running bind-time accumulation monitoring has that data by construction, which raises an uncomfortable question for reserving and capital actuaries: if the December 31 R-CAT charge is computed against a portfolio that peaked materially higher in September, is the resulting RBC ratio actually representative of the capital adequacy the carrier carried through the season that mattered, or only of the capital adequacy it happened to be carrying on the one day regulators measure? Nothing in the current RBC formula requires disclosure of the intra-year peak, but an actuary with access to a continuous accumulation feed now has the information needed to test that gap internally, and Own Risk and Solvency Assessment narratives are a natural place to document the finding even where the statutory formula does not require it.

The Over-Optimization Risk in a Soft Market

The economics pushing carriers toward continuous monitoring are real. Global reinsurer capital reached a record $790 billion as of March 31, 2026 (Aon, midyear 2026 report), Gallagher Re put dedicated reinsurance capital at $648 billion at the end of 2025, up 11% year over year, with alternative capital in ILS and cat bond structures growing 18% to $135 billion over the same period (Gallagher Re, First View: Options and Opportunities, July 2026). First-half 2026 catastrophe losses ran approximately $38 billion through June 15, below the ten-year average, and Gallagher Re estimates 2026 reinsurer returns on equity in the 14% to 15% range (Gallagher Re, First View: Options and Opportunities, July 2026), a combination that gives reinsurers every incentive to keep deploying capacity aggressively into a buyer's market.

A carrier with a genuinely accurate, continuously updated view of its retained accumulation has a defensible case for trimming reinsurance purchase to match its actual exposure rather than a conservative batch-lagged estimate padded for uncertainty. That is the intended benefit of the technology. But the same tool used carelessly creates a specific failure mode: a carrier that reads its real-time accumulation dashboard as license to buy reinsurance down to the exact modeled PML, with no margin for model error, data latency between the dashboard and the actual bound book, or a peril the model underweights, has replaced a conservative approximation with a precise number that is still wrong in ways the dashboard cannot show. Continuous monitoring reduces estimation lag; it does not reduce model risk, and treating the two as the same thing is how a carrier ends up under-protected the one season the softening cycle turns.

There is a related exposure-visibility problem specific to managing general agents. MGA-fronted programs typically report bound business to the carrier or reinsurer through periodic bordereaux, often on a 30- to 45-day cycle rather than at the moment of bind, which means a carrier's continuous accumulation system is only as current as its slowest-reporting delegated authority partner. Gallagher Re's March 2026 report on AI-related aggregation risk, "Smart Systems, Blind Spots: Rethinking Insurance for the AI Era," developed with MIT and Testudo, focuses on a related but distinct concern, correlated losses from shared AI model failures rather than bordereaux lag specifically, but the underlying theme is the same: aggregation blind spots do not disappear just because a carrier's own book is monitored continuously if a material share of bound risk arrives through a channel still running on batch-cycle reporting.

Documenting Continuous Accumulation in Rate Filings

Most cat loads embedded in rates currently in force were derived from batch-mode accumulation data, the weekly or monthly snapshot methodology described above, because that was the state of the art when the underlying filing was prepared. Actuaries preparing rate filing support in jurisdictions that scrutinize catastrophe load methodology should treat a shift to continuous monitoring as a disclosure item, not a silent internal upgrade. The actuarial memorandum should state plainly which accumulation methodology, batch or continuous, produced the exposure data behind the filed cat load, note the effective date of any transition between the two, and address whether continuous monitoring changed the modeled accumulation materially enough to warrant a load adjustment independent of any change in the underlying catastrophe model version. A regulator reviewing a filing that cites a lower cat load has a reasonable basis to ask whether the reduction reflects genuinely lower modeled risk or simply a more precise accumulation estimate replacing a conservative batch-lagged one, and the filing should answer that question before it is asked.

Why This Matters for Actuaries

From reviewing cat program structures at regional carriers before and after AI accumulation tool deployments, the shift from weekly to continuous monitoring typically reveals 3% to 7% more correlated exposure in the retained layer than the weekly snapshot implied, a material gap at exactly the moment carriers are trimming reinsurance in a soft market. Pricing actuaries should not assume a lower observed accumulation under a new continuous system means genuinely lower risk; it may mean the old batch system was simply blind to concentration the new one now sees. Reserving and capital actuaries should treat the December 31 R-CAT snapshot as a floor on disclosure, not a ceiling, and use intra-year accumulation data, where available, to test whether statutory capital adequacy held up during the season it actually mattered. Ceded-re and enterprise risk actuaries evaluating a proposed reduction in catastrophe reinsurance purchase should ask whether the case for trimming rests on a genuinely more accurate exposure view or on a lower number produced by a faster but not necessarily more conservative measurement process. The tools are real and the operational benefit, catching a zone limit breach at bind instead of a week later, is not in dispute. What is still unsettled is whether the capital and pricing frameworks built for a batch-cycle world have caught up to a book of business that now updates every time someone signs a binder.