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

A cohort of primary carriers is replacing the weekly batch probable maximum loss cycle with accumulation monitoring that updates retained catastrophe exposure at the moment a policy binds, rather than one to three weeks later once geocoding and model runs catch up.

The timing is not incidental. With the July 1 renewal delivering risk-adjusted reductions of 20% to 25% or more on the strongest North American programs, knowing live accumulation rather than last Friday's changes how much reinsurance a carrier buys.

Key Takeaways

• One to three weeks is the conventional lag from bind to a refreshed zone PML, longer when geocoding queues back up around a renewal season.
• Three capabilities hide behind "real-time": bind-time selection, zone limit enforcement in the workflow, and event-response reallocation against a live storm footprint.
• The R-CAT charge is a December 31 snapshot, the modeled worst year in 100 net of reinsurance plus a 10% surcharge on the ceded portion, and says nothing about the September peak.
• MGA bordereaux typically report on a 30 to 45 day cycle, so a continuous system is only as current as its slowest delegated authority partner.
• $790 billion of reinsurer capital against first-half catastrophe losses near $38 billion, below the ten-year average, is what makes trimming reinsurance tempting.

The Batch Cycle and Its One-to-Three-Week Lag

The conventional rhythm has changed little in two decades. New business and endorsements leave the policy system as flat-file extracts on a nightly or weekly schedule, enter a geocoding queue that resolves each address and assigns peril zone, distance-to-coast band, construction and occupancy, then run against the licensed model to produce updated 100-year and 250-year PMLs by zone.

The lag is an artifact of independent design rather than a technology failure. A policy system built for transaction throughput, a geocoder built for address-match accuracy and a catastrophe model built for peril science were never meant to share one pipeline.

The consequence sits with the underwriter. Someone binding a large coastal commercial risk on a Tuesday cannot know whether it pushes a zone past its internal limit. The breach surfaces in next week's batch, with the policy already in force and the remedies reduced to non-renewal at expiration or facultative cover after the fact.

Real-Time Means Three Different Things

Bind-time risk selection is narrowest and most mature: the underwriting system queries a live accumulation surface when a quote is generated, so the marginal effect on the zone PML is visible before binding. Zone limit enforcement hardcodes thresholds into the workflow, so a risk past the cat budget triggers a referral or decline rather than relying on someone consulting a report. Event-response reallocation re-runs accumulation against a firming storm footprint so capital and reinsurance decisions shift inside the event cycle.

Each rests on different tooling, and they mature at different speeds. Selection needs property-characteristic inference resolving in seconds rather than batch hours. Enforcement needs anomaly detection that flags a bind pattern clustering risk in ways the guidelines did not anticipate. Most carriers have selection in production and enforcement in pilot; full reallocation is rare.

Vendor	Core Product	Architecture	What to Ask About Refresh Frequency
Moody's RMS	ExposureIQ on the Intelligent Risk Platform	Unified data lake; live peril-event feeds	Is the "near-real-time" exposure report event-triggered or on a fixed polling interval during active storms?
Verisk	Synergy Studio; Model Exchange	Cloud-native modeling engine with API access to multiple third-party models	Does the accumulation surface update at bind, or only when a scheduled model run executes?
Guidewire	DataHub with ICEYE and Intelligent AI integrations	Event-driven, API-first, embedded in policy and claims workflow	Is accumulation aggregation native to DataHub, or dependent on a third-party accelerator's own refresh cadence?

The vendors sit at different points in the pipeline. Moody's RMS built ExposureIQ on a unified data lake with live peril-event data feeding exposure reporting during an active storm. Verisk centers on Synergy Studio for footprint tracking alongside a Model Exchange running several third-party model views on one governed financial engine. Guidewire embeds accumulation inside DataHub, fed by ICEYE satellite insights and high-resolution rebuild-cost data.

Across all three, real-time in current deployments usually means the data lake can be queried on demand, not that every underwriting decision triggers a full recompute. What matters to an actuary is what event fires a refresh: bind, endorsement, a scheduled interval measured in minutes, or an external forecast update. That answer determines whether the tool closes the batch lag or compresses it.

It also governs how to read the output: a lower observed accumulation under a continuous system may mean the batch version was blind to concentration rather than that risk fell.

A Point-in-Time Capital Charge Against a Moving Book

The risk-based capital catastrophe charge is computed once, from year-end data: the modeled loss at the worst year in 100 net of reinsurance recoveries, plus a 10% surcharge on the ceded portion, for earthquake and, from the 2025 data year, wildfire and convective storm.

That is a December 31 snapshot. It says nothing about retained accumulation on September 10, mid peak Atlantic season, when the in-force book, the treaty attachment already used and the concentration itself could all have looked different from either end of the year.

Continuous monitoring makes that gap measurable for the first time. A carrier on weekly snapshots never captured enough intermediate points to reconstruct the intra-year curve; one monitoring at bind has it by construction. That raises a question for capital actuaries: if the year-end charge is computed against a portfolio that peaked materially higher in September, does the ratio describe the adequacy carried through the season that mattered, or the adequacy on the one day regulators measure? Nothing in the formula requires disclosing the peak, and an Own Risk and Solvency Assessment narrative is where the finding can go.

The soft market makes this bite now. Reinsurer capital hit a record $790 billion at March 31, dedicated capital was $648 billion at end-2025 and up 11%, alternative capital grew 18% to $135 billion, first-half catastrophe losses ran roughly $38 billion below the ten-year average, and reinsurer returns on equity are tracked at 14% to 15%. A carrier with an accurate live view has a case for trimming its purchase toward actual exposure.

The failure mode is buying down to the exact modeled PML with no margin. Continuous monitoring removes estimation lag; it does not remove model error, latency between the dashboard and the bound book, or a peril the model underweights. Precision is not conservatism, and the difference shows up in the one season the cycle turns.

The delegated channel undercuts it further. MGA-fronted programs typically report through bordereaux on a 30 to 45 day cycle rather than at bind, so a carrier monitoring its own book continuously still carries a blind spot the size of its delegated authority volume. A cat load filed off continuous data when the prior filing used batch snapshots is a disclosure item rather than a silent upgrade, because a regulator seeing a lower load can fairly ask whether risk fell or the measurement merely got sharper.