Integrating CAT Models (AIR, RMS, CoreLogic) into the Underwriting Workbench

CAT model outputs — AAL (Average Annual Loss), PML (Probable Maximum Loss), and EPS (Expected Probable Severity) at various return periods — have historically lived in a workflow that's separate from the commercial underwriting submission review. The CAT modeling team runs portfolio analytics. The underwriting team reviews individual submissions. The two workflows intersect at accumulation reports and treaty renewal, but not at the submission intake stage where individual risk selection decisions are made.

That separation made operational sense when CAT model queries required batch processing and human interpretation. It's less defensible now that probabilistic CAT model outputs can be queried at the individual risk level, at submission time, with return periods and exceedance probability curves available as risk-level signals rather than only as portfolio aggregates. This piece covers what it takes to connect CAT exposure signals to the underwriting workbench — and where the technical and operational challenges in doing so actually lie.

How AIR Worldwide, Moody's RMS, and CoreLogic Differ in Practice

Carriers writing commercial property typically license at least one probabilistic CAT model — and increasingly two or three for key hazard types, particularly wind and earthquake. The three major vendors — AIR Worldwide (now part of Verisk Atmospheric & Earthquake Solutions), Moody's RMS, and CoreLogic Property & Hazard Intelligence — each produce loss estimates from different underlying model methodologies, different vulnerability functions, and different stochastic event sets.

The practical implication is that for the same location and building characteristics, the three models can produce materially different AAL and PML estimates. This isn't model error in the pejorative sense — it reflects genuine scientific uncertainty about hazard frequency and severity in many territories, particularly for secondary perils (severe convective storm, winter storm, wildfire) where historical data is thinner than for named windstorm and earthquake. Model spread between the three platforms on U.S. Gulf Coast hurricane AAL for a specific commercial property location might be 15-25%. On California earthquake, model spread can be larger.

For underwriting workbench integration, the model selection and blend methodology matters. Using a single model for all perils produces a point estimate with embedded model uncertainty that isn't visible. Using two or three models and presenting the range — or a blended AAL using explicit weights per peril — provides the underwriter with the uncertainty signal, not just the central estimate. For individual commercial property accounts, the spread between model AAL estimates is itself an underwriting signal: high model uncertainty on a coastal Florida account is a risk attribute, not just a data quality issue.

Risk-Level CAT Queries at Submission Time

Connecting CAT model output to the submission workflow requires two technical capabilities: geocoding the submission location to the precision required by the CAT model, and querying the model's API (or a data service layer over the model) at the time the ACORD 140 submission is processed.

Geocoding precision matters because CAT model hazard grids have finite resolution — typically 30-meter to 1-kilometer cells depending on the model and peril. A submission with a street address geocodes to a specific lat/long, which falls in a specific CAT model cell, which has specific hazard parameters. An address-level geocode is adequate for most commercial property CAT screening; a property centroid or parcel centroid is more precise and preferable for high-TIV risks. The difference between an address geocode and a parcel centroid can matter for coastal locations near a hazard gradient.

ACORD 140 provides the location-level inputs needed for the CAT model query: street address for geocoding, construction code, occupancy code, year built, and TIV. For commercial properties with multiple locations on a single ACORD 140, each location requires a separate CAT model query and the results need to be aggregated at the account level. The aggregate AAL and aggregate PML at a specified return period (100-year, 250-year, 500-year) are the account-level outputs that get incorporated into the underwriting risk score.

What CAT Signals Should Feed Into Appetite Rules

Not all CAT model outputs are equally useful at the submission triage stage. The most useful signals for appetite pre-screening are those that produce binary or threshold-based decisions without requiring a full actuarial review of the CAT modeling output:

AAL relative to TIV. AAL expressed as a percentage of TIV (the "AAL ratio") provides a normalized CAT cost signal that's comparable across accounts with different TIVs. An appetite threshold of "AAL > X% of TIV" can be evaluated automatically at submission time — accounts above the threshold route to underwriter review, not auto-refer. The threshold value should be calibrated against the carrier's reinsurance cost structure and target loss ratio, not set as an arbitrary industry benchmark.

PML at the treaty aggregate limit. If the carrier's per-occurrence reinsurance program attaches at $2M (for example), an account with a 250-year PML above $1.5M is consuming a significant fraction of the retainable capacity. The CAT model query at submission time can produce this figure, and the appetite rule can route high-PML accounts to senior underwriter review regardless of how the risk scores on other dimensions.

CAT zone accumulation signal. The query can also return the carrier's current aggregate exposure in the risk's CAT zone — triggering a capacity warning if adding the new risk would push the zone accumulation above a defined threshold of the reinsurance treaty limit. This requires the CAT system to have current portfolio data, not just risk-level model outputs.

A Scenario: Integrating CAT Signals Into a Commercial Property Workbench

Consider a commercial property carrier in the Southeast writing primarily light commercial and mixed-use accounts in the $500K to $5M TIV range. The carrier's reinsurance structure includes a per-occurrence program attaching at $2M and an aggregate CAT program with a Florida wind sublimit. The underwriting team reviews individual accounts without real-time access to CAT model estimates — the CAT modeling work happens quarterly as a portfolio exercise.

A submission arrives for a commercial retail strip in a coastal Pinellas County, Florida ZIP code: TIV $2.8M, Masonry Non-Combustible construction, 1987 vintage, commercial retail occupancy. The rate charged is adequate at the actuarial level if the loss cost is what the ISO territory and construction class suggest. But a CAT model query on this account would show a 100-year return period PML of approximately $1.5-2.0M depending on model — roughly 54-71% of TIV — reflecting the high-hazard coastal Florida wind exposure at that location.

At current TIV, this single account would consume a material fraction of the per-occurrence reinsurance capacity for this loss scenario. Without the CAT signal at submission time, the underwriter is rating this as a standard commercial retail account with territory and construction adjustments — not evaluating its reinsurance structure implications. With the CAT signal in the workbench, the routing logic would flag the account for senior underwriter review specifically on the CAT exposure basis, where the reinsurance cost of writing this risk can be explicitly evaluated against the premium.

The Portfolio Accumulation View at the Workbench Level

Individual risk CAT signals are valuable, but the portfolio accumulation view is what makes the CAT integration operationally meaningful. A carrier's reinsurance program is structured around expected aggregate losses by zone and peril — and the individual risk selection decisions during a policy year determine whether the actual accumulation stays within treaty limits or creates retentions the carrier didn't intend.

Connecting the underwriting workbench to a live CAT accumulation view — showing current aggregate TIV and modeled AAL by zone, updated in near-real-time as new submissions are bound — makes the accumulation constraint visible to underwriters making daily individual risk decisions. When the coastal Florida zone is approaching the treaty aggregate limit, that information should change how the underwriter evaluates new coastal Florida submissions, not be discovered at the next quarterly portfolio review.

We're not saying that every underwriter needs to be a CAT modeler — they don't, and the complexity of CAT model outputs is such that surfacing the right simplified signals matters more than providing raw model access. What we are saying is that the separation between "CAT modeling team" and "underwriting team" as organizational silos, with quarterly information exchange, is a workflow design from an era when real-time CAT queries weren't feasible. They're feasible now, and the workflow should reflect that.

Implementation Practical Considerations

Integrating CAT model queries into the submission workflow requires API access to the CAT model vendor's services or a data intermediary. AIR Worldwide, Moody's RMS, and CoreLogic all offer API-based risk query services — the format, latency, and cost structure vary by vendor and contract. For a carrier licensing one or more of these models for portfolio analytics, the incremental cost of enabling API-based risk-level queries is generally much lower than the full portfolio modeling license.

Latency is a practical constraint. A CAT model query for a single commercial property risk takes 1-5 seconds depending on vendor and complexity. That's acceptable for an ACORD 140 submission intake workflow where the goal is pre-screening before underwriter review — there's no reason the CAT query needs to be synchronous with the submission receipt. Running the query asynchronously and attaching the result to the submission record before it hits the underwriter queue is the appropriate architecture.

The data governance requirements for CAT integration also need attention. Submission data — specifically location and TIV data — is being sent to the CAT vendor API for risk-level queries. The carrier's data governance policy for submission data handling should explicitly address CAT model queries, including whether the submission data is retained by the vendor and how it's used.