Observed Appetite: A Computational Framework for Measuring Commercial Insurance Carrier Underwriting Behavior at Distribution Scale

Determining which commercial insurance carriers will quote a given submission is a prerequisite to every commercial-lines placement, yet the signals available to independent agents – carrier-self-reported appetite guides, application-programming-interface (API) responses from a small set of integrated carriers, and individual agent experience – have not been formalized as a computational object, nor measured against each other at scale.

We call this measurement task observed appetite and distinguish it from the three established categories: stated appetite (carrier-published), API-revealed appetite (returned by integrated platforms), and verified appetite (confirmed directly with the carrier). Observed appetite is the empirical distribution of a carrier's quote-or-decline behavior computed from real submission outcomes through that carrier's own portal, partitioned by industry classification, geography, and risk size.

We formalize observed appetite as a five-stage computational task (Submit → Classify → Aggregate → Score Confidence → Reconcile) with an explicit four-way ontology of appetite sources. We describe the QuoteSweep system, which implements the observation infrastructure through AI-powered browser agents operating on authenticated carrier portals without requiring API partnerships. We evaluate the stated appetite baseline against itself across a corpus of 509 commercial property and casualty (P&C) carriers. This paper presents the framework and the stated-appetite baseline; production-scale observed-appetite measurement is left to future work.

Only 2.2% of carriers publicly disclose any industry × state × size interaction, only 1.2% annotate appetite at six-digit NAICS resolution, and only 4.7% of line-of-business commitments specify a revenue threshold. When the same carrier's published PDF appetite guide and its own public appetite web page are compared at the NAICS-2 sector level (n = 189 carriers, 3,780 cells), the PDF asserts 2.14× more sector inclusions than the web page (Cohen's κ = +0.25, 95% CI 0.22–0.28 – fair agreement under Landis & Koch). Two surfaces of the same carrier describe systematically different coverage scopes. We release the corpus as an open dataset for reproducibility.