Computer Vision Finds New Holes in Commercial Property COPE Data

ISO loss cost relativities differ 40-60% between fire-resistive and combustible-frame commercial construction, yet industry surveys consistently find that 20-40% of commercial COPE inputs are stale or inaccurate by renewal (CAPE Analytics, 2025). Computer vision tools extending from residential hazard scoring into commercial building analysis now surface the occupancy changes and construction modifications that self-reported data misses, raising a direct question for actuaries: when AI-derived COPE data reclassifies a book, does the carrier need to refile?

COPE Data and the Six-Class Construction Problem

The COPE framework organizes commercial property risk into four dimensions that together produce the loss cost estimate underwriters apply to a given location. Construction covers ISO classification across six classes, from Class 1 fire-resistive construction in steel and concrete frames to Class 6 modified combustible wood-frame buildings, with a premium spread that can reach 40-60% between the extremes for an otherwise identical building profile (ISO Commercial Property Program). Occupancy captures the hazard classification of the building's use: a dry-goods retailer, a restaurant, a warehouse, and a light-manufacturing tenant each carry materially different frequency and severity expectations even within the same physical structure. Protection encompasses the public protection class assigned to the fire district plus private suppression, primarily whether a sprinkler system is installed and maintained. Exposure covers proximity to adjacent hazards, total insured value relative to replacement cost, and for schedule risks, how risk concentration and neighboring occupancies might amplify a loss.

Each dimension feeds into the ISO loss cost selection at rating. A building in a Class 9 fire district without a sprinkler system carries a very different base loss cost than the same construction in a Class 4 district with a wet-pipe suppression system, and the multiplication of those factors means a single misclassified element can shift the indicated loss cost by 15-30% before any company adjustment. Accurate COPE data is what converts the filed rate manual into a loss cost that actually reflects the risk on the policy.

The self-reported data problem compounds at every renewal. An insured submits COPE information at application or at the last field survey. That data then ages. Tenant turnover shifts occupancy class: a dry-goods retailer becomes a restaurant with high-BTU cooking equipment, carrying substantially higher fire frequency exposure. Unreported renovation work adds a wood mezzanine to a masonry building, changing its ISO construction class without triggering a policy endorsement. Sprinkler systems fall into partial impairment from unmaintained zone valves, removing a protection credit that was factored into the original rate. None of these changes generate an automatic update to the carrier's underwriting record. The loss cost applied at renewal reflects a risk profile that may no longer exist.

From reviewing commercial property loss cost analyses for carriers across multiple states, the COPE data override rate, where underwriters manually correct submitted occupancy or construction classifications during renewal review, typically runs 15-25% of accounts. One in five to six commercial risks in a standard book is priced against a classification that a reviewing underwriter would change on inspection. Scale that across a $200M commercial property book and the loss cost error embedded in the unreviewed portion is not a rounding adjustment. It is a systematic understatement concentrated in the accounts where the most material changes have occurred and gone unreported.

COPE Element	Common Error Type	Primary Change Driver	Loss Cost Channel
Construction	Frame reported as masonry; unreported additions lower the class	Renovations without permit pull or policy endorsement	Direct factor in Group I/II base loss cost selection
Occupancy	Hazard class shifts with tenant turnover	Retail-to-restaurant, warehouse-to-light-manufacturing	Occupancy loading on the base loss cost
Protection	Sprinkler impairment; PPC rezoning unreported	Deferred maintenance, fire district boundary changes	Multiplier applied to the loss cost table
Exposure	Adjacent occupancy changes; TIV not updated for inflation	Neighboring tenant changes, unreplenished insured values	Coinsurance and limit adequacy factors

Commercial Buildings Seen from Above: What Imagery Can and Cannot Infer

Aerial and satellite imagery tools have scored residential property risk for several years. Roof condition, age of the covering material, proximity to wildfire fuel loads, overhanging tree canopy, and maintenance quality are all legible from high-resolution overhead imagery when the model is trained on labeled examples. The residential application is bounded: a limited set of physical characteristics on a property type with highly standardized structural logic. Every house has a similar frame.

Commercial buildings do not cooperate in the same way. The occupancy inference problem is the hardest. A 10,000-square-foot commercial structure can house a dentist's office, a light manufacturing tenant, a dry cleaner, or a restaurant kitchen with high-BTU cooking equipment, and from the exterior these tenants can look nearly identical. Current commercial computer vision tools attempt occupancy inference from signage, rooftop equipment configuration, loading dock presence, parking lot layout, and permit history drawn from parcel data. The inference accuracy drops sharply for generic commercial buildings where exterior signals are ambiguous, and multi-tenant strip centers present the additional challenge that the hazard class of the highest-risk tenant, not the average, typically sets the occupancy loading for the entire structure under ISO rating rules.

Construction class inference faces a related constraint. ISO Class 1 fire-resistive construction and Class 2 masonry non-combustible construction look nearly identical from aerial imagery: both present solid exterior walls, flat or low-slope roofs, and similar building footprints. The distinction lies in the interior frame and the roof deck materials, which overhead imagery cannot directly observe. A joisted masonry building (Class 4) and a non-combustible steel-frame building (Class 3) can occupy the same parcel and present the same overhead signature while carrying loss cost estimates that differ by 20-30%. Fire suppression presence is similarly opaque from the exterior; the only direct visual signal is a sprinkler riser and backflow preventer assembly near the building entry, which current models detect inconsistently across imagery resolutions.

EagleView launched Horizon in April 2026 as an agentic geospatial intelligence engine drawing on more than 25 years of property intelligence and a 3.5 billion-image library covering 96% of the U.S. population (EagleView, April 2026). The platform enables time-series change detection, comparing current aerial captures against historical imagery to identify structural modifications that appear in newer images but not in images from prior policy periods. For commercial COPE validation, the most useful application is not point-in-time classification but change detection: identifying buildings where the current roof structure or exterior profile has changed materially since the last underwriting survey, then flagging those locations for human review rather than auto-reclassifying them. That distinction, flagging versus reclassifying, is where the actuarial and regulatory consequences diverge.

CAPE Analytics has extended its commercial property intelligence to derive COPE attributes including roof condition, occupancy signals, and exposure characteristics from aerial and satellite imagery across large portfolios (CAPE Analytics, 2025). The practical workflow for most carriers is that AI-derived attributes generate a flagged review list, not an automated COPE update, precisely because the construction and occupancy inference limitations require a second layer of confirmation before a classification changes and the rating algorithm recalculates the applied loss cost.

What Verisk's Commercial GenAI Underwriting Assistant Ingests

Verisk launched its Commercial GenAI Underwriting Assistant in September 2025 as part of the Augmented Underwriting Suite, integrating Touchstone for catastrophe model exposure and Rulebook for ISO loss cost data into a single submission-processing workflow (Verisk, September 2025). The system automates submission intake, summarizes complex property schedules, surfaces risk flags from AI-derived property attributes, and delivers actionable insights to the underwriter at the point of decision rather than requiring manual data lookup across disparate systems.

The actuarially significant integration is the one between AI-derived property attributes and the ISO loss cost data. When the system's imagery analysis produces an attribute, a construction class signal or an occupancy indicator, that diverges from what the insured submitted, the underwriter sees a flagged discrepancy. The system does not automatically update the COPE record or recalculate the loss cost; it surfaces the divergence for human resolution. That workflow design is deliberate: automated COPE reclassification without human review would create a gap between the filed rate methodology and the basis on which individual risks are actually rated, a gap the current regulatory framework has not addressed.

Early deployments showed meaningful accuracy gains from AI-assisted COPE validation. AIG reported that underwriting data collection and accuracy rates increased 15% after deploying generative AI platforms in early 2025 (CAPE Analytics, 2025). The gains flow primarily from submission normalization and data aggregation across fragmented sources, not from the construction-class inference problem at the center of the COPE accuracy question. The hard part, inferring interior construction characteristics and tenant occupancy from exterior imagery without a field survey, remains a limitation that the current generation of commercial AI tools manages through flagging and escalation rather than automated resolution.

The Rate-Indication Arithmetic When 15% of the Book Moves Construction Class

Consider a carrier with a $500M commercial property book where AI-derived construction analysis flags 15% of insured locations as potentially misclassified, suggesting the actual ISO construction class is higher than what appears on the policy. The actuarial question is whether that potential misclassification, if confirmed, produces a loss cost change material enough to affect the rate indication and, in states with prior approval requirements, to require a new filing.

The arithmetic is not complicated. A construction class error that moves a risk from ISO Class 2 to Class 4 carries a loss cost difference of approximately 20-30% for the affected locations (ISO Commercial Property Program). Applied to 15% of a $500M book at a 60% loss ratio, the expected loss cost error on the misclassified portion runs in the range of $9M to $13.5M in annual expected loss. Whether that figure is material to the rate indication depends on the carrier's credibility standards and the size of other adjustments in the filing, but it is not a number a certifying actuary can treat as noise in the development or trend analysis.

The soft market context sharpens the stakes considerably. Global commercial property insurance rates fell 9% in Q1 2026, the seventh consecutive quarterly decline (Marsh, Q1 2026 Global Insurance Market Index), with large-account U.S. property recording double-digit decreases. Aon's Q1 2026 data shows an average property rate change of -15%, slightly recovered from -18% in Q4 2025 (Aon, Q1 2026 Property Market Dynamics Report). When rates decline at this pace, carriers that carry stale COPE data face an adverse selection dynamic: competitors deploying AI-derived construction and occupancy intelligence identify the well-constructed, well-protected risks and price them more aggressively. The carrier with stale data writes a disproportionate share of the risks where the actual loss cost exceeds the filed rate. The loss cost error that is diluted in a rising-rate environment becomes the primary driver of adverse loss experience once price competition compresses margin.

Advanced data analytics applied to commercial property risk can raise premiums up to 15% and improve loss ratios up to 5% when the AI-derived inputs are accurate and properly integrated into the rating algorithm (McKinsey, cited in CAPE Analytics, 2025). The loss ratio improvement requires that the reclassification flow through to the applied rate, not just to an underwriting flag. A flagged discrepancy that does not change the charged rate accomplishes only the identification step; the loss cost correction requires that the updated COPE classification propagate to the renewal pricing on that specific account.

AI-Derived Classifications and the State Filing Clock

ISO commercial property rates in most states are filed as advisory loss costs, with company loss cost multipliers applied on top. The filed rate uses ISO construction class, occupancy group, and protection class as inputs to the loss cost lookup. When a carrier begins applying AI-derived COPE classifications that differ from insured-submitted classifications, it is rating risks against inputs that may no longer correspond to the definitions in the filed rating manual. The regulatory question is whether that constitutes a departure from the filed methodology requiring a new filing, or a data quality improvement within the existing methodology.

State rate filing requirements differ by state and filing law: prior approval, file-and-use, and use-and-file jurisdictions each have different timelines for when a changed methodology must be filed versus when it can be deployed while under review. The core obligation is consistent across all three: if the basis used to determine the applicable rate changes materially from what was filed, a new filing is required. Whether AI-derived COPE inputs constitute a methodology change or a data quality improvement depends on how the filed rate manual describes the source of those inputs. Most commercial property rate filings reference the insured's reported construction, occupancy, and protection data without specifying whether that data must come from self-reported sources or may be supplemented by third-party AI-derived attributes. That ambiguity is the filing interpretation question carriers are navigating now.

The NAIC AI Model Bulletin, adopted in December 2023 and in force across 24 states as of August 2025 (NAIC, 2025), requires carriers to maintain documented governance of AI systems used in the insurance lifecycle, including third-party data quality and model validation. The NAIC Third-Party Data and Models Task Force, formed in 2024, is developing a regulatory framework for evaluating AI-derived data used in rate filings, with particular attention to whether external AI attributes require a separate filing disclosure or are covered under existing data quality governance provisions. That framework has not been finalized, leaving carriers that deploy AI-derived COPE attributes in rating today to make a filing interpretation judgment in a regulatory gap.

The actuarial certification obligation is the more immediate issue. An actuary certifying a commercial property rate filing in a state that has adopted the NAIC AI Model Bulletin should document whether AI-derived property attributes were used to classify risks in the experience period and, if so, whether the AI-derived classification was validated against the filed classification definitions before application. The same obligation applies when AI-derived COPE flags were used to identify and reclassify accounts during the experience period: the representativeness of the rating data relative to what was filed is a material question for the certification, even if the current filing form does not explicitly ask it.

Carriers that use computer vision to identify COPE discrepancies and then update the COPE record after underwriter confirmation before renewing the policy are doing what the rating framework expects: correcting the classification to match the actual risk. The obligation is to document that the correction was validated, that the updated classification uses the same construction class definitions as the filed rate manual, and that the resulting rate change was applied consistently rather than selectively. Selective application, correcting misclassifications only in the direction that benefits the carrier's book composition, would itself constitute a rating methodology problem regardless of whether the reclassification was AI-assisted or human-initiated.

Carriers that treat AI-derived COPE attributes as a data quality input to the existing COPE classification process, requiring underwriter confirmation before any record updates, are on firm footing under current filing standards. Carriers that allow AI-derived attributes to automatically update the ISO classification used in rating, without underwriter confirmation or documented validation against the filed class definitions, have created a methodology divergence that a market conduct examination will eventually surface. The commercial property book that looks well-selected and accurately priced today, because the AI is catching the worst COPE errors, can develop adversely once an examination concludes that the applied rating basis departed from the filed methodology.