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, and industry surveys put 20-40% of commercial COPE inputs as stale or inaccurate by renewal (CAPE Analytics). Computer vision tools built for residential hazard scoring are now surfacing the occupancy shifts and construction changes self-reported data misses. That raises a filing question before it raises a pricing one.

Key Takeaways

40-60% loss cost spread between ISO Class 1 and Class 6 on otherwise identical building profiles, with a single misclassified COPE element able to shift the indicated loss cost 15-30% before any company adjustment.
20-40% of commercial COPE inputs are stale or inaccurate at renewal, because tenant turnover, unpermitted renovation and sprinkler impairment generate no automatic update to the underwriting record.
ISO Class 3 and Class 4 buildings can present the same overhead signature while carrying loss cost estimates 20-30% apart. The distinguishing features are the interior frame and roof deck.
$9M to $13.5M of annual expected loss error on a $500M book where 15% of locations are misclassified by one construction class at a 60% loss ratio.
Property rates fell 9% in Q1 2026, a seventh consecutive quarterly decline, with Aon recording an average property change of -15%. Stale COPE data costs more when there is no rate margin covering it.

What a Misclassified Element Does to the Applied Loss Cost

COPE feeds the ISO loss cost selection at rating through four dimensions. Construction runs six ISO classes, from Class 1 fire-resistive steel and concrete to Class 6 modified combustible wood frame, with a 40-60% premium spread across the extremes. Occupancy carries the hazard class of the building's use. Protection covers the public protection class plus private suppression. Exposure covers adjacency, total insured value against replacement cost, and concentration on schedule risks.

Because those factors multiply, a single misclassified element can move the indicated loss cost 15-30% before any company adjustment. A Class 9 fire district without sprinklers is a different base loss cost from the same construction in a Class 4 district with a wet-pipe system.

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

The decay is structural rather than accidental. An insured reports COPE at application or last field survey, and the data ages against a building that keeps changing. A dry-goods retailer becomes a restaurant with high-BTU cooking equipment. A wood mezzanine goes into a masonry building and changes its ISO class with no policy endorsement. Sprinkler zone valves go unmaintained into partial impairment, removing a protection credit already factored into the rate. None of it triggers an update, so the renewal loss cost describes a risk that may not exist.

What Imagery Infers, and What That Is Worth

Residential scoring works because the structural logic is standardized: roof condition, covering age, wildfire fuel proximity, canopy overhang, maintenance quality. Commercial buildings do not cooperate.

Occupancy is the hardest inference. A 10,000-square-foot structure can hold a dentist, a light manufacturer, a dry cleaner or a restaurant kitchen and look identical from outside. Models work from signage, rooftop equipment, loading docks, parking layout and permit history, and accuracy falls sharply on generic buildings. Multi-tenant strip centers add the rule that the highest-hazard tenant, not the average, sets the occupancy loading for the whole structure.

Construction is nearly as constrained. ISO Class 1 fire-resistive and Class 2 masonry non-combustible present the same solid walls, low-slope roofs and footprints from above. A joisted masonry Class 4 and a non-combustible steel-frame Class 3 can sit on the same parcel with the same overhead signature and loss costs 20-30% apart, because the distinction is the interior frame and roof deck. Sprinkler presence shows only as a riser and backflow assembly near the entry, detected inconsistently across imagery resolutions.

So the useful output is change detection rather than classification. EagleView's Horizon, launched April 2026 on a 3.5 billion-image library covering 96% of the US population, compares current captures against historical imagery to flag buildings whose structure changed since the last survey.

Verisk's Commercial GenAI Underwriting Assistant does the same at the attribute level, surfacing a flagged discrepancy for human resolution rather than updating the COPE record and recalculating the rate.

The arithmetic behind the flag is what matters. Take a $500M book where analysis flags 15% of locations as potentially one construction class off. At a 20-30% loss cost difference and a 60% loss ratio, the expected loss error on that portion runs $9M to $13.5M a year. Whether that reaches materiality depends on the carrier's credibility standards and the other adjustments in the filing, but it is not noise in a trend selection.

The soft market removes the cushion. Property rates fell 9% in Q1 2026 and Aon recorded an average property change of -15%, recovered slightly from -18% in Q4 2025. A carrier on stale data in that environment writes a disproportionate share of the risks where actual loss cost exceeds the filed rate, because competitors with the imagery are pricing the genuinely well-constructed risks away from it.

Correcting the Record Is Not the Same as Changing the Method

The filed rate uses ISO construction class, occupancy group and protection class as lookup inputs. Applying an AI-derived classification that differs from the insured-submitted one means rating against inputs whose provenance the rating manual does not describe.

Most commercial property filings reference the insured's reported construction, occupancy and protection data without specifying whether it must be self-reported or may be supplemented by third-party AI-derived attributes. That silence is the interpretation question, and it decides whether this is a methodology change requiring a new filing or a data quality improvement inside the existing one. Prior approval, file-and-use and use-and-file states differ on timing, not on the underlying obligation: a material change in the basis used to determine the applicable rate requires a filing.

The regulatory answer is not finished. The NAIC AI Model Bulletin, in force across 24 states as of August 2025, requires documented governance of AI systems including third-party data quality and model validation, and the Third-Party Data and Models Task Force formed in 2024 is still developing how AI-derived data used in rate filings should be treated. Carriers deploying these attributes today are making the judgment inside that gap.

Where it goes wrong is selectivity. Correcting misclassifications only in the direction that improves book composition is a rating methodology problem whether the correction was AI-assisted or human-initiated, and the flag list makes selective correction easy in a way a field survey program never did.

A carrier that confirms each flag with an underwriter, updates the record against the same class definitions the filed manual uses, and applies the resulting change consistently is doing what the framework expects. One that lets the attributes update the ISO classification automatically has a divergence between applied and filed basis that will look like accurate pricing right up until an examination reads it the other way.

Computer Vision Finds New Holes in Commercial Property COPE Data

Key Takeaways

What a Misclassified Element Does to the Applied Loss Cost

What Imagery Infers, and What That Is Worth

Correcting the Record Is Not the Same as Changing the Method

Further Reading

Sources