// Commercial property
The SOV is a coverage document, not a rating document.
Merged cells. Footnotes inside value fields. TIV subtotals sitting in data rows. Secondary modifiers missing entirely. Then a margin clause caps recovery at a multiple of whatever number ended up in the file. On this line the schedule is not an input to underwriting. It is most of the underwriting.
We work the schedule as an artifact — the merged cell, the multi-building row, the blank modifier column — because that is where the money is decided.
// The situation
Rate is going back, and so are the terms.
Property rates fell 6.3% in Q2 2026 and 75% of broker respondents reported increased capacity; Marsh puts US property at −13% over the same period, an eighth consecutive quarterly decline.
CIAB Commercial P/C Market Index, Q2 2026; Marsh Global Insurance Market Index, Q2 2026 — same direction, different magnitude, and both worth printing.
Terms loosen with the rate: broader definitions, higher sublimits, lower deductibles. Rate is easy to give back later. Sublimit and deductible erosion is what hurts, and both live in fields on a schedule nobody has reconciled. That is where we start, and the first output is a list of the fields that disagree with each other.
// The stack you actually run
The artifacts and the models they feed.
Two details that separate people who have done this from people who have read about it.
First, a percentage wind/hail deductible may apply to the building limit of insurance or to the loss amount, which is a large difference on a partial loss — and it typically applies per building rather than per occurrence, so one hail event across a 40-location schedule can trigger 40 separate deductibles. Second, the modifiers that most change modeled loss are the fields most often blank: roof geometry, roof anchorage, roof age, wall cladding, window protection, cripple walls, soft story, first-floor elevation. They are usually not SOV columns at all. They live in narrative appraisals and engineering surveys, and when they are absent the model substitutes conservative regional defaults.
// The build
Four builds, and what each one is measured on.
SOV normalization with conflict surfacing
Arbitrary broker schemas mapped to your canonical location model, with merged cells, embedded subtotals and multi-building rows handled explicitly. When the SOV total disagrees with ACORD 140 TIV, the output says so and shows both. This is reconciliation, not extraction — an independent read of each document produces a confidently wrong file.
Surface: broker email and document store, out to the canonical location model in your policy administration system.
Geocode correction and building-to-policy matching
Geocode resolution decides everything downstream: a ZIP-centroid geocode on a coastal risk produces a fundamentally different loss estimate than a rooftop geocode. Addresses corrected, resolution recorded as a field rather than assumed, multi-building campuses reconciled to policy records.
Surface: your address and location service, written back to policy location records and the model exposure set.
COPE gap filling with provenance
Aerial imagery and property intelligence fill roof condition, geometry, material and degradation. Every filled field carries its source and its confidence, so an underwriter can see what was measured and what was inferred. Blank fields make the model reach for a conservative default, and neither side can then say how much of the quote is risk and how much is ignorance.
Surface: imagery and property-intelligence APIs, written back as tagged fields, versioned per refresh.
Valuation drift flagging
Surface schedules where reported values have not been refreshed against replacement cost, and quantify the coinsurance exposure at partial loss. Flag where a margin clause or an Occurrence Limit of Liability Endorsement turns a stale schedule number into a hard recovery cap.
Surface: policy administration values against the exposure set, on the renewal calendar.
What we measure
Hours per submission from broker file to model-ready schedule, baselined against your own range rather than an industry average. Percentage of locations at rooftop resolution versus ZIP centroid, before and after. Share of fields extracted, inferred and defaulted, reported per schedule rather than in aggregate. Share of TIV sitting in the top 2, 6 and 16 locations, because underwriter attention is spent evenly across a schedule where value is not. Count of schedules where reported values have not been refreshed against replacement cost, with the coinsurance gap quantified at partial loss. These are definitions and baselines, set in the first week.
// We ran one of these
The part of running a carrier that transfers here is not property underwriting — it is having owned the pipeline a rating decision came out of, where a blank field became a conservative default, the default became a price, and somebody had to defend the price to a regulator.
The operating record →// What's hard about this
Two limits, and what we do about each.
Imagery only sees what is visible from above.
Sprinkler type, alarm grade, central-station monitoring, interior occupancy, equipment bracing and business-interruption dependencies are invisible to aerial data. The highest-value underwriting judgment on a property risk is precisely what computer vision cannot reach. So the build fills what imagery can genuinely fill and stops at the boundary, marking the unreachable fields as unreachable instead of inferring them — and routes those fields to the survey, the appraisal or the loss-control visit that can actually answer them.
A cleaned schedule looks authoritative whether or not the values were ever right.
This is the specific failure mode of AI in exposure work: normalizing the schedule raises confidence without raising accuracy. Insurance-to-value drift is a human and appraisal problem, and no model can tell you an insured’s reported value is stale — while the margin clause or the Occurrence Limit of Liability Endorsement will cap recovery at a multiple of that stale number. At 90% coinsurance, $2.3M carried against $2.88M required pays roughly $400K on a $500K loss. The refutation is structural: the confidence signal ships alongside the data, every value states whether it was extracted, inferred or defaulted, and the drift flag names the schedules where the number was never re-based.
// What ships with it
The governance file, scoped to exposure data.
A model inventory entry for every filled-field model, data lineage from imagery vendor to schedule field, pre-deployment testing results, drift thresholds with remediation triggers, and a named human decision-maker specification. For property the load-bearing artifact is the provenance record on inferred values, because that is what an examiner asks about when a modeled loss estimate is challenged. Colorado’s outcomes testing is enforceable and Connecticut requires an annual AI compliance certification attested by a named officer.
NAIC and state bulletins · current at September 2026.
Bring us one schedule.
We will tell you what your schedules are missing and how much modeled loss each gap moves.