Computer Vision in Commercial Insurance Underwriting in India

Commercial property underwriting in India has long relied on a model that is expensive, slow, and geographically uneven: physical risk engineering surveys conducted by empanelled surveyors who visit a site, inspect construction, assess fire protection, and complete structured risk reports. For major industrial accounts in Mumbai, Pune, or Chennai, the model functions reasonably well. A qualified fire engineer with manufacturing sector expertise can be dispatched within days. For a mid-sized factory in a tier-2 or tier-3 city, a sawmill in Assam, or a cold storage facility in rural Rajasthan, the model breaks down. Qualified surveyors are scarce outside major urban centres, travel time is significant, and the cost of a full risk engineering visit for an account generating INR 5 to 15 lakh in annual premium is often difficult to justify.

The result is a coverage and pricing problem. Insurers writing commercial property risks in remote or semi-urban locations have historically either charged a loading to compensate for survey uncertainty, required the insured to self-declare construction and occupancy details that are never independently verified, or simply declined risk in geographies they could not survey. The sum insured inadequacy problem in Indian commercial property, where independent studies suggest that 35 to 45% of industrial policies are underinsured relative to current reinstatement costs, is partly a consequence of surveys that are either superficial or never conducted.

Computer vision applied to satellite imagery and drone-captured data changes the economics of this problem. A satellite image analysed by a trained CV model can produce a roof construction classification, a building footprint measurement, a preliminary fire load assessment based on visible storage patterns, and a comparison against historical imagery to identify changes in the risk profile, all at a cost measured in tens of rupees rather than the INR 20,000 to 80,000 a physical risk engineering visit costs. The question for Indian commercial insurers is not whether this technology works in principle but how to integrate it into underwriting and survey processes that are regulated by IRDAI.

Satellite imagery-based risk assessment for commercial property uses a combination of high-resolution optical imagery, synthetic aperture radar, and multispectral data to extract risk-relevant features from a target site and its surroundings. Commercial satellite operators including Planet Labs, Maxar Technologies, and Indian providers such as Pixxel and SatSure can now provide sub-metre resolution imagery of most Indian industrial locations, updated at frequencies ranging from daily (for Planet's SkySat constellation) to weekly or monthly for standard coverage.

The CV models applied to this imagery in insurance underwriting contexts are trained to identify and classify a range of features. Building construction type classification (reinforced concrete frame, steel portal frame, load-bearing masonry, pucca or kuchha, mixed construction) is among the more mature capabilities, and is generally most reliable for large, clearly bounded industrial buildings. Accuracy declines for smaller or partially-obscured structures and for mixed-construction facilities common in Indian SME manufacturing. There is as yet no published, industry-wide accuracy benchmark for these models in the Indian context, so any specific accuracy figure quoted by a vendor should be treated as a pilot-level claim and validated against the insurer's own ground-truth data for its portfolio.

Fire hazard indicators visible from satellite include: roof-mounted equipment that may indicate industrial processes with fire risk (paint spray booths visible as isolated structures with ventilation equipment, chemical storage yards with distinctive container patterns, timber yards identified by the characteristic stacking patterns of rough-sawn lumber). Proximity risk factors are assessed by measuring distances to neighbouring occupancies, identifying shared wall constructions, and mapping water bodies, fire station locations, and road access that affects firefighting response time.

Probable maximum loss estimation from aerial data

PML estimation, the core output of property underwriting, has traditionally required detailed site surveys to assess construction quality, fire compartmentalisation, sprinkler system coverage, and salvage probability. CV-based PML models integrate satellite-derived construction classification with building footprint measurements, occupancy classification (often derived from a combination of visual features and business registration data), and exposure data from the insurer's own portfolio to produce preliminary PML estimates.

In practice, CV-based PML estimates are less accurate than a well-conducted physical survey but materially more accurate than estimates based solely on policyholder declarations, which is why they are useful for early risk screening and mid-market pricing rather than as a substitute for engineering judgment on large or complex accounts. The size of the gap depends heavily on construction type, image quality, and the maturity of the model, and there is no published industry benchmark to anchor a single figure, so insurers should validate accuracy on their own book before relying on CV-derived PML for pricing. As a matter of treaty and facultative practice, reinsurers may accept CV-assisted assessments as supporting documentation when a licensed surveyor or risk engineer has professionally validated the data and the risk falls within defined parameters.

Drone surveys occupy the middle ground between satellite imagery (remote, lower resolution, no access to vertical faces or interiors) and physical site visits (high information quality but expensive and geographically constrained). A drone equipped with an RGB camera, a thermal imaging sensor, and a LiDAR unit can capture a commercial property in 2 to 4 hours of flight time, producing a georeferenced 3D point cloud, thermal maps identifying heat signatures from electrical equipment and hot processes, and high-resolution images of roof condition, drainage, facade integrity, and visible fire protection equipment.

The CV models processing drone survey data are more capable than satellite-based models because the imagery is higher resolution and captured at angles that expose features invisible from orbit. Roof condition assessment is an area of relative maturity: CV models can classify roof condition across categories such as excellent, good, fair, poor, and critical, and Indian insurer pilots have reported usable accuracy on standard industrial shed constructions, though no industry-validated benchmark has been published and results vary by roof type and image quality. Skylark Drones and Asteria Aerospace are among the Indian providers supplying drone survey and geospatial data-processing services that the insurance sector can use for risk engineering data capture.

The question of whether a drone survey satisfies IRDAI's requirements for surveys of commercial risks above defined thresholds is not yet formally resolved. IRDAI (Surveyors and Loss Assessors) Regulations, 2015 and subsequent amendments define the scope of work for licensed surveyors and loss assessors. The survey requirement for large commercial risks (generally above INR 5 crore sum insured) specifies that a licensed surveyor must assess the risk. Whether a licensed surveyor using a drone and CV analysis tools satisfies this requirement is different from the question of whether a drone alone, without a licensed surveyor's involvement, satisfies it.

The practical resolution adopted by most Indian insurers integrating drone surveys has been a hybrid model: the drone survey is conducted by a specialist operator and produces a structured data package; a licensed risk engineer reviews the data package, adds professional judgment on the elements that require it, and signs the survey report in their own name. This model satisfies IRDAI's surveyor requirement while capturing most of the cost and time savings from drone-based data collection. Survey cycle time drops from 10 to 15 days on the traditional model to 3 to 5 days on the hybrid model, and costs fall by 40 to 60% for accessible sites.

The CV models deployed for fire hazard detection and building condition assessment in Indian commercial underwriting are not general-purpose image classifiers. They are purpose-built models trained on labelled datasets of Indian industrial and commercial properties, with labels generated by qualified risk engineers who have tagged images with fire hazard indicators, construction deficiencies, and occupancy characteristics.

Fire hazard detection from aerial and drone imagery works through a combination of object detection (identifying specific high-hazard equipment or material types), scene classification (categorising the overall risk profile of a facility), and anomaly detection (flagging departures from what a similar facility would typically show). Specific detectable hazards include: liquid petroleum gas cylinder banks identifiable by characteristic shape and placement patterns; external chemical storage tanks with insufficient bunding visible from above; electrical transformer rooms lacking fire separation from adjacent occupancies; and roof penetrations that suggest ad-hoc electrical installations creating fire ignition risk.

Building condition assessment focuses on indicators of deterioration that affect both the fire resistance and the structural integrity of the insured property. From drone imagery, CV models assess: roof sheet condition including visible corrosion, missing panels, and improper repairs; skylight condition including cracked or missing glazing; downpipe and drainage condition; structural column condition where columns are externally visible; and the condition of fire doors, roller shutters, and external escape routes. Each of these assessments maps to IRDAI-aligned risk rating factors used in the insurer's tariff or rating model.

The integration of CV model outputs into the underwriting workflow requires that model outputs be expressed in terms that the underwriter's rating system can consume. A CV model that outputs 'roof condition: poor, probability 0.87' needs to be translated into a loading factor on the sum insured or a rate adjustment that the underwriter's system can apply. Indian insurtech vendors such as Artivatic.ai have developed middleware that performs this kind of translation for property underwriting systems, mapping model outputs to the standard rating factors used by de-tariffed commercial property policies. Brokers and insurers should confirm any specific vendor's current capabilities directly rather than rely on market descriptions.

The central regulatory question for computer vision in Indian commercial property underwriting is whether and how CV-based assessments interact with IRDAI's survey requirements. The answer differs depending on whether the context is pre-inception risk assessment (underwriting survey) or post-loss assessment (claims survey).

For pre-inception underwriting surveys, IRDAI (Insurance Products) Regulations, 2024 and the associated product guidelines for fire and engineering insurance do not specifically mandate physical surveys for all policies. The survey requirement is primarily a risk management and rating tool. For accounts above defined thresholds, most Indian insurers require a risk engineer's report as a condition of coverage, but this is an internal underwriting requirement rather than a mandatory regulatory requirement in all cases. The use of CV-based assessment to support or replace physical surveys in the underwriting context is largely within the insurer's discretion, subject to its own underwriting guidelines and reinsurer requirements.

For post-loss surveys (claims assessment), the position is different and more restrictive. IRDAI (Surveyors and Loss Assessors) Regulations, 2015 require that claims above INR 1 lakh be assessed by a licensed surveyor. The surveyor's report is the primary evidence basis for claim settlement. CV and drone-based tools can assist the licensed surveyor in collecting and processing evidence, but the survey report must be signed by the licensed surveyor who takes professional responsibility for its contents. IRDAI has not indicated any intention to allow AI-based assessments to replace licensed surveyors in the claims context, and industry expectation is that this position will persist through the current regulatory cycle.

The interaction with reinsurance requirements is a separate constraint. International reinsurers providing capacity on large Indian commercial property risks often have their own survey requirements embedded in treaty or facultative conditions. GIC Re, as mandatory reinsurer of first cession under IRDAI (Reinsurance) Regulations, 2018, has accepted CV-assisted survey reports when produced by licensed surveyors who have professionally validated the data. Several international reinsurers have similarly accepted hybrid drone-plus-engineer survey reports for accounts within defined sum insured bands. Above INR 200 crore sum insured, most international reinsurers continue to require traditional risk engineering surveys by accredited engineering firms.

The economic case for CV-based risk assessment in Indian commercial insurance is clearest for the mid-market: accounts with sum insured between INR 2 crore and INR 50 crore, where the traditional surveyor visit cost is material relative to premium income and where the insurer's risk selection is currently impaired by survey coverage gaps.

Consider an illustrative mid-sized general insurer writing a few hundred crore of fire and property premium with a meaningful commercial segment. On the traditional model, the annual survey cost (surveyor fees, travel, administrative coordination, and underwriter time spent chasing delayed reports) is a material line item. A hybrid model that replaces a large share of physical surveys with drone-plus-CV assessments on accessible mid-market accounts, while retaining physical surveys for large industrial accounts and specialist risks, can cut that cost substantially. These figures are illustrative; the actual saving depends on portfolio mix, geographic spread, and how much of the book is genuinely accessible to remote assessment.

The saving is real but so is the implementation investment. Building or buying the CV model infrastructure, training staff to interpret model outputs, integrating those outputs into the rating system, and establishing the operational model for drone survey coordination all require upfront spend in year one, plus ongoing maintenance and model-improvement costs. Whether the payback lands in one, two, or more years depends on the volume of surveys displaced and the insurer's starting cost base, so each insurer should build its own business case rather than rely on a generic payback claim.

The risk selection benefit may ultimately exceed the cost saving. Insurers that can survey more risks more quickly and more accurately have better information for pricing, better ability to identify risks that have deteriorated since inception, and better portfolio monitoring capability. An insurer that re-screens its entire commercial property book with CV-based assessments every 12 months (a frequency that is economically impossible with physical surveys) can identify sum insured gaps, occupancy changes, and condition deterioration before they manifest as claim frequency or severity shifts. This information advantage compounds over time and is difficult to replicate without the technology.

The most visible operational benefit reported by insurers adopting a CV-first approach is faster policy issuance, because removing the physical-survey scheduling bottleneck collapses the dead time between risk submission and a usable risk view. For mid-market accounts where a full engineering survey is not strictly required, this can take turnaround from a couple of weeks down to a few days. Brokers placing such accounts should ask insurers directly how CV-based assessment changes their quoting and issuance timelines, rather than rely on generic market claims.

Indian insurers deploying computer vision in commercial underwriting have followed different paths depending on their starting position and distribution model, and three broad patterns are visible across the market. None of the specific configurations below should be read as a vendor-style performance claim; brokers should confirm any individual insurer's current approach and results directly.

The first pattern is satellite imagery used as a pre-survey signal. Some larger general insurers run satellite-based analysis to generate a preliminary risk score for new commercial property accounts above a sum-insured threshold, providing the underwriter an early read on risk quality and flagging accounts that warrant an accelerated or more thorough site inspection. Here the model supplements rather than replaces the physical survey on large accounts, reducing the underwriter's information vacuum between risk submission and survey completion.

The second pattern, favoured by insurers with a higher proportion of SME and mid-market commercial business, is drone surveys with CV analysis as the primary survey mechanism for accessible accounts within a defined sum-insured band, with a licensed risk engineer reviewing and countersigning every report. Insurers on this model report meaningful reductions in survey cycle time and cost, with the exact magnitude depending on how much of the relevant book is genuinely accessible to drone capture.

The third pattern is CV-based condition monitoring used as an at-renewal triage tool. Instead of a fresh physical survey at every renewal, the platform flags accounts where imagery analysis detects changes in building condition, occupancy, or neighbouring risk beyond defined thresholds; only flagged accounts receive a physical inspection request, while clean accounts renew on the CV assessment plus a policyholder declaration. Done well, this materially shrinks the share of the portfolio needing physical renewal surveys.

The accuracy benchmark question has not been resolved at an industry level. Individual insurer pilots report varying accuracy figures depending on building type, geography, and the specific CV model architecture. An industry-level benchmark dataset for CV-based risk assessment accuracy, which would allow meaningful comparison across providers and models, would be a significant step forward, but no such standard has been published, so each insurer must validate accuracy against its own ground-truth data.