Photogrammetry is the science of extracting dimensional and geometric information from photographs. In its traditional form used in surveying, architecture, and engineering it involves multiple calibrated cameras and complex 3D reconstruction algorithms. In insurance claims assessment, the approach is adapted for the practical constraints of a customer submitting photographs from a mobile phone: the images are 2D, the angles are inconsistent, the lighting is unpredictable, and the customer has no training in what to capture.

The AI photogrammetry engine deployed in advanced motor claims systems does not attempt to build a precise 3D model from mobile phone images. Instead, it operates through a classification and estimation framework trained on hundreds of thousands of historical claims images, each labelled with the actual assessment outcome produced by experienced surveyors. What the model learns, at scale, is the mapping between image features and assessment outcomes not the geometric reconstruction of the damage, but the damage categorisation that a trained surveyor would apply.

Stage 1 — Image Capture Protocol

The process begins with a structured image capture request sent to the customer via the insurer's app or WhatsApp channel. The customer is asked to photograph the vehicle from a defined set of angles typically a minimum of four exterior perspectives plus targeted images of the specific damage area. The key architectural decision here, documented in Bajaj Allianz's deployment, is the requirement for 360-degree coverage rather than isolated damage images. This serves two simultaneous functions: it provides the model with sufficient contextual information for an accurate assessment, and it makes fraudulent staging significantly harder, since a 360-degree set exposes the full vehicle condition and surroundings in a way that a targeted single-image submission does not.

Stage 2 — Image Validation and Quality Scoring

Before the damage assessment begins, the submitted images pass through a validation layer. This layer checks image resolution, blur levels, lighting adequacy, and coverage completeness against the required angle protocol. Images that fail quality thresholds are rejected and the customer is prompted to retake specific shots. This step is critical: the assessment accuracy of the downstream model is directly dependent on input image quality, and allowing low-quality images through degrades output reliability in ways that are difficult to detect post-hoc.

The validation layer also runs initial metadata analysis examining EXIF data for timestamp, geolocation, and device information. Anomalies at this stage images with edited metadata, timestamps inconsistent with the reported incident time, geolocation that does not match the reported accident locatio are flagged as fraud risk signals before the damage assessment begins.

Stage 3 — Damage Classification and Estimation

The validated image set enters the damage assessment model. This model operates through a hierarchical classification framework:

Damage extent classification: Is the damage surface-level (paint, trim, minor dents) or structural (panel deformation, frame damage, mechanical component exposure)? This classification drives the first branch point in the settlement workflow surface-level damage routes to rapid settlement; structural damage routes to an enhanced review queue.

Component identification: Which specific vehicle components are affected? The model identifies damaged components against a taxonomy of standard vehicle parts bumpers, bonnets, doors, fenders, mirrors, lights, windscreens and maps each identified component to a repair cost range calibrated to the specific vehicle make, model, year, and variant.

Repair vs. replacement assessment: For each affected component, the model estimates whether repair or replacement is the appropriate resolution, based on damage severity indicators trained against historical surveyor decisions for the same component type.

Estimate generation: The combined component assessment produces a settlement estimate range, expressed as a lower bound (repair scenario) and upper bound (replacement scenario), which is presented to the customer for consent before any payment commitment is made.

Stage 4 — AgentVerse Workflow Orchestration

The photogrammetry output does not exist in isolation it feeds directly into SubVerseAI's AgentVerse workflow orchestration layer. Based on the damage classification and fraud risk score, AgentVerse routes the claim to the appropriate downstream workflow: instant settlement for low-value, low-risk, clearly surface-level claims; enhanced review queue for structural damage or ambiguous image sets; fraud investigation escalation for high-risk submissions with anomalous metadata or damage inconsistencies.

Throughout this routing process, DataVerse provides the customer intelligence context prior claim history, policy coverage details, vehicle service records where available, and overall fraud risk profile that informs the confidence thresholds applied at each decision point. A first-time claimant with a clean policy history and consistent image metadata has a different risk profile than a customer with three prior claims in 18 months submitting images with edited timestamps.