What Did NxGN Deliver?
NxGN delivered the AI claims photo review solution in two phases over approximately three months. Specifically, the team deployed it as a REST API on PMD’s own infrastructure, integrating directly with their Vertigo line-of-business system.
Phase 1: Image Classification and Quality Checking
NxGN built a set of deep learning models using Keras that classified each incoming image into one of eight angle categories. In addition, the models assessed image acceptability by detecting missing metadata, cropped vehicles, blur, and overexposure. The final architecture used an ensemble approach. Specifically, a multiclass angle prediction model fed into binary model checks, processed through an angle arbitration algorithm, followed by angle-specific validity models.
Phase 2: Data Extraction and Policy Matching
This phase addressed the most resource-intensive manual step. NxGN enhanced the API to extract structured data directly from images using computer vision, OCR, and barcode decoding. For example, vehicle photos yielded licence plate numbers. Similarly, licence disc photos provided VINs, engine numbers, and vehicle details. The system also extracted South African ID numbers from driver’s licence photos.
When Vertigo sent a batch of images through the API, any identifying data found in one photo could consequently match the entire batch. As a result, only when no image in the batch returned usable data would a human operator need to step in.
Technical detail: Model architecture and performance
The classification models trained on PMD’s existing photo library. This required extensive data preparation: downloading, cleaning, transforming, and organising images into properly labelled training, validation, and test sets. Furthermore, each model training epoch took approximately 20 minutes. A full run of 40 epochs therefore required around 13 hours before the team could assess results.
The ensemble classification architecture achieved individual angle accuracy ranging from 92.8% (engine) to 98.8% (licence disc) when tested against 40,541 real production photos. Of these, the models correctly classified 14,431 as invalid and routed them for review. Moreover, only 5.7% of images were incorrectly classified as valid when they should not have been.
For data extraction, the API handled multiple plates in frame (for example, another vehicle in the background) by prioritising the primary plate. Additionally, licence disc extraction served dual purposes: policy matching and cross-checking against existing policy data for capture errors or fraud indicators.
The API processed each image in 25 milliseconds to 2 seconds on PMD’s infrastructure. It consequently handled JPG, PNG, BMP, and TIFF formats.
What Were the Results?
| Metric | Before NxGN | After NxGN | Improvement |
|---|---|---|---|
| Photo angle classification | Fully manual — call centre staff classified every image | 91% automated accuracy across 8 categories | 88% reduction in manual AI claims photo review workload within first month |
| Image quality checking | Manual visual inspection for blur, cropping, overexposure | Automated flagging with confidence scores at point of upload | Instant quality feedback; invalid images caught before reaching the queue |
| Policy matching (non-app submissions) | 100% manual for 85–95% of weekly volume | Automated via licence plate, VIN, and ID number extraction | Auto-matching eliminated the single most resource-intensive manual step |
| Processing speed per image | Minutes (human review) | 25 milliseconds to 2 seconds (API) | Orders-of-magnitude speed improvement |
| Evaluation consistency | Variable — reviewer fatigue caused inconsistent quality judgements | Identical standards applied to every image, every time | Eliminated reviewer-to-reviewer variation in underwriting quality |
| Data validation | No cross-checking against policy records | VIN, engine number, and vehicle make extracted and compared to policy data | Capture errors and potential fraud indicators surfaced automatically |
Key Performance Metrics
The photos team’s role shifted fundamentally within the first month of production. Instead of spending hours on repetitive classification and matching, call centre staff refocused on genuine exceptions. For instance, these included images flagged as uncertain by the models, batches that could not be automatically matched, and quality issues requiring customer contact. In short, the API handled the routine volume, while humans handled the edge cases.
Compounding Gains
The phased approach created a compounding effect. Because Phase 1’s angle classification was already in place, Phase 2’s data extraction became more accurate. Since the API already knew which type of image it was processing, it could extract licence plates, VINs, or ID numbers more reliably. Consequently, the combination of classification plus extraction enabled the policy auto-matching workflow. This therefore delivered the largest operational efficiency gain.
Decoupled Growth
Critically, the system decoupled operational cost from growth. As a result, PMD could absorb increases in application volume without proportionally expanding the photos team. Moreover, with photos processed in seconds rather than queued for manual review, the bottleneck that had delayed policy issuance disappeared. This consequently improved customer turnaround times.
What Happened Next
Following the AI claims photo review deployment, NxGN embedded analytics resources within PMD’s team. The goal was to upskill their internal data science capability, transitioning from a delivery engagement to a knowledge transfer model. As a result, this ensured PMD could extend and maintain the AI models independently. However, the engagement wound down in early 2020 when the COVID-19 pandemic disrupted operations across the insurance sector.
Frequently Asked Questions
How accurate is AI-based vehicle photo classification?
NxGN’s deep learning models achieved 91% overall accuracy across eight photo angle categories. Specifically, the team tested them against 40,541 real production images. Individual category accuracy ranged from 92.8% to 98.8%. Furthermore, only 5.7% of images received an incorrect “valid” classification.
Can NxGN’s claims photo API integrate with existing insurance platforms?
Yes. The API deployed on PMD’s own infrastructure and integrates directly with their Vertigo line-of-business system via REST endpoints. Moreover, the architecture is platform-agnostic and can consequently connect to any system that supports standard API calls.
How long does it take to deploy an AI photo classification solution?
NxGN delivered PMD’s full two-phase solution in under three months. Specifically, this included image classification, quality checking, and automated policy matching. The primary time investment goes into data preparation and model training using the insurer’s actual production photo library. As a result, this ensures the models are tuned to real-world conditions rather than generic image datasets.