Key Sections
Introduction
In a groundbreaking collaboration, MyCover.AI joined forces with Omdena to transform the insurance landscape through the power of artificial intelligence. Together, we embarked on an AI Innovation Challenge that revolutionized the way vehicle insurance claims are handled, achieving success in two key areas:
- Automated Vehicle Image Validation: Utilizing Exploratory Data Analysis (EDA) and advanced AI techniques, our project streamlined the process of validating the authenticity of vehicle images submitted in claims. This step was crucial in ensuring that the data fed into our systems was accurate and reliable, setting the stage for precise damage assessment.
- Enhanced Car Damage Assessment System: Our two-stage model not only improved the accuracy of identifying and assessing car damage but also significantly bolstered our fraud detection capabilities. By automating these processes, we were able to enhance the efficiency of claims processing, reducing both the time and cost associated with manual assessments.
The success of this initiative led to several transformative outcomes for MyCover.AI:
- Talent Acquisition: The project’s success facilitated the hiring of a highly skilled engineer, further strengthening the partner’s team and enhancing their technical capabilities after the project completion.
- Secured Funding: The demonstrable impact and innovation of our project attracted significant investment, securing a major funding round that will support continued innovation and expansion in AI within the insurance sector.
Discover how we did it!
The Problem
Inefficiency and Fraud in Assessing Damage Insurance Claims
The problem of inefficiency and fraud in assessing insurance claims poses significant challenges for both insurance companies and their customers, creating a cascade of negative consequences across the industry.
Currently used Traditional Systems cause several negative consequences both for Companies but also for Clients:
- Financial Losses: Insurance fraud is a major financial drain, with companies incurring losses amounting to billions annually. These losses stem from paying out fraudulent claims, which inevitably leads to higher premiums for all customers to offset these costs.
- Operational Delays: Inefficiencies in the claims process can lead to slower claim handling and resolution times. This not only ties up resources but also impacts a company’s ability to serve other customers promptly, affecting overall productivity and operational efficiency.
- Worker’s Time Misallocation: A significant amount of resources is spent on manual checks and assessments to counteract fraud and manage inefficient processes. This diverts resources from other potentially more impactful areas such as customer service, innovation, or strategic planning.
- Damage to Reputation: Companies that consistently suffer from inefficiencies and fraud may gain a negative reputation, which can deter potential customers and affect business growth. Trust is crucial in the insurance industry, and once it’s compromised, it can be difficult to rebuild.
- Decreased Satisfaction and Trust: Delays, increased costs, and the complexities involved in the claims process can lead to decreased customer satisfaction. This dissatisfaction can cause customers to switch providers or dissuade potential customers from signing up.
The Background
Traditional Processes of the Industry
The vehicle insurance evaluation industry has traditionally been heavily reliant on manual processes, which are both costly and prone to human error. This reliance stems primarily from the need for physical inspections by adjusters who assess damage to vehicles firsthand. These adjusters often travel to the location of a vehicle or require it to be brought to an assessment center, which adds significant time and expense to the claims process.
Furthermore, the task of estimating repair costs is typically performed manually, involving the input of detailed damage assessments into estimation software or paper forms. This manual entry can lead to errors in data transcription and judgment, which may result in inaccuracies in repair cost estimations and insurance payouts.
The Goal
Our goal was simple to create an innovative automatic system that would put our Partner MyCover.AI ahead of the competition reducing time, costs and inefficiencies thanks to AI automations in Car Demage Insurance Claims, and more comprehensive analysis at the same time.
The final solution needed to be able to conduct among others these tasks:
- Validating vehicle images
- Identifying damage
- Classifying its severity
- Determining repair costs
How many tasks can be enhanced or replaced through AI?
How much time we can save?
Let’s find out!
The Challenges
Challenge 1.
Fraud Detection Complexity
While AI significantly enhances fraud detection capabilities, training models to identify fraudulent activities accurately is inherently complex. Fraudsters continually evolve their tactics, thus AI systems must be dynamic and adaptable to new patterns of fraud. This requires ongoing learning and adjustment.
Challenge 2.
Regulatory Compliance and User Trust
Implementing AI in the insurance claims process presents the dual challenge of adhering to stringent regulatory standards while also building trust among users who simply want to.
Effective strategies for facing that challenge include enhancing transparency through detailed explanations of AI processes, regularly updating systems to keep pace with regulatory changes, and conducting thorough audits to ensure compliance.
Additionally, engaging with customers and staff through educational initiatives helps demystify AI technologies, alleviating concerns and fostering a supportive environment for its acceptance and integration into everyday insurance operations. A big part of Omdena’s mission is providing equality of awareness in AI therefore we are supporting our partners with multiple solutions for increasing awareness and upskilling in the sphere of AI.
The Partner
Mycover.ai is the insurtech startup building Africa’s digital insurance infrastructure, recently has announced the close of a $1.25 million pre-seed funding, led by Ventures Platform.
Founded in 2021, MyCover.ai is dedicated to tackling the challenges in the African insurance market, including limited access, inadequate coverage, high costs, and poor customer experiences. The company uses technology to transform crucial aspects of the insurance value chain, including underwriting, product development, distribution, and claims processing.
Since its launch, MyCover.ai has quickly established itself as the clear market leader for innovative insurance solutions, with over $1M in gross written premiums, strong partnerships with leading insurance providers, and the launch of multiple innovative insurance products to the market.
Our Approach
To this problem
Stage 1.
Deciding on the Methodologies and Technologies
Aspect 1:Building an Automatic Fraud Detection System
Feature: License Plate Detection
Utilized YOLOv8 model for real-time, accurate detection of vehicle license plates.
Feature: Text Reading and Analyzing on License Plates
- Reading and Interpreting Text: OCR carefully reads and interprets the text on the license plates.
- Verification of Vehicle Records: This step is crucial for ensuring that the text on the license plate matches the vehicle’s official registration records.
Feature: Image Comparison
- We used two special tools—VisionTransformer and Cosine Similarity. Think of VisionTransformer as a smart system that can look at pictures and understand different details about cars, just like a human would. Cosine Similarity is a way to measure how similar two pictures are.
- Played a pivotal role in the fraud detection system by spotting discrepancies and similarities that indicated possible fraud.
Aspect 2: Car Damage Assessment
The second aspect of our tool needed to be able to perform the type of car damage assessment.
Feature: Defect Detection Model
The first model in this stage was designed to meticulously identify defects on the car’s exterior, which is essential for accurate damage assessment.
Feature: Damage Severity Evaluation
The second model evaluated the severity of the identified damages. This evaluation was crucial in calculating the total estimated repair cost, a key component in processing insurance claims.
Stage 2.
Putting the methodologies into practice
Step 1: Starting with a Strong Foundation – Exploratory Data Analysis (EDA)
- Challenge: High-quality data is crucial for accurate AI training but challenging to ensure.
- Process: We began with Exploratory Data Analysis (EDA), a thorough examination of our image datasets. This step involved checking each image to make sure it was clear and relevant, helping us select the best examples of vehicle damage.
- Technical Term: Data Verification and Validation
- Impact for Businesses: This careful selection means our AI models are trained on realistic, applicable data, leading to more accurate damage assessments and fewer incorrect claims.
Step 2: Harmonizing the Data for AI Accuracy – Image Standardization
- Challenge: Differences in image size and shape can affect AI accuracy.
- Process: We standardized these variations by analyzing and adjusting the images to common dimensions. This means making sure all images are reshaped to match a standard size and shape, which helps our AI process them more effectively.
- Technical Term: Image Standardization
- Impact for Businesses: Standardized image input leads to consistent and reliable AI assessments, crucial for businesses looking to scale operations smoothly.
Step 3: Refining Details for Precision – Semantic Segmentation (Annotation)
- Challenge: Precise image labeling is crucial but can be error-prone.
- Process: We meticulously checked each image label to confirm it accurately described the damage shown. This detailed labeling is crucial for training our AI to recognize and understand different types of vehicle damage accurately.
- Technical Term: Semantic Segmentation
- Impact for Businesses: This rigorous process reduces mistakes in AI predictions, ensuring that assessments are dependable and useful in real-world settings.
*Understanding Image Labeling and Annotation
When we talk about transforming vehicle insurance with AI at MyCover.AI, two terms that often come up are “image labeling” and “annotation.” Let’s break down what these mean and why they’re crucial for making our AI system effective.
What is Image Labeling?
Image labeling is the process of identifying what is in an image and marking it with a label. For example, in the context of vehicle insurance, if we have a picture of a car with a dented bumper, labeling the image would involve indicating that there is a dent and noting its location on the bumper.
What is Annotation?
Annotation goes a step further by adding more detailed information to the labels. Using the same example of the dented bumper, annotating the image would involve not just labeling the dent but also describing its size, shape, and perhaps even estimating the severity of the damage. This could include marking the exact edges of the dent on the image so the AI can learn exactly what a dent looks like and where it is located.
By carefully labeling and annotating images, we ensure that our AI models are not only learning continuously but are also refining their ability to assess and process claims with high accuracy. This meticulous preparation ultimately leads to faster, more reliable claim handling and enhances the overall customer experience in vehicle insurance.
Step 4: Ensuring Fairness – Class Rebalancing and Data Augmentation
- Challenge: AI models might favor more common damage types, skewing results.
- Process: To prevent this, we enhanced the variety in our training data using techniques like image flipping and cropping. We also used a method called mean average precision (mAP) to make sure our AI evaluated all types of damages fairly, no matter how rare.
- Technical Term: Class Rebalancing and Data Augmentation
- Impact for Businesses: This approach ensures that all types of damages are recognized and assessed accurately, making the AI system fair and effective across the board.
Step 5: Optimizing Detection – Bounding Box Optimization
- Challenge: Detecting damage accurately depends on how well the AI can identify and outline the damaged area.
- Process: We refined the settings for these outlines, known as bounding boxes, to ensure they accurately fit different sizes and shapes of damage. This fine-tuning helps the AI recognize and evaluate damage more effectively.
- Technical Term: Bounding Box Optimization
- Impact for Businesses: Precise damage detection speeds up claim processing and increases the accuracy of damage assessments, enhancing efficiency and customer satisfaction.
Benefits and other Applications of these Methodologies
Applied Methodology 1:
Semantic Segmentation (Annotation)
The term “semantic” in semantic segmentation specifically refers to the context or meaning of what’s being analyzed within an image. Unlike basic image processing that might simply detect shapes or objects without understanding their function or relevance, semantic segmentation classifies each part of the image according to what it actually represents—making it incredibly useful for applications that require deep understanding and interaction with visual data. Here’s why this added semantic layer is crucial.
Can also be utilised in:
Healthcare
How It Works: In medical imaging, semantic segmentation can classify different tissues, organs, or anomalies in scans such as MRIs or CT images.
Benefit: Enhances the accuracy of diagnoses by clearly delineating affected areas from healthy ones, helping doctors to detect diseases early and plan treatments more effectively.
Automotive Industry
How It Works: Applied in driver assistance systems, semantic segmentation can distinguish roads, pedestrians, vehicles, and other elements in real-time video from vehicle cameras.
Benefit: Increases driving safety by improving the performance of automated driving systems, helping prevent accidents by providing precise information about the vehicle’s surroundings.
Agriculture
How It Works: Uses drones equipped with cameras to capture images of crops and then applies semantic segmentation to identify areas of disease, irrigation issues, or pest infestation.
Benefit: Optimizes crop management by pinpointing problem areas that need attention, thereby reducing waste, maximizing yield, and increasing overall farm efficiency.
Retail
How It Works: In retail environments, semantic segmentation helps in analyzing customer behavior by segmenting different areas of a store to see where customers spend the most time.
Benefit: Provides valuable insights into customer preferences and store layout effectiveness, enhancing store management and marketing strategies.
Manufacturing
How It Works: Used in quality control processes to inspect products. AI can examine images of products on the assembly line and identify defects by segmenting images into distinct regions and analyzing them for inconsistencies.
Benefit: Ensures product quality by identifying defects early in the production process, reducing waste and improving customer satisfaction.
Urban Planning
How It Works: Semantic segmentation can be used to analyze aerial imagery of cities and classify areas by land use, such as residential, commercial, or industrial zones.
Benefit: Aids urban planners in making informed decisions about city planning and development based on detailed analysis of how space is being used.
Applied Methodology 2:
Image Standardization
Image standardization is the process of adjusting and modifying images to ensure they all meet a specific set of criteria regarding size, resolution, color, and format. This consistency allows AI and other automated systems to analyze and process images more accurately and efficiently.
Can also be utilised in:
E-Commerce and Retail
How It Works: Standardizes product images across your online platform, ensuring that every item appears uniform in size and lighting.
Benefit: Enhances customer experience by providing a visually cohesive catalog, making it easier for customers to compare products and make purchasing decisions. It also improves the performance of visual search tools, helping customers find what they want quicker.
Real Estate
How It Works: Applies consistent quality and dimensions to property photos uploaded to listing websites.
Benefit: Creates a more attractive and professional-looking website. Standardized images help potential buyers better assess and compare properties, potentially speeding up sales cycles and increasing client satisfaction.
Healthcare
How It Works: Standardizes diagnostic images such as X-rays, MRIs, and ultrasounds for analysis.
Benefit: Improves diagnostic accuracy and reliability. Consistent image quality ensures that medical professionals are making decisions based on the pathology present, rather than variations in image quality, enhancing patient care.
Manufacturing
How It Works: Ensures that images used for quality control, such as in automated inspection systems, are uniform in lighting and scale.
Benefit: Increases the detection accuracy of defects and irregularities in products. Consistent images reduce errors and false rejections in automated systems, improving product quality and reducing waste.
Security and Surveillance
How It Works: Applies consistent formatting to images captured by surveillance cameras, particularly in facial recognition systems.
Benefit: Enhances the accuracy of facial recognition and other security protocols, reducing the risk of false identifications and improving overall security measures.
Time Frame
- The whole project from start to finish took less than 3 months!
The Outcome
- Faster Claim Processing: Our AI speeds up the assessment process, leading to quicker claim resolutions.
- Improved Customer Satisfaction: Accurate and efficient claims handling boosts trust and retention among customers.
- Reduced Fraud: Enhanced detection capabilities decrease the chances of fraudulent claims.
Further Possibilities
Integration of Blockchain Technology:
Logic: Blockchain acts as a secure, decentralized ledger that records data across a network of computers, making the data nearly impossible to alter without consensus.
Value: Implementing blockchain can drastically reduce fraud by ensuring all data related to repairs, damages, and claims are permanent and transparent. This boosts trust and integrity in insurance processes, making each transaction verifiable by all parties involved.
Artificial Intelligence in Predictive Analytics:
Logic: AI can analyze vast amounts of data to identify patterns and anomalies that may indicate potential fraud or predict future claims based on historical data.
Value: This proactive approach helps insurance companies prevent fraud before it occurs and better assess risk, leading to more accurate premium settings and healthier insurance portfolios.
Real-Time Damage Assessment Tools:
Logic: Mobile apps that allow customers to immediately upload images of damage, which are then analyzed by AI to assess severity and estimate repair costs on the spot.
Value: Such tools enhance customer engagement and satisfaction by providing quick feedback and initiating the claims process faster, reducing wait times and improving service delivery.
Collaborative Data Sharing Platforms:
Logic: By securely sharing data between different insurance providers, the amount of data available for analysis increases exponentially.
Value: This collaboration enhances the accuracy of AI models used in fraud detection and damage assessment, and promotes a broader understanding of vehicle damage trends and repair needs, fostering industry-wide improvements and innovation.
