Damage assessment is a critical component in the automobile insurance claim process. Currently, it’s manually executed by third parties and entails considerable cost and resources. The inherent difficulty in capturing images in an uncontrolled environment is the biggest challenge in automating the process, 

In his talk at MLDS 2021, Chirag Jain, AVP, Data Science and Insights – Augmented Intelligence Practice at Genpact, elaborated on using AI state-of-the-art deep learning architectures and data enhancement approaches to generate millions of images to train the system. He dilated on the lessons learnt and how their solution was able to segment out body parts and damaged areas at the pixel level to generate robust insight for decision-making in multiple scenarios including total loss, repair-replace, cost estimates etc.

Manual Intervention Vs AI

Manual auto claim processes take anywhere between 5 to 10 days to review the claim. The process involves steps such as the customer care agent registering and sending a claim request to the adjuster, assigning the case to the mobility inspection team, analysing images captured through self serve or field representatives, reviewing the estimate with loss details, and more. AI can streamline the entire process as the computer vision classifies total loss, detects damage automatically and provides inputs for line-item estimates such as part prices, vehicle history and overall estimate cost — all of which could be done in 24 hours.

Genpact’s Claims AI gets First Notice of Loss (FNOL), followed by images of damaged vehicles being analysed by computer vision to identify attributes such as part analysis, part damage likelihood, the severity of the damage, total loss and cost estimates for the vehicle’s repair. 

Genpact’s Claims AI is trained on millions of images supported by auto-labelling and synthetic data generation and can detect six damage types such as scratch, dent, crush, missing parts, misalignment and tear. It can further handle the detection of 38 parts in SUVs, Sedans, trucks, hatchbacks, minivans and coups. 

Pixel-level Detection

While the techniques mentioned above cover the surface-level detection, Jain said they are exploring pixel-level part detection to extract detailed information on the damage. “Pixel level detection is critical for finer part boundaries,” said Jain.

Some of the techniques Genpact uses to understand the pixels include class activation, training data augmentation such as synthetic data generation, transfer learning, etc. 


While the state of the art technology has been able to help detect damage type, damaged parts, and make a decision on repair and replace, it is not without challenges. Jain laid out the obstacles they face while using AI in claims settlement including:

  • Zoomed in images which makes it difficult to distinguish between parts
  • High reflection of surroundings making it difficult for damage detection
  • Random angles of capture which might lead to ambiguity for part detection
  • Exposed engine parts or interiors coming in the way of precise damage detection
  • Low-quality images
  • Uncontrolled count of captured images

Jain said the decisions are not solely based on AI. There is a fair amount of human-intervention involved. “AI-powered line estimation is definitely our focus, but if in certain instances AI is found to be making incorrect decisions, humans take over,” he concluded. 

The post How Genpact Uses AI To Automate Vehicle Insurance Claims Process appeared first on Analytics India Magazine.