Leveraging Artificial Intelligence for Automatic Image Analysis
Mar 25, 2018
Christian van Leeuwen
Insurance companies consume terabytes of information every day in the form of digital data. Valuable data is being provided by more and more sources. All this data quickly paints a reliable picture of a risk or a submitted claim. By doing this, it is possible to filter out unwanted risks and fraudulent claims. State-of-the-art technology is required to process this influx of information and convert it into applicable knowledge and insights. This requires intelligent software capable of processing information quickly, learning independently, drawing smart conclusions, and making recommendations. Similar to a human, but smarter and more efficient. We're talking about artificial intelligence (AI).
3 types of Artificial Intelligence (AI)
We can identify roughly three types of artificial intelligence.
The first benefits everyone and is known as 'general' AI. Examples of this type of AI include natural language processing, facial recognition, and augmented reality. In the latter, the physical world is overlaid with elements from the virtual world. We use elements of general AI to support specific process in the other two forms of AI.
The second form is called product-based AI. It involves product with specific tasks. Just think about systems designed to identify and deactivate computer viruses, spam filters, and programs that detect fraud patterns. The latter is extremely valuable to us. If we feed AI with information about fraud cases, the better it becomes at this specific task. AI systems learn by doing and by drawing on feedback from many different users.
The third type of AI is custom or domain-specific AI. In short, AI that is trained for usage in a specific niche. An example from the insurance sector is AI that was trained to identify fraud patterns. Straight-Through Processing (STP) was integrated, aswell as the characteristics of a single insurance company: the specific products, target groups, distribution channels, and claims processes. This allows us to determine how AI can best be integrated into the business operations of an individual insurance company.
AI and Image Analysis
Combining all three types of AI often leads to the best results; for instance, using Artificial Intelligence to analyse images in an STP environment while simultaneously checking for fraud. Insurance companies receive huge amounts of visual material. This is used to illustrate claims and support the claims process (e.g. determining the legitimacy of a claim). The photos are submitted by policyholders or thirth parties to illustrate damage (e.g. car damage) or as evidence that lost or stolen items were indeed in the possession of the owner (e.g. jewellery, clothing, or cameras).
In order to optimize the use of photographs, it must be determined that the right photos were submitted and that the damaged object was photographed from every angle. It's also important to determine that the object in the photo is indeed the insured object; that older damage is not being included in the claim; and that the photo has not been used before, downloaded from the internet, or photoshopped. The next step is to determine the damage and the costs of repairing the damage. Is the car a total loss? How bad is the fire damage? Are there any visible indications of fraud?
These analyses can be made with the help of the three AI types listed above. General AI is trained with millions of images to identify general objects like cars, windows, buildings, etc. The huge quantity of data and the impressive computational power means these analyses are carried out at lightning speed and with great accuracy. Product-based AI is trained using specific claims (e.g. glass damage) to determine the amount of damage and to identify potential fraud.
As a final step, AI can also be trained with various machine learning algorithms in the specific application domain; in this case, the process used by the relevant insurance company. Combining these three types of AI ensures smooth and efficient claims processing. The longer AI is used and the more data and feedback it receives the better and faster it will work. This will speed up the claims assessment process. And it may even lead to a fully automated process in some cases, such as STP. An added benefit is that fraudulent claims are filtered out more efficiently. In addition to image analytics, there are dozens of AI types currently available and under development. Insurance companies can use these technologies to improve and streamline their risk analysis and fraud identification processes. In the coming period I will elaborate on this subject using some of our fraud prevention cases as an example.