AI in Insurance: Managing risk in Commercial Buildings Insurance.

In a previous post, I mentioned the work I did last year as part of an Internship focused on AI in the Insurance Industry.

One of the ideas and concepts we explored focused on Site Inspection for commercial buildings insurance based on some use cases I’ve observed in the utilities sector. Utility companies usually manage large infrastructure that is critical to their business of supplying a service to their customers. The infrastructure they manage is usually complicated, large and widely distributed. The consequences of infrastructure failure can be pretty severe. However, doing this at scale is dangerous work and just isn’t economically feasible. Utilities companies are looking at using drones to record videos and do the work of an engineer who might ordinarily scale dangerous and hard to reach infrastructure.

Is this an example of automation replacing people?

Well, no, it isn’t. An engineer needs to fly the drone. They will understand key parts of infrastructure that are prone to failure so that knowledge won’t be replaced. Its a case of reducing the time taken to inspect infrastructure, reduce the risk of injury and reducing the risk of infrastructure failure. In theory, this should result in more regular inspections.

So, what does this have to do with AI?

Lets assume a site inspection using drones might generate 40–60Gb of footage. Somebody will need to review that footage and create a report detailing their findings. However, introducing an object detection model to do the work of the human means that the report could be automatically generated. Again, this doesn’t replace the human - it puts the human at the centre because they effectively audit the report with the net benefit of only seeing what they need to see and significantly reducing the time to create the report. The end to end process of site inspections can be significantly further reduced meaning more inspections and less infrastructure failures.

And, what does this have to do with Insurance?

Whilst we were working in Hursley, we thought about the sheer size of the facility and how it might be insured. Could you pivot the Utility site inspection approach to the work of an Insurance Site Inspector who is tasked with classifying risk for a policy? This was quite an exciting thought, because this leads to more services that insurance companies could potentially offer their clients such as recommendations to lower premiums, or just generally advise where disaster might be imminent or things generally might need fixing to stave off long term damage.

A simple object detection model to explore the art of the possible.

This idea remained conceptual throughout the Internship, but we ran it past people currently doing this function and they thought it was feasible. I believe in the ability of the technology to deliver the use case, so it was a case of finding the right time and inspiration to explore it further.

This happened during a particularly quiet trade show. One of the things I get to do in the Systems AI team is go to industry events and show our technology in action. In this case, I was showing PowerAI Vision and my work on the simple car damage classification model I reference in my last AI in Insurance blog post.

During a quiet period, I started to use a popular search engine (I bet you can’t guess which one) to find some pictures of farms. Now, farms vary wildly. I gave myself the best possible chance of building a successful model by focusing on farms in the USA and decided to build a model to detect buildings, silos and trees. Why farms I hear you ask? Well, I thought there would be a great deal of footage and farms need insurance right? There’s probably lots of risk on a farm that insurers need to manage with their clients.

This also coincided with a new feature that was just incorporated into PowerAI Vision: image segmentation, so I was keen to have a play and this was the perfect excuse! I uploaded my images, started playing with the labelling tool to draw polygons around areas of interest and labelling the data. I then trained a model. All point and click. I’m no data scientist, but I’m a keen and fearless explorer of technology and I understand the process and outcomes. Honestly, my model training results were pretty modest, but this was a proof of concept. I understand that to do this in production, you would need a lot of data. I didn’t have the luxury of time and an endless supply of quality data.

Next step was to find a video that could be abstracted as a drone flying over a farm. Again, at the mercy of publicly available data, this was a challenge, but I eventually found something I could test my model with.

In the space of 2 hours during the quiet moments, I had created a rudimentary model and tested it out using PowerAI Vision. Even though the accuracy scores were pretty poor, testing the model with the video footage was pretty cool and given the time and scale of the problem, I am convinced that doing this kind of classification could be achieved at production scale.

So, I can detect buildings, trees and silos, I could then parse the json output from the inferencing API to look at objects in close proximity like trees overhanging buildings that might be a risk or maybe going more granular on the buildings and detecting materials that should be replaced like asbestos. It takes a bit of imagination I admit, so I’m looking for forward thinking pioneers who are willing to give it a try! If this is you, please get in touch!

If you are interested, my project materials are here.

I hope you enjoyed reading this article. I work with some great people in the IBM Systems AI team, so make sure you check out what they are up to in our team blog area.