AI in Insurance: A simple use case.
Life is never dull working in the IBM Systems AI Team! We have the privilege of meeting so many great and interesting people and working in all kinds of different industries. The maturity level of AI adoption and the type of AI being applied varies significantly between industries.
We work with Healthcare professionals and researchers looking at Deep Learning Computer Vision for Image classification and segmentation, Financial Services companies building risk models incorporating machine learning and exploring deep learning techniques, Manufacturing companies using Deep Learning Computer Vision for high speed fault detection and Utilities companies tackling proactive maintenance of infrastructure and anomaly detection using video and monitoring data. And, listing these off…it really is the tip of the iceberg!
Late last year, I was given the task of running an Internship for Insurance Companies interested in AI at IBM. The plan was to educate them on our Systems capabilities and look at what they are doing to innovate by applying AI in Insurance. We put the call out and Zurich, Allianz and Travelers Insurance sent representatives to work with me and develop a white paper.
We spent 2 weeks at our Lab Campus in Hursley, UK and then a week over in Austin, TX at the Client Centre. It was a really enlightening process and great fun to work with these three individuals with different domain expertise within Insurance.
As part of this engagement, we ideated and explored use cases that we thought we could develop given the tool sets available and our level of skill. As none of us are practicing data scientists, we quickly zoned in on a use case to classify car damage as an example of using AI to accelerate Claims payout and potentially provide an early First Notification of Loss.
To prove the use case, three of us were assigned a category of car damage and scraped the internet to find images that matched our categories: “Fender Bender”, “Pristine” and “Write off”. We collected about 30 images for each category and uploaded this to an IBM PowerAI Vision instance we had running at our internal lab. We then trained a model that would be able to classify new images, deployed it and tested it. No coding. All point and click. Caveman simple and took about an hour, end to end!
This was just a simple proof of concept and the data was pretty sympathetic, with all images from a front angle and very distinct categories, given the time available and the fact this was a bit of a leap into the unknown.
When we tested the deployed model API (again, point and click, drag and drop), we were pretty amazed that given the limited data set, our test images for each category worked beautifully. Of course, this isn’t a production grade model, but it shows what’s possible.
The main idea driving this (pun intended?) was “what kind of data do insurance companies have that they don’t currently utilise that we can use AI to turn into something of real value”. Car insurance claims probably generate a great deal of image data that could be used in this way.
The project data repository is here. You can use this to recreate the model we built using IBM PowerAI Vision.
So, what did we learn?
- Getting started with AI needn’t be the exclusive domain of a Data Scientist. There is software out there that can lower the barrier to entry and put the AI in the hands of the data domain experts.
- Start small. Think about what you can do to prove your idea without a huge investment in time and resources. Adopt a “Fail Fast” mentality.
- AI is built with data. Think about the kind of data that you collect and archive because it is of little apparent use beyond storage for regulation. It could be a ticket to a new offering to delight your customers and offer you a competitive edge.
- Austin has a great craft beer scene. It’s the capital for live music and the barbeque is awesome. Its also very hot in September and the IBM team running the client centre will make you very welcome and have some great people of tap to educate you about enterprise AI!
The White Paper itself is pretty expansive and discusses much more beyond Use Cases.
I hope you enjoyed this article. I’m planning a new blog post on another use case we explored in the White Paper so watch this space!
