At an If P&C Insurance event in Stockholm, representatives from Ping An provided inspiration and insights into the possibilities artificial intelligence and machine learning have to offer for insurance companies and their customers.

Established in Shekou, Shenzhen, in 1988, Ping An Insurance (Group) Company of China, Ltd. covers the entire financial services spectrum, with a full range of financial products and services, including insurance, banking and investment. By the end of 2018, with over 1,800,000 employees and life insurance sales agents, the Group’s net profit attributable to the parent company hit RMB124.245bn, up by 39.5%.

Technology-driven business

At Ping An P&C Insurance, customer service representatives receive over one million telephone calls per day. To manage this volume, it is vital to have a robust service model that utilises technologies such as artificial intelligence. To successfully serve every customer, the business model requires the latest smart technologies.

Whether managing claims or answering customer enquiries, Ping An is benefitting from technologies such as human face recognition, voiceprint recognition, image/video recognition, speech recognition, as well as natural language processing.

As of June 30, 2019, the company boasts a first-class technology team of 101,000 technology employees, 32,000 R&D employees, and 2,200 scientists. Moreover, Ping An has established eight research institutes around the world and is partnering with leading universities and research institutes to develop the technologies needed to increase efficiency and better serve their customers.

Dr. Rong Xiao, Director of AI Department at Ping An P&C Insurance

Managing the customer ecosystem

Dr. Rong Xiao, Director of AI Department at Ping An P&C Insurance, provided a deep dive into how artificial intelligence and machine learning play an integral role at Ping An.

“Actively collecting information will help improve efficiency, this example is based on creating the best possible automated customer experience for online users. For example, when a customer calls – these discussions are recorded. These recordings are reviewed to assess how specific calls should be handled, and what is required from the customer service to complete the call successfully.

When there is enough data, you can build a model which will support the chatbot to always work towards a positive outcome on a particular call. The collected data provides insights into what type of response from a customer service representative on a call is required in order to reach the best customer experience.

Actively collecting information will help improve efficiency.

ʺDuring his workshop, Dr. Xiao spoke about at the development of ‘a one-stop customer ecosystem’. This involves integrating all the available data and information into a single portal. Creating just one online app to cover the full customer journey, whether this involves purchasing a home or a motor vehicle or something else.

Dr. Xiao explains, “if you look at cars, our application helps customers select the right car for their purposes, it will then locate such a vehicle for sale in the nearby area and offer financing for the purchasing of the vehicle. The same app is then also used to purchase insurance for the vehicle, schedule any maintenance, and even locate a parking space when you are shopping for groceries.

To this end, the app can be used to offer real-time coupons, such as free parking, to support an unparalleled customer experience. Effectively all the difference stages, up to making an accident claim or selling the vehicle online, after which the cycle can be started over again.”

Building a technology tool-kit

Technologies like image recognition (for example, OCR, face/object recognition and image searching) can help assess auto collision damages, identify livestock, as well as fight fraud.

Although the scenario for using technology to recognise a specific pig or cow recognition is quite limited. Dr. Xiao provides an example of how this technology can be used to confirm the identity of insured farm animals. While it may be easy to assess that the dead cow in an image is the same animal that was insured (based on its distinct markings), it can be difficult to confirm if a dead pig is indeed the insured pig.

“Using facial recognition and other technologies, artificial intelligence can be used to not only recognise the pig, but also calculate the volume of the animal based on photos taken and filed with the claim. So, we use a computer vision algorithm to estimate the weight of the animal by calculating its volume using deep learning algorithms. The estimation error is usually under 20%, which helps reduce the risk from claims.”

Dr. Xiao notes however, that the accuracy level with regards to animal recognition is not as high when compared to human facial recognition. “There are three reasons for this,

  1. pig/cow face images do not contain as many skin texture patterns.
  2. pig/cow recognition also needs to be compared the face images from the same animal, thus images are needed of the animal both before and after its death. Collecting such a data set can be pretty hard to do.
  3. Moreover, the appearance of the dead animal changes a lot (view angle, stain, skin tone, pupil shape).”

Dr. Xiao concludes, “collect your own data, create your own models. In that sense you will have to build your platforms from scratch, though you can get some basic solutions off-the-shelf, but usually these will only help you get started. To make a stable model, you will need massive amounts of data. The best off-the-shelf products are built with massive amounts of data.” He highlights that “In the end, there are no quick solutions in AI.”

Did you know?

Chatbots need five capabilities:

  • The bot must be able to listen, utilising ASR for example
  • Chatbots must speak to the customer, for example by using TTS
  • The chatbot will have be able to look and see, e.g. using OCR or image/video recognition
  • The chatbot will have be able to think, e.g. using knowledge graphs to do inference, using models to make decisions
  • Last but not least, the bot will have to create transactions and complete the work, this requires APIs

Written by

Kristian Orispää