Where does the European insurance industry stand in terms of advanced analytics, artificial intelligence and automation? Do we see that traditional methods of data analysis are now being labeled by the term "machine learning"? Maybe the industry is more advanced than that: Are real chatbots, for example, already ubiquitous? Let's have a closer look.
I had the opportunity to visit the "Insurance AI and Analytics Europe" conference in London. About 80% of participants came from European insurers, with a high percentage of executives. In this blog I will summarize the most important impressions of the two conference days in short form. If you prefer an overall status across EMEA for all industries, however, this study gives fresh insights.
Using machine learning to approach familiar processes - and to promote cultural change
Insurance companies have a very long history of using analytics. Actuaries have traditionally been in charge of this task. Their methods are often based on so-called Generalized Linear Models (GLM), which are designed for small amounts of data and not too complex requirements. These methods, according to Kenny Holms, Head of Predictive Analytics at Argo Group, are plain vanilla techniques from a bygone era - before the advent of Big Data and machine learning. Newer methods and more data often lead to a higher accuracy of predictions. And only those who have unstructured data under control (today still more than 2/3 of the data available) can exploit the full potential.
Modern data scientists are taking a very pragmatic approach to using these methods. We see them more and more as part of analytics teams in insurances. However, this also leads to the necessity to actively drive the cultural change. A Chief Digital/Data Officer should coordinate all insurance activities around digitization, AI and data-based decisions. But even more important, he should ensure freedom, allow data-based research and experimentation - instead of being satisfied with what the organization has always been successful with. This alone can help to attract young talents, who typically perceive Facebook, Google, Apple and such to be much more sexy than the good old insurance industry.
Machine learning and regulation - a contradiction?
In insurance companies, the inevitable question is, of course, how machine learning with its black box approaches (e. g. neural networks) can be reconciled with regulation and explainability. This, according to the discussion participants, is primarily a question of presenting results: Move away from formulating the decision model as a rule. Be able to explain why you are convinced that the algorithm works! Have your audit trails ready and be able to tell which model led to which decision based on which data and at what time.
What is the right use case?
Where do you start if you want to establish machine learning to improve predictions in insurance? Where should AI systems be used first? Fraud is generally seen as a good starting point to try out new models and gain experience. Anthony Barker, Head of Claims Operations EMEA at SwissRe Corporate Solutions, clearly favours claims processing ("This is where the promise becomes real!"). Here, it is the business departments who are called upon to promote innovative ideas with AI applications. Do not wait for IT and all the technical prerequisites to be up and running to try out ideas. Those who apply this approach have high differentiation potential. Barker asked the audience to imagine a completely different way of processing claims than typically experienced today: Assume that an AI, by reading the local news, Twitter and other social media, concludes that you had a fire in your house. Why not just pay the insured client 50% of the possible damage immediately and then see how things will continue from there? Of course, this would not only include operationalization of AI and many changes in business processes: "Culture bites!"
The insurance ecosystem is both excited about the potential of AI in insurance, and concerned about how far the disruption will go. Read this report by SAS insurance experts on the scale of change ahead.
Operationalization is still difficult
Operationalization was indeed a major issue. A survey of the audience, which is certainly not representative, revealed that only a few organizations use machine learning for scoring purposes, but almost no one actually implements the results operationally. Generally speaking, the automated execution of analytical decisions is considered to be very difficult for a total of 50% of the survey participants in the audience. But the automation of underwriting and pricing, the claims process, fraud prevention and, last but not least, customer interaction with a personalized customer experience are of particular importance.
Most insurance companies are at the very beginning of the path to fully or largely digitized insurance products. And this requires a complete rethink, as Peter Ohnemus, CEO of Dacadoo, put it in a provocative wake-up call: A dramatic change is imminent for the health insurance industry, which has its back against the wall. The biggest driver to avoid health care costs is personal lifestyle (40% of costs are dependent on lifestyle factors). Health insurers should therefore see themselves as partners for lifestyle in future - instead of simply handling costs of resulting diseases. Healthcare becomes wellcare. And this leads to using wearables to implement a platform economy with data-based products and services.
Visiontrack's idea is a similar one: by actively influencing behavior, the supplier of surveillance cameras for vehicles wants to avoid damage. A camera monitors the driver and the driving situation in real time and sounds an alarm when the driver yawns, looks distracted to the side, makes a phone call, or closes his eyes. The aim is, of course, to avoid damage - and make an insurance the driver's "guardian angel" - which provoked controversial reactions in the audience.
Emotion instead of technocracy
Of course, chatbots are a topic you can't ignore at such a conference. Insurance companies plan the digitalization of all of their processes with customer contact, not least for reasons of efficiency. They also try what is technically possible. Nevertheless, according to the clear feedback from the auditorium, almost all applications of chatbots have failed so far. Is this due to the fact that speech understanding and automated decisions are simply not yet mature enough? David Stubbs, founder and CEO of RightIndem, emphasized the need to better understand the customer's basic needs when digitizing processes, for example in damage assessment. A pure technocratic approach will fail. If an insured person has suffered a loss, he has had a bad day - and would like to be able to speak to someone.
Conclusion: Even with AI and machine learning, bad processes remain bad processes. Only new thinking on many levels will bring success.
How ready is the insurance ecosystem for AI?
Where does the insurance industry stand today in terms of advanced analytics, AI and automation? Are we seeing traditional methods of data analysis now being labeled by the term "machine learning" or is the change more than skin deep? We hosted a digital panel discussion where these topics were covered. You can read the highlights of the discussion here: How ready is the insurance ecosystem for AI?