Artificial intelligence in insurance (Part 2): Implementing AI
03/11/2017

AI installations are complex, and not without their pitfalls for insurers. But the end benefits more than justify the means.

By Eileen S. Burns, FSA, MAAA, Consulting Actuary; James Dodge, Practice Leader, Advanced Analytics & Data Solutions; and Tony Huang, Data Scientist, Advanced Analytics & Data Solutions, Milliman.

Part 1 of our series on artificial intelligence (AI) explored its likely applications in the insurance sector. Here we look at some of the issues for insurers when it comes to implementing AI and machine learning.

The decision on whether to implement an AI project should follow the same rules as any other analytics project. The essential questions to ask are:

• What is the business context for the project?
• What is the business problem you are trying to solve, opportunity you wish to pursue, or outcome you are trying to improve?
• Are you confident that your algorithm, data, and process are all fit for the purpose, and will deliver what you require?
• What is the probability of a successful outcome?

People and data
Over and above these aspects, there are two crucial elements to carefully consider: the skills of your staff and the availability of data.

Data science is often described as a team sport, and with good reason. An understanding of the business is as vital as knowledge of the techniques to manipulate data; but it’s rare to find an individual who knows both in enough detail. Putting together a multi-skilled team will almost certainly be the way forward.

Any analytics project requires data—and lots of it. The first investment to make when implementing AI is in the data. In the digital era, data must be seen as a strategic asset, and must receive the attention it deserves.

With this in mind, it makes sense to consider transitioning to a cloud-based storage and processing platform. Investing in an on-site solution may not prove worthwhile given the need for scalability in data storage and data querying.

The risk factor
There are other risks worth bearing in mind. A key challenge with predictive analytics in the insurance business is the consequence of getting it wrong.

A deep learning algorithm might be used by a movie-streaming service to suggest film choices based on users’ past viewing habits. If the algorithm directs a customer to unsuitable titles, the result is unlikely to be more than mild bemusement on the part of the user.

But using machine learning to determine an appropriate rating, and then filing it with the regulator, has more far significant risks.

As a result, it will require additional education to put senior executives and regulators at ease with this type of analytics. There are parallels between traditional actuarial models and predictive analytics that provide guidance in this regard.

A tenet of actuarial training is that mathematical tools and models should be understood and used responsibly. That means being comfortable enough with the inner workings of a model to assess whether it is producing flawed information. There are some standard tests that can help: looking at extremes, simulation tests, scenario tests, and so on.

There are parallels, however. A tenet of actuarial training is that mathematical tools and models should be understood and used responsibly. That means being comfortable enough with the inner workings of a model to assess whether it is producing flawed information. There are some standard tests that can help: looking at extremes, simulation tests, scenario tests, and so on.

More intensive testing is warranted with more complex models. Predictive analytics can be biased despite not being human, so the outcomes need to be checked.

Rather than trying to explain all of the interactions in the model, a smart approach is to try to prove that the model isn’t discriminating, and isn’t causing adverse effects. Even if the machines are doing most of the hard work, people need to be kept in the loop.

The upside
None of these challenges should dissuade insurers from pursuing artificial intelligence projects. The advantages of AI, especially in terms of handling unstructured data, are considerable.

When it comes reducing costs, AI has the potential to optimize and automate of all kinds of insurance activities. By setting better assumptions, you are more likely to get pricing and reserving right, and to hold the right amounts of capital for the risks you’re exposed to.

AI offers revenue opportunities, too. For example, by enabling services that help de-risk policyholders. This is already happening to an extent in auto insurance, with customers receiving scores based on how safely they drive. Similarly, health insurance products may offer benefits for healthier lifestyle choices, such as going to the gym, or taking a certain number of steps per day.

In the current environment, we see AI being used in certain defined tasks: it will be sprinkled onto areas where it can work its magic. The future, however, will see insurers designing business models that unleash the power of AI from the outset.

ADD COMMENT

no comments

Thank you for your submission.