Machine Learning Cuts To The Core Of The Insurance Industry

Subscribe to Our Blog

The insurance industry has always been among the early adopters of advanced data analytics, from the first mortality tables in 17th-century England to the vehicle-tracking devices monitored today by auto insurers such as Progressive.

The latest wave of sophisticated math being adopted by insurers is artificial intelligence and its offshoot, machine learning. Interest in AI is surging. Last year 32% of insurance industry CIOs said AI and machine learning were their top game-changing technologies, up from 6% in 2017, according to a Gartner survey. The International Data Corporation projects that insurance industry spending on AI will grow at a 48% annual clip to $1.4 billion in 2021.

Machine learning (ML) is just one branch of AI, but it’s the one most insurers are focused on. ML enables computers to learn automatically how to make predictions and classifications by processing vast quantities of data and finding subtle and sometimes non-intuitive patterns. ML is different from other approaches to AI because, in ML, the algorithms provide their own answers instead of the answers being programmed in by experts.

There are lots of ways insurers are starting to take advantage of AI and ML, including in customer acquisition, claims servicing, and fraud detection. But the hard problem in insurance, the one we think is most worth solving, is improving the central challenge of underwriting: who do you extend policies to, and at what price?

The predictions made by a traditional underwriting model (which likely uses decades-old logistic regression math) might consider an applicant’s income, loss history, and information on the asset being covered (in the case of P&C), and yield a probability of acceptance. Such a model would not be very predictive as, on its own, income is not a very good indicator. A collection of probabilities from a variety of variables, however, can be much more predictive. Machine learning allows insurers to consider a much larger number of variables. ZestFinance’s ML models, for example, typically use 10 to 100 times more variables than traditional models. And when those ML models are used in combination with each other — what’s known as an ensembled model — they can perform significantly better.

There are some drawbacks, of course. The more data and types of models that you use, the harder it gets to interpret your results. It would be irresponsible to run an ML underwriting model without being able to understand and discuss the contribution that every variable makes. Lack of interpretability is especially a non-starter in highly regulated businesses such as insurance, where regulators require carriers to inform applicants why they were denied or approved.

Here are three key things to keep in mind when thinking about adopting ML in your insurance business (and we think it’s good advice even if you’re not in insurance):


Act boldly.

AI and ML models will have a transformative effect on insurers who are willing to embrace the technology. But many organizations are understandably hesitant at first. As a result, they tend to engage in one-off “Random Acts of ML” rather than incorporating it into their broader business strategy. You may consider building a good core of talented people within IT to support these technologies and align them more closely with business goals. Or you may decide to partner with a fintech or two that have insurance-specific tools for business projects with clear value, such as fraud detection or underwriting. Those who think bigger — and act boldly — will be amply rewarded.


Recruit champions at every level of the organization.

Our experience suggests that a successful ML strategy requires buy-in at every level of an organization. It often takes leadership from the top to drive the overall strategy, but it will also require input from business owners as well as IT, legal, compliance, analytics, procurement, information security, and data science teams. Making sure all of these stakeholders are heard while keeping the project on track can be challenging. Often the process can benefit from assistance from a neutral, outside advisor with deep change management expertise.


Make sure you consider both the power and explainability of your model.

Even as the business case for machine learning in insurance underwriting becomes increasingly clear, you won’t be able to put your model safely into production if you aren’t able to explain it — to your regulators, board members, or other outside stakeholders. That means being able to validate the model’s inputs, monitor its ongoing performance, and understand how each variable impacts the decision-making process. Make sure the ML tools you’re using come with real explainability built in from the start. 

Let’s discuss how machine learning can change your business outcomes and improve your business profitability. Contact us today.

Talk to Sales