Machines Learn Policy Review

Going back, you were working as an insurance attorney, so what led you into insurtech?

I got really tired of seeing the same documents over and over during litigation. One of the first cases I ever worked on, I was told to find the words “Montgomery Point” or variations of those words in probably more than a million emails, which took me about three years to review. I started paying attention to technology and different ways of doing things. I started noticing machine learning—in the litigation context, it’s called technology-assisted review. I realized machines could do what I was doing and they could actually do it much better.

How is technology able to tackle something as seemingly complex as insurance policies?

I would argue the policies are complex because of the convoluted way they are put together. A lot of times, a policy is a compilation of forms and endorsements, which is another way of saying it’s a handful of documents jammed together. We realized we could leverage machine learning technology to break an insurance policy down into its individual clauses and we could break the clauses down into their numerical values or the wordings to help people understand them faster. We could do that because of lot of the time the industry was using the same documents over and over and over. Machines are very, very good at sorting out repetitive documents.

You started with a focus on claims. How did the shift to policy analysis come about?

We started working with one insurance company and then another and collecting claim documents at the source and then uploading them to our software. The pivot point was when an underwriter called us on a very large claim. She said, “I’ve been reviewing the underwriting file in your claim document management software and really like it. Can I use this software for policy review?”

Like any good entrepreneur, I said, “Sure! What’s policy review?” I really didn’t know a whole lot about policy review at that time. Slowly but surely, we realized that this was a much bigger problem than the claim document problem we were solving and that we should go solve this policy review problem.

How are policies similar, and how do you find the differences that make a difference?

There are a couple different ways I look at a policy. There is the language, and then there are the numerical values.

Insurance language could be a clause that excludes any damage by war or terrorism or government. Another clause might say, “We do not accept responsibility for anything damaged if there is an attack, which involves missiles by a government person.” Those two different clauses may essentially mean the same thing.

The first thing we did on machine learning was we trained our machine learning platform—we call it Johannes—to understand that both clauses mean the same thing. Both should be labeled “Exclusion—War.” We’ve done this work on more than a million clauses now and identified about 3,000 categories. As of now, we can take wording from one carrier and wording from another carrier and line up similar clauses, even though these clauses are written differently.

Step two was to go deeper into the policy and not only pull out the policy language but also pull out the numerical values.

Imagine you have a war policy deductible for $10,000 and on another policy you have a deductible for a missile shot by a government person and that deductible is for $100,000. In the end, the broker needs to be able to line those up and tell their clients that both of these are deductibles for war and this one’s for $10,000 and this one’s for $100,000 and advise them which one’s better. In that case, you have to be able to apply a common label across both deductibles and show how the data compare. That’s our Policy Check tool we recently launched. It’s standardizing the language and the numerical values no matter what an insurance policy calls it.

Why Johannes?

Johannes is named after Johannes Gutenberg, the guy who invented the printing press. The more we have worked with insurance policies and artificial intelligence, the more I have come to appreciate old document systems. And the printing press was the first document system. Johannes is basically our attempt to pull apart existing insurance documents and turn them into usable data.

When did you launch and how has it been going?

We started building in late February 2015. We launched a beta version in February 2016. That was fun in that we brought people in and they broke the software immediately, which is how it should go. We launched a full version in late 2016 and signed two of our large enterprise customers. One of them is a partnership with QBE. They’re bringing 126,000 insurance forms into RiskGenius so that they can review, research and bind those forms a lot faster.

How do you see machine learning changing the industry?

A lot of the work that brokers have to do is highly repetitive, and they struggle to even get it done on time because there is so much of it. I compare it to Uber, which realized the supply of rides was not satisfying consumer demand. They created more supply by finding more drivers and cars.

On the broker side, there is a lot of demand for the best insurance product possible. A lot of time brokers don’t have the time to provide that analysis to their customers, particularly small and medium-sized customers. Brokers can’t put in the time needed to do that type of work because they need to focus on the big accounts. Our goal is to free up the brokers to do more policy analysis faster so they’re doing less of that repetitive work. In the long run, it’s going to create better relationships between brokers and their customers because they’re going to spend more time on risk management and finding what’s in the customer’s policy and what the best coverage is.