Rousmaniere: Artificial Intelligence Has Entered The Building

Artificial intelligence is already at work in the workers’ compensation industry, through the use of what is called machine learning. It will likely take years before machine learning becomes a staple of paying claims, issuing insurance policies, and personnel training. But right now is the time to understand what it does. 

Surprisingly, we can intuitively grasp what this technology is about. And we have instances of actual use today. Keep in mind that this is not hypothetical, or experimental, or “way out there where people fear to tread.” 

Claims staffs can use AI to more accurately predict claims outcomes and select interventions, such as case management or subrogation reviews. Clara Analytics was founded expressly to apply machine learning to workers’ comp claims prediction. The firm started five year ago as a division of a larger company; two years ago it became independent. 

Senior vice president Tom Ash, a veteran of the workers’ comp industry, told me that most  people expect that AI will do thing faster and more accurately with less people. It does those things. But, he told me, “what AI really does is to report something that we don’t already know. AI will determine on its own, solely by data analysis, what the claims drivers are without bias.” 

Clara discussed with industry experts the drivers of high claims costs – Ash uses the term “high complexity claims.” These experts listed a handful of influential drivers. After analyzing two million claims, Clara found that the most influential driver had been completely overlooked. 

That driver is the replacement rate between the take home pay of the injured worker, before their injury, and the indemnity payment for wage replacement. The higher the replacement, the greater likelihood of a highly complex claim. 

I discovered few years ago that this relationship varies according to unique circumstances of the worker, and also by jurisdiction. Ash says that Clara found that to be the case. 

He did not share which states tend to have very high or very low take-home wage replacement rates. But my own analysis tells me that the following states, among others, are likely to have high replacement rates, where the indemnity payment may be over 90% of pre-injury take home pay: New York, New Jersey, Illinois, Hawaii, and California. Other states may tend to have take-home wage replacement rates below 80%: New Mexico, Rhode Island, New Hampshire, and Connecticut. Clara’s analysis is far more subtle than mine. The firm analyzed millions of claims. This is machine learning. 

Valen Analytics provides to workers’ comp insurers tools to improve risk assessment and pricing of policies. It can now help insurers accurately price policies even when there is no loss history data. The firm uses a variety of techniques including a machine learning process called ensemble learning. Its test on 650,000 policies showed the superiority of its new approach.    

For its Unavailable Loss History Model, the data include insurance policy, exposure and premium information along with third party information from regional, public and rating bureau sources. 

Let’s now look at what is under the hood in machine learning. I am pretty confident that even data-adverse people will grasp the following description. 

Start with the incomplete sentence: “The quick brown fox jumped over the lazy [word].” What is that missing word? Assume that this is written in Turkish so that we, with no knowledge of Turkish, have no idea what any word means. And the computer makes no effort to find word meanings such as by referring to a dictionary. Yet it suggests what the word is with a very high level of confidence. 

The computer, without any idea of word meaning – in that regard the computer is stupid -- will produce a numerical score for each word relative to the probability of each word coming before or after each of the others. Hence, for the 8 stated words, “the” will have a relatively high value for its relation to “quick” and “lazy” and “lazy” will have a lower value relative to “brown.” 

Next, add this sentence: “To be or not to be, that is the question.” These are 10 words, one of which (“the”) was in the first sentence. The computer will rescore each of the 18 words with each other. 

Next, add 100,000 more sentences, one at a time. With each new sentence, the computer will recalibrate scores for each word relative to its probability of coming before or after each other word. It is searching eventually for strings of words. It will, for example, soon predict that “you” usually comes after “well thank.” It will at some time calculate an extremely low probability that the final word in any sentence is “the.” It will assign a very high probability, with a great deal of confidence, that the missing word is “dog,” much more likely than it is “cat” and far more likely than it is “cloud.” 

Recalibrating by claim after claim, the computer comes to predict ever more accurately the outcome of the next claim fed into it. Insurance policy by insurance policy, it is ever more accurate in predicting the most appropriate price that meets the insurer’s underwriting criteria. 

Now, a question. Is it correct to describe this process is mimicking what the mind does? That is, is it “artificial intelligence”? According to a concept of the brain, it is. The brain is constantly updating best guesses about sensory inputs and cognitive phenomena.   

What to do? Learn more about how you can use AI.

Share This Article: