9November 2018too few of the latter to derive dependable models. Indeed, AI has once more proven the old adage that every cat claim is a unique world.Analytic tools are only as good as the data they rest on. We have found AI to be the most demanding analytic discipline we've worked with thus far in terms of requiring clean data. Indeed, AI is strongly intolerant of inconsistent data. Anyone thinking about a new AI project should start with a deep dive into the quality of the available data and a realistic assessment of what will be required to clean up a less than ship-shape data heap. If your project deadlines are going to slip anywhere, data quality will be the first milepost you miss.We see AI as a decision support tool in our industry. Handling claims is all about making a series of often complex, multi-factor decisions. The question is always "how can we help our adjuster make the right decision on this issue concerning this claim today?" Success in deploying AI is thus defined in terms of making better decisions. For example, can we reduce the average number of reserve changes the adjuster makes while handling a complex claim? Can we help the adjuster reliably select which claims to refer for specialist intervention as early in their development as feasible?What this means is that the output of the AI process has to slip easily and naturally in the adjuster's workflow. This is an important caveat because the raw output from a series of AI models is not user friendly--unless you happen to be a statistician. In our experience, designing the AI to end user interface was the most challenging part of the project. This required us to simplify the output into a few very basic numbers, in the case of reserve advice, or to run the output through a group of specially trained nurses for medical management recommendations before it could be socialized. The human factors engineering became the trickiest part of a massive AI project because, ultimately, everything has to compute in the same engine--Homo sapiens V 1.0.In our industry, insurance, using AI as a decision support function turns out to have multiple applications. AI becomes a way of approaching many kinds of problems. For example, can we help our adjusters become even better at detecting opportunities for subrogation recoveries? Or possible claim fraud? Which claimants are most likely to litigate? AI helps us make better, more consistent decisions about all of these matters. AI came in the door four years ago as a stand-alone application but it is becoming a platform, the underpinning for a wide range of decision-making processes at the heart of claims handling.But--we do not let AI powered chatbots replace our adjusters. The key word in decision support is support. The AI modeling process represents one way of analyzing a problem: compare the present instance to the thousands of similar claims we've handled in the most recent ten years (our cutoff for data relevance) and predict how this case is likely to develop. For all the hundreds of data points we capture, the adjuster still knows important things that are invisible to the AI system. The adjuster talks to the claimant, for example, and while the AI system uses advanced text analytics to extract clues from every note and document that hits the claims system, it doesn't hear what an experienced adjuster hears. The AI system misses both nuance and immediacy in understanding the claim and the claimant. Turns out that good old Homo sapiens V 1.0 still does these things better than our most finely tuned AI.In our model, AI and the adjuster work together in a collaborative effort, pooling their conclusions. All of this is documented in the claim file, of course, and over time the adjuster learns how to work with the AI system, learns what it does well and learns when to trust his or her own conclusions, should they disagree. Ideally, the AI system becomes one more trusted colleague, helping to make the claims process quicker, more accurate and more expert in serving the claimant. What have we learned about AI?Analytic tools are only as good as the data they rest onGary Anderberg
< Page 8 | Page 10 >