In today’s marketplace, countless vendors and technologies claim to leverage artificial intelligence (AI) or automate manual claims processes.
One could argue that starting with the very first computer, this is what everyone has been working towards. The current technology is in many ways a continuing maturation of this process — focusing on the steps in the claims process that previously were thought too complex or opaque to automate.
For the last 50 years, claims professionals have been provided the functionality (and associated limitations) of applications using traditional programming languages. By now, we are all intimately familiar with how traditional technology departments work to collect requirements, document and define “rules” for decisions, and create programs that coexist with claims examiners to improve claims processing. Now, with the introduction of AI, there are several new components that play a role in what can be programmed:
Artificial intelligence opens the scope of human activities that can be automated. While humans follow documented instructions and rules in claims management, they are also optimized to find solutions based upon patterns they have encountered previously. AI introduces pattern matching to machine programs — allowing the creation of software that no longer requires a discrete breakdown of all the decision-making rules ahead of time. Instead, AI offers the ability to leverage historical data to map out and perceive those rules in an abstract manner. If experienced claims examiners looked at claim X and decided action Y in the past, then AI systems can be trained to do the same.
Not to be limited by the methods used by humans to make decisions, the machine learning AI can be even better. In fact, AI can identify patterns in data that should be used for decisions even if the examiners either did not use those criteria or did not even realize they were using those criteria. This advancement allows for the creation of intelligent decision engines to optimally automate tasks in the claims process that previously were thought to be too complicated or that need a subjective decision to be made.
Designed to look at early claim data elements, auto-triage systems combine the data with historical patterns that may be difficult to otherwise see to accurately place a claim into a particular “bucket” for processing. There are many ways companies can choose to accomplish this. Some may want to train their AI to place claims into differing cost estimation buckets — allowing them to distinguish levels of risk by cost. Others may prefer a more abstract definition of “complexity” that can align with costs, but sometimes simply helps identify claims that might require more time, more human contact or a higher level of experienced oversight.
When auto-triage ends, auto-adjudication takes over. Usually if a claim has been triaged as simple and able to be processed without human intervention, the remaining steps of the claims process can be automated. Using AI combined with more traditional methods, the claim can be moved from intake into eligibility and finally, through payment and closure. These auto-adjudicated claims are not only less expensive to manage, but often result in greater claimant satisfaction as they are not delayed awaiting review by adjusters — thus providing closure to the claimant quickly and easily.
Usually discussed as a stand-alone technology, predictive modeling often represents a key component of artificial intelligence solutions. Combining machine learning and complex statistics, it uses historical data patterns to predict the future. When automating decisions or steps in the claims process, predictive models can be used to choose the actions that provide the optimal predictive outcomes based upon cost, customer satisfaction or risk.
While these models leverage modern technology, they can also be used to optimize the amount of human intervention on critical claims. Targeted prescriptive recommendations to add experienced resources can be made where nuances in a claim point to several scenarios:
Where fraud likelihood is high, the claim can be escalated to investigators or provided to human adjusters for further research and action. AI models have the ability to not only look for fraud on individual claims, but to look at the entire book of business and see patterns of fraud in otherwise innocuous looking individual cases.
Where litigation likelihood is high, special actions on the claim can be taken. These can be tasks such as saving and archiving retail security video, interviewing witnesses and other participants while their memories are still fresh or otherwise providing improved care and communication with a claimant to avoid dissatisfaction.
Where special expertise could benefit the claim, either based upon current “triggers” or prediction of future triggers, AI can escalate the claim for involvement of various teams. These teams might consist of nurse case managers, subrogation investigators or those specializing in other complex or unique claims.
The next realm in machine learning AI lies in text mining of unstructured data. While ultimately providing similar outcomes to traditional structured data predictive modeling, the text mining of documents, notes, witness statements and other elements of the claim offer up to the system a significantly increased amount of information. Having claim systems capture all comorbidities or dangerous drug combinations is unreasonable. However, having AI learn the risks associated with seeing them mentioned in notes, medical documents or correspondence is possible.
Key challenges and risks
While all of this sounds like a utopia for claims management technology, it still comes with its own set of challenges and risks. Many of the events we are asking AI to learn about are considered rare, making them often difficult to predict — even with the best data and statistical modeling software. They hardly ever occur as a percentage of the overall book of business. This does not mean they are impossible to work with, but it behooves the builders or buyers of automation/predictive software to understand how these can still have challenges with accuracy. Beware the vendors who claim to have models that have 95% accuracy in prediction of rare events. If those events occur on 3% of the claims, then greater than 97% accuracy is the bare minimum that any basic model should produce.
Ultimately, the effectiveness of any AI based automation or prediction models will be dependent upon both the quality of the data and the quality of the actions taken by a claims team — should prescriptive actions be recommended or escalated to them. Like most technologies, it often works best when not fully manual or fully automated, but instead combined in just the right amounts to minimize risk while maximizing customer satisfaction.