Automating Processes For The Insurance Sector Using Artificial Intelligence
Like any new technology or business approach, the term “claims automation” means different things to different people, depending on how you look at the opportunity and ways to achieve it. The origin of this ambiguity is the fact that claims handling processes themselves — the processes to be automated — are somewhat unique to each insurer.
Furthermore, in many cases, there are no instructions that explain the system, the flow of information or the claims handling rules. Instead, the knowledge of how to handle claims is embedded in the organization.
However, if we think of claims automation as the process of automating the tasks of a claims handler, the opportunity becomes more concrete. As important, by putting the claims handler at the center of the picture we can see how AI techniques can be used to automate the tasks involved in resolving claims, ultimately improving the customer experience, optimizing claim outcomes and reducing claims handling costs.
Breaking down the claims handling process
Despite a lack of industry standards when it comes to claims handling, insurers do tend to follow a de facto pattern in how their claims handling teams work. The key areas that these teams are responsible for are most often identified as:
- FNOL: The first notice of loss is the moment that the customer registers a claim with the insurer.
- Claim updates: The customer seeks information about the claim, or has additional information to provide to the insurer about the claim.
- Policy analysis: The claims handler analyzes the policy to understand the applicable coverage and exclusions.
- Claim circumstances: The claims handler builds a picture of what happened in the incident leading to the claim, identifying characteristics that are important from an insurance point of view (such as liability).
- Document analysis: The claims handler validates documents that are provided by the insured.
- Loss estimation: The claims handler uses his or her knowledge, tools or the advice of an expert to generate an estimate of the loss.
- Coverage decision: The claims handler applies the policy to the circumstances to decide whether the claim is covered.
- Fraud decision: The claims handler asks for clarification from the claimant, or refers suspicious claims to the SIU.
- Resolution decision: The claims handler applies business policy and their judgment to decide what type of proposal(s) to offer the customer.
The interesting thing is that these tasks — as heterogeneous and complex as they are — all fall under the responsibility of the claims handling team. In many insurance companies, a single person is capable of accomplishing them all. Hence, we can think of claims automation as the challenge of building an AI claims handler. This goes far beyond Robotic Process Automation (RPA) or automating a single task.
In this definition, an AI claims handler must:
- Provide effective interaction with the customer.
- Understand the claim.
- Make claims decisions.
But, an effective AI claims handler must also be capable of making what can best be defined as a “process decision.” And this is where RPA and static business rules fall. The AI must be able to determine when it doesn’t have the right data and information necessary to move the claim forward. This is something that a real claims handler does naturally, yet is a non-trivial problem when building an AI solution. In this situation, the AI must be able to either seek more information (from the customer or another party involved in the claim) or refer the claim to a human claims handler. Essentially, we’re adding a fourth requirement an AI claims handler needs to fulfill:
4. Be able to make process decisions.
So what does this look like in practice? Let’s explore a little deeper.
Interaction with the customer
The primary motivation for implementing claims automation is to enable better customer experiences, which typically begins with the first notice of loss and establishes customer expectations for everything that follows. For example, many consumers now expect that FNOL can be done online — similar to how they interact with other businesses.
Yet, it’s not enough to simply build an interface that adheres to modern user experience guidelines. This is where AI comes into play and supports true claims automation. For example, submitting a claim online and hearing back a week later that you’ve uploaded the wrong document creates a terrible experience. The right AI can tell you instantaneously that an error has been made and how to correct it, keeping the process moving and ensuring customer satisfaction. But as we’ve already highlighted, FNOL is only the beginning of the customer experience around a claim.
With claims understanding, we can focus on both the challenge and opportunity related to AI and claims automation. Each of the tasks in this area requires both a model of the data required to handle the claim and the ability to extract that data from the claim or supporting documents. The following examples should give a flavor of the approaches that are necessary to fill a data model with the required information.
In the case of travel insurance, many insurers issue one policy per trip which tends to follow standard forms but often does not include a structured record of the policy terms, only the policy document. The result? Claims handlers must refer to the policy document for each claim to understand coverage and exclusions.
For AI to be successful in these cases, the AI should be able to apply OCR and natural language processing to understand which terms and exclusions are present in a given policy based on the knowledge of the standard forms. Naturally, this can be applied to other types of insurance as well.
Accident reporting forms help provide an understanding of what happened when an accident occurs. But, standard forms can never capture the entire account of the accident, which is why it’s crucial to allow the FNOL to capture the claimants’ descriptions in their own words.
Here, two different techniques can support automation. Document analysis can extract the standard set of circumstances from the accident report, eliminating duplicate efforts, saving time and reducing the potential for inconsistencies. Adding the use of natural language understanding to analyze the claimant’s account of the accident helps avoid asking the claimant for the same information in another place later in the FNOL. The aim is to build a rich model of the circumstances with just a little burden placed on the claimant as possible.
Claims handlers spend a lot of time reviewing documents which can be used to validate the circumstances presented by the claimant or which provide information from third parties. Often, they cover key aspects in the claims handling process, such as information from health care providers or an estimate for the cost of repairs.
Document analysis is an area where the AI claims handler can play a critical role.
And since the AI claims handler is reading documents only with the end goal of automating claims, the reading should be limited to only what is necessary for the situation. For example, in some claims, key attributes of the document must be validated without necessarily needing to understanding the entirety of the document. A travel delay is a good example in that the system must only recognize a boarding pass, confirm the passenger is insured, and that the flight number corresponds to the claim declaration. All other fields represent unnecessary information.
However, in some situations, reading key information from the document, without having any prior knowledge about what to expect, is critical. Invoices in health claims are a perfect example. The AI claims handler must understand the billing code, the procedure and the cost for each item, and validate key patient information. In many cases, the structure of the document itself can help ensure that extraction is coherent (for example, that items add up to a total).
Today, the most effective approach is to train models to understand specific types of documents. Yet, claims automation requires that we build models that can make high confidence predictions. This can be done by enabling the model to say, “I don’t know,” in cases where a high level of confidence is not obtainable. This is a critical requirement in the design of the system and feeds into the decision-making tasks that are also part of the process.
Despite the diversity of how loss is estimated and reimbursed across lines of business, there are still tremendous opportunities to apply AI to optimize claim outcomes. Systems can use invoices from the customer as the basis for a loss estimate — with the caveat that inconsistencies must be able to be spotted and raised as appropriate. Estimates can be based on known industry norms, or even on information from third-party sources such as online listings. Finally, “low-fidelity” cost estimates — is a car lightly or heavily damaged for example — may provide enough to know what to do next to without needing to know the exact repair costs.
The claims process exists (and fundamentally insurance itself exists) to help customers to resolve problems caused by events outside of their control. To keep them moving towards the resolution, insurers must make decisions about how to handle their claims. This is the defining responsibility of a claims handler. All other tasks in the process support the ability to make the best decision about a specific claim.
To be successful in resolving customer claims, the AI claims handler must be able to make:
- A coverage decision — is the claim covered?
- A fraud decision — does the claim have merit, or is there a high probability of fraud?
- A resolution decision — what is the best way to resolve the customer’s claim?
To make these decisions, the AI claims handler must have a deep understanding of the insurer’s business combined with a wide experience of claims. Decisions must combine probabilistic decision-making with the top-down implementation of business requirements. And, we need to think about fraud. Without effective fraud detection in place, automation may simply multiply the number of fraudulent claims. As a result, a working and effective fraud detection system is a prerequisite to implementing claims automation.
These considerations lead to the final set of decisions required for effective claims automation which relate to the interaction between the AI claims handler and the real world.
A good claims handler follows the process. A great claims handler knows when to break out of the process. This is particularly important for an AI claims handler which is otherwise capable of repeatedly and rapidly making mistakes without realizing there is a problem.
Fundamentally, the AI needs to know when it doesn’t have enough information to make a good decision. When this is the case, it can either ask someone — the customer, an expert, a human claims handler — for more information, or it can refer the case to a human claims handler (the AI equivalent of asking someone more experienced to step in).
So where does that leave us? Claims automation is a long-term effort and represents an exciting vision for insurers, particularly because it can enable a transformative customer experience. However, because of the scope, claims automation can be viewed as being fuzzy and ill-defined.
By starting with the claims handler, we have a concrete definition of claims automation: automating the tasks performed by a claims handler. We can use this definition to enumerate the capacities that we need to build an AI claims handler. This makes it clear that some OCR (Optical Character Recognition) or a few business rules are not enough. We need a system that can make decisions about the claim and how to handle it — including handing over to the human experts when needed.
Some of the pieces are already in place. The industry has seen tremendous benefit from applying AI to the problem of fraud, for example. Other pieces are in their nascent stages and will take time to develop. Insurers need to have a long-term vision for what claims automation can mean for their business and organization, and find partners who can help them turn that vision into reality.