Artificial Neural Network to evaluate quality in insurance operations

The major challenge facing insurance operations is one of human capital. Despite the omni-channel push, we still face high call volumes since an insurance policy is unlike any other retail product and consumers require human validation for most decisions.

Training an Artificial Neural Network (A.N.N) to evaluate call quality: High volume insurance call centers have a “signal-to-symbol” problem, the complex product makes it difficult to design machine learning paradigms that evaluate quality. Below, I propose an Artificial Neural Network with layers designed to allow for meaningful dimensionality reduction of the signal as it pertains to quality from an insurance standpoint.

A high level overview of algorithm design for each layer is also presented. It would be better to let a Natural Language Processor deal with the voice data signal and feed the transcript it produces as a signal to the A.N.N.

Artificial Neural Network Paradigm:

INPUT Training data — Signal would be an annotated transcript of the call, labeled by a human after listening to the call.

Binary Classification Question: The quality rating from the learned classifier guesses whether the call needs to be reviewed by a human.

1st Layer: Learn policy type classification: Training data sets in the form of annotated transcripts must be able to able to help a neural network to learn to distinguish home vs auto policy calls by prevalence, frequency, and density of certain words like “home”, “car” etc.

2nd Layer: Learn transaction type classification: There are too many types of transactions that could happen within the context of a home or auto policy, so it is better to have a secondary classification layer to further reduce feature dimensionality.

Within home policies, the type of transaction could again be learned through pattern recognition of prevalence and frequency of certain words from call transcripts. For example, words like “closing”, “moving”, followed by policy change words like “new address”, “heating type” can be modeled to classify home policy sale calls.

3rd Layer: Learn to check for underwriting compliance: Nodes in this layer know the call was about “home policy” and “buying a single family home” as shown in the image above. These nodes are specific to a particular type of policy, and they predict if whether the call was compliant of underwriting rules hard coded in them. This is done by noting procedural keyword presence in the transcript, validating agent’s actions, and checking for “exceptional” response keywords that need to be flagged.

For example, the call was to buy a standard home policy, enterprise tools that allow binding of policies prompt agents to ask the client standard eligibility questions like “do you have an alternate heat source?”, and other such eligibility questions which determine the underwriting compliance of the policy bound. Since there is standardization of the call from the front end, we should be able to setup logic to test if such questions were indeed asked. We can also set alerts for exceptional answers that require additional processing, such as a client with a “wood stove”.

4th layer: Learn to process exceptions: Exceptions to the standard flow of a procedure are very common in insurance processes, and a layer that learns how each exception is processed correctly would help further reduce dimensionality. Exception processing algorithms are specific to “exceptional” words, like “wood stove”, running analysis of chunks of transcripts around these exceptional words to evaluate if transaction was processed correctly.

Example: IF client has a “wood stove” AND words likes “WETT certified” and “need proof” are found around the exceptional word it in the transcript, THEN REWARD, as exception was probably processed correctly by agent.

5th layer: Learn to evaluate Quality: The final quality layer will include various “perspectives” on call quality. This is because different perspectives will apply based on the call.

Example: A home policy sale may need to be reviewed from a call and policy quality perspective as shown in the image above.

A “policy quality” function might focus on comparing the coverage on the policy sold during the call to an “ideal” policy model, to compare and rate this policy. This could be a measure of the agent’s salesmanship.

OUTPUT: The learned signal classifier should output a numerical quality rating which when below a threshold limit, will trigger a human review. It should also output classified symbol object that correlates with the labeled signal.

In conclusion, this approach allows for modular layers that can be iterated along with the organization’s perspective on call quality and weightage of features. Separating underwriting compliance into its own layer allows for the constant change in underwriting rules to be accommodated easier.