Classifying claims correspondence at an insurer
Insurance, regulated industry
Industry and type of organization
A mid-sized insurer operating in a heavily regulated environment. Data protection and processing security are a condition of entry to that market. The engagement is ongoing.
Business problem
The customer service team and the document-processing team were working at the edge of their capacity. Every letter entered the queue in the same form, regardless of whether it concerned an urgent case or added nothing new.
Rising handling times began to threaten deadlines, and with claims a delay means the risk of a claim lapsing and lower customer satisfaction. The backlog kept growing because the queue grew faster than it could be cleared.
Approach and scope of work
We designed and deployed an NLP (natural language processing) model that reads incoming correspondence and assesses it on two counts: whether it is relevant to the customer's case and what risk it carries. Letters irrelevant to the case are set aside from the main queue so they do not slow down cases that require a response, and nothing is deleted in the process. The rest are classified and passed on with a ready category, so the next steps of the process can run automatically. Underneath, text classification with supervised learning does the work: the model learns from labeled examples of letters and attaches a confidence level to every decision, so uncertain cases go to a human.
Business effect
around 25%
fewer documents in handling
the system sets aside correspondence irrelevant to the case
90%
recognition accuracy
correctly identified risk category
30%
fewer backlogs
reduction of old correspondence backlogs
About 25% of documents stopped burdening the main queue, because the system sets aside correspondence irrelevant to the case. The risk category is identified correctly in 90% of cases, which lets certain cases be routed automatically and leaves doubtful ones for a human to check. Correspondence backlogs dropped by 30%.
Scale of the implementation
The solution works on hundreds of thousands of documents a year. We maintain and keep developing it, so the model is retrained as the character of incoming correspondence changes.
Security and confidentiality
We ran the project under a GDPR-compliant regime. The data the model worked and learned on was anonymized, so at no stage was it tied to a specific person.
The whole thing runs in a separate, secured environment built for this project. The architecture was dictated by regulatory requirements: the data does not leave the agreed environment, and every processing step can be demonstrated to an audit. In a regulated industry this is the condition for putting a solution into operation at all.
Reviewing documents by hand eats up time that no one plans into the budget. If it looks similar at your company, write us a few sentences about your process. We will tell you whether we see room to take work off your team, including when, in our view, there is none.
Related articles in the cm-opti knowledge base
Service area: Process automation and document processing.