Sagacify A green "S" rounded shape with the text "Sagacify" Sagacify A white "S" rounded shape with the text "Sagacify" Contact-us

Automatic Email Processing

Challenge

Sagacify is helping DAS, a leader in the Legal Protection insurance market, to optimize its incoming email treatment processes.

As in many other sectors, insurance companies are dealing with partners and clients almost exclusively by email. In the case of DAS, this results in several thousands of incoming emails per day. DAS already has automation systems in place according to the intent mentioned in the email. For example, once an email is identified as mentioning a renewal of insurance police, it will automatically trigger an update of the police contract.
However, today, tagging emails with the right intent, the right insurance police, and the right status in the conversation is a manual process that is done entirely by humans. The tasks involves reading the email and attachments, figure out the intent, and encode these labels in a user interface. This tasks is repetitive, time consuming and prone to human errors.



Solution

Sagacify developed an artificial intelligence system that assists the DAS email labeling teams. Every incoming email is analyzed by the Sagacify email intent classifier. From the content of the email, the algorithm decides what is the subject of that email, who sent it (customer, lawyer, broker, … ), what is the intent and where it should be transferred. When the algorithm has not enough confidence in its interpretation, the email is sent to the DAS labeling team for validation and used as a new example for the model who gets more and more accurate with time.
Our automated emails treatment system also detects key information from the emails, such as

    • which emails contain an invoice and the invoice content
    • client number, policy number, license plate numbers (old and new if there are several), contract number, …

These extracted information can then be added to the email in a structured way, and used by the right system. The integration is transparent.
With such a workflow, instead of having to manually process every single incoming emails, DAS teams can focus on complicated emails only.

The graph below illustrates the results of our solution on real world data. We can see that if we accept to process an email with 90% of precision, we can process 90% of the total volume of the emails (recall), Which means that about 20% of the emails only will still need a manual intervention.

Relevant for this assignment:

    • Experience in deploying, producing, and maintaining an artificial intelligent system that is handling thousands of prediction per day.
    • Experience in following good practices in machine learning (including CRISP-DM) to develop a qualitative classification system for text.

Associated keywords:

    • Experience in deploying, producing, and maintaining an artificial intelligent system that is handling thousands of prediction per day.
    • Experience in following good practices in machine learning (including CRISP-DM) to develop a qualitative classification system for text.