About
First implementation of a fully automated AI emails classification system in the Belgian Insurance Industry.
Innovation presentation
AXA decided to launch an email handling system (Malta) to sort emails automatically thanks to AI. The NLP project within AXA has already generated a successful POC. Since the results were very positive, AXA decided to extend it to ALL inbound emails.
The system is simple, you train the MALTA Tool with all emails from previous years (at AXA 6 million emails for year 2017) and then you analyse if it works properly or not. Then you retrain again. And so on. Until the model reach the desired outcome.
The tool will allows approximatively 30% of automatic email classification with a precision level of 98% (percentage of correctly classified emails). An experienced human being can only achieve 95%. So it allows people from that department to spend more time on value added tasks such as answering to clients but also it ensures continuity in the handling of this task with better results and less rework.
Moreover, to train the solution, data has been stored on the data lake in order to have distributed work. This is also a totally new approach since the data lake is a very upcoming technology too.
On top of that, an automatic OCR is performed to read all the text on the attachments and perform better classifications.
Uniqueness of the project
First of all, MALTA Tool has never been used on Inbound emails in Insurance before.
Secondly, email handling for such quantities of data is quite unique. NLP with AI on 6million files a year is quite a big number.
Finally, the timing from the first line of code to production is only 8weeks! A revolution in the current way projects work. From scratch, a full model can be trained in 2-3 weeks and implemented 5weeks later. It allows full flexibility to the business and the “Fail fast” technique is definitely applied.
In-Depth Analysis
The Reason Behind
AXA wants to give more added-value tasks for his employees. Time spent on email classification is heavy and thanks to NLP, they will be able to focus on more added value tasks such as complex classifications and support in GDPR regulation.
Moreover, AXA Belgium is considering using Artificial Intelligence solutions in order to increase automation and efficiency. Actually, average inbound document is taking more than 8 hours to reach the operational departments what impacts client satisfaction and with NLP it occurs instantly.
Results
The pilot has reached an overall precision of 98% and an overall automation of 27% and we are about to beat these results. We can conclude that AI works for AXA when applied to inbound email.
Moreover, it is becoming more effective with the quantity of data increasing! That’s why, we are able to split the emails in 6 categories and 266 sub-categories in MALTA tool for all inbound requests.
The NLP project at AXA has a lot of added-value, especially:
• Automatically classify emails into categories (efficiency increase)
• Pre-populate emails classification even for the cases where precision is lower than the target
• Rapid classification
• Increasing customer service
• Formulate recommendations on how to create a better classification
Key Dates
- Launch date 12 January 2024