Malakoff Humanis - Leveraging AI to Democratize Insights From Customer Feedback

Team members:
Nikola Lackovic,  Data Scientist (NLP & voice technology specialist)
Gauthier Lalande 
Layal Saad-Koubeissi
Zhijie Zhou

Country:
France

Organization:
Malakoff Humanis

Description:
Malakoff Humanis is one of France's leading social protection groups, covering all the insurance needs of people in supplementary pensions, health, welfare, and savings.

Awards Categories:

  • Organizational Transformation
  • AI Democratization & Inclusivity
  • Value at Scale
  • Alan Tuning


Challenge:


Speech Analytics aims at analyzing the category of calls within the CRM framework, so as to enable different internal stakeholders to leverage oral feedback received to improve our product  and customer experience.

We therefore needed a solution able to receive, treat, analyze, classify and output the data to a visualisation tool, from an external server to a PowerBI interface.  Dataiku enabled us to overcome the main challenges encountered:

  • It helped us integrate the fully scaled solution with AWS S3 containers in order to store the data.
  • The entire pipeline was then set  up without using any additional components and everything was built using the user graphic interface, apart from Python recipes which were needed for various reasons.
  • The dynamic and adaptive type handling was a feature which eased the process of implementation all along the way.
  • Data preparation and several painful jobs were done using the in-built recipes and permitted to bypass the weight of coding everything in Python 3.
  • The graph-based solution is very nice to grasp the entire workflow at a glance, also easing the process of metacognition over the entire pipeline.
  • Dataiku was then exposing data back to S3 from which the PowerBI was then linked in order to display the data.


Solution:


Speech Analytics is a horizontal product available at everyone's fingertips - from a technician looking to solve product issues thanks to client feedback, to a high-level manager visualizing the interactivity of the client with multiple teams within the organization. 

Input data consists of different types of client data: conversational transcripts of calls, metadata from IVS, and CRM knowledge-base. 

Call transcripts are established thanks to a Speech-to-Text external partner, along with several description metrics to facilitate data comprehension - hence integrating multi-model data presented an interesting challenge for the project. 

A state-of-the-art fine-tuned transformer for French language, called camemBERT, was implemented for Natural Language Processing. We also leveraged a tonal (positive, neutral, negative) model built by Dataiku Data Scientists in order to predict the sentiment of a conversation.

All along the process, every step was built within a design node in order to make a prototype that was therefore tested within the pipeline. When the use case was working under the design node, we built the scenario to run every hour within call centers hours, and migrated to the automation node.

The automation node-design is up 24/7. It is a sanctuary on which we migrate the data workflow from the design node.

The recipes used in the flow are: SQL recipes, Python Recipes, Data Preparation and Machine Learning Recipes. The entire flow built within the Dataiku platform is now running every day from Monday to Friday during working hours, 9 AM to 7 PM (GMT+02).

As a latest development, a retro-feedback loop based on call center helpers has been implemented to feed the transformer -  this will be pushed to production in the next internal release.  This is also permitted with the integration of the EKS clusters technology within the DSS framework with one of our Data Engineers, which will enable us to scale to the maximum monitoring of data (85% of all calls).


Impact:


The benefits are multi-faceted:

  • Cost savings

The solution will enable us to automatize 45 seconds per call over 12 millions calls handled every year, which leads to tremendous cost saving in terms of human resources dedicated to answering those calls. 

  • Improved customer satisfaction

Customers are getting faster answers to their questions, and more valuable interactions as they are directed  to our most relevant team members for their requests - who will provide them with support and guidance, beyond the usual transactional interaction. 

  • Data science democratization

Through making conversational data available to the broader organization, Speech Analytics empowers people to gain insights from customer feedback. 

In order to display the results, we leverage a visual stack in Microsoft PowerBI, which is a very easy and affordable way to enhance capabilities of. gathering information.

The next development is to trigger actions within other components of the informational ecosystem. For instance, we're looking into developing in Dataiku a suggestion trigger for tele counsellor allocation, so that for every traffic group within a IVS cluster, we will be able to predict the t+1 call volume in order to hourly adapt the tele counsellor presence in call centers.

Public