Jérémy Jawish: “That’s right. Machine learning is involved in all three phases of product development. Firstly, during the configuration phase, when we set up a number of workshops for the purpose of better calibrating the solution. Typically, we adjust our loss compensation thresholds to those of La Banque Postale Assurances IARD, once we’ve integrated the APIs. This was followed by a testing phase. Only once it proved to be successful, we transition to the production phase.
In the next phase, responses proposed by machine are reviewed by claims management advisers. The initial production launch took place under human supervision, which means that we don’t allow the tool to directly offer solutions to policyholders. We let it gather information and then add the results of its decision-making to an evaluation interface. Policyholders will get regular reminders during this phase. The results are reviewed by very experienced claims management advisers. This allows us to measure how relevant they are, which is also referred to as the compliance rate. We then move into automation, but only when this rate is over 92%. We successfully exceeded this threshold within the three months allotted. Throughout this period, we closely monitor the managers’ feedback to optimise the solution. Regular exchanges of ideas are organised with the claims management teams to discuss and interpret this feedback.
The last phase is when there is no more human control. Decisions are no longer reviewed and approved by managers. Instead, we let the solution do its thing. However, the key performance indicators are closely monitored. Through this, we identify a list of improvements and measure their impact on the results.”