Automating the claims intake for Belgium's legal protection leader

The challenge 

When a client ends up in a legal dispute, DAS takes the case on: legal advice from legal experts, finding a solution, and covering the legal costs. Around 200 experts claim managers and legal experts resolve most disputes without ever going to court. Most of that work arrives by email: over 100,000 inbound emails a month, from clients, brokers, and attorneys.

Each of those emails passed through the same manual step: an operator opened the message, read it along with its attachments, chose from more than 100 categories, and routed it to the right team. As volumes kept growing, that manual step could no longer keep pace.

The categories themselves added to the problem. They had grown over the years, and some sat so close together that even experienced operators disagreed on where an email belonged, which meant emails regularly ended up in the wrong place and had to be sent back and rerouted by hand. Gradually, response times stretched, and the pressure on the service team grew.

The solution 

DAS came to us with a clear request: automate the classification of inbound claims emails. The goal was to free its team from the manual handling and to protect the quality of service that volume was putting under pressure.

DAS’s category set was structured into clearly separated groups. On that set, our data and AI team trained, developed, and deployed two AI models, built on DAS's own labelled email history. 

One predicts each email's primary and secondary categories from the message and its attachments; the other detects policy and claim numbers. Together, they let an incoming email be classified, matched to the right case, and forwarded to the right department.

The system splits the inbox into two streams. Emails classified with high confidence are processed entirely automatically, with no manual validation needed, while everything below the confidence threshold is queued for review by an operator. Each correction an operator makes feeds back into the models, so performance keeps improving with use.

Alongside the system, we delivered a monitoring dashboard giving DAS's administrators a real-time view of volumes, success rates per category, and overall performance. On that basis, DAS took the decision to go live.

Today, the system processes 1.2 million inbound emails per year: 100% automation rate for underwriting classification and reference extraction, and 70% for claim management. DAS's claims team shifted from manual triage to handling the cases that actually need them, on infrastructure that runs securely and adapts through regular updates without disrupting live operations.

The approach

Before training the model, we needed to understand how email treatment at DAS actually worked. That meant sitting with the people doing it instead of reading a process document. From those conversations came the metrics that would define success: precision, recall, and the share of emails still needing human treatment.

The work then ran in phases, each with its own deliverable. A small cross-functional team of AI and software engineers worked across those phases together rather than in sequence, with weekly contact keeping decisions close to the work and a steering committee across budget and priorities.

DAS evaluated the results before committing to production. Leadership had the full performance picture before any next step was agreed. The rollout was gradual: a test environment first, acceptance tests, then the switch to production.

The claims inbox was the second system we built for DAS. The first had handled email classification for its production department. After claims, DAS came back a third time for invoice control, where abnormal billing amounts are now flagged automatically before manual review.