An AI agent triages over 15,000 claim emails a year, cutting case-opening time from 5 minutes to a few seconds for an insurance broker.
A major insurance broker manages, on behalf of a French property developer and its contractors, more than 15,000 emails related to damage claims and ten-year warranty coverage (DO/CRAC). About 2,500 claims must be recorded in its claims management system — an operation that, when done manually, takes nearly five minutes per case and involves three claims handlers.
The process includes sorting emails, opening the case in the management tool, sending acknowledgements of receipt, and checking numerous heterogeneous data points: contract, construction project reference, dates, claim details, attachments, and more.
The company wanted to automate this step in order to focus teams on higher-value cases. The chosen solution relies on an AI agent connected to the shared mailbox. It classifies each email (claim or otherwise), automatically extracts data and documents from messages, checks completeness, determines which coverages may apply, and injects structured information into the management system. Data is retained for a maximum of 30 days, in line with GDPR and the CRAC agreement.
All incoming messages related to claims are redirected to a shared inbox monitored by the AI. The agent reads the subject, body, and attachments and classifies the email into three categories: new claim, additional information, or other. Out-of-scope emails are forwarded to the relevant teams.
Attachments — letters, reports, contracts, quotes — are processed using a combination of OCR and a language model. The AI extracts required fields such as policy number, site reference, date of loss, description of damage, location, and type of work, as well as declared amounts and supporting documents.
If information is missing, the agent sends an email request specifying the documents needed. Based on damage codes (category, location/part affected, cause) and dates, the AI also determines relevant coverages — such as property damage or liability — to guide processing. Code lists (101–112 for categories, 201–219 for locations/parts affected, and others) are integrated directly into the model.
Once the data is complete, the AI automatically creates the case in the internal information system via an API or an import file, attaching normalized documents. Emails and files are stored on a dedicated server for 30 days, then purged in accordance with GDPR.
Volume, turnaround time, and completeness statistics are collected to continuously improve models and processes. Everything is handled in Outlook: the AI applies tags to emails to indicate processing status — claim received, extraction in progress, awaiting additional information, case injected — enabling clear tracking and very easy deployment.
Classifies each email as a new claim, a follow-up, or out of scope, and routes it accordingly.
Combined OCR and a language model pull policy numbers, dates, and damage details straight from attachments.
Once complete, the case is created automatically in the claims system — fully structured, no re-entry.
Opening a case went from 5 minutes to a few seconds. With nearly 2,164 cases in 2024, the solution saved more than 22 working days over the year — about 11 days of productivity per handler.
The DO/CRAC email automation project illustrates how AI can transform a traditionally heavy workflow.
Before the project, three handlers sorted about 15,000 messages per year to identify 2,500 claim declarations. Each case required opening a file in the tool, creating a reference, sending an acknowledgement of receipt, and entering information. Heterogeneous attachments — scanned PDFs, photos of reports, forwarded emails — made the task harder.
The solution relies on a Microsoft Graph or SFTP connector to retrieve emails, an advanced OCR engine, an insurance-specialized language model, and a rules-based engine for business validation. The 30-day retention rule and data classification are applied by default. Automated actions are described across ten flows in the plan: initial triage, extraction, completeness checks, injection, notification, and reporting.
Some messages contain partial information, such as a missing policy number, or illegible documents. The AI then triggers a request for additional information and puts the case on hold. Human intervention is limited to disputed situations or detected anomalies.
The same architecture can be deployed for other claim types, such as liability or property damage outside construction, or other channels, such as web forms and instant messaging. Continuous learning improves document recognition and coverage detection.
In summary, this case shows that by combining intelligent triage, semantic extraction, and business validations, it is possible to automate the majority of claim declarations received by email while meeting regulatory requirements and improving team satisfaction.