The Document Overload Challenge
A Geneva wealth management firm managing assets for wealthy families and foundations received dozens of complex legal and financial documents daily: annual fund reports, investment contracts, tax legal notes, regulatory documents, and banking correspondence. Wealth managers had to read and analyze these documents to identify relevant information for their clients, an extremely time-consuming process.
A manager could spend up to 40% of their time simply reading documents, non-billable time diverted from value-added activities like advising and client relations. Moreover, some important documents sometimes went unnoticed in the daily flow, creating compliance risks or missed opportunities.
The firm sought a solution to automatically generate structured and reliable summaries of these documents while preserving absolute confidentiality of client data subject to Swiss banking secrecy.
The Document Intelligence Solution
We developed an automatic summarization system based on Azure AI Document Intelligence and Azure OpenAI Service. The technical architecture combines several Azure services hosted in the Switzerland North region to guarantee data sovereignty.
The process begins when a document is placed in a dedicated SharePoint library. A Power Automate flow detects the new document and determines its type using Azure AI Document Intelligence, which identifies whether it's a contract, financial report, legal note, or other. The document is then sent to a GPT-4 Turbo model specifically configured for legal and financial document synthesis.
The system prompt was meticulously designed with managers to extract critical information according to document type. For a fund report, the summary includes performance, main risks, strategy changes, and notable events. For a contract, parties, duration, termination conditions, main obligations, and legal points of attention are extracted. For a tax note, the summary highlights implications for wealth structures, action deadlines, and recommendations.
The model generates two summary levels: an executive summary of 150-200 words for quick reading, and a detailed summary of 500-800 words structured by thematic sections with extraction of key passages from the original document. These summaries are automatically stored as SharePoint document metadata and indexed in Azure AI Search for later semantic search.
A notification email is sent to the concerned manager with the executive summary and a link to the full document and detailed summary. The manager can quickly decide whether the document requires thorough reading or if the summary suffices.
Operational Benefits
After nine months of deployment, the firm measures spectacular gains. Average time devoted to document review decreased by 70%, with managers now processing in one hour what previously required three to four hours. Document processing capacity increased tenfold, enabling coverage of more information sources without increasing headcount.
Analysis quality also improved because managers can now read more relevant documents instead of limiting themselves to those they have time to process. The system identified several tax and investment opportunities that would probably have been missed in the daily flow.
Clients also appreciate receiving regular summaries of news impacting their wealth, a differentiating service the firm can now offer thanks to freed-up time. Client satisfaction measured by NPS increased by 12 points since introducing the service.
Reliability and Quality Control
Summary accuracy is closely monitored. Each week, a random sample of summaries is manually verified by a senior lawyer. The fidelity rate to the source document reaches 97%, with the 3% deviations being mainly minor omissions rather than factual errors. The system is configured with a confidence threshold: if the model estimates the document is too ambiguous or technical, it flags the summary for human validation before distribution.
A complete audit log traces all processed documents, generated summaries, and user actions, meeting FINMA compliance requirements for management companies.
Architecture and Security
The architecture is entirely private and sovereign. All Azure services are deployed with private endpoints via Azure Private Link, eliminating all Internet exposure. Data never leaves Swiss datacenters. The GPT-4 Turbo model used is deployed with Azure OpenAI Enterprise contractual guarantees that client data is never used for training third-party models.
Access is controlled by Azure AD with mandatory multi-factor authentication and conditional access policies based on geolocation and device status. Each access is logged in Azure Monitor for audit.
The total monthly system cost represents approximately 800 CHF, including Azure resources, Power Automate Premium licenses, and houle support. This cost is quickly amortized by productivity gains equivalent to several days of work saved each month.
Conclusion
This solution illustrates how generative AI can transform a profession with a strong intellectual component like wealth management. By automating preliminary document reading, we enable professionals to focus on analysis, strategy, and human relationships with their clients, where their true added value lies.