The Initial Problem
An independent financial advisory firm managing mandates for Geneva and Vaud SMEs faced a classic but costly problem: tracking unpaid invoices. With about a hundred invoices issued monthly, the collection process required an administrative staff member one day per week. Standardized reminders sent in batches lacked personalization and effectiveness, with some clients systematically paying only after the third reminder.
The average payment delay reached 52 days while contractual terms stipulated 30 days net. This situation impacted the firm's cash flow and required a bank credit line to smooth out timing gaps. The firm sought to automate this process while maintaining personalized and professional communication, suited to the trust relationship with its clients.
The Developed Solution
We designed an intelligent automation system using Power Automate, AI Builder, and Azure OpenAI Service. The architecture relies on several interconnected flows.
A main flow runs daily and queries the Sage accounting system via its REST API to identify all outstanding invoices, classified by age and amount. Invoices are categorized as slight delay (up to 15 days), moderate delay (16 to 45 days), and significant delay (over 45 days). For each invoice, the flow retrieves the client interaction history from the CRM (calls, emails, meetings) as well as previous reminders.
This contextual information is then sent to a GPT-4o model deployed in Azure OpenAI Service with instructions to generate a personalized reminder email. The prompt includes parameters such as client name, mandate nature, amount owed, delay duration, relationship history, and examples of tones to adopt based on delay category.
For a slight delay, the tone is courteous and assumes a simple oversight. For a moderate delay, the message reminds of contractual terms with firmness but goodwill. For a significant delay, the text mentions potential consequences such as service suspension or transfer to a collection service, while proposing a payment plan if necessary.
The model also generates an optimized subject line for each email. Before sending, a human can validate generated emails via Power Automate approval, or let the system automatically send first-category reminders. Emails are sent from Outlook with automatic tracking, and any client response creates a task assigned to the accounting manager.
The Measured Results
After six months of use, the impacts are significant. Average payment delay dropped from 52 to 31 days, a 40% reduction. Time devoted to reminders fell by 85%, going from one day per week to one hour of validating complex cases. The response rate to reminder emails increased by 60%, with clients appreciating personalized messages rather than standardized letters.
The firm also noted improved client relationships, with several mandators spontaneously mentioning the quality and courtesy of communications. The system enabled earlier identification of clients in financial difficulty, allowing payment terms to be adjusted before the situation deteriorated.
Financially, the 21-day payment delay reduction freed up approximately 180,000 CHF in cash flow, allowing the firm to reduce its credit line and save several thousand francs in annual interest.
Technical Aspects
Integration with Sage required creating a custom API layer since Sage 50 doesn't natively have a modern REST API. We developed an Azure Function that securely queries Sage's SQL database and exposes accounting data via OData-compliant REST endpoints.
Calls to Azure OpenAI Service are controlled by Power Platform DLP policies to ensure no sensitive data transits to unapproved services. The GPT-4o model used is deployed in private mode with commitment not to use data for training.
The total monthly system cost, including Azure and Power Automate Premium consumption, amounts to about 150 CHF, equivalent to two hours of salary cost saved each week. Return on investment is achieved in less than two months.
Future Developments
The firm plans to extend this automation to other processes, notably automatic qualification of incoming leads and generation of personalized mandate progress reports. This initial success demonstrated the value of AI applied to repetitive administrative tasks that still require some degree of personalization.
Conclusion
Automating reminders perfectly illustrates how AI can transform a time-consuming process into a competitive advantage. By combining the text generation power of large language models with Power Automate orchestration, we created a system that not only saves time but also improves the quality of client interactions.