Automating Manual Processes
Manual data and invoice reviews are a daily reality for many organizations. The time spent reviewing text fields and categorizing entries for accounting purposes can easily add up to dozens of working hours every month.
Insurance company TM was no exception. Each month, TM specialists had to review electronic invoices containing around 5,000 payment lines in a complex format — reading them line by line and manually entering the information into their system. This time-consuming process not only created significant workload, but also increased the risk of errors. Altogether, the task required 80–100 working hours every single month.
At the same time, TM was encouraging more vendors to submit invoices electronically, meaning the volume — and the workload — was expected to grow over time.
Until recently, there was no practical way to streamline this cumbersome workflow, and automation seemed nearly impossible. However, advances in artificial intelligence made it possible to simplify the process and build a reliable automation solution using modern technology.
Custom-Trained AI Model
Data security and reliability were top priorities throughout the project. A large language model (LLM) was trained exclusively on TM’s data and is used only within their environment. The model is fully closed and hosted internally at TM, ensuring that sensitive data never leaves their infrastructure. Despite being a tailored solution, it remains cost-efficient to operate.
The solution reads each entry, analyzes it, and classifies it into cost categories such as labor, materials, or services. The information is then automatically entered into TM’s claims system.
Over time, the model continues to learn and improve. If data is incomplete or unclear, it can flag entries for review. It also learns from corrections, allowing it to quickly and reliably recognize new or previously unseen data.
Close Collaboration
The project was developed through close collaboration between TM and Kolibri. Alongside the AI development, a dedicated web interface was built to allow specialists to monitor the data and make adjustments when needed. The interface was designed specifically around the needs of TM’s specialists.
Having visibility into the process helped build confidence early on. Since the AI solution was introduced, trust in the system has grown significantly as specialists saw firsthand how reliable the model is.
99% Accuracy and Measurable Results
Since implementation, the AI solution has achieved 99% accuracy in data analysis and classification. While the system performed reliably from the start, it has continued to improve over time.
The data generated from the process has also delivered meaningful insights and led to significant operational savings for TM.
AI-driven automation doesn’t have to be complex — as long as it’s built in close collaboration with stakeholders and domain experts. In this case, people guided the development, with AI serving as a powerful tool to support them — not replace them.