Visma’s AI team is transforming document processing in Europe, handling over 18 million documents monthly with specialized models. Initially reliant on traditional deep learning, the team pivoted to transformer-based technology around 2020, positioning Visma ahead of trends like ChatGPT. Their flagship product, a document extraction engine, efficiently processes invoices and receipts across multiple languages, extracting critical data directly into workflows.
The system employs about 50 specialized models, ensuring optimal performance through modularity. Visma prioritizes high-quality data, training large general models with documents from over 100 countries. This enables efficient model fine-tuning while maintaining a low error rate of 1-3%.
Despite challenges such as language and format variations, Visma’s AI is evolving towards greater autonomy, enhancing its ability to provide real-time insights. As businesses seek practical AI solutions, Visma’s journey illustrates the potential for intelligent automation, making processes more efficient and supporting informed decision-making.
Source link