Extract invoice information with AI
Our company in short
Basware provides services in e-invoicing, accounts payable, and financing across 175 countries worldwide. As a leader in the Accounts Payable automation sector, Basware expertly manages the receipt of invoices from suppliers, checks invoice details against purchase orders, stores invoices in the accounting system, and ensures payments are scheduled by their due dates. Annually, Basware processes over 170 million invoices.
Our business problem and machine learning
The SmartPDF service extracts invoice information from PDFs sent to Basware via email and converts them into the e-invoice standard format. Initially, SmartPDF required manual template creation and validation to extract data from PDFs, which proved costly and unscalable. To address this, Basware implemented SmartPDF AI, utilizing supervised learning to parse and process invoices automatically, eliminating the need for manual intervention. The goal is to replace traditional templates and manual validation entirely with AI.
This AI service operates on the AWS cloud, employing running SmartPDF AI to extract text from the invoices. It utilizes supervised machine learning to predict field values from the extracted text, trained with 30 million samples, covering 800,000 invoice types, and utilizing 2TB of raw data. Maintaining high accuracy in this process is critically important.
Our research and solutions
The problem: Human errors in training data. If data is wrong, AI learns wrong. Objective: Enhance data quality.
Phase1:
A committee of six AI models employs a voting system for robust decision-making. Features include anomaly detection and the use of hierarchical or pivot-based clustering for efficient invoice extraction. Additionally, a grammar-based invoice signature creation method is implemented, designed to be resilient to OCR inaccuracies. This approach has successfully reduced errors by approximately 30%. While accuracy for crucial fields is satisfactory, it is not uniformly high across all fields.
Phase2:
Machine learning-based embeddings are automatically discovered in an unsupervised manner through a fully ML-based solution. This system autonomously identifies the signatures of various fields without relying on heuristics or human-labeled training data. While the model achieves high accuracy across all fields, there is a need to improve its coverage.
Future direction
OpenAI's GPT-4 offers significant potential for enhancing invoice data extraction by leveraging its advanced natural language understanding and generation capabilities. This could represent an opportunity rather than a threat, providing more sophisticated tools for recognizing and processing complex invoice information, particularly at the line-item level.
A study initiated to explore alternative methods for extracting invoice data with GPT-4 could involve comparing results from various engines and potentially integrating them to achieve superior recognition results. Key factors to consider in this exploration would include the costs associated with implementing and maintaining such technology, its performance in real-world scenarios, and the accuracy of the data extraction compared to existing methods.