Advanced Multilingual Natural Language Processing in Industrial Digitalization: A Case Study of Autonomous Document Classification and ERP Integration in Manufacturing
TL;DR
The design, implementation and results obtained through maAIGENT, an Explainable AI solution developed for a manufacturing organization, with the role of classifying enterprise communications and supporting their integration with the ERP system, show a reduction of the average processing time and an estimated saving of 1,650 hours per year.
Abstract
Abstract The purpose of this paper is to present the design, implementation and results obtained through maAIGENT, an Explainable AI solution developed for a manufacturing organization, with the role of classifying enterprise communications and supporting their integration with the ERP system. In the analyzed company, the treatment of sales related documents was made manually, although these documents represented the starting point of several important operational processes. According to the internal analysis, approximately 11,000 documents are processed every year, and each of them required, before the implementation of the system, about 10 minutes of human work for identification, classification and ERP registration. Thus, the manual process consumed around 1,833 man-hours per year and had a measured error rate of 4.6%, errors that were reflected in production delays, missed commercial opportunities and administrative frictions. maAIGENT was built as a hybrid architecture, in order to combine the reasoning capacity of a LLM with the security requirements of an industrial client. The local application is deployed inside the customer’s LAN and performs the operations that are connected with sensitive data and communication with the ERP system. The cognitive component of the system uses Azure OpenAI GPT-4o, accessed through a private subscription, while only the sanitized textual payload is sent outside the local environment. In this way, the original emails, attachments and operational records remain under the control of the organization. An important element of the implementation is represented by the explainability of the classification. For each document, the model returns not only the class, but also a short rationale in Romanian language, based on the content of the email and on the text extracted from attachments. This explanation was used to support faster human validation and to make the automatic decision more transparent for the operators. The results obtained after the stabilization of the system show a reduction of the average processing time from about 10 minutes to 1–2 minutes per document, which represents approximately 90% reduction and an estimated saving of 1,650 hours per year. This reduction includes both the automatic classification and the transmission, through API, of the routing decision to the ERP system, replacing the previous end-to-end manual workflow. In the validation made on 650 operational documents, maAIGENT obtained 98.62% model-only accuracy and a macro-F1 score of 0.986, compared with Human-in-the-Loop validated labels, while human corrections were necessary in 1.38% of the cases.