IT_One: A solution for classifying corporate mail based on a large language model

The main articles are:

• Speech Recognition (Technology, Market)
• Speech technology: On the path from recognition to understanding

2025: Introducing an Automated Corporate Mail Routing Solution in Large Companies

IT ONE has developed a smart solution for automatic routing of corporate mail in large companies. A large language model (LLM) analyzes letters and forwards them to the relevant unit to the desired specialists. The solution speeds up the processing of incoming messages, reduces the number of errors due to the human factor, reduces the cost of routing letters. The company announced this on January 17, 2025.

In large companies with an extensive structure - a large flow of internal correspondence. Letters from different departments are sent to a single sorting center, and its operators send incoming requests for execution to the corresponding service. The manual process is often accompanied by errors that lead to deadlines for tasks and financial losses. In addition, companies are forced to maintain a staff of operators, which is not always easy to staff. As a result, a request for automation of corporate mail routing appeared on the market.

The company has IT_ONE developed a method for further training an open LLM to classify letters in a large holding. The test version of the model is trained on 33 thousand messages with 55 thousand attachments having different formats (documents, archives, scans, images, etc.).

The task required teaching the model to process about 500 thousand potential class combinations in the presence of 2.2 thousand unique combinations of target records in the data.

The developers further trained the model in conditions of significant imbalance of the provided data - from 1 to ~ 8,500 copies per class. In addition, model training was hampered by the complex composition of the data - message marking errors and similar class names for different objects.

According to Anton Shalin, the leading AI developer of the IT_ONE, the finished model classifies one letter in 3-12 seconds depending on the size of the message and the attachment. The classification accuracy is 90%.

The solution can be further trained on the datacet of any company and deployed inside its perimeter in a short time. At the same time, the customer receives full local control over the solution and data protection, as well as independence from external factors.

In the process of creating the solution, we used several know-how when working with LLM itself, its further training and interaction with the completed model in the final classification mode. As a result, we created a solution that significantly speeds up the routing of corporate emails, reduces the number of errors and reduces the cost of maintaining this function in the company, - comments Anton Shalin.

According to him, the automation of only one incoming case routing function allows companies to get the advantages associated with optimizing business processes and increasing the speed and quality of decision-making.