Classifying nonconformance reports

In a manufacturing company, nonconformances are recorded in several ways: an operator fills in a Microsoft Forms questionnaire or a paper form at the end of the shift, an inspector writes an e-mail about a rejected batch, a customer sends a complaint, a foreman notes it in the shift log in SharePoint, and some entries end up in the quality module of Business Central, Odoo or Monitor ERP as a free-text comment. The quality manager gathers all of it into an Excel sheet where everyone describes the problem in their own words, so the same cause goes under five different names. The monthly management report takes days to prepare, and corrective actions start when the problem has already cost several batches.

We implement ChatGPT through the API with a Power Automate flow. The flow collects nonconformance records from Microsoft Forms, SharePoint shift logs, a designated mailbox and, if you use one, an export from the ERP quality module into a single SharePoint list. The model classifies each record according to your structure: product or product group from the Business Central or Odoo item list, operation or workstation, cause category (material, machine, method, people, measurement), consequence (rework, scrap, delay, customer complaint) and severity. The free-text description stays, but uniform categories appear next to it, so the records become comparable. Unclear records are flagged for the quality engineer to clarify. Every Monday the quality manager receives a summary in Teams: which nonconformances recur, in which products and operations, how that changed since last week, with links to the specific records. People make the decisions on corrective and preventive actions and register them in their own quality system, while the solution only shows where to look. The model runs on the OpenAI platform without training on your data, and if product information is confidential, we implement the same solution with a private model in MDP CLOUD's GPU cloud in Lithuania (EU).