Mining / Industrial minerals
Sibelco: Classifying purchase orders in customer emails and attachments
- Company
- Sibelco
- Country
- Belgium
- Adoption stage
- In operation
- Source published
- Date basis
- The date the source was published. It can differ from the date adoption started.
- How the source was checked
- Read the full source text
The work problem
Sibelco's central mailbox received purchase orders alongside transport requests, complaints and invoices. Customer-specific product names and codes meant Customer Service had to identify actual purchase orders first. Supplier CROPLAND explains that some accumulated business rules existed only in employees' experience, requiring joint work to document exceptions and decision paths.
Technology and data
CROPLAND and Sibelco built an order intake agent using an OpenAI language model. Incoming emails become Salesforce cases, and attachments, including technical specification sheets, are analysed individually. A second model filters irrelevant files and a check prevents duplicate processing. Uncertain documents go to an employee for handling. The agent integrates with existing systems to change classification first; direct actions in the ERP remain a subsequent phase.
Results
CROPLAND reports that classification accuracy rose from 80 to 85% to above 95%, saving about 30 minutes of manual work per order. It says Customer Service can redirect time previously spent identifying documents to order processing, follow-up and quality control.
Limits and open questions
The source does not disclose the evaluation sample, measurement period or share of documents handed to staff. It does not establish that the next phase of direct ERP actions is already operational.
Sources
Compiled from public sources. These are not results from ATF Works customers.