Geminor

About Geminor

Geminor is a pan-European waste resource management company, operating in several European countries with an annual volume of several million tonnes. This entails a high volume of incoming invoices related to completed assignments, which must be verified and matched to the correct assignment before they can be approved.

The challenge

Geminor spent a significant amount of time compiling and matching completed work with incoming invoices. With many parallel assignments and often detailed invoices, this was a time-consuming manual task. The company already had the information in its own system. The challenge was to automate the actual matching between invoice and assignment.

The solution

Miles built a solution using Anthropic's Claude APIs, based on a combination of Sonnet and Haiku, to automate the matching between incoming invoice lines and registered assignments in Geminor's internal system.

Before the solution was put into production, the team conducted thorough testing and benchmarking: different models and prompts were tested against invoices that had already been handled manually, and the results were compared against the ground truth. Many of the AI's assessments were not only correct but, in practice, more precise than the original manual records, because the solution considers multiple factors simultaneously, such as matching the correct assignment to the correct price at the correct time.

Results

In testing/benchmarking: The solution was able to automatically handle around 70% of all incoming invoices (for the defined, largest customer type).

After nearly 2 months in production: The proportion handled fully automatically is around 20%. The reason is a volatile market situation where many invoices have included extra cost lines (such as fuel surcharges) that are not entered as cost elements in the system. When the price discrepancy exceeds the tolerance level set by Geminor, the solution automatically stops the process and sends the case for manual review – exactly as designed.

In short: Less time is spent on manual processing of incoming invoices, and the system is built with a built-in safety mechanism that ensures discrepancies are caught manually rather than being incorrectly approved.

‍