We employed our own state-of-the-art Deep Learning based anomaly detection models to infer anomalies in past transactions or predict potential future anomalies like duplicate payments.


Being a large corporation with thousands of suppliers and hundreds of thousands of payments every year, Societe Generale European Business Services (SGEBS) asked Lummetry to help them automate and make the work of their invoice payments department more efficient.


About 1% of all the payments value are lost due to operational error like duplicate payments. Such duplicate payments occur due to alteration of the determinants of each invoice: Vendor name, invoice number, invoice amount, invoice date. For example, due to errors in OCR, the invoice number was altered and instead of one letter (“l”) we might have a number (1). Therefore, the invoice number is seen as different by the payments department and another payment is done. Example can continue.

For the high value of payments made by SG EBS, this 1 percent means a lot of money. As a consequence, they’ve built a department focused on verifying manually all the potential duplicated invoices provided by their existing invoice monitoring tools. Unfortunately, such traditional tools return an immense number of invoices that would need to be analyzed manually by humans (it looks for combinations of 3 invoice criteria like same invoice amount, same vendor name, same date). Moreover, the false positive rate of these returned invoices is about 98%. It means that for each 100 invoices analyzed only 2 of them prove to be duplicated after all. In conclusion, the huge effort done by all the people in this department is extremely inefficient and costly. Our aim was to reduce the human effort by at least 40% using Machine Learning.


We created complex Deep Learning models capable of analyzing all combinations of invoices and determining the similarities between invoices in order to provide the probability that 2 invoices might be duplicated. We even created ensemble models based on 3 neural graphs and designed a voting system based on which each neural graph can vote for each pair of invoices if it considers it to be a duplicated one or not, with different weights.


Our solution managed to reduce the human effort with an impressive rate of 60%.

In conclusion, we provided a successful solution capable to reduce the human costs by 60% and increase the satisfaction of the people from the invoice payment department as machines helped them make their work more efficient.