Duplicate Payment Analytic with AI

70% time savings through the use of AI

In this blog post, we present our improved duplicate payment analytic. It now uses artificial intelligence to improve the quality of the results.

In the mentioned analytic, we put great emphasis on not overseeing any true duplicate payment, that’s why a liberal search strategy is necessary. This has the disadvantage that many so-called false positives are found. These are found duplicate payments, which actually turn out to be none.

Various attempts to filter out false positives based on rules have failed because the defined rules were too static. An example: If one would like to exclude invoices of recurring payments (e.g. for rent or garbage fees) from the result, it would be reasonable to apply the following rule: If for several invoices the vendor and the amounts are identical and the dates (in the document date) are either 30 or 31 days apart, then it is not a duplicate payment. In short:

Vendor identical & amounts identical & document dates 30 or 31 days apart -> No duplicate payment­­

The problem: In general, real duplicate payments exist that meet these conditions. The example illustrates how difficult it is to define generally valid rules. Whenever defining rules causes problems, it makes sense to think about using an AI.

The result of our analytic so far has shown a ratio of true duplicate payments to false positives of

­1 to 10

So on average only every 11th finding is an actual duplicate payment.

In order to provide our customers a better analytic, we have added an AI to the analytics. It reduces the number of false positives by more than 70%. This results in a ratio of

1 to 3

On average, every 4th finding is a direct hit. Thus, we save our customers more than 70% of work, because most of the manual reviewing is eliminated.

The mentioned AI learns on a training pool which consists of previously found and labeled duplicate payments and is located exclusively at the customer’s site. That is, each customer of ours trains its own AI on its own data. The mentioned pool is continuously growing, so that the training takes place on an increasingly larger amount of data. As a result, we will be able to improve the stated ratio of 1 to 3 once again in the future and thus offer an even more comfortable result, because a decreasing number of false positives will have to be reviewed. 

Are you interested in an analytic that reliably identifies double and multiple payments? Then please feel free to contact us.