Machine Learning in Concur Mobile: How Does it Work?

Many SAP Concur users will recognise the ExpenseIt service or app as one that, in the background, automatically extracts information such as amount, date, and expense type from your uploaded receipt. To perform the task, ExpenseIt uses an array of machine learning (ML) algorithms, but in the latest release of Concur Mobile, ExpenseIt users may notice something different when they capture a receipt.

SAP Concur has led the way in developing ML algorithms to reduce time on task and help minimise the amount of user entry needed to create an expense report. Cloud computing is attractive for many reasons, such as seamless updates of the service. But on-device, or Edge computing, offers the benefit of real-time processing.

Throughout the last year we’ve been working on a new algorithm which is able to predict (or extract) the total amount from a receipt image, and it runs – on device – in the SAP Concur mobile app, which will reduce turnaround time for users and help with extracting handwritten total amounts as seen below.

A prototype of the new algorithm in the SAP Concur mobile app

 

More data helps, and the data scientists at SAP Concur are fortunate to have the largest data set in the T&E (travel management and expense) industry to pull from, with more than 51% of the total T&E market share according to IDC. In the third quarter of 2018 we processed more than £87 billion in customer expenses.

 

Millions of images used for training

The new algorithm for extracting the amount only needs the image as input and has been trained using millions of receipt images. When you use the app you will quickly see the result from the algorithm on the screen and have the option to correct any inaccuracies right there. ML isn’t a 100% accurate process, and blur, shadows, and other factors can occasionally result in errors. Overall, users will have a better experience with the new functionality inside the SAP Concur mobile app.

 

From text to images

The ExpenseIt machine learning system works by first sending the images uploaded though a process called optical character recognition (OCR), which extracts the text from the receipt images. All token extraction algorithms developed to-date by SAP Concur have used only the text as input, but the new amount extraction algorithm works by processing the image directly.

 

A solution for handwritten receipts

The new algorithm can deal with printed text receipts as well as receipts containing handwriting. This really helps with some meal receipts, which often contain a handwritten total amount when a tip is added to the bill. This class of receipts has been notoriously difficult, since the text extracted sometimes didn’t contain the handwritten portion, and as a result the algorithm often couldn’t extract the amount correctly. We hope that you will find this new algorithm very useful.

 

ExpenseIt Pro standalone app sunset

The ExpenseIt Pro standalone mobile app will not receive the on-device algorithm and was sunsetted December 31, 2018. ExpenseIt users are now able to access it inside the SAP Concur mobile app as a value-add service that turns receipts into expense line items and sends them directly into Concur Expense.

ExpenseIt will continue to process receipts in the background and automatically extract information such as amount, date, and expense type, leveraging the best of ML and human auditors while helping minimise the level of data entry by the user.

 

Data science at SAP Concur

Data science, machine learning, and deep learning are rapidly changing the software industry, and at SAP Concur we’re taking advantage of the opportunities this presents. We have a team of brilliant engineers working on improving the software and services we deliver every day.  We hope you will enjoy the new Concur Mobile ExpenseIt experience and if you have comments please let us know next time you submit an expense report!

 

Download the whitepaper to learn more about why you should be excited about ML and artificial intelligence. 

Loading next article