Black Box Or White Box? Machine Learning For Finance And Risk Processes

Birgit Starmanns

Consider two different approaches to logic and system-driven decisions, commonly known as a white box and a black box approach.

With a white box approach, a term that comes from software testing, the logic and steps are known and can be traced. The logic may be mapped in flowcharts, rules, or code; most importantly, the logic and the steps of a program are very clear.

Enter the black box approach. In this case, the inner workings are not known, for example, to a software tester who sees only the inputs and the outputs. In terms of automation, this approach is also relevant when a process is too complicated to allow rules to be defined. Instead, the internal mechanism is hidden; and the subject here is that an artificial intelligence system makes specific inferences.

By extension, the argument can be made that this is similar to the way that humans learn. We make decisions every day, in general without consulting a list of rules one by one; we have learned through experience. And additional experiences influence the decisions that we make in the future.

Now let’s apply this to business systems, specifically finance systems. First, let’s look at different types of automation. As finance organizations undergo digital transformation, the need to become more efficient through automation is key, to allow finance departments to reduce errors, and to improve their speed of processing and the financial close. Such efficiencies also allow finance teams to transform their own organizations, to focus on more strategic tasks instead of handling a myriad of exceptions on a manual basis.

Types of automation

Let’s first take a look at the different types of automation, and why machine learning is different from other types of logic.

  • Rules engine. In this scenario, business users define specific rules. Sometimes these are set up in configuration; sometimes they need to be coded by an IT department. These rules are then executed on a periodic basis, most often during a period-end close. However, rules engines often become less effective over time, because they are rarely revisited. A company may expand into a new customer base for which different rules are relevant. If the rules are not adjusted – and they rarely are – finance departments must manually deal with more and more exceptions, especially as transaction volumes grow.
  • Robotic process automation. Robotics is essentially automating a manual task in a consistent way, similar to writing a script. Examples of robotic processes could be loading data into a system, where the same fields are populated, often from an uploaded file. It can also be thought of as writing a macro in Excel to execute a certain set of tasks – tasks that never change, such as manipulating data or generating a graph.
  • Machine learning. Machine learning can identify patterns in knowledge-intensive processes, without explicitly defining the patterns by rules or macros. The machine learning engine learns from historical transactions during an initial training period. It then continues to learn as finance teams make decisions based on exceptions; think of this as continuing education. Therefore, as an organization defines new business models, and additional exceptions are generated, the actions taken by finance teams on those exceptions allow the machine learning engine to incorporate these decisions into its learning. Since machine learning is based on an algorithm, it does not actually generate rules, but continues to learn – yes, a bit of a black box approach, and again, similar to the way that humans learn.

Finance applications that leverage machine learning

With the effectiveness of machine learning as part of SAP Leonardo, finance and risk applications are already leveraging machine learning in several scenarios, and the number continues to grow. These include solutions supporting:

  • Cash application. A cloud solution that learns from historical transactions of applying customer payments to invoices for open accounts receivable items. Based on the preferred tolerance level, cash can be applied automatically, leaving finance teams free to deal only with the most complex exceptions.
  • Intelligent goods receipt/invoice receipt reconciliation. A cloud application that learns from historical data and decisions of the finance team in handling exceptions to propose decisions on clearing differences.
  • Business integrity screening. A solution that employs a hybrid set of rules plus predictive analysis. Predictive analysis leverages machine learning to scale across thousands of predictive models to find patterns related to fraud, compliance failures, and other exceptions, thus reducing related financial losses.

Learn more

If you are attending SAPPHIRE NOW and ASUG Annual Conference, visit these sessions for additional information and interactions:

For additional information, you can also visit these sites:

And read this blog: The Evolution of Modern Receivables Management with Machine Learning

Follow SAP Finance online: @SAPFinance (Twitter) | LinkedIn | Facebook | YouTube


Birgit Starmanns

About Birgit Starmanns

Birgit Starmanns is a senior director in the Global Center of Excellence for Finance and Risk at SAP. She is focused on the go-to-market for new solutions, and the business benefits they can bring to organizations, such as cloud solutions for finance and applications based on SAP Leonardo technologies such as machine learning. Birgit has over 27 years of experience across solution marketing, solution management, and strategic customer communities. Prior to SAP, she was a principal in management consulting organizations, including Price Waterhouse and several boutique firms. Birgit holds a BA and MBA from the College of William and Mary. She is the author of many articles for the Financials Expert, the coauthor of the SAP Press book Accelerated Financial Closing with SAP, and the SAP Labs guidebook Product Costing Scenarios Made Easy. In addition, she is a regular presenter at various SAP events.