Extending Robotic Process Automation Capabilities with Machine Learning

By: Gautam Bhattacharya, Partner & Leader Analytics, EY and Dr. Rohit Lotlikar, Director & Data Science SME, EY

As Information Technology (IT) advances, manual or labor-intensive work gets progressively automated. However, much of manual labor remains without being automated because it is difficult to get disparate enterprise applications to talk to each other while keeping the end-to-end implementation flexible enough to deal with changes in business requirements. However, Robotic Process Automation (RPA) changes this.

Robotic Process Automation (RPA) refers to process automation where computer software drives existing applications through their standard user interface in the same way that a human does, in other words it orchestrates applications via presentation layer integration in contrast to integrating via the application back-end.

RPA is especially suited for situations where the process logic or data formats have not reached a level of stability or standardization that is likely to make backend-based integration feasible or worthwhile.

While the majority of back-office business processes have clear rule-driven nature making them candidates for RPA, we frequently encounter processes either involving free text and/or judgment calls. If the proportion of cases involving free text and/or judgment calls is significant, the use of text analytics and machine learning technologies should be considered.

The role of machine learning in Business Processes Automation

Machine Learning is the science of learning from data. It involves processing of historical data records for discovering patterns in the data, and describing the relationship in the form of rules. These rules can be then programed into RPA to automate the decision making. The machine learning algorithm must be provided with the historical data consisting of decision calls and the snapshot of the relevant data fields at the time the decision was taken.

In many situations, the experience and intuition of the human will impact the decision. In such scenarios these algorithms will, in effect, learn the “wisdom-of-the-crowd”, which describes combined experience and intuition of the humans.

How can I apply machine learning?

In the context of automation, there are two modes in which machine learning can function:

· Decision support: Here the focus is on improving the quality of decision making while the final decision is taken by the human, considering the guidance from the algorithm. An example of this scenario is guided cross-selling where the product recommendation is made by algorithms, based on discovering how demographic and behavioral attributes of the customer relate to the propensity to purchase a product.
· Decision automation: Here the focus is on productivity gains and makes sense for high volume decisions. For this to work, the algorithm must have a high level of accuracy equal to or better than that of a human on the cases that are processed by it.

Deciding whether to apply machine learning

For determining whether machine learning is feasible and can add value, we first check on whether it is a “local” decision or a “global” decision.

· If a decision on a case involves looking beyond the case at hand to look for trends/patterns across the cases (i.e., a “global” decision), then using machine learning is essential. Common use cases include credit approval and fraud detection.

· If a decision on a case can be made by looking at the data associated with that case alone (a “local” decision), then it is easier to start with the reverse question and determine if and how much of the cases can be handled by SME defined rules. If a significant proportion (in the sense of FTE costs) of cases cannot be handled by SME defined rules, then that process makes a good candidate for machine learning. If the rules need to be tuned very finely, which would happen if there are a large number of factors or it involves noisy free text, it becomes difficult for a human to encode the rules accurately and machine learning is likely to be a good prospect.

When a process is very rule driven, automation will ensure consistency, since human error is eliminated. The machine learning algorithms output a set of rules that are then employed in the process, and therefore, machine learning ensures consistency as well. When there are process or policy changes, in how the decisions need to be taken, or new situations are encountered, these must be monitored, i.e., the response of the model should be observed.

Considerations when machine learning is used with RPA

Since machine learning algorithms infer the logic by observing human responses over a considerable number of cases, they are most ideal when the underlying logic is “stable”. When there is a change to the underlying logic, the algorithm needs to observe a sufficient number of human responses under the new logic before it can take over from the human. The RPA implementation must include a checkpoint to demarcate those cases to which the new logic must be applied and reroute them to a human until the algorithm has been adequately “trained”.

When applying machine learning based RPA, it is prudent to be concerned about the correctness of the response that the algorithm would output when it encounters situations that are unlike anything it has observed in the past. This can happen even when machine learning is not employed; say if the RPA programmer has not strictly specified the scope outside of which a robot should throw an exception. In either case, the system needs to be “stress-tested” on simulated examples, its response observed and compared with the desired response.

In recent years, RPA has evoked significant amount of interest in the Shared Services Delivery and BPO communities, and the production deployments of this technology are growing at an accelerating pace. RPA vendors and partners have been fairly clear in stating that this is a technology for automating standardized, rules-driven, non-judgmental processes. At the same time, they are alerting customers to the next wave of automation — cognitive RPA — which combines RPA with machine learning technology to extend the boundaries of automation even further. While machine learning technology has been around for a while, it has been hard to deploy in service centers due to the need for backend access to the data. The developments in RPA technology has made machine learning technology more amenable to use in shared services centers thanks to the ability of RPA to access data and drive the process via application user interfaces.