Integrating Machine Learning with Robotic Process Automation

In June 2018, a report from McKinsey & Company predicted the transformational power of artificial intelligence (AI) and automation combined. AI and its subsets have since become intelligent additions in the world of automation to attempt to solve the issues companies are facing beyond problematic repetitive processes.

Machine learning (ML) is one of many forms of artificial intelligence (AI) being applied in the RPA industry today. Machine learning is the application of AI to provide systems the ability to learn and make decisions without being explicitly programmed to do so. In RPA, ML can advance robots beyond rote process execution and allow them to take on tasks that traditionally required human decision-making. Artificial intelligence capabilities can also be applied in a variety of ways to improve data integrity, add structure to unstructured and semi-structured data sets, enhance business insights, or improve automation execution.

Broadly categorized, there are three different ways that machines learn:

  • Supervised, in which the algorithm is trained on a labeled data set and given desired output values. The goal is to find specific relationships or structure in the input data that can produce the correct output. A structured data set would show the ML model information (photos, words, numbers) and the model would remember this information for later use. When presented with new data, the model compares it to examples it learned while being trained. Supervised learning is helpful for projects like classification problems, which ask the algorithm to identify input data as a member of a specific category or group.

  • Unsupervised, in which the algorithm is fed unstructured data and is not given the desired output values. Instead, these algorithms independently detect patterns and rules within data. The most common algorithms within unsupervised learning are clustering, representation learning, and density estimation.

  • Reinforcement, in which the algorithm is continuously learning and reiterating based on feedback. The best example of reinforcement learning in RPA is remote desktop automation robots, also referred to as attended robots, which are trained by human inputs as they work alongside them.

How Machine Learning is Applied in RPA

Let’s explore the various ways machine learning is being applied in automation.

Data structure and quality

There are 2.5 quintillion bytes of data created each day. Over the last two years alone 90 percent of the world’s data was generated. On average, companies with less than 1,000 employees have an average of 22 applications deployed. For enterprise companies, the average number of applications deployed jumps to 788. Given these factors, it is no surprise that the majority of companies have more data than they know what to do with. Bringing data together in logical ways to glean meaningful consumer and business insights is an ongoing challenge that is slowly being eliminated with advancements in machine learning.

Unstructured data accounts for roughly 80% of the data that companies process every day. Examples of unstructured or semi-structured data include images, audio, image-based PDFs, paper forms, text files, or customer service emails. Machine learning, and other cognitive capabilities like optical character recognition (OCR) and natural language processing (NLP), can be applied to these data to extract and structure it for use in automation. OCR engines can be used to identify, extract, and categorize data from scanned images. NLP can be trained to understand sentiment in free-form text, like customer service emails, chats, and voice inputs.

Improving Automation Execution

Machine learning algorithms can also improve the delivery of automation services. For example, algorithms can be used in computer vision to train robots how to recognize and interact with onscreen fields and components. Recursion is often used to reduce code complexity and optimize robot runtime, and machine learning models are also used for exception handling.

Task mining is another emerging application of machine learning in automation. In this instance, robots are trained to analyze daily task information gathered from employees to produce process maps and suggest processes for automation based on the highest return on investment (ROI.) This application can be a mixed bag as it requires a great deal of training to strike the appropriate balance between ROI, level of effort, and general fit for automation.

Attended Automation

Attended automation, sometimes referred to as remote desktop automation (RDA), is where robots work alongside humans to supplement their work or aide in better decision-making. Machine learning can be used to ingest data from various sources in real-time, which allows robots to help the human determine the next best step in their workflow. Machine learning can also be combined with other cognitive capabilities, like NLP, to allow robots to replicate the simpler decision-making within a human’s workflow to move even closer to achieving end-to-end automation.

Industry-specific Applications

Each industry has its own unique challenges that are being solved, in part, by applying machine learning.

  • Healthcare - The industry that is perhaps most overwhelmed by the volume of data, and the potential insights to be found within it, is healthcare. Healthcare data is a wealth of information that can enable providers and payers to proactively provide individualized care to the people they serve. However, to achieve a proactive, and even preventative, care model requires a complete understanding of the patient through data. Non-healthcare factors such as demographics, environmental, lifestyle, diet and exercise are equally important in understanding risk factors, personalizing treatment options, or aiding in disease identification and prevention. Gaining actionable insights from the myriad data required to complete the patient profile would be impossible without machine learning.

    Within the context of RPA, machine learning can aide health plans in cleaning and structuring the data they receive from providers to be used in automation. ML is also ideal for detecting anomalies in claims and identifying opportunities for process improvement within claims. For example, robots can be trained to observe factors that cause a claim to be manually processed and determine if those factors could be resolved.

  • Insurance - Insurers have already seen success by deploying machine learning to massive data sets to forecast loss and calculate risk. With regards to RPA, ML is primarily applied in two key areas: data intake, through the extraction and structuring of semi-structured or unstructured data for use in automation, and advanced analytics, through the analysis of historical data to drive better real-time decisions, eliminate leakage, and aide in forecasting.
    Primary examples include:

    • Claims intake, which can happen online, or via email, fax, or phone. ML, OCR, and NLP can be deployed to extract information from FNOL submissions, which allows robots to determine intent and route the claim accordingly.

    • Subrogation, where claim notes, diaries, or police reports can be examined to highlight or eliminate subrogation candidates. NLP can be used to analyze the text and identify related phrases that could be used to determine potential fault. For example, “a rock hit my windshield” (not a candidate), or “other driver ran stoplight” (candidate.) This “intelligent” first pass on claims would surface subrogation opportunities to ensure they won’t be missed by humans, a problem that costs the insurance industry millions every year.

    • Determining things like Cause of Loss or Bodily Injury, where robots are trained on what to look for in claims information to enable more timely and accurate registration of claims.

  • Banking and Financial Services - AI capabilities are being deployed in the banking and financial services sector in a variety of ways, from detecting fraud to chat bots to processing credit agreements en masse. Anti-Money Laundering (AML) and Know-Your-Client (KYC) are two compliance processes that are popular candidates for machine learning. Forensic accounting to detect anomalies within transactions is time-consuming, prone to error, and incredibly costly for banks. However, robots can ingest data from a variety of sources and be trained to identify the markers that indicate risk and potential fraud.

    Banks have also adopted cognitive capabilities to improve service to their customers. For example, JPMorgan Chase has implemented a virtual chat assistant to provide personalized customer support 24/7, and several competing banks have followed suit. Banks are also using ML to examine consumer behavior online to predict growth opportunities. Robots can be inserted throughout the customer service process to supplement for humans where possible, or deliver issues to humans for resolution.