Recently, WorkFusion hosted Craig Le Clair, vice president and principal analyst at Forrester Research, in a webinar on building an automation framework for digital transformation. At the end of the discussion, there were some very strong questions around scaling automation programs. Our guest blogger and director of market enablement, Kyle Hoback, gives us his insights below.
What is the shelf life of RPA?
I expect RPA as a technology to be around for years to come, but its definition and use will likely evolve.
At its core, RPA is useful in scenarios where programmatic methods are unavailable and the user interface (UI) is the best — or only — method to retrieve and enter data. Further, RPA is useful for empowering business users to automate mundane tasks, like basic copy/paste work, which they would rather pass off to a robot.
These scenarios will continue to exist as long as legacy systems lack upgrade schedules; when people who can leverage an API are tied up on other projects; and where tasks without a large business case may still have automatability.
But the definition of RPA is already starting to change shape. Is optical character recognition (OCR) — a technology in use for many decades before RPA existed — part of the RPA definition? Are use cases with unstructured and semi-structured data that require machine learning within the RPA umbrella? Or, because capabilities such as OCR and machine learning allow RPA to be applied more broadly, are they required capabilities, thus making them part of the RPA definition?
Likely the answer to all of these is “yes,” which means the definition of RPA will expand, while the original needs of the technology will continue.
How do you think about adding machine learning to RPA?
In a digital transformation sense, the most powerful way to think about machine learning is having it perform tasks where you cannot define all the logic.
In the RPA world, many automation programs have hit a brick wall. They have included all the rules and defined all the exceptions possible, yet ROI is not yet apparent. Often, this is because the scope of the automations is too simplified, focused on low-value aspects of current operations, or that entire processes were skipped because they lacked digital, structured data. Machine learning can often be the missing link, especially on data sets that contain emails, long documents and varieties of layouts.
There can be a common misconception, though, regarding what the machine can learn. Often, people will take the approach that they’ll just “let the machine learn” and “it’s not RPA, so it must be solved by machine learning.” While machine learning is powerful, it is not complete magic.
Machine learning is best focused on defined tasks. Although you may not be able to write down all of the logic required to complete each of those tasks, you should be able to define the desired outcomes. Particularly in regulated environments such as Banking, Insurance, and Healthcare, it is better to direct the machine learning toward determining particular data fields rather than whatever it thinks is best.
Machine learning often will open up stalled RPA use cases to further automation, or pull previously out-of-scope use cases back into scope.
How do you do machine learning without data scientists?
At WorkFusion, we love our data scientists and are constantly amazed at what they can do. However, we don’t think our customers should require a data scientist every time they want to do a machine learning task.
The core of our machine learning approach is our Process AutoML capability and our ML SDK. Why define your algorithm upfront, when you can experiment with many? Why has a person to think through all the relevant features of the data, when the system can identify these for you? Why require Ph.D-level skills when you can positively impact production without them?
With our Process AutoML and ML SDK approach, we have empowered technical RPA developers to apply data science without being a data scientist — yet business users also can add machine learning without even knowing how to code. We have simplified the automation well enough that it can achieve value for customers in shorter timelines, yet still give the opportunity for high-powered analysis in situations where that is required.
There are many applications where data science is required, but just because machine learning is required doesn’t mean you can’t get good-enough results with the right tools and an engaged user.
Also published on Medium.