By Jen Schellinck
While Robotic Process Automation (RPA) proponents tend to emphasize that RPA is not artificial intelligence, and most AI/ML practitioners would agree, it is nonetheless very possible that RPA will act as a gateway both to new applications of artificial intelligence and to new forms of artificial intelligence. In this article I will discuss exactly why RPA has the potential to be such a powerful force in the growing adoption and development of AI technologies.
RPA supports the automation of routine computer-related tasks, freeing people to carry out the more complex activities that make best use of their skills and expertise. The technologies behind Robotic Process Automation (RPA) have been available for some time but, like many technologies, these relevant component technologies (e.g. screen-scraping, work process automation) first developed and matured steadily in the background for a few decades before combining to form what in the early 2000’s came to be called ‘Robotic Process Automation’. At that point RPA began what promises to be a fairly meteoric rise in popularity, aided by the increasing role that digital technologies play in the workplace and a growing societal appreciation for the potential of data science, artificial intelligence and other data driven technologies.
Roughly speaking, RPA works by connecting a special piece of computer software to the front end of either another piece of software (e.g. a word processor or spreadsheet) or an entire computer system (e.g. an operating system). By intercepting the user interface signals that are sent out by the software or system, and then responding with signals through that same user interface, the RPA software can behave in almost exactly the same way a human does – clicking windows, opening files and copying and pasting information from one place to another. Supported by the RPA software, organizations first study or even directly record the human actions required to carry out the digital elements of an activity, and then automate these elements via RPA.
To understand why this represents such substantial potential for AI, it helps to first have in mind a picture of how we currently think human intelligence works (although though this, too, is constantly evolving). If we think of organic intelligence in terms of layers, we can call the environment the first layer of intelligence. This might seem odd, but there is an entire area of research (situated cognition) devoted to the study of this – think of how we might drop crumbs on a trail to find our way back to our house in the forest, thus effectively placing our ‘intelligence’ in the forest. The second layer is the sensory layer, where our bodies interact with the environment, and receive sensory information from that environment. Very importantly, as it relates to RPA, we are not passive recipients of this information. We move about and act on our environment, determining what type of information we get.
The third and fourth layers of intelligence are more familiar, perhaps. These are the perceptual layers and the cognitive layers. In the perceptual layer, raw sensory data gets transformed into something more sophisticated. For example, raw visual data might get transformed into a series of connected lines. In the cognitive layer, we carry out even more sophisticated intelligence activities – e.g. reasoning, planning, deciding – before then taking action, and the cycle begins again.
In the early days, cognitive researchers had a simplistic view of the layers, and assumed that information could really only flow in one direction – in from the senses, through the perceptions and onwards into the cognitive processing areas. We now increasingly appreciate, however, that information can flow in both directions, for all of the layers. Our ‘deeper’ processes can feed back into what we perceive and what we sense. For example, if we have been bitten by a snake, we may even see the stick sitting in the grass as if it were a snake.
AI technologies work (or have the potential to work) in very similar ways – in fact, different types of AI technologies have been developed to replicate the functionality of each of these different layers: machine learning was originally envisaged as the perceptual layer and good old fashioned AI (GOFAI) the cognitive layer. But up until now, technology itself has not supported the ability to place AI in environments where these layers can really perform to their maximum potential. Rather, up until relatively recently, we have been placing our AI technologies in what is the equivalent of a dark room, with only a few flashing diodes to help them determine what to do next.
I would argue that the richer the digital environment we can provide to our AI, the more intelligent they have the potential to become. Because of this, almost as a side-effect of being attached to an already developed human computer interface (an HCI), RPA has the ability to provide this richer environment, because the RPA is essentially acting as a sensory and perceptual layer, which can then feed information into the cognitive layer. Closing the loop, this cognitive layer can in turn influence the behaviours, environment and subsequent information being received by the RPA tool, creating a powerful feed-back loop that can be used for learning and adaptation.
To make this somewhat general discussion more concrete and practically relevant, let me talk more specifically here about how each of the main categories of AI/ML – supervised learning, unsupervised learning, reinforcement learning – can potentially interact with RPA technologies.
First, let me distinguish between the two different ways in which RPA can work with AI/ML technology. The first way is to include, in the RPA processes, the capacity to collect data. For example, while the RPA process is answering routine e-mails, it can also be doing some basic processing of these e-mails (e.g removing headers) and then storing the resulting e-mail data in a database for future use by AI/ML algorithms. This approach alone does have tremendous potential to increase the data available to the current generation of AI technologies, but it is still the equivalent of keeping the AI in the dark room. The difference here, is that instead of two diodes, the AI is receiving, perhaps, a package of e-mails to read, or a stream of videos.
Using the right algorithms the AI can still learn quite a bit from this input, but without additional work on the data, it’s likely that the AI would need to default to an unsupervised approach, which is mainly simply pattern matching. This is because supervised learning requires very specifically formatted and structured data – called training data – in order to work. It’s not impossible that the required format of the data could be generated by the automated task in certain specific circumstances, but this would by no means be a default case. As a result, there would likely need to be a separate solution developed for the AI to be practically useful in this context. And it’s unlikely that the result would directly allow the RPA to improve in its automated tasks in real time.
The third main type of AI technology, reinforcement learning, typically requires either specially formatted and structured data or for the AI to be directly placed in a digital environment. This brings us to the second, and potentially most game-changing way in which RPA can work with AI technologies. The RPA technology essentially takes the AI out of the cave and into a still limited, but much richer environment – that of the HCI. This means that the AI can start to receive data in the correct format for the application of reinforcement learning, in addition to unsupervised learning (e.g. basic pattern recognition). It also means that the AI can start to easily interact with its human ‘colleagues’ through the HCI – for example, the AI could potentially get the right sort of training data needed for supervised learning by incorporating a step in its automated process where it sends an e-mail to a human expert, and receives a properly structured reply, which it can then incorporate into its training set.
These simple examples really only scratch the surface of the ways in which RPA can facilitate both the use of current AI/ML technologies and also additional interactions with humans that can directly facilitate learning by enhancing the feedback loop of information. Over time, the HCI itself could be modified to be more suitable for both humans and AI bots to use – in fact, the HCI itself could even begin to learn and adapt, in the same way that our environment reacts to and adapts our interactions with it. There are many exciting avenues that have yet to be explored and developed. I look forward to seeing what becomes possible when we have RPA and AI working in tandem with people to provide a truly augmented intelligence environment.