kirill_makarov - stock.adobe.com
Intelligent process automation is increasingly seen as a complement to robotic process automation, extending the scope of robotic process automation with AI technologies. CIOs and other IT leaders should consider some of the important distinctions between the two automation technologies as part of their technology roadmaps.
RPA and IPA
IPA covers a larger scope of work than RPA -- it can handle more types of data formats and can promise to enable new types of more intelligent decision-making. Getting the best results from an IPA strategy requires IT and data science teams to collaborate more deeply than is necessary with RPA.
"RPA is purely robotic in nature and doesn't require intelligence to operate," said Banwari Agarwal, global market leader of intelligent process automation at Cognizant. Consequently, it is a good technology for well-defined, rules-based processes.
In contrast, IPA is used for more complex processes that benefit from AI capabilities. This involves combining smart data intake, natural language processing, machine learning and operational analytics with RPA. "Both RPA and IPA are valuable in different types of situations," Agarwal said.
Start with RPA for simplicity
Agarwal believes the main attraction of RPA platforms is that they are easy to use and don't require deep technology skills. Many RPA technologies can be implemented with low-code or no-code. However, the ultimate value and outcomes of RPA projects are limited.
IPA development and implementations are significantly more complex. The technology requires data extraction and classification, machine learning and AI to foster decision-making. Businesses using IPA will need experts on hand who have an in-depth understanding of an evergrowing set of tools and capabilities in the space.
Agarwal said technical skill requirements for users are key distinctions IT executives should be aware of upfront. The technical skill required for RPA ranges from basic to mature, whereas the technical skill required for IPA ranges from mature to advanced. RPA, not surprisingly, has considerably more traction as a result of this ease of use. "There are more processes being automated with RPA than IPA," he said.
Process efficiencies associated with RPA, however, are not as high as the potential efficiencies realized by IPA. Agarwal said in RPA deployments, humans continue to play a significant role in data extraction and decision-making alongside the rules-based processing handled by RPA tools. IPA, in contrast, promises greater value in reducing manual labor costs, because it automates much of the human decision-making.
According to Agarwal, supplementary technologies can help companies migrate from RPA to IPA deployments, including process mining and refinement, smart intake tools, machine learning, AI and an operational analytics platform.
The RPA and IPA continuum
Deven Samant, director of enterprise solutions at Infostretch, a digital engineering solutions company, sees the move to IPA as a continuum, with RPA serving as the foundation for the AI, machine learning and analytics IPA brings to automating business processes. "You can't have IPA without the foundation of RPA," he said.
Samant views this spectrum as having three key phases. More and more enterprises are creating digital workforces and automating business processes that are very defined. At the next level, machine learning helps the system understand and operationalize decisions. The third level is AI, where machines can start making decisions typically made by humans.
The first two phases are more process driven -- they are about automating very defined and deterministic processes. In the third phase, machine learning and AI enable the bots to handle more nondeterministic in behavior. It's about moving from getting a machine to think about a task to getting a machine to think about the process, Samant said.
Support for semistructured data
Angelo Poulikakos, managing director in Protiviti's internal audit and financial advisory practice, an IT consultancy, said IPA is all about combining RPA with complementary technologies, such as optical character recognition, natural language processing, data analytics and chat interactions, to bring the robot to life. These capabilities extend the robot's work, allowing it to read unstructured data, interpret human speech, pay attention to trends and predict outcomes.
Poulikakos agreed that most organizations typically start with RPA before embarking on IPA-oriented use cases. For example, Protiviti has helped several clients build RPA robots that automatically provision or deprovision access to systems based on a well-defined access request form and approval workflow. These workflows are commonly specified using things like checkboxes and drop-down menus to identify the user, level of access and current status.
After an RPA robot has stabilized in an environment, it can be extended via IPA, such that a chatbot can facilitate the provisioning or deprovisioning of access. The chatbots can interpret a user's intent to drive actions that may not have been spelled out. For example, if someone said, "Mary left the organization. Please remove her access," the bot would gather the input and subsequently trigger the RPA robot that would initiate an approval workflow and perform a defined action. At the same time, it would save the conversation history to serve as an audit trail.
Learning from humans
Eldon Richards, CTO of Recondo Technology, a healthcare revenue cycle automation platform, said that one of the key differences between RPA and IPA is IPA's ability to learn from experience. This skill matters most when there is a high degree of variability in a process or in the data used to support the process. With RPA, the implementor must handle the variability in programmed algorithms or rules ahead of time. With IPA, handling the variability can sometimes be learned automatically from experience.
There are two key ways that these differences play out in practice. First, IPA can be used to automate certain processes that are too labor-intensive for RPA tools. When there are large numbers of edge cases -- for example, when unexpected circumstances occur such as missing or inaccurate information or where numbers exceed a typical threshold -- implementing RPA requires developing logic to handle each one. IPA may be useful in such situations if it is possible to learn from an experienced human actor performing the processes as long as the IPA tool can observe enough of those edge cases.
Second, IPA can be used when higher-level cognition is required to make a decision in a process. For example, RPA can be effective at filing emails if the filing is based on attributes such as sender, key words found in the subject line or whether the email has an attachment. In contrast, IPA would watch which emails a human put into their spam folder and which ones get immediate replies. This would allow it to make a more sophisticated decision, Richards said.
Collaboration required for RPA and IPA
IPA projects can also affect workplace benefits, such as facilitating the collaboration between the data science teams and the line-of-business professionals that have the necessary subject matter expertise about the document-based business processes being automated. This leads to better implementations and the identification of additional high-value use cases, said Tom Wilde, CEO of Indico, an IPA platform for unstructured content.
Adding a layer of intelligence to RPA can have a transformative effect on processes, as well as when organizations collaborate on finding better feedback loops for training the AI models. "Suddenly the bots can cope with high-value, decision-making tasks, as well as repetitive ones," said Arvind Jagannath, director of product management at AI Foundry, a mortgage automation platform.
The AI models driving RPA decisions can be improved when business users and data scientists can identify which sets of data to use for ongoing training. This might include evaluating the performance of models over different time scales. Short time scale may look at what loans human experts approve or deny -- a longer time scale could consider which loans human experts approved, but subsequently defaulted, to further refine the model. "With more data, the models for making decisions can become more accurate and reliable," Jagannath said.