Turning big data into big content: business process management is resurging with robotics process automation

Article Featured Image

Consummate AI 

Digital agents make AI accessible for automating business processes because they invoke total AI, which pairs its statistical foundation (typified by machine learning) with its knowledgebase. Virtual agents leverage multiple levels of consummate AI, including the following: 

♦ RPA: According to Joe Bellini, COO of One Network, the template, rules-based approach of conventional RPA is an AI dimension involving “the simplest, repetitive functions. It’s crazy to have your people spend a lot of time on repetitive functions when you can automate it.” 

♦ Rules engines: Expert systems, rules engines, heuristics, and optimizers spawned from AI’s knowledgebase are critical for balancing the different needs of diverse parties in business processes for supply chain management, for example. Digital agents may apply these approaches with static algorithms. 

♦ Dynamic algorithms: When equipped with the mutable algorithms of machine learning models, virtual agents and RPA are endowed with a host of new functionality to support intelligent automation processes. “With the new stuff around machine learning and neural networks, you’re really providing an improved dataset and the machine learning is determining how to make better decisions through the pattern recognition,”  Bellini noted. 

Intelligent process automation 

The wonders of today’s digital agents are predicated on synthesis with cognitive computing capabilities. These technologies transform RPA’s enterprise viability into intelligent process automation (IPA). Digital agents can perform astonishingly sophisticated use cases, many of which exceed traditional RPA’s “screen scraping” while turning big data into digestible content. Hyland senior product evangelist Carolyn Kane described a use case in which AI-enhanced digital workers onboard new customers in a lengthy process involving parsing emails and attachments, querying federal databases, and opening, using, and closing apps so that what would take a human 10 or 15 minutes can be done in seconds. Themes for common applications of IPA include the following: 

♦ Text analytics: The pattern recognition of deep neural networks is formidable for simplifying and expediting text analytics when activated by bots. Conventional text analytics often involves exhaustive procedures for identifying concepts and entities. Conversely, “deep learning comes along and says, ‘I’m going to invert that problem,’” Wilde said. “Just show me the outcome you’re trying to achieve, show me a sample of inputs, and I will backward-solve for that.” This approach gives bots contextual understanding of the text analyzed. 

♦ Performant resilience: The brittleness of traditional RPA usually manifests when items shift on screens. Contemporary RPA uses computer vision to understand what’s on screens for dependable resiliency across multiple variables. “We need to run the same bot on different people’s machines that may have different evolutions, like the screen evolution, as well as different lag [time] in internet connections,” Kakhandiki explained. “Somebody might have high bandwidth, or somebody might have low bandwidth and the bot might need to wait longer for the drop-down menu. There can be application changes as well.” 

♦ Data engineering: Virtual agents can handle some of the basic, time-consuming facets of data engineering tasks such as integration, which is particularly important for the decentralization of data sources with big data. Kakhandiki referenced a life sciences use case in which a single bot was responsible for communicating with more than 60 applications for “getting the data in the right format that makes sense” for business users. 

♦ Task designation: Furnishing bots with cognitive computing technologies enriches their ability to compartmentalize individual components of a larger process into smaller ones handled by separate bots. With this methodology, “each of these bots connect to NLP,” Mahalingam explained. “Each of these bots has the data modeling ability to call an API and get predictive analytics from an ML engine, and each has its own small decisioning engine to move processes forward.” The premier boon of this approach is that instead of building dedicated bots around a particular cognitive computing technology, bots can access multiple AI dimensions. 

Human intelligence, machine intelligence 

Human input is critical to maximizing the overall utility digital agents provide business processes. The timely applicability of bots to cyclical business procedures inherently modifies the roles of humans who once performed them. This difference includes humans concentrating on more knowledge-oriented tasks but also involves dispensing human knowledge to digital workers for improved machine intelligence. Humans effectively become supervisors of digital workers in a mutually beneficial relationship in which the former leverages the latter’s labor, while the bots learn from human expertise. During the data capture phase for auto insurance claims, for example, bots can implement OCR and machine learning-based classifications so that the human “verifier only has to make decisions on exceptions,” mentioned Hyland CPO John Phelan. This method systematically decreases the number of exceptions to boost bots’ overall learning. 

Statistical confidence 

While intelligent systems powered by digital agents can learn from human decisions, in other instances, according to Kakhandiki, virtual agents explicitly “have to reach out to the right human to be able to get their input and learn from that.” Digital agents can solicit human input via a variety of channels, including emails and internal communication systems. Their ability to incorporate that input into future iterations of a process such as adding structure to an unstructured PDF document, for example, is instrumental for creating “scores that reflect how confident a bot is that it found the right answer,” Wilde said. By learning from humans, digital agents can “learn black and white use cases rather easily, and over time they can start covering some of the gray areas also,” Kakhandiki noted. 

KMWorld Covers
for qualified subscribers
Subscribe Now Current Issue Past Issues