AI goes one step further. For a given workflow, AI provides cognitive automation — imitating how a person thinks and learns. It can make decisions on its own. For example, when coupled with computer vision technology, AI can read unstructured data, such as a handwritten invoice, and decide how to respond and process it.
“AI has more autonomy and can handle a wider range of tasks,” Spruill says. “It can ‘understand’ a document in a way that algebraic rules simply cannot. AI use cases are plentiful, from finding errors or patterns in finance and supply chain to assistance for pilots, surgeons or maintenance techs.”
The Challenges of AI
The role of human beings in the decision-making loop is one of the key differentiators between RPA, AI-enhanced RPA and AI technologies. Across the spectrum of these technologies, RPA requires less direct human involvement, while AI requires human oversight.
“If it’s just RPA that’s automating a process that is very defined, it is automated. But if AI is included, then humans should be in the loop,” Halvorsen says. “If we look at the cybersecurity example, we might be using AI, but it is still an analyst that is making the final decision. Where in the loop do they fit? That’s the question for many agencies using RPA with AI.”
Another challenge of using AI with automation is managing the data that it is trained on.
“With AI-enhanced RPA and AI, the quality of the data going in is important. You need to have humans check that,” Halvorsen says. “If there’s a deviation or you have to apply a recommendation to the process, you really need to make sure the data sources are transparent, accurate and auditable, and you need to understand how that data could be impacted by attackers.”
RELATED: As artificial intelligence evolves, so do threats.
RPA vs. AI vs. AI-enhanced RPA
For agencies struggling to fill positions or looking to drive greater efficiencies in their operations, RPA, AI-enhanced RPA and AI offer different paths to meeting these needs. Part of the challenge is determining the right technology for the task.
If you can define workloads into set rules or tasks, then it’s probably suitable for automation via RPA,” Spruill says. “AI helps when you need intelligence or advanced logic, but all of this requires having a clear strategy from top-level management. Start building teams that can gather user needs and use cases, then develop functional requirements.”
In thinking about where to apply these technologies, it’s also helpful to look beyond the individual task and tech. Understanding where a single task fits into a larger process or use case can help determine the best technology option for the job. This requires management to think more holistically, with an orchestration mindset.
DISCOVER: Platform engineering is the “natural evolution” of DevOps.
“We are now seeing a shift toward process orchestration,” Shah says. “With intelligent process automation, we can now think at an orchestration level. I have 20 tasks to automate, not a single task. That’s big-picture automation. You end up with a richer set of processes rather than simple tasks. Long-running processes take the idea a step further by no longer depending on something happening immediately. They can wait until the necessary action occurs, such as the delivery of a package, before moving forward with the process.”
Measuring Success with Automation and AI
At GSA, recent findings from an inspector general review found that the agency lacked evidence supporting the purported work-hour savings it had claimed for its ongoing RPA program. Being able to measure and define success accurately should be a key component of any RPA, AI-enhanced RPA or AI implementation.
“I’d use three rules to measure success with automation and AI,” Spruill says. “How much time is saved from the task? How much time is saved from an employee’s week? And how much time is saved by the team? If you want more precise metrics, bring in a lean engineer to perform value stream mapping, but these three make great starting benchmarks.”
