The Quest For End-To-End Intelligent Automation

The pandemic has seen accelerated interest in process automation as organizations have scrambled to overhaul business processes and double down on digital transformations in response to disruptions brought about by COVID-19.

And for IT leaders stepping into or already steeped in such modernization efforts, artificial intelligence — mainly in the form of machine learning — holds the promise to revolutionize automation, pushing them closer to their end-to-end process automation dreams.

But for now, AI-powered process automation remains a piecemeal approach, in which AI is involved in individual tasks but not across the entire process chain. Regardless of how vendor’s spin it, fully intelligent automation has not yet arrived — but organizations working to fill the gaps are finding innovative ways to this promising concept closer into being.

The current state of intelligent automation

A typical use case for AI in automation includes the following: instead of requiring someone to manually re-key information from a PDF into a form, an AI is trained to do it for them. Or, when an employee would normally hunt through corporate documents to answer a customer question, an AI suggests possible answers.

As for the rest of the process, humans are at the core. A human business analyst figures out what goes into a particular process. Developers use robotic process automation (RPA) systems to create process flows. More business analysts monitor the performance of the process, seeking to find bottlenecks and to come up with ideas for other steps that could be automated either through traditional scripting or with AI augmentation.

In other words, AI, thus far, has acted as a tool that can fill niches in the greater scheme of automation.

“One of the well-kept secrets about AI is exactly how narrow each use case is,” says Elena Christopher, senior vice president of research at HFS Research.

Still the technology to knit together end-to-end intelligent automation may already be out there, at least in bits and pieces. Challenges remain, including the fact that gaining visibility into business workflows isn’t always simple, because employees often switch systems to carry out part of a task, or they complete an action that is difficult to capture digitally, undercutting AI’s ability to fully comprehend a process from end to end.

Using computer vision to gain process insights

Genpact, a global professional services firm with nearly 100,000 employees, runs thousands of processes for hundreds of clients, including many Fortune 500 companies. It is using intelligence to match up individual transactions in core systems with the bigger processes they’re part of. But intelligence falls short when, say, an employee leaves the core system to pull up a page in a web browser, says Sanjay Srivastava, chief digital officer at Genpact.

For example, one step in a process may involve looking up a pricing page to check whether a particular item is more or less than $100. To complete the task, an employee might pull up the relevant page and visually scan for the price before deciding what to do next. That action — seeing something on the screen — is hard to capture digitally.

That’s where computer vision comes in, using cameras on workstations to track where employees are looking. “It absolutely has to be done with the consent of people,” Srivastava says. “Typically, the companies we serve have pre-existing policies that govern this, so this works for some companies and doesn’t work for others.”

Using a combination of techniques, including computer vision, Genpact’s automated system collects all actions related to business process, not just those by employees in a particular department or job role. Identifying all tasks and linking them into a workflow is called process mining, which Genpact has been doing using a homegrown AI-powered, automated solution for three years now, Srivastava says. In the past year, the company has added computer vision to make the technology more comprehensive.

Once the business process has been identified and mined, Genpact can then monitor and troubleshoot individual instances of that process, or fine-tune the process based on ongoing feedback.

“Let’s assume that 100,000 laptops were bought last week from my company,” he says. “We can pull out 100,000 end-to-end processes. We will monitor all the versions of that and can track it right down to any specific one. Why is it that this one process deviated? You can fix business problems or adapt to the new normal.”

That “new normal” could include automatically instrumenting changes in the process. For example, if approval is required for changes of more than 10% to an original invoice amount, and approval is given 100% of the time, perhaps it is time to adjust the business rule from 10% to 20%, Srivastava says. Robotic process automation can then skip the approval step for invoice changes under 20% or deliver a pop-up note to remind employees they no longer have to forward the invoice to finance for approval and can instead make the final decision themselves.

AI-generated automation scripts

Digital transformation consultancy UST has been using AI for the past three years to help automate those tricky to digitize business process tasks, says Adnan Masood, chief architect of AI and machine learning at the company.

Masood, who collaborates with both the Stanford and MIT AI labs, has a patent on discovering business processes using unsupervised learning, connecting individual tasks into an end-to-end process.

Take, for example, that same website that an employee scans to look up a price. It’s one thing for an automated system to identify that this is happening as part of a process; more challenging is to be able to duplicate the action, given that websites change all the time.

“We vectorize the inputs that come in and use deep learning to identify what is an input,” Masood says. Then, the next time the employee goes to that website, the AI can pull the data automatically, without someone manually scripting the task. Then the AI can improve over time, with reinforcement learning.

“There’s a human in the loop who reinforces it, or can tell the computer that the data field didn’t validate properly,” Masood says. “Typical RPA platforms just automate the grunt work. But cognitive AI is self-optimizing, self-learning.”

Within certain parameters, the system can also go on auto-pilot, he says. “That’s where it can directly take action based on what it has learned.”

But Masood’s system still requires a base set of logs to work from, and finding and pulling together all the process logs remains a time-consuming manual process, he says. “The data is never in a single system. It comes from multiple sources and different modalities. Someone has to do initial configuration.”

If data is in legacy systems, this process might require custom coding, opening firewalls, or getting regulatory or cybersecurity approvals, not to mention data engineers to set up the data flows, a process that can take months.

Once data is collected, the AI starts to analyze the workflows. It can take a few more months to validate the workflow map, depending on the number of people involved in the process and the frequency of transactions, he says.

With the process now mapped, human beings must still validate it. “A subject matter expert looks at it and says, ‘No, this step is wrong, here are the right data sources for it,’” he says.

The challenge of deep integrations

For companies that have turned to RPA, tasks outside core systems can still present challenges. Workarounds such as web-scraping and OCR document scanning are short-term fixes that can introduce errors and cause processes to break. A better solution is deeper, machine-to-machine integrations via APIs.

“When you use RPA against a website, the automation isn’t as stable,” says Megan Amdahl, senior vice president of partner alliances and operations at technology consulting firm Insight.

RPA can watch what users do on a website, she says, but there’s no way for the RPA system to know the website may offer an API that can be used to get the data directly. As a result, Insight went back to manual scripting for these kinds of automations.

“The IT department now hand-codes the API data transfer requests,” Amdahl says. “With it being within IT, they can operationalize it faster, and it could scale faster. If we hard-code via APIs, it doesn’t affect the automation. API standards change, but they change more slowly.”

Other obstacles to full automation are business partners who haven’t digitized their processes yet, or who have incompatible systems.

“Customers want to do things the way they do it, and they will go to a company that will receive the information the way they want to give it,” she says. “So, for example, they could order the majority of things they want off of our website — but their internal processes require that they create a purchase order.”

If a client is able to connect systems directly, it does require manual coding to make the connection, she says. “And then you have to have maintenance around it, so it doesn’t break,” she says.

As more companies move to SaaS platforms for core business systems, and those SaaS providers work with one another to set up data integrations, all this may soon change. “Having the pre-built connectors is the absolute ideal state,” Amdahl says.

The specific over the general

PricewaterhouseCoopers has found some success with AI-powered process mining, but to a limited extent in specific, narrow cases, says Anand Rao, PwC’s global AI leader.

“We have put together quite a few vendor tools to do it,” he says. “Given the disparate type of jobs that people do, if we just had a background bot that is looking at everyone, we wouldn’t be able to make any sense of it.”

As such, Rao warns against cure-all automation sales pitches. “If they’re saying I can install process mining software and replace ten people, I wouldn’t believe that,” he says.

Moreover, most of the AI used in RPA is for specific, individual tasks, says Chida Sadayappan, lead specialist for data cloud and machine learning at Deloitte Consulting.

“Everything is AI-enabled,” he says. “There are documents that can be read, data can be extracted, PII redacted. There are imaging activities, like recognizing the damage on a package.” But that’s not AI-powered RPA, he adds. “AI is just replacing some of the mundane tasks.”

There is AI and machine learning being infused into process workflow automation, he says. “But there are not a lot of use cases.”

Large insurance companies may be infusing AI into process workflow automation, and financial institutions may use it to process mortgage applications, Sadayappan says, “but as for the rest, there are very few large-scale workflow automations.”

But Dan Diasio, global artificial intelligence consulting leader at Ernst & Young, sees potential for growth. “AI has a lot of point solutions. But there’s an opportunity now to go from point solutions to AI-powered platforms. And automation is looking to jump into AI-powered tasks. The two things are converging,” he says, adding that EY has “actually brought the team that traditionally focuses on process automation into the AI team” because of this.

The purpose of the process

AI for process mining and workflow automation is still in the early stages, says

Gartner analyst Marc Kerremans, who authored a report about processing mining in April. “But it can only improve.”

But what isn’t going to be automated, at least in the immediate future, is context awareness. It will still require human analysis to ascertain whether a process should exist in the first place, or if it needs to be replaced with something else.

Meanwhile, vendors are investing heavily. All the major RPA vendors are building or buying process mining functionality and investing in AI, including Celonis, UiPath, Automation Anywhere, Blue Prism, and Livejourney.

Major vendors are also in the fray, buoyed by acquisitions. IBM’s recently purchased process mining company myInvenio, while SAP has acquired process mining vendor Signavio and Microsoft snapped up workflow automation vendor Softomotie.

Eventually, Kerremans says, process mining will be a common feature in enterprise platforms.

Still, Kerremans advises not sleeping on process mining. “Even if you don’t have all the information available, start with the tasks that have readily available information,” he says. “You will still create insights, visibility, and value. If you don’t do this but wait three years, you will be behind.”