SAP Seeks RPA Bot Breakthrough, But Will Workflows Cooperate?
The concept is simple enough: Use robotic process automation (RPA) bots trained with machine learning to do the digital busy work of humans, such as approving invoices. SAP already has a product that ostensibly does that, and now it’s hoping to make it even better through the application of user behavior mining and machine learning. However, as SAP customer Merck discovered, getting ML-powered RPA to work in the real world can be tough.
RPA is seeing renewed interest as companies look to automate as many business processes as they can. The world of enterprise computing seems poised for a business process automation breakthrough, and the combination of machine learning plus chatbot technology seems like fertile ground for a money-saving breakthrough.
SAP CTO Juergen Mueller touched on ML-powered automation during his one-and-a-half hour keynote address during SAP TechEd in Barcelona this week. In addition to new SAP HANA Cloud and SAP Data Warehouse Cloud announcements (which you can read about here), Mueller announced that by the end of the year, SAP will have more than 200 machine learning use cases live on SAP products and services.
Clearly, SAP is looking to push ML into its own products. There’s a lot that SAP can do to utilize ML behind the scenes, such as to prioritize items in a list or to surface recommendations to a user. But SAP also wants to enable its customers to use ML and BPA in a more direct manner, including through the creation of RPA bots to automate certain white-collar jobs.
SAP already offers SAP Intelligent RPA, an offering within its SAP Leonardo line that allows users to build bots that “mimic humans by replacing manual clicks, interpret text-heavy communications, or make process suggestions to end-users for definable and repeatable business processes,” according to the Intelligent RPA Web page.
However, that offering will see some improvement, according to Mueller, who announced the start of a limited beta for a new program that will augment Intelligent RPA by using machine learning to collect and refine user data. The program is being spearheaded by Spotlight, an internal startup for SAP and a part of the Hasso Plattner Institute.
“We learned it doesn’t make sense to just automate business processes,” Mueller said during his keynote. “Before you apply machine learning, you should understand how your end users use the system. What do they do? What is their job to be done? You could go out and ask all your employees in your company what are they doing?”
But perhaps a better way of gathering this data is to use automated user behavior mining. That’s the gist of the new Spotlight for SAP tool that was launched this week.
“Imagine you have been asked to improve your business efficiency through automation, but how to find out where to start?” SAP’s Amrit Merz said on stage at TechEd. “We built a tool called Spotlight for SAP which is available today. It provides you with an overview of those processes and transactions that require high manual effort.”
Merz demonstrated how the tool could be used to mine user behavior in the accounts payable process. The software presents information about the users, such as how much time they spend working within the user interface, in addition to different navigational paths through the software to accomplish a given task.
“To drive deeper into these processes, we developed the idea of user behavior mining, which is essentially in a pilot stage currently. It quantified the number of steps taken and the time spent in certain process across SAP UI technology,” she said. “When you know how your employees execute a job and how they work, we can help them automate!”
SAP’s software will help close the loop on user automation by creating a personalized RPA bot that’s trained on the data about the user’s product navigation. The bot generation process itself is automated (naturally), and deployment takes just a matter of minutes, we’re told.
We’re at the very beginning of the age of custom RPA bots. The technology is quite young still, and there are bound to be bumps along the way, particularly for higher order tasks that humans are naturally good at, but which thwart even the most sophisticated AI bots man can create today. We were reminded of just how far we have to go by German pharmaceutical giant Merck.
Merck had looked to SAP and Google technology to automate one of its most painful tasks: Invoice reconciliation. The company tried using earlier Leonardo software (not the new Spotlight tool discussed above) during a proof of concept, and elected not to go live when the POC failed to deliver good results. (The image recognition technology was sourced from Google.)
Merck’s new ML-based workflow was not as accurate as the rules-based invoice matching product that it was designed to replace, according to Oliver Moessner, a UX architect at Merck.
“We tried to automate this assignment and we reached about 60% to 70% accuracy and it was not sufficient to say we want to have it live,” Moessner told Datanami at TechEd this week.
One of the challenges of the workflow is the fact that invoices arrive in all different forms. SAP transforms everything into a PDF, but that’s the easy part. The hard part is figuring out where the pertinent pieces of information actually reside on the form.
As Moessner explained, the invoice reconciliation workflow is easy when the billing party includes the audit number. But when that magic number is absent, it’s up to the user (or the bot) to discover who the entity is. And that’s where it gets tricky.
There may be multiple names on an invoice, and humans have the ability to discern the relative importance of names depending on where they appear on the form. There are additional variables that decide who the invoice should be forward to for approval. These are things that humans do relatively intuitively, but it’s proving tough to translate that knowledge to a machine.
“You’ll find a couple of names on the invoice, but as a human you know, everything below this line, ignore it because it’s rubbish,” Moessner said. “Inside the text you need to find some keywords or events where you can assign it to a person who’s related, and this assignment is not so easy.”
To be fair, Merck’s humans aren’t perfect at this either. That means the company falls back on rules that require the invoice to be above a certain monetary threshold to require approval; all invoices below that amount are paid automatically. It’s no different than any major company that can’t afford to get bogged down in the weeds every time a vendor seeks payment.
As the company worked its way through the project for automating the invoice matching, it discovered there are often problems within other business processes, including the ordering process. Now it was trying to devise a way to handle these upstream inconsistencies within the new workflow, and it was becoming too much.
“It’s a little bit complicated, this process, and it’s a nightmare, to be honest,” Moessner said. “If the starting process is wrong, then you can’t approve or make it better. It’s building a house on sand. Even if you have a nice elevator, the house will smash on itself.”
The good news for SAP is companies around the world, in every industry, still have work to do on refining the existing workflows for human-powered enterprise applications that already exist today. This is fruit ready to be picked, although it isn’t necessarily low-hanging or easy to harvest. RPA will arrive eventually, but the big harvest that people are expecting might be a little later than some initially thought.