Building artificial intelligence into business process management isn’t easy. Many companies add AI to processes by building or buying single-task bots, such as natural language processing systems or vision recognition tools, and adding them to processes using traditional, non-AI methods. For example, engineers write scripts, and business analysts create automated workflows using process visualization tools.
But it still actual human intelligence to tease out the processes, to connect disparate systems into a single coherent process, to change processes as the business evolves, and to spot and fix problems.
Now, AI, machine learning, and related technologies are making inroads into this territory via robotic process automation (RPA). This combination of AI and RPA adds up to intelligent process automation (IPA), according to McKinsey. In addition to RPA and machine learning algorithms, IPA also includes process management software, natural language processing and generation, and cognitive agents, or “bots.”
According to McKinsey, IPA can translate to 20 to 35 percent improvement in efficiency, 50 to 60 percent reduction in process time, and returns on investment in triple-digit percentages. It’s still early, however, as most companies are in early stage development, using individual pieces of AI, but rarely connecting them into a complete end-to-end automated process, much less into a process flow powered by AI.
“There are no use cases which will go all the way across yet,” says Gartner analyst Moutusi Sau, referring to RPA adoption in the financial services industry. “There have been some chatbot engines out there, and AI decisioning tools, but you cannot build momentum on one particular solution. Banks want to do more than one thing.”
The humble bot
For many companies, the journey to intelligent process automation will start with a single intelligent bot, often a chatbot that answers questions posed by customers or employees.
That was the case for Germany’s ZF Group, the third-largest automotive supplier in the world, which began applying intelligence to its business processes just over a year ago.
“In our corporate communications area, we have a lot of repetitive work,” says Andreas Bauer, the company’s IT manager. “We have a lot of emails coming into our inboxes with a lot of repetitive questions.”
The first step was to create a bot, a basic tool to answer the most repetitive questions.
“It was very lightweight as the first step,” he says. “If someone asked if they can apply for a job, or to find out the open positions. We tackled first one use case, then another. We’re now working on a finance bot, where customers want to know about the status of their invoices or bills.”
But once most of the steps of a business process are automated, then a new level of intelligence can be applied — intelligence about the process itself. So, when choosing vendors for its bots, the company had an eye towards that future.
“We are heading in the direction of automating the whole process chain,” says Bauer. “We weren’t looking for just a bot. What we have been looking for is an orchestration and integration platform, where we could easily adopt these technologies and combine them with intelligence.”
ZF Group was looking for platform capable of learning from experience, he says, while avoiding unintended consequences. “Everybody’s heard about the Microsoft bot that went crazy,” he says.
So while automated integration and orchestration is the end goal, the company also wanted a platform with built-in checks and balances. “There was the fear of something going crazy and us not being able to control it,” he says. “You have to be careful, you have to keep an eye on the technology. It’s not like the technology maintains itself. You have to put effort into it.”
ZF Group chose Vizru, a bot platform that offers a management, governance, and language support layer underneath the bots, called stateful network for AI process (SNAP), which will stop a bot if it demonstrates anomalous behavior. The SNAP layer can also flag or halt a transaction if there are compliance violations or sensitive data is being shared inappropriately between processes, according to Vizru.
The platform also offers built-in support for intelligent routing, whereby the system can, for example, automatically “fast-track” an approval process for a transaction that is always approved at the end of the business process.
Another approach is to add intelligent decision points into a traditionally-automated business process.
That is what American Fidelity Assurance is doing. The Oklahoma City-based company provides 2.5 million insurance policies to 1.5 million policy holders. One challenge American Fidelity faced was automatically routing the many emails that come in each day to the correct destination. In the past, a human would decide where each email should go.
“Is there a way to get advanced machine learning to learn from past data, from past decisions, and make the same decision that a human would make?” asks Shane Jason Mock, the company’s vice president of research and development, who was inspired by a tour of Amazon to do just that.
“I know I was really challenged going into a warehouse at Amazon,” he says. “I realized that there are folks who are pushing the envelope and they’re doing amazing things. Maybe it’s not what others in the insurance space are doing. But the benchmark isn’t really what others are doing, it’s how it will help our customers.”
American Fidelity turned to UiPath, an enterprise RPA vendor, and AI platform DataRobot, to add intelligence to its processes.
“In the new email process, we combined the RPA component with the machine learning component, and the combination of the two decides where the email needs to be routed,” he says.
In many cases, traditional approaches to RPA will hit a decision point that is too complex for a simple automation.
The company is also looking at using artificial intelligence for process mining to automate process discovery, rather than have business analysts figure out what happens in the company.
“We have some proofs of concept going on,” he says. “But it’s too early to comment on that.”
The traditional approach to business process management involves business analysts talking to managers and employees, conducting audits, then creating charts that illustrate the organization’s various business processes.
“Many client engagements where we go in, there’s a process workflow on the wall,” says Sumeet Vij, director in the strategic innovation group at Booz Allen Hamilton. “But is that how things actually happen? You’ll find that how things actually happen is different, the bottlenecks are different. Using machine learning to do process mining helps people get a picture of how things are actually happening.”
In addition, these tools can update the processes as the business evolves — even spot anomalous behavior in real time.
One company that already has an intelligent process mining system in place is Chart Industries, a manufacturing firm serving the energy industry, headquartered in Ball Ground, Georgia.
A couple of years ago, the firm was struggling. The energy industry had been hit hard by a drop in oil prices, the company’s stock price dropped, and the top executives were replaced. New leadership wanted to make changes. For example, Chart had three main divisions, and even though they shared a single ERP system from Oracle and J.D. Edwards, there were multiple back offices handling accounts payable, accounts receivable, and other back office tasks — each with their own processes and procedures.
“We were finding that our customers were effectively taking advantage of paying us later than they should,” says Bryan Turner, Chart’s executive vice president of IT.
There were also other opportunities for affecting cash flow. For example, in some instances, the company could take advantage of discounts for paying vendors within a certain period of time; in others, it could take advantage to holding on to cash longer. The benefits of better efficiency here can go into the millions, Turner says.
Chart turned to Celonis, a process mining vendor, to help uncover opportunities such as these.
“We have it running on a few custom systems today. As long as it has a database and transactions and time stamps, then you can punch it into Celonis,” Turner says. “A lot of the heavy lifting was how to move data between our organization and the SaaS application or the Amazon back end of Celonis.”
Celonis goes through and identifies the business processes — not as they should be in theory, but how they are actually done in practice. Then, it uses machine learning to identify patterns and anomalies.
The business process can be viewed in the form of charts, such as Visio diagrams, and managers can drill down into the process, down to the level of individual transactions.
“Just in one example of late payments, we had annual savings of $240,000,” says Turner. “The software has paid for itself several times over and we continue to see that the cost opportunity is definitely working with both our suppliers and our customers.”
How much data does it take?
AI systems typically require millions of points of data to make usable predictions. Few companies have that much internal data about their business processes.
According to Ray Wang, principal analyst and founder at Constellation Research, Celonis is in a good position to use AI to help companies with intelligent process automation, due to its business process mining platform. Other transactional vendors, like Workday or Salesforce, may also be well positioned to help customers use their historical data to automatically discover and manage business processes, he says.
They may be able to get to a point where they’re able to orchestrate a process and figure out the next best action, he says. “But they will take a while.”
Some vendors providing enterprise ERP, CRM and similar platforms are likely to begin including intelligent process automation tools in the future, if they haven’t already. Salesforce, for example, is making intelligent tools available via its Einstein platform.
In these cases, enterprises are benefiting from AIs trained on the data sets of all of the vendors customers. In other cases, companies may be able to buy pre-trained models and tune them to their own needs, or find open source or commercially available training data sets.
In addition, enterprise data can also be augmented by external data that helps to inform business process, such as weather data, or financial markets data.
“More data helps make your algorithms more robust,” says Booz Allen Hamilton’s Vij. “But we also realize that many times when we walk in, customers don’t have all the data.”
And the whole business process doesn’t have to be fully automated for intelligence to be applied, says Vij says. Many corporate tools don’t have a digital interface or API, and some business processes require a lot of human labor. An intelligent process might know when to route a task to a particular person for handling. And sometimes, a step that might look like it requires a human might not.
“People have things in SharePoint and Drupal, where it’s unstructured, and you need a human to look at it and find it,” says Vij. “But you can apply advances in natural language processing to extract structured information instead of having to have people read it.”
Processes that are ripe for intelligent management include HR processes such as onboarding and financial processes such as claims processing, he says.
Business process analytics
Seann Gardiner, senior vice president of business development at DataRobot, an AI platform provider, says that some of the most advanced companies have enough business process data that they can now look at the overall picture of what’s happening, and make analyses and predictions.
“They’re taking the exhaust from the RPA process and trying to capture that and learn from it and make those processes smarter,” he says. “I wouldn’t say that we’re seeing it very broadly in organizations, but we’re starting to see it.”
If a company has a strong focus on process-level automation, and can unsilo that data, then it might be ready, he adds. “But you have to have business leaders who believe in automation and in an AI-first mentality, and can make the organizational changes needed.”
Companies in the Fortune 5000 are ready, he says, and have processes in place where they can adopt a combination of AI and RPA, he says. “The question is, do they want to put the work in to be able to make those wholesale changes to the organization.”