“We have found about $100M in value through hours saved that we can unlock in the next 5 years for our businesses by using IPA on unstructured data.” – Sean Nicolello, VP of Intelligent Automation for MetLife.
Companies embarking on process automation projects generally face a learning curve. Most organizations start with simple tools, like robotic process automation (RPA), before graduating to solutions that can handle more complex tasks involving unstructured content.
That was the approach MetLife took on its automation journey. MetLife is more than 150 years old and has no shortage of unstructured documents to deal with. It got its feet wet in process automation by using RPA for structured documents and has since graduated to intelligent process automation (IPA) tools to handle unstructured content for use cases ranging from automated insurance contract analysis to customer onboarding.
We recently sat down with Sean Nicolello, VP of Intelligent Automation for MetLife. He shared the company’s approach to intelligent automation, the importance of educating the user base and setting expectations, how to choose automation use cases, why federation is essential to scaling automation efforts, and his process for selecting an IPA tool.
Indico: What has been the toughest part about deploying automation in insurance use cases across MetLife, maybe starting with RPA?
Sean Nicolello: In the beginning it was setting expectations. Later on, it was scaling projects effectively.
In terms of setting expectations, we first had to dispel the notion that a robot could replace any task that a human can do. In reality, robots are really quite dumb, although they can perform valuable work. Organizations that experimented early on with RPA learned how to optimize the use of the tool and got lots of value. But those that came in looking for robots to do the thinking for them were underwhelmed and not convinced. We had to do roadshows, be transparent and work extra hard to set expectations about what we can and cannot do.
With RPA, results were black and white: either it worked or it didn’t work. Over time we got good at predicting what to expect in terms of output and utilization.
When we moved on to more pure artificial intelligence solutions like intelligent process automation, it was back to square one in terms of resetting expectations. It was a similar situation to people thinking the robot could do everything for them. Now they think the AI solution can think for them. That may be true but it’s only 5% of the thinking that can be replicated. So, it’s small, but it’s powerful.
Then trying to scale projects was difficult because it requires everything that worked on a small scale to continue working on a larger scale. You need to really focus on the operational model and ensure you’ve got adequate funding and executive support. It takes a big push to make it happen.
Can you talk about how MetLife originally approached the idea of automating processes that included unstructured content?
We started off with a good foundation of structured content digitization. We had a tool that could digitize templates and structured forms, so we knew what the industry was capable of doing and what it lacked.
Our team had been speaking with businesses and collecting their automation use cases. We began to compile a graveyard of opportunities because our tools, primarily our OCR tools, couldn’t support them.
At the same time, we have an innovation team whose job is to work with all our LOBs globally and compile a list of strategic wants and needs so they can find startup vendors to fill the technology gaps. The innovation team came to us saying they were seeing an emerging common problem that was being articulated in different ways. One was, I have these contracts and I just want to find out “if x then y” but I have to go through every word to figure it out. Or, two, I have all these invoices and we have to manually review them because they’re different every time. Three, I have an email inbox with a small army of people that read hundreds of emails a day just so they can move the email to the right team to take action.
Because of our experience with OCR we recognized that the common problem was the structure of the data. From there, we set out to find the best in class tools that could help us address unstructured data. Two broad categories emerged – pure play data digitization/extraction/classification and more nuanced NLP [natural language processing] based text analytics.
We ran an RFP to find vendors that excelled in both. Looking across the market, we asked consultants, asked Forrester, Gartner, and so on, and ran an RFP with over a dozen vendors to see what they could do.
Plus we gave the vendors a pretty nasty pair of use cases to test the DD and TA capabilities. Long story short, spoiler alert, Indico emerged as a top provider across both.
With Indico, we saw two things. First is the power of the platform in terms of accuracy of output and speed of output. Second, working with Indico from a customer perspective has been the best vendor to buyer experience I’ve had in my whole career and all the teams that work with Indico say the same. This isn’t a marketing pitch either and I really appreciate the customer focus you bring to the table.
How is IPA different from RPA? Why can’t you just use RPA?
The way I describe RPA vs IPA is like this: RPA simulates human action and IPA simulates human intelligence. RPA is the brawn but IPA is the brain. This resonates well.
IPA is the natural evolution of RPA. It allows us to do a lot of the same things in RPA like moving data from point A to point B, but it allows us to do it with data we previously were unable to do it with. RPA is limited and will be commoditized. Because it’s bound by rules it can do anything a rules-based system can do. That doesn’t mean RPA will die out. It means RPA will get better in collaboration with IPA technologies.
IPA is boundless. It can get smarter forever and get better over time with learning. It can change form and evolve, and it can exist in a million different ways; a machine will train and learn a bit better and a bit differently each time. This is why we have hundreds of use cases in our graveyard waiting for a companion to RPA to add intelligence.
For example, we’re deploying a use case with Indico for one of our LOBs that will automate the complex and manual process of reviewing inbound new business documents. Indico will be ingesting the documents, digitizing the data, supporting the associates’ review and compare of the data and accelerating the decision timeline for new business. This is happening across five parts of the process and is expected to reduce the manual burden by thousands of hours per year. We’re expanding this to include RPA in areas that help to move the data before and after it goes to indico. It’s a good example of how IPA plus RPA creates the best value.
When you conducted your POC evaluation, what surprised you about the vendor products and the vendor results you observed?
I was surprised by how little data was needed to train a model to get very high accuracy. I was expecting a machine would need thousands or high hundreds of sample documents but to see results from a data set of 50 to 100 is amazing.
The other surprise was the variability. Because of its probabilistic nature you can really see competitive differentiation emerging in the IPA space. Some vendors had an accuracy score of 80 out of 100 and some had 5 of 100 on the same data set. Indico was in the top set
Why is unstructured content so difficult a problem to solve and why is it important that MetLife has a solution to it?
MetLife celebrated its 152nd birthday this year. We have documents that go back over 100 years and we have highly paid data scientists in our company but there’s a mismatch – those data scientists can’t unlock insights from that data; no one can. But we’d like to be able to use that data to help us better predict mortality and morbidity. A small adjustment to our actuarial models can result in billions of dollars of savings and revenue generation over the next 5 to 15 years.
Second, our CoE [Center of Excellence] is in the business of scaling. It’s not enough for us to have a point solution to a single or small number of use cases. We want to continue expanding our automation scope. To that end, we need a solution that covers the broadest number of problems in the organization. The biggest, broadest and highest-value opportunity we see right now, by far, is around extracting value from unstructured documents. We have found about $100M in value through hours saved that we can unlock in the next 5 years for our businesses by using IPA on unstructured data.
What are some of the most promising use cases in insurance for IPA and why?
Contract analytics, an area where Indico excels, is one. We have so many contracts. Our business is a contracts business. We write contracts to create evidence of our promise to insure. With individuals and group customers, we have millions of them. On our investments side it’s the same thing – we wrote contracts for execution of all of our various investments so were looking at hundreds of thousands of contracts. We want to know what’s inside without manually looking inside.
Group customer data processing like invoices is another. With invoices there’s dozens of insurance-specific forms that vary from customer to customer. We pride ourselves on customization to build the best customer experience but that means manual processing when we want to do something with the data, because the invoices are different from one another. Now we can process these different data elements automatically.
A third use case is customer onboarding. The onboarding process for insurance is awful. We are making big promises so, for something like a mortgage, we need to make sure we have everything in good order. The onboarding process may mean accumulating dozens of documents with hundreds of pages. We want to be able to process these more effectively, find fraud, categorize risks, and more.
How do you evaluate and engage with line of business on new use cases?
Our CoE has a three-prong model: relationship management, delivery, and production support.
For relationship management, or engagement lead, we have a robust, mature process for approaching new or existing LOBs. First, we look to understand their full strategy – what they want to do, what they need to do, gaps, problem areas, opportunities, etc. Then we understand their processes – what are they, review any documented processes and have working sessions to help them identify areas for automation. We use this as a catalyst to start building a pipeline of use cases. The most important thing is that before we formally engage, we identify a healthy pipeline of use cases so it isn’t just a one-time thing.
Then we prioritize their use cases for impact and time to value. Our first attempt is to find use cases where we can quickly and easily provide value – quick wins that we try to produce in a matter of two weeks to production. This shows the technology is real and it’s not a year-long turnaround. Then we go to higher value use cases with greater complexity and keep the wins coming.
What are some of the key trends in automation you see that are important to MetLife?
Federation. As we grow our CoE, we recognize that maintaining control over the building and delivery of solutions cannot sustain itself if we want to reach the whole organization. To that end, we are investing in federating our solutions. What I mean by that is upskilling our business users to learn the tools and be able to run their own automations. We could keep a degree of control over the technology, licensing agreements and such, but we would give power to the business to build what they want.
That is one of our bigger reasons for going with Indico. There are two approaches we could have taken. One is to onboard a platform and delivery solutions in the CoE. That works but will reach a limit in terms of our ability to deliver or we will have to grow the team size too large. The other is to find a tool that business users can learn and use. That’s why we selected Indico. We were able to train our business users first, and they took to it fast. This is a major factor for us and where we see automation going – to the hands of the business users.
When you look out into the future of IPA, what do you see as the next horizon?
In terms of a discrete feature, it’d be dealing with unstructured handwritten data. That’s the holy grail in my opinion.
With respect to the overall roadmap, attaining real end-to-end automation across any process. With RPA as the brawn and IPA as the brain we’re closer now than ever to replicating human knowledge work. Getting to the point where we can automate 95%+ of a process every time is the next great horizon.
Watch the video to see the entire fireside chat from IA Week Digital: