We Need Smarter Bots: The rapid evolution of the automation ecosystem

00:04 [music] I hope I didn’t lose any more folks than where we are in the room. [laughter] Okay. I’m Mahesh Vinayagam, founder and CEO of qBotica. You stole a question from me, so maybe can I– I didn’t actually look at the show of hands. So how many know about RPA? Yeah, almost everybody. So it’s kind of– it’s no big deal because it’s a hot trend right now. So whether it is Gartner, HfS, Forrester, almost everybody is calling out as a trend. And Gartner calls that as a hyper-automation. And they call that as the fastest-growing approach– now, I don’t want to say technology, but the fastest-growing approach that enterprises all around the world. There is no specific industry. There is no specific region. There’s no specific department. It’s kind of across all enterprises, across all geographies. It’s being adopted. Why is it being adopted? Right. That’s what I’m going to talk about. And then talk about what we are doing and where RapidMiner plays a part. So that’s how my presentation today is going to be. So to introduce our firm, we’re about three years old, very new to the industry. I used to work for a company called Syntel. I worked for them for 20 years. And around 2015, I caught this trend of RPA and figured it out that it’s actually going to make a big impact in the next five years. And I decided to kind of throw my kind of prediction and– I wanted to throw my career behind the prediction, see how it takes us as a company. So I did that in 2017 with a couple of my colleagues, we founded qBotica. And in the last three years, we made significant strides, so much so that Gartner recognizes us as part of their landscape today among some really big names in the industry. And this week, HfS announced us as a hot vendor so in the space. That’s about us. So that end of informational.

02:17 But in terms of going back to the topic, every company, every department wants RPA. And what is RPA? So we want to kind of obviously define it. I know some of you know it, but it’s kind of level set and kind of to understand what is being automated through RPA. Simple tasks are getting automated today, whether it’s opening applications, performing calculations like tabbing from one screen to another. And basically any human task on a computer following a set procedure or a process. That’s what is getting automated using RPA. Right? Here, we are not touching the end system. We are not changing a process or a procedure. What’s being approached is automation of work. So you have heard of Future of Work. Future of Work redefines how work is being performed today, whether it’s in the physical world or in the enterprise world in the– behind the computer or on the computer. Right? So that’s what is being automated. Right? So automation of human work on a PC. Right?

03:38 And then let’s kind of go deeper into that. Right? So what does that automation of work on the computer means? It means that somebody is being given a task to perform by their manager, then they follow a particular process, which is like a word document, say it’s, okay, this is what you do on a [visio?] diagram, whatever it is, right, that somebody is being trained. And they use many systems. So systems such as simple as Excel, Office, and then they use the RP systems to perform that work on a computer. So that’s what they do. And they [can?] be green screens, black screens, whatever you call, they use that to perform that work. So is that automation, automation that’s being performed at that level, is that enough to make progress? Right? That’s the question that’s been– and are being asked. The RPA as a technology is very powerful. The reason it is so hot and many industries are adopting is because of the reasons that we saw in the previous slide. Right? It’s simple. It’s simple. It just makes sure that if somebody is processing orders– I know that a very large customer that we deal with in Phoenix, a Global 10 customer, has about 300 people sitting in India performing a job, and they just do one task. Right? And then they just follow that and keep doing that. They are allocated work on a queue, they look at it, they [inaudible], and they don’t make any decisions. They keep doing what they are doing day in, day out. Right? Three shifts. That’s what they do.

05:14 So is that automation of that work is important. That’s important because then only you could actually understand what’s happening behind the process. Today there is a lot of data that’s being missing because human is actually taking some short steps, shortcuts. Whatever they are performing, how they are being performed, the speed at which that transaction happens, it’s all very based on the human action rather than what’s actually required to perform the task. So is that enough is the question right now. Right? So no, it’s not enough because it needs to be advanced. Humans, when they do repetitive work without thinking, it’s actually that that’s robotic. Right? But that’s not how almost every other people do it. There maybe there is some class of people that does it, but then as you move up the organization, there’s a lot of decisions being made. Sometimes there’s empathy that’s being applied for the task. Right? It’s not like your garage opener, it just [inaudible] [thinks?] and go. No, it’s like the person that’s sitting at customs, he figures you out, like, okay– although he has data to go behind who you are, he just tries to make some decisions there. And similarly, I was at a airline club two weeks ago, and I showed them, “Okay, here’s my boarding pass, I want to use your lounge.” And they were like, “No, sir, we have a rule, it’s three hours before flight only you could use it.” And I look at my clock and I have like three hours and 15 minutes ahead. And I had a standby flight, which is actually one and a half hours away. I said, [“I’m going on a?] standby flight, you could actually let me in.” “No, no, you have to go back to your gate, get your boarding pass, and come back and show me if it’s one and a half hours, and I’ll take you in.” And I said, “If I go and come back, it’s 15 minutes already.” Right?

07:17 So there are very, very, very few people that will act like that robotic. Right? So she was so robotic about it, she didn’t even think about the fact. Right? But then that’s not what’s many times. Many times, I had great experience at the very same club, people will actually go out of their way to help you. That’s what human does. Human actually does not follow the process as is. They’re actually able to give some empathy, make some quicker decisions to impact the process. How do we actually make that happen? Right? And the other question is, there are actually a lot more system’s integration challenges than what a simple web screen stuff is. Right? There is a lot more to be integrated with. So as these tasks become complex, then there’s going to be a lot more system degradation challenges. And finally, the big question, right, data. 80% of the enterprise data is not in those lakes and oceans of data that’s been created. Right? So they are actually outside of the structured data. Right? They are in forms of emails, documents. Actually, there are some process logs which doesn’t even go into the data lakes. Right? So it’s basically, what’s being stored is the final outcome of a transaction, the decisions. How it went there, how it actually passed through the enterprise systems, that’s not at all recorded. Right? There’s very, very minimal information that’s available. And these are things that needs to be conquered as we go to the higher level of automation. And in fact, like – I missed the important point – third party systems. Right? So a lot of your data is not even within the company. You actually have to go into a third-party portal, like a government portal or somewhere else, to get that information. So it’s not even with us. So these are things that needs to be taken care of because today, the tasks being performed requires you to actually address a lot of these things.

09:24 So what we want to kind of lead to is an intelligent automation ecosystem. I call it an ecosystem because it just not one product that can actually solve this problem. We heard that this morning, right, when Mike from Forrester talked about adjacencies. Right? So we call it the ecosystem. You need to have multiple tools or products to solve that problem and the different types of technologies to attack that issue. Right? So if you think about human, what a human does on a computer, human actually reads stuff. Right? They actually understand things, and then they perform an action. So you need technologies that will actually help do that different actions that the human does. So you need the ecosystem for having some good dialogue. Right? So you have an informed dialogue. So today, if you go to many chatbots that’s out there, they are FAQ bots. They are pre-programmed. You just ask a question, it’ll just keep repeating the same answer. So that’s what it does. How do we actually make that intelligent? Then it needs to have relevant on-demand connections to the enterprise systems. It needs to make an action at the point of service. Right? So how do you actually make that happen? These are actually the next steps of the evolution from an automation perspective. And then how do you actually bring context and sentiment to that interaction or the transaction that’s being automated?

10:59 And more so – I covered this before – one of the key things of automating the ground level processes, is to collect that valuable information of how it’s being performed, what are the transformations it takes. Right? As simple as somebody entering that data, it actually creates an action, like something is happening to that after the data entry is done. Say if it’s invoice approval process or a order entry process. So from that point, how is that data kind of flows into the system? So you need to have that entire intelligence built-in. So actually, computerization of those manual processes will help us gather that data, right, to get that intelligence. And more importantly, the exceptions that’s happening. Today, humans actually make, like I said, shortcuts. There are some exceptions that are not even recorded. And in fact, the most famous field that stick in a front-office system is, other. Right? So they actually have a classifications of so many things. For every transaction type, they’ll actually have, okay, a dispute because of the reason A, reason B, reason three. Then there is, other. So people just do other because it’s easy. Do other, write a comment, and then pass a transaction through. Right? And just get an approval from a manager and get it through. And don’t we see that? Right? So when those things happen, then there’s valuable information about why are the exceptions approved is missing. Right? And if we actually automate those processes, you are forced to look at the exception parts, you are forced to make sure you select a valid reason for those exceptions or whatever actions have been taken. So these are critical things that could actually come out when we automate those processes.

12:52 So let’s look at some use cases. So I kind of gave a very high-level [raw?] review at this point, the reason why automation RPA needs to be done or intelligent automation needs to be done. Let me look at some use case. So the first use case is a very typical use case [scan?]. You would see this in every RPA forum, invoice processing as the use case. We are also going to look at it, but with a little bit of a spin for this conference. And this is what we have been doing for actually a year and a half. So we have been partners with RapidMiner from the get-go. Right? In 2017, we became partners of RapidMiner. And for the first few year, we are working with the product to understand more about process discovery, process mining. We are trying to use a open-source tool, that’s what on the platform called RapidProM. We’re experimenting with that. And then we moved on into using the actual machine learning models. And this particular use case is how we are actually using it in this context. So look at the invoice pricing use case. Right? It’s about somebody that opens an email, they then download the attachment, install it into a folder, open that attachment, look at it like, do a three-point validation for a match. You do a three or four way match. That’s what a invoice is about. Right? Look at your ship-to address, your line items or your totals and all that stuff, and the P.O. number and so on so forth. So that’s what people do, look at it. And then they understand the context sometimes, “Okay, why is this invoice being presented at this time?” Blah, blah, blah. They kind of understand some context around that invoice. Then they actually see if something wrong with that– if something is wrong with that invoice, they will call a manager or they themselves do some investigations before they actually do the approval on the system. That’s what a human does. Right?

14:49 So in terms of if you are to automate that process like we’ve figured before, we could automate it using four different methods. So you automate the actions of them doing the keystrokes, the mouse strokes. All that stuff could be automated using RPA. And looking at the invoice, right, could be automated using OCR. And then to an extent, if you want to understand certain aspects, like I said, what does it mean, is like for one of our largest clients, we process over 700 to 800 thousand invoices a year using this approach. For them, the vendor name doesn’t match always. The vendors could be really big names like Oracle or Cisco and all that stuff. They have people put some department codes next to the vendor names. Vendor name is always not the same on your system. A task as simple as this becomes complicated because it’s not actually uniform, the data is not always right, it’s not a straight match. You can’t actually automate that. Right? So we had to use NLP to determine what this [name?] could be. And sometimes we actually determine the line items. Somebody said, “Samsung television, 50 inches.” Somebody might just say, “Samsung, apostrophe 50,” and, “TV.” Right? And would that match because the P.O. was issued on one thing and the invoice comes on another? So how do you actually match those? NLP gives us a better way to do that.

16:22 Then finally, machine learning. Right? So you need to, if you are to understand the context, the patterns, the anomalies of those invoices, you need some machine learning. Now, let’s look at the products that we used. So when it comes to the ERP, our– currently, we actually work with about 10 to 15 clients, and many of them SAP. No surprise on that. Then Oracle. One of them uses Epicor. A couple of them uses other smaller ERPs. And then obviously you work with Excel, you work at the Office systems and so on, so forth. And then from a RPA perspective, we are partners with all the big names in the industry, UI Path being our biggest partner. And we actually kind of created this framework on all of these platforms. So whether it’s UI Path, Kofax, Blue Prism, EdgeWave, and automation anywhere. And I think some of them is already a partner of RapidMiner as well, I believe. And on the NLP side, we use NLPBots, but we could always use our open-source platforms like Dialogflow or Rasa to do that NLP that I talked about. And then, finally, for the machine learning, we just use RapidMiner because we actually feel that– we’re convinced with what we have seen so far. And we are a big user of it’s API online cloud version as well.

17:47 So let me just run that demo for you. So in this demo, what we show is that whole process that I described, how it’s being executed. So for this particular demo piece, we’re using UI Path, ABBYY, and then NLPBots, and RapidMiner. So the bot, actually, at the background has logged into SMTP and download the invoice attachments. And now it’s actually trying to kind of get into the ERP. Again, we are using your open-source ERP for this demo. So now, the demo actually, the bot launches the ERP system. Now it’s downloaded the invoices, taken all the data, and now going to match the invoice data against the pure data on the ERP. So that’s what’s happening right now. Then at this point is when– okay, let me just pass here. So once it understands– it understands, extract the data from the invoice, it matched with their business rules against the ERP. Now it’s all good to go. Right? The three-way match or the four-way match is good. But then we thought a lot of fraud happening– we were actually doing some study on where does invoice fraud happen. Invoice fraud doesn’t happen by any of this four-way match or three-way match, it actually passes all that rules. They actually pass the vendor name, the ship-to-customer name, the dates; everything matches, even the P.O. numbers matches. Right? They do the fraud at the bank account details. They change the background details on the invoice. They change the address on the invoice. So that’s when the fraud happens. It’s a very subtle thing and there is no automatic way of doing that. Right?

19:52 Today, humans can’t find it. When you’re actually having a bunch of folks that’s doing the same thing, they’re just looking at those four-way match because they’re being paid by the hour to kind of process the number of invoices. They just keep doing that. And such anomalies are not even watched. Right? And then we ask people like, “How do you actually catch these anomalies?” They said, “Either you look at those numbers–” again, you will have to go through the invoice, take that value, which is actually at there somewhere printed on the invoice. Or you’d actually call somebody, the vendor themselves, to take this information. And that’s another option. That option is they find it that these fraudulent invoices come at odd times. They don’t come in at exactly during the heavy period. Sometimes the under-the-floor invoices happens at normal times, people don’t even realize it and it goes through. So we kind of took those use cases, and we created a model based on, “Okay, where does the anomaly happening with the dates, with some information that’s been changed from the previous submissions from those minutes?” We excluded the line items, the variable values on the invoice, and then took only the values that actually doesn’t change as much, which is, we talked about the bank account details. And then the pattern, we can always look at the time pattern when they’re being submitted from a date perspective and all that.

21:19 So those are the things that we took, we created a model, and we created a model against– a customer allowed us to take around 300, 400 thousand worth invoices to create that model. So we actually took that data, which is [inaudible], but we just needed this information. So for that, we don’t require the customer details, only the other information was required. So we actually took that data, ran the model, and we use that model in this flow. So after the business rules are passed, now you send it to RapidMiner, and RapidMiner then gives us the score back, saying what the score look like, whether we can approve it or not. So once the score is low, we send it to a human in the loop. We kind of pass it on to human to approve it. If it’s above the score, it’s automatic approval happens. So this is how we actually created a fraud detection model within the flow. It’s simple. It’s not actually difficult. In fact, we presented to one of RapidMiner’s customers in Sweden, they were like, “Wow, I didn’t know that, that we could do this.” Now it’s very, very simple to just use the API to call at runtime within the flow of a bot. And it’s very simple. All of these bot technologies allow for an API call during the execution. And then we did a little bit more creative things, like if when the human in the loop comes in, we actually send the email to the human, we can actually set up a calendar invite based on data that we have using an LP and all that stuff. It’s actually using natural language generation. We can at least say, “Hey, this invoice is not actually correct,” when they come in to a meeting and pull [who?] is needed into that meeting to make sure that it’s the right invoice. So those are stuff that we created.

23:11 So going back to– let’s look at another use case. So this is a– this is a use case that we are working with a client right now, is a insurance broker based in Scottsdale, Arizona, where we want to help that broker, cover a lot more areas, geographies within the US, and also a bigger customer base than what is happening today. So they are like a sub 20, 30 million dollar company broker. So they really want to go big. They’re like only six, seven years old. They’re very aggressive. They’re growing their business. They really want to take on big. So to them, what we are creating is a chatbot, obviously based on NLP, Rasa chatbot. And when the customer comes in and talks to that chatbot, the chatbot will ask some questions, you take that answers– with those answers, we use RPA to go to about 14 different carriers who can actually provide a quote back to that bot. Right? So it could be like Hartford– many such business and insurance companies are there. They are all in a different format, so we’d actually have to treat the data a little bit, massage the data before we go on to these 14 different insurance companies. Get the quote. The easiest thing would be to say, “Okay, let’s take the lowest quote but that’s not right. Right? So we want to make sure that we actually give the right quote to the right customer. We are now training a model to understand, based on the customer demographics, very basic demographics: as their revenue, their location, the type of business that they are doing right now. But we could actually go even more granular in terms of how do we actually address it. But now we are just taking those three things and building the model so that at runtime, based on the customer demographic, we can get the quote, and then decide which of the quote is actually most relevant. And that will be decided by the machine learning algorithm. And we are looking to use RapidMiner for that [piece?].

25:18 So we have now done the top part. We’re building the bottom part at this moment. And then we hope that once that we decide which one to serve the customer, we could actually engage the customer back on chatbot. And then when they say, “Yes,” the RPA kicks back in, binds the insurance, which is basically, fill the application back and sends the email, information, all that. So this is say, a kind of– this is how an integrated, intelligent automation looks like whether on invoice side. Then this is on the front office, the invoice is more on the back office. This is on the front-office side, how we could actually create systems that are more meaningful, takes decisions at runtime. And those automation go beyond just task automation into a comprehensive automation. So that’s kind of my end of my presentation. And thank you for listening. And if there are any questions, I am willing to take right now. Thank you. [music]