• 4-8 hours of the marketing responsible to identify current issues
• Some images and videos to use in their chat campaigns
• Approvals such as operations and security teams’ approvals which becomes the main bottleneck in most cases.

We started off with the founding story of WorkFusion:

Max: I was part of the research team at MIT CSAIL Lab that was studying Artificial Intelligence and we came up with machine learning technology that could watch people perform knowledge work and mimic that work and over time learn it and become much better than humans. This is sort of like how self-driving cars work.

Then I took a sabbatical from the research and I went out to India, just to travel, and I found there are many firms in India that offer outsourcing services for knowledge processes. They call them Business Process Outsourcers (BPO) and I went to visit them like Wipro and Infosys and Cognizant. I went to visit them and I saw that these were pretty laborious operations. A lot of people working in almost like on sweatshop conditions I would say, working 24/7 doing mortgage processes, back office processing for these companies and all of them were smart people but they were wasting their talents doing this repetitive labor.

And that’s how we came up with the idea to take the research out of MIT and apply it to automating repetitive knowledge labor. So instead of outsourcing to humans across the globe and sort of taking away their humanity if you will, you can outsource it to AI robots. That’s the genesis and we started the company in 2011, so I was in research from 2009 to 2011.

Cem: How is the current landscape different from this beginning state you described and what sets WorkFusion apart?

Max: The space that we occupy is called Intelligent Automation, which is also sometimes known as Robotic Process Automation. These are software robots that primarily take over human repetitive work in knowledge industries.

What are knowledge industries? These are banking, financial services, insurance, healthcare, telco, anywhere people sit in the office and do work behind computers with documents and data. I mentioned all of these different verticals and we are active in all of them but if you think about banking you know the use cases are all across different parts of the bank. For example, account opening.

We have a customer in South Africa, Standard Bank of South Africa, it’s a $170B enterprise, almost 200 years old. It took them 20 days to open an account. They deployed the robots, it actually took about 5 minutes. Obviously, all the compliance checks and document processing are done by robots. Back office, compliance, operations, loan processing, mortgage processing are other major areas benefiting from our robots in banking.

In insurance, it’s claims processing. Processing a car damage or house insurance claim are all things that people usually do. It’s fairly repetitive and boring and the robots after learning can do these jobs much better.

In terms of the market, we are differentiated by providing two types of robots. One we call Robotic Robots which primarily do cut and paste operations. If you think about a company there are people who take information from one system and paste it into another. There are many systems that enterprises have and sometimes the information needs to be copied from one to another. A good example is when you go to a bank and you open an account usually information will be entered in several different systems. It’s the same information. So we called these copy-paste robots Robotic Robots. There’s not much judgment going on, so it’s robots just doing this mind-numbing work just over and over again.

We also have robots that we call Cognitive Robots. Essentially they are robots powered by machine learning or artificial intelligence. Those robots watch people apply judgment to their knowledge work which is should I grant this claim, what should I do with these documents, things like reconciliation. Let’s say I am getting insurance claims that claim a certain amount of items were damaged, we need to check if that correlates to the insurance policy you signed up for. Or data entry. When you look at documents like loan documents, that come as faxes or emails and then enter that information into the system. That requires human judgment and this is where WorkFusion is the strongest. MIT research that I mentioned really focused on this part which is watching humans perform that work and actually learning that over time and becoming much better. Sort of like the Go player, AI playing Go first is not as good as humans but over time becomes much better than humans, so that’s really the differentiated technology that we invented, patented and over the last seven to eight years developed with sort of $75M worth of VC funding.

Cem: You said that robots are essentially teaching themselves by watching people but there is also an interface for people to explicitly program the robots, right?

Max: Yes, that’s the role of the Robotic Robots. You can program those robots and you can give them rules. That’s the simplest case and we have a product that’s free that’s called RP express that allows you to create simple robots and program them yourself. The complex part comes when the robots have to learn because those rules are not easy to program. That’s why artificial intelligence and machine learning have created a lot of noise in the industry. It’s because instead of programming computers you can make them learn now. This allows them more freedom of decisions and it’s a much easier thing. So, you can use simple robots and program them. But there are more complex operations for which you simply cannot program all the rules. They live in people’s heads, think about driving, it’s a complex operation. Then you need to apply what is called cognitive robots that actually watch by learning.

Cem: You have mentioned some really critical cases for the bank like opening an account, if you use a Cognitive Robot for such a use case, since there is no explicit programming there can be cases where robots sees an edge case so what happens then, someone needs to check the logs to make sure that those edge cases are dealt with or how does that work?

Max: Part of our invention is something that’s called human in the loop. The edge cases are always learned by robots through watching humans perform you know more and more of edge cases right. So if you don’t program a robot, the robot learns from examples. You eliminate edge cases because examples are really the work that humans are doing. Over time the robot learns most of the examples. If there is some example the robot cannot handle because it has never seen it, it will delegate it to a human and by human actually handling that edge case one more time the robot will learn. So it works like a human in the loop.

Again think about a self-driving car, it doesn’t start out driving autonomously on its own but it watches the driver and over time learns how to handle different situations until it becomes completely autonomous. So it’s the same concept where over time WorkFusion’s technology Cognitive Robots become completely self-sufficient because they’ve learned.

Just like you would teach your new employer, you will bring him on and you would show him the easy stuff, then hard and harder, in the end, you would rely on them to do most of the things and if they don’t know how to do it they would probably give it to their supervisor but in the end they will become very proficient. Same thing with robots they become more and more proficient until they’re able to handle what humans do and much more actually right, they become superhuman in a way.

Cem: And this is quite interesting indeed, what is the current limit? For example, how do you choose processes to use?

Max: Yeah, so it’s anything that’s repetitive and lives inside the computers so any type of data entry, reconciliation, routing, chatting, these are good use cases. We actually provide a framework with our consulting partners that grades different processes. We also issue heat maps for different industries that show best types of processes that are good for intelligent automation. I would estimate that usually about 70% of knowledge-based companies’ work can be automated using intelligent automation.

Cem: That’s quite impressive. What are the percentages that you have seen with your clients? Are people really pushing it that high or they are more cautious?

Max: Most big companies are on the way. They want to achieve and I don’t know if you’ve seen the recent announcements for instance from a CEO of Deutsche bank or ex-CEO of Citigroup and then P & C Bank, they’re all saying that within the next five years or even less, thirty to sixty percent of all the jobs in the bank will be done by robots. Most of these announcements are coming because these companies are customers of WorkFusion and they have a program that over a span of five years plans to achieve that level of automation within the bank. Or insurance companies, they start smaller but of course over time that’s where it’s going to get where most knowledge work will be done by robots in the enterprises.

Cem: That’s indeed ambitious and impressive and with so much work automated, the maintenance of the bots and potentially dealing with the edge cases becomes important. On the maintenance side do you rely on the company itself or is a work split between you and the company?

Max: So obviously we’re software a company so we provide the technology and the technology has a lot of capabilities for security, oversights making sure that robots are well managed, so the software provider itself has a responsibility to provide a software platform that treats robots just like employees from a perspective of security and oversight. Of course, you have implementation partners such as Deloitte, Cognizant, Cap Gemini and others that help companies to do the implementation and first level of support. Most companies themselves that have ambitious goals of automating 69-70% of the jobs build what is called Center of Excellence internally and then WorkFusion trains their employees. We provide online training, you can go to automationacademy.com and actually take training courses in WorkFusion’s technology and so they train their employees and form a Center of Excellence which looks after the software.

Cem: And what percentages of the cases in your experience companies use these outsourced service providers?

Max: I think it’s always a blend, I’ve seen most companies use partners and their own people as the center of excellence and blend those two together. Because if you think about the scale of automation in a large company it is just a lot. Think about 100K person enterprise where 70% of those jobs can be automated, so there are many many processes that can be automated. So it’s not enough just to have your own center of excellence, you need partners to be able to help you to drive the program. So, we see a blend of these just because of the scale of the implementation.

Cem: If you shed a bit more light on pricing and ROI that would be great.

Max: You know we set out to provide a very clear and commercial model for our customers to achieve ROI. They expect at least sixty percent savings if not more so we charge what we call expert process right not robots because I see a lot of companies they think about charging for a robot which doesn’t really make sense because you actually don’t know if one robot is going to replace one person because it doesn’t work exactly that way. Maybe you need more robots because you want to do more work, maybe two robots will end up doing one person’s job because of the technical limitations. So the way we charge is, we charge per process, the process could be what I describe it could be the account opening or compliance process or invoice processing, order to cash whatever you want the process to be and we charge $25K per process per year. And usually, the process is at least work of 10 people. So it’s like 5 , 10 or 100 people so if you calculate their salaries $25K per year is probably less than one full-time person doing process that usually ten people or maybe even 100 people would be doing it, so we don’t differentiate between a process that 100 people do or 10 people do we just charge a flat fee for process.

Cem: Then of course companies start with the largest processes, I mean financially it makes much more sense that way, right?

Max: It makes sense almost at any scale if you are spending more than $25K a year on that person because you know the other benefits are not only the savings right, but if you think about it, the robots will do it at better quality than humans. Humans get tired of this repetitive work so their average quality is 85%, so they do miss keying and mistakes, robots quality is 98+% and robots work 24/7 they work every day, don’t take breaks, they don’t look at Facebook so they can process a lot more information and also you don’t have to spend money on real estate, software licenses and you can save on computers and so forth. So, if you think about it, it’s a much bigger benefit than just you know eliminating somebody who is doing this work, you’re getting better quality, a lot more work so it’s really a bargain.

Cem: It also becomes a leaner, easiest to manage, a more responsive organization as well. We have been talking about the current state of RPA so what’s in the pipeline, how do you think the industry will evolve and what will be the next phases of automation?

Max: We spend a lot of time thinking about the future. I think directionally, the robots will get smarter and smarter. Because I always say WorkFusion’s robots start from the mail room and work their way up. That means you know as they work in the enterprise just like any apprentice if you think about how humans promote their people, you start at a lower level of organization and then one day you might be CEO you never know right because you get smarter and smarter. So we see the same progression where robots are now able to take on more and more complex tasks. So you know the progression of that is I would say is theoretically infinite right, where robots will gain more and more knowledge and will be able to do more and more complex tasks so the initial estimate of sixty percent of current banking jobs being done by robots will go much higher from there and of course at that point and enterprises can become not only a lot agiler because they can make decisions faster but also much smarter, so it will only increase the capital growth of the economy.

Cem: And it also creates huge potential payout for the RPA industry as well. Do you want to talk a bit about the competition?

Max: Yes, sure so the competition is really when you think about other companies in RPA space, they focus only on robotic bots which is what you describe like that.

Cem: Not the cognitive bots you are saying? I understand

Max: Yes, because our roots are in MIT as AI company first and you know RPA is second.

Cem: Clear, but aren’t they throwing now millions into the cognitive space as well to build smarter bots?

Max: I think when WorkFusion entered RPA space a few years ago cognitive and AI was very far on the horizon, now every vendor talks about AI. But if you think about the space in general, there’s AI washing that’s going on in every part of being industry, you hear companies every day talking about AI this and AI that, in actuality it’s a very you know me being a scientist, an applied scientist at that, it is a very difficult problem to solve to be able to learn any process using AI to do learning of an arbitrary thing so it’s the problem that takes many years to solve not only on theoretical level but also on a practical level because it’s just not only money but there is time because it takes a long time to be able to train these things, so there is a time gap of several years that will need to be closed for anybody outside of us. We spent $75M not only creating the intellectual property in research but also building this. So it takes time and as you know we continue to invest more so we’re going to stay ahead of the competition but we’re happy that most vendors now woke up to the fact that you need both cognitive and robotic capabilities to be able to help enterprises automate their work.

Cem: Definitely. And a final question about automation because you have been seeing so many enterprises essentially automate a significant portion of the businesses, have you seen any successful patterns and companies that are able to upskill the personnel and manage this change in the company?

Max: Yeah, funny enough outside of like outsourcing companies, we see that most big enterprises are struggling with, doing more with less. It’s not that they’re saying we’re going to try to fire people that are doing those work, they are saying I have these people that I would rather deploy on more revenue-producing work that requires a higher level of human intelligence. Let’s robots do sort of this grunt work that’s non-differentiating for me. Most of the activities that we are seeing within our customers when they automate processes are around reskilling people and moving them on to smarter work because they’re very important to the enterprise in terms of the revenue production.

Cem: Yes, definitely. I was also thinking along the same lines but as you said outsourcing companies that’s a different situation.

Anant: AppZen has been around for about four years now. We started off building a platform for the enterprise back-office automation beginning with expense reports, initially on the employee side. We aimed to help employees easily create expense reports using technologies like NLP, machine learning and so on.

As we were progressing down that front it dawned on us that there was a bigger problem. Talking to CFO’s of our customers, we learned that there was a bigger problem around compliance and auditing, where not much had been made and the existing solutions had a human process around them. We started exploring that and about two years ago launched the AppZen AI audit, a purpose-built solution for auditing expense reports or detecting compliance misuse, fraud issues in expense reports.

It’s a platform that works with existing expense management systems, companies do not have to change what they’re using, our customers use everything from Concur, Oracle, SAP, Coupa or whatever they want. Behind the scenes AppZen works in real-time extracting data from these systems, analyzing it, augmenting it, applying machine learning to it and clearing out misuse and fraud and compliance in real-time. That’s the real power of the system. We have been out with the expense automation audit platform, the platform will be extended to offer the same services for invoices, procurement and so on.

Cem: It is interesting indeed. So that then really increases the scope a lot.

Anant: Yeah because those are really ancillary so our platform has three key segments to it.

The first one is data extraction. Essentially, we look at all the transaction documents in case of expenses, we look at everything from the expense data and the structured data all the way to any receipts, emails, portfolios, travel documents, credit card data, whatever we can get our hands on. We leverage databases that belong to the company itself, it could be the card swipe data or their HR data. That’s the set of data which belongs to the transaction, belongs to the company.

Our second piece of the platform is a data augmentation platform which extracts data from public sources, it could be reading social reviews of merchants, finding out pricing information, finding out the true nature of business to see if there any gentlemen’s clubs and things like that which are masquerading as a restaurant, finding out information on individuals to see if they’re politically connected and so on so forth. That’s a huge piece of information that we have been collecting over a period of time for every analysis that we are doing and it all gets back into the database back into the expense report. This piece mimics the human research like you and me if you were doing this manually, we would Google some stuff if you don’t know, this is what the machine is doing but at an unprecedented scale.

The third piece is the behavioral aspect of It. To figure out whenever there is a violation, what is the pattern behind it? Tells us whether it is an intentional violation like somebody misusing expenses, intentionally committing fraud or an accidental violation. It’s very important to differentiate between those patterns to figure out which ones you want to focus on, which are the groups of people or which are the individuals or to focus on, which are the vendors to focus on. So that’s the third piece or our platform.

This is how auditing or compliance is done for any kind of transaction, right now the whole system is optimized, and used for expense reports but it’s the exact same process for procurement and invoices and so on.

Cem: That’s very clear and the companies that are providing software, for example, the expense management software, are they trying to get into this business as well?

Anant: No actually, this is a completely different technology but also an application of focus that a company needs to have for providing what we are doing. All the expense management applications are focused on easing and improving the employee experience. So if you look at who uses AppZen and they’re probably a handful of people in the company who will use AppZen. They might have 10,000 employees but the people who are aware of AppZen, using AppZen are in the back office, maybe 10, 20 people who work in finance or auditing. When employee expense management companies are focused on the front end, making sure that those 10,000 employees who are using the expense management product are able to do it very easily, are able to spend the least amount of time reporting the expenses importing the travel itineraries and so on.

While ours is a back-end application, it’s not just related to expense reports but it takes care of the holistically the overall policy and compliance of expenses. It’s a different approach it requires a different mindset.

Cem: Can you explain a bit the ROI, from the company’s perspective? For example how big of an operation is it? And if you could also add in some cool things that you caught along the way, I’m sure there are plenty of expensive improprieties.

Anant: What we find in an expense report is always a surprise to the companies. Most common are people going to strip clubs expensing that as entertainment. We have seen bigger compliance issues around companies or the employees entertaining guests from companies which have been banned by the US government. We have seen violations where managers are not looking at what their employees are doing, expensing everything from yoga classes to spinning classes and nobody knows. All the way to companies or individuals intentionally duplicating their expenses, claiming the same expense multiple times.

So, these are the kinds of things that happen. It’s very normal and we end up finding that in many companies. It’s always a surprise to them but for us, it’s part of the business that our system is this, is getting smarter and smarter to figure out different kinds of fraud patterns. We see even things like on travel itineraries, people are booking their spouses and flying on that, canceling and claiming the money and things like that right? So, there are plenty of violations, I won’t name any companies or get into details of that. In terms of ROI which is key here, why would somebody buy our software and use that?

So, the key things are basically 3. One is using software like ours allows them to have 100% auditing in real time. If you see today most companies, if they want to do any kind of auditing, any kind of review of expenses it has to be manual and it’s very expensive. I mean employing 10 people to look at your expense reports is very expensive. So, what most companies do is, they routinely audit only about ten, twenty percent of their expenses. Most of the time to reduce costs, they offshore that function sending it to India, Philippines and so on.

So as a result of AppZen doing that audit, we take that company directly from a 10-20 percent audit to 100% audit. Instead of having a week turnaround or two weeks later somebody calling you telling you that hey, there is an expense report problem out here, this is happening in real time within minutes of an expense report submitted. So, in terms of ROI, a reduction in the headcount or number of people required, we have seen anywhere between 50 to 80 percent reduction in a number of people that are required to process expenses, as a result of using AppZen.

The second one is a much higher ROI point and is about T&E expenses. As a result of AppZen doing 100% audit, taking your expenses from a 10-20 percent audit to 100% audit is a significant fivefold increase in the number of expense reports that are audited. But more importantly, our detection rates are so far higher. We are able to find stuff on things that have already been reviewed by their auditors, reviewed by managers and paid off.  Because machines can do so much more research on every expense, compared to what a human can. So, our detection rates of far higher and we have been able to find out violations anywhere between one to five percent of expenses. So, if you look at like at T&E expenses that happen in a company, taking a few points of that expense can save company millions of dollars. That’s probably most in terms of value, highest value in terms of ROI.

And the third is compliance. Now, you can’t measure that. What would happen if you found your company to be non-compliant with regulatory requirements? We have regulations for financial companies, pharmaceutical life sciences companies, we have gender regulation like FCPA compliance, anti-bribery compliance, all those kinds of things. So those you cannot measure them, most companies are not doing anything around it in expenses because it’s impossible to look for every attendee, find out their political exposure, whether they work for the government and things like that. But if you are caught in some violation then the cost is immense to the company.

So that’s the third one, where we are protecting things by having an active compliance. It is real-time so the if at all there is a problem happening around compliance we know right away.

Cem: I expect when you first move in, there are many cases you discover but over time I expect people to get smarter and understand the how the current system works. How do you cope with that?

Anant: The premise is that also employees are honest, they’re going to do the right things. But if you don’t detect anything and they if they find that there is nobody policing these expenses, then they’re going to have people who are trying to be opportunistic sometimes like let me slip this one through. Our goal is to change the behavior of the employee, if you change the behavior of the employee makes them more compliant, then automatically your T&E expenses are going to get reduced because they’re going to be less leakage, fewer violations. So, the way we do it is by first real-time auditing it and then measuring employees’ behavior.

One of the key aspects of the AppZen platform is to detect the behavior and metricize it. So, we have something called as an Appzen Behavior Index (ABI), think of it like a credit score for employees. Just like we have a credit score which shows creditworthy we are, an AppZen customer has an ABI score for every employee. The ABI score is very simple, it is between zero to hundred and basically tells you if this employee is compliant or not. If they are compliant then, they’ll have a score which is lower, the lower the score the better. If their score is more than 50, then it there is something wrong in that compliance. And there is a complex model behind the scenes which calculates that index for the company and typically what we see is that, in most companies, 90% of the employees will be in the 0-10 range, in terms of scoring and then there will be a few outlier employees in the 50-plus range.

Now, why are we doing that? Because like you said, if we start finding violations in real time the employees are going to know about it. The employees won’t know how somebody is detecting doing this. However they will see that any kind of violation that they do, they are being detected as they get messages from my system which warn them like we found that you upgraded yourself to a business class, our policy does not allow it, don’t do it again. Most companies will not go back to the employee for a refund unless it’s a very serious violation. Most companies will say don’t do this again, otherwise, we are not going to pay you. But the goal here is to police it in real time and to let them know. When the employee is told all the time that this is not allowed, we are letting it go a couple of times, they will stop doing that behavior. However here’s some who will not stop, that’s why we want to measure it.

So, as we look at a score every month, you want to see a reduction in the score. A reduction in the score tells us that this employee is changing his or her behavior which means the violations are going down. We come to a point where the scores come down because the employees know about it and the employees are not indulging in that kind of behavior. But then there may be a few outliers who continue to do that and then those have to be stopped either terminated or the manager talks with them and so on.

Cem: One thing that I hear or think most companies are, how they use their data and yeah, I find that exciting when companies can use most of all customers’ data in an anonymized aggregated way. I think you have an opportunity to do that as well, right?

Anant: Absolutely, that’s key. Our system always learns different patterns around expenses. So, why are there anomalies in these expenses, in these business meals or entertainment? How often do they upgrade? What kind of upgrades are consistent? The system is figuring out where the anomalies are vs just finding out there’s an anomaly. Some insights are standard across the industry where our system continues to learn them while some insights are specific to a company’s culture. That remains within the company’s domain itself.

So, we have two kinds of learning: one learning across the industry and a second which is just for that company which tells us that though this policy states certain things, this company is very lenient, you don’t want to find certain things, company is okay with certain kinds of behavior. These are unwritten rules, we have to find and learn them because we can’t keep on detecting things which nobody cares about. Those are the two things that we always look for. One is across all our customers’ data, new fraud patterns, new tolerance limits which keep on getting aggregated and some things which are very specific to the customer itself.

Cem: And can you talk a bit about the set-up, how long it takes, what’s the pricing?

Anant: We work with almost every major expense management company out there. As I said Oracle, Concur, Coupa, Expensify, Chrome River, we are integrated with their expense software. Since AppZen in is already connected to these software, customers don’t have to do anything different, there is no IT involvement, they simply turn on the button and the data starts coming from these expense management software into AppZen. From an implementation point of view, there’s no friction at all, it can be turned on within a few minutes.

Cem: No fine tuning at all, from your part?

Anant: That’s the installation or connection part. We start off with a template of patterns that we already know. So, we don’t start off with saying what do you want us to find? Most customers don’t know what they want to find. We know what to look for, we already have a very good pattern detection template. So, we start off with that template for every customer and we calibrate it based on the policy that they have published. So, we configure the company’s policy into our template.

Once we do that, we have a period called a parallel run period. It typically goes on for about four weeks. For smaller companies, it is about two weeks, for a very large company it might go on for four weeks. Essentially that’s implementation because, during that time, AppZen is running behind the scenes, real-time detecting everything while the company is still following its current process. All that they’re doing is validating what the AppZen does. Are we finding few things? Are we finding too many things? Do we need to tone down some of the findings because they are good but not something that is important right now? We basically figure out what are the top three things that we want to look at after we go live, which is up in four weeks and then slowly go deeper and deeper, turn the knob, so that even the other things which we initially didn’t look at those things get looked into. So that’s our typical go-live period, it takes about four weeks for a customer to go live.

Cem: And how do you price it?

Anant: Pricing is based on transactions, so we only charge for the number of expense reports that the customer is processing. So, we look at hey, are you doing hundred thousand expense reports a year, then here is your price. So, depending on the volume, we will give them a discount.

Cem: Can you share the price for a small customer? Just to give me an idea.

Anant: Yeah so for a small customer, it could be around four dollars per a processed expense report. We’ll discount it based on the volume that they have.

Cem: Okay then for a large company with like hundreds of thousands, there will be significant discounts, right?

Anant: We have customers processing million expense reports, Fortune 50/100 companies. The fact is that the system has such a good ROI because there are no soft savings, this is all hard dollar savings that the company gets within a few months. Companies easily get a very high ROI with us.

Cem: What do you think of competition and what do you think sets you guys apart?

Anant: We don’t have any competition today. We compete with essentially company’s internal T&E auditors. Such teams normally have a team of 10-20 people in T&E, we are basically doing that function. We don’t look at them as competitors because we are essentially giving them a tool to do this job much better and use their time much more productively. We haven’t come across anybody attempting this, there are lots of robotic process automation (RPA) tools, who try to mimic how human auditors work and we have customers who’ve tried these leading RPA tools. Nobody can build a generic tool and use it for a specific business function. It just doesn’t work. You can’t take an RPA tool and train it to do certain things. This requires a purpose-built solution. We have seen numerous companies take off-the-shelf products and try to use it for T&E and it never works. Our detection is so far higher. That’s the kind of thing we see, it’s not a real competition but it is people trying generic tools and trying to build teams in-house who can do limited learning so it never works.

Cem: It sounds like such a good business because you’re saving hard dollars and for you, once we have the product, the variable cost is relatively limited. On top of that these hard dollar savings, you deliver labor savings to the company. It looks like such an attractive market to enter if you have some data to train on and if you can be able to build a good product.

Anant: It’s not just the product, it’s all the surrounding things around it. We are product agnostic in terms of what we use. The product has analyzed so many different industries and customers of different sizes that the data that we have is a big part of our value. The patterns that we have built is not based on 100, 200, 300 thousand expenses. They are built on tens of millions of expense transactions that have been classified. So, as we process more and more we are getting better and better. We are able to achieve things which are impossible for those to replicate and we will always be at the forefront of things, we are going to be the holistic solution for auditing everything in a company. Starting with expenses, we’re just taking it around to all the other verticals in the company.

Cem: Yes, procurement also sounds very similar.

Anant: It is something that our customers keep asking us, it is just matter of time when we start focusing on it but that’s where every customer keeps asking us, well if you can provide the same tools that we have today for invoices for procurement.

Cem: Great, thank you very much for your time.

Oliver: I’m one of the four founders of ViSenze, which is a combination of two words: visual and sense. The objective is simply to bring together the visual web and try to make sense out of it using computer vision. We spun out of the National University of Singapore about five years ago, where we worked in a lab that was set up by National University of Singapore as well as Tsinghua University in China – one of the top three universities in the region. We built technologies that focused on areas that Google wasn’t looking at, one of which was computer vision, a very nascent field at that time. We taught machines to process pixels and understand concepts within images themselves, and we played around with a lot of social data that we gathered from the visual web in China as well as in US. This allowed us to understand not just objects, but also concepts in images. The problem statement that we focused on at that point in time was very simple – if we can extract intelligence from pixels and images, what can we do to help online shoppers who are basically telling us that they are searching but they’re not finding? Why is it that they are searching and not finding? Is it because they’re using the wrong keywords or is it because the keywords that they used in the first place did not exist in product taxonomies that merchants and retailers had in the first place? The answer is actually both, so therein lies the huge disconnect between the way we describe things and the actual products themselves that are being tagged by merchants and retailers.

If I could show you a very nice cage-high heel ladies sandal or shoe, how would you even try to visualize that? You can see it but if you were to show the same shoe to ten women in a room, they’ll all use different descriptions like “cage lady sandal shoe,” “strappy lady sandal shoe,” or “gladiator lady sandal shoe.” We all use different terms that are most familiar to us but not necessarily the most appropriate ones in order to match with the product taxonomy. So, we figured if we could take away this hassle of keyword guessing, and instead just use pure images, we would be able to shorten the path towards search as well as discovery. Now I always like to say that a picture is really worth a thousand words, but you just don’t want to use a thousand words to describe the picture. Instead, you should let the image speak for itself. So, that’s exactly what we do. For example, across the border, products in Japan are tagged with Japanese metadata and keywords. How do I even try to look for a cage lady sandal in Japan if I don’t speak Japanese, right? But images themselves are a universal language, so using visual search, we can take that image as an input or as a query, process the metadata identifying what exactly is in that image, extract the visual attributes (in apparel, that would be things like color: is it a mandarin collar? Is it a standard collar? Etc.). We extract all of these visual attributes that allow us to approximate the actual result as if you were the real shopper in the real world looking for that product. It’s kind of a like an experience when you walk into a store and you say, “Hey, I like the pattern very much but I don’t like the cut – I want it shorter.” You can describe that in the real world, but in the online world, you don’t have that benefit. So, by training machines, we’re able to approximate that – we call it our “nearest neighbor” algorithm. Because you may find a dress that’s short but the store also has something that’s much longer, so we try to approximate the entire product as if you were looking for the real stuff in the real world.

With that, we’ve been able to apply ViSenze’s technology fundamentally starting in the fashion vertical because fashion is the most visual vertical amongst all of the e-commerce verticals that we looked at, and it’s very easy to identify that the technology helps to be able to search better. About three years ago, we transitioned from machine learning to deep learning and that basically gave us the quantum lead in terms of not just the accuracy but the overall relevance of results that we’ve been able to generate. We’ve been playing around with different metadata, which is not just visual, but rather input that we extract from pixels and images, in order to recommend better and more relevant items. For instance, if you’re looking at a $50 dress, does it make sense for me to show you a visually-similar dress that is worth $500? That’s not relevant, even though it’s visually similar, so we’ve been able to take in a lot more signals as compared to just visual signals and generate more relevance in contextual recommendation.

Moving forward very quickly, we’ve moved beyond fashion. ViSenze is now into home and décor, furniture, CPG (Consumer Packaged Goods), schematic drawings, artwork, product designs, stuff like that. They’re very visual in nature itself and have a distinctive pattern to the image, so we use that and we’ve been able to apply that. Currently, about 80% of our clients are in the retail space and about 20% are in a non-retail space. So, to give you a sense of the kind of customers we have on the retail side, they’d include brands like Rakuten in Japan (the number one marketplace), we have Uniqlo Japan (Japan’s largest fashion retailer) that uses us on their mobile platform. We’ve worked with three of the top five e-commerce players in Southeast Asia, two of the top five e-commerce players South Korea, and currently have two of Europe’s largest fashion brands using us. I can’t name this because we are under strict NDA specifically for them, but I’ll be very happy to show you some sites that use image search without putting names to it. And of course, we’re working with some major retailers, as well as market places, in North America right now.

The team is still very young. We have about 40 people on the R&D side, and we probably have one of the largest concentrations of computer vision and machine learning guys under one roof in Singapore, where we are headquartered. Our underlying stack is completely built on deep learning today. And of course, the other 20 percent of different use cases (outside of retail) that we looked at – trademarks, logo detection, artwork and product design, schematic drawings – those are very interesting areas that visual search can be applied to. So, what I just described to you, it relates to one branch of our solution that is visual search, on the image recognition side. Image recognition is basically transferring or converting from pixels into key words. We have our product tagging APIs which is in the beta stage right now and is commonly used by some major retailers to enhance or enrich the keywords in their catalogs. For instance, if I’m able to detect a mandarin collar, we can then expose that color as a keyword description which allows for me to search for a jacket with the exact phrase “mandarin collar.” We use that in order to expose what we call “fashion attributes” in the fashion vertical. When used in fashion, they enrich the catalog of all these major marketplaces, allowing them to have their products found better using the natural keywords that everyday shoppers use.

Cem: Thank you. I think we put you under image tagging and visual search and those are the two topics that you mentioned so far, right?

Oliver: Yes, that’s right.

Cem: But I’m sure we missed some things because it’s difficult to categorize the whole space but it’s good to hear that the we’ve got the important ones right.

Oliver: These are the two most important verticals that you put us in, and the other two verticals you might want to take note of. We also offer product recommendations. Visual search and product recommendations are different engines, fundamentally; the underlying logic is different. We work with the different partners and different players who use our technology as part of their product recommendations. Instead of a consumer uploading an image to search for something, while he or she is looking at something on the site itself, the product recommendation system uses visual inputs to make a recommendation to the consumer.

The last area that we’ve been working on is what I call “video commerce.” The idea is very simple; if we can process text, we can also process videos. Since videos have a lot of intelligence and signals, the chances of extracting signals out of videos, like from fashion video shows, are a lot richer. The ideas can be exposed, and product recommendations can be shown back within the video stream as we detect them. Therefore, I call it a video commerce experience; instead of creating shoppable videos, we make videos shoppable using machine learning.

Cem: That’s a pretty interesting use case, and there’s plenty of existing material to use. So, that’s interesting.

Oliver: Yes, it’s a very new era, only a few guys are out there in the market that I know of that are doing it. We have a lot of people creating shoppable videos the old way: embedding products within videos and designing videos for products to be embedded inside and exposed. But the newer way, using machine learning to expose shoppable moments within videos, is few and far between.

Cem: You are essentially doing image tagging on the video, right?

Oliver: It’s a combination of both image recognition as well as visual search on video. Not only do we detect, we generate an exciting shopping use case, so we recommend, or we do a visual search against a product database in order to find relevant products. It doesn’t just have to be fashion, it could be other stuff as well. It could be a branded product, such as a logo, or it could also be used for contextual advertising purposes. We call it “in-video contextual advertising”, and it’s a very rich area that has yet to be fully explored.

Cem: And can you talk a bit about industry landscape, about the competition, large and small players. We can also discuss pricing and roll out times for your products.

Oliver: Okay, let me take the first one which is the landscape. Visual search is already appearing on mainstream search today where we have major players like Rakuten implementing visual search on their platform. It really heralds the mainstreaming of visual search as part of the natural search experience for any consumer. My vision is to see democratized visual search and have visual search as common a feature as keyword search on any shopping site or any shopping platform.

Now with that, I see the market place has all sorts of visual search solutions out there, and even platforms like Google utilize it. But here’s the challenge; whereas general image search has brought itself to a sufficient level that stretches right across every vertical, how does one get deeper, sharper data and understand consumers’ behavior better? This is where companies like ViSenze come into business, because fundamentally when we train data, we train much sharper data, and we train on real data that is provided by merchants and retailers that we work with. There is a difference because working on real world data and understanding the way that people shop and whether they convert things like optimized prices, or don’t convert, signals the optimization that is required to understand actual consumer behavior. General visual search as a worldwide search engine may find the product you’re searching for, or may not find you that product. Fundamentally, they’re not optimized for conversion, and that’s where the foundational disconnect between companies like ViSenze and worldwide web search engines like Google, happen. Since we are vertical, we call ourselves “Vertical AI agents” that are heavily optimized for a specific function, like engagement, conversion, or discovery. In terms of the landscape, you will see that there are a few companies like ViSenze as well, and we are not alone in focusing on key verticals. We look deeply into things such as visual attributes of fashion apparels. Earlier,  I talked about mandarin collar, that’s an example of how we do visual attributions. When we were in India, we thought that we understood modern fashion wear, but didn’t realize there was a niche for ethnic wear. For instance, what’s a sari?

Cem: I’ve worked in India for a bit so I can imagine the challenges, completely different fashion wardrobe.

Oliver: Yes, that’s right. Your standard outfit concept in ethnic wear doesn’t really work because an ethnic wear or sari is considered a full outfit, there’s no upper body or lower body concept, it’s just the entire outfit. You have to retrain for that domain understanding. That’s what gets us further and deeper than most other people when we look at other specific verticals.

Cem: Yes, totally. What do you think about Amazon? They have plenty of data for sure.

Oliver: Of course, the big guys have the advantage of data. Amazon just a couple of months ago launched Echo Look, have you tried that?

Cem: No, no. How is it?

Oliver: It’s alright. Amazon’s Echo Look is a device that you place in your bedroom and every morning before you go to work, you are supposed to ask Echo Look, “How am I dressed today?” It’s meant to give you an answer, or a recommendation based on how you could dress better. Now, I honestly wonder how useful it is as a real-world product. If someone has been dressing well every day for 35 years of their life, why would they ever need that product to tell them to dress better moving forward?

I see this product as a fashion statement that Amazon launched, but I think the real use case will be one in which we use AI to help people discover things that they don’t already know. For instance, if I bought a new scarf, is there a better way for me to wear this scarf? What color combinations does it go with? It’s about sharing good suggestions and then letting the consumer decides at the end of the day. It’s also about equipping and empowering consumers with visual knowledge that they would not otherwise have. I see this as a large opportunity for AI given the data that we have, and we will be able to use that at scale for consumers.

Cem: You work with plenty of these big e-commerce companies but you also have a solution that is possible to use as self-service; you have the API, with these large e-commerce companies is it more like exposing the API to them or more like consulting work where you work with them to make sure that their solution is tightly integrated with the systems they are providing?

Oliver: I’ll give you two answers to that. Most of our clients use our APIs and these are easily integrated and implemented. Some clients use our STKs that we provide to them as well, and it’s pretty standard.

At ViSenze, what we built is a highly trainable and highly configurable algorithm model. On top of the models that we have, we can take in feedback and data directly from clients and retrain that model extremely quickly. We can use a client’s data in order to optimize to their environment and their use case. For instance, in North Asia unlike in the U.S., a lot of their images are visually noisy. It’’s not uncommon to see their catalogue images having two or three models in the same image itself, so we are optimizing for their environment.

Although ViSenze is a cloud-based solution right now, we’re moving towards an on-premise solution where we can actually deploy or transfer our entire algorithm model within a closed data environment. This will let clients feel more secure having their data updated and shared within that environment. So, we have two models to go with.

Cem: Very clear. And once you have the model set up, one thing that I am wondering is how frequent is visual search use by the customers of the e-commerce websites? I’m sure it depends a lot on culture maybe or about the time it was introduced, but you know so much more, so whatever insight you want to share, I’m curious.

Oliver: That’s a very good question, I’m glad you asked that. When we started ViSenze five years ago, of course there were natural skeptics; does it actually work? Is it a natural consumer behavior? Five years later, when we have guys like ourselves and even Amazon using visual search on their platforms, it’s almost as natural as keyword search.

I’m going to share with you some stats. In 2016, we processed more than 350 million queries in the whole of the year. This is almost equivalent to one million searches a day from all of the clients, who are then exposing our APIs to their end consumers. In the first half of this year, we have seen that volume already increase three times. I am seeing a lot more visual search happening in one form or another, whether it’s an upload search experience or a product discovery experience, or recommendation experience – it is happening at scale.

Cem: And how do you see it changing in the future? Are there any trends you’re already seeing?

Oliver: Yes, in fact I’m seeing a lot more people using social media images – Instagram and Pinterest. In China, we have images from Sina Weibo, which is one of the top three social media platforms. People are not just searching clean catalogue images that they come across on other competitive websites. They’re using social media images – from Instagram, Facebook – that are being shared by their friends or Pinned, reposted, and liked. I see a big influence coming from the social media space and companies like Instagram and Pinterest are catalyzing and propelling visual search forward faster.

Cem: Well, we’re almost running out of the time we allocated so if you have any thing that you wanna share as a final message to potential end users we can talk about that.

Oliver: My favorite line is, “the future of search is visual,” one way or another it’s coming. Mary Meeker talked about it her 2017 Internet Trends Report, to be fair she said “visual as well as conversational” but it is there. We believe that visual search will become even more mainstream moving forward, but it is not just in shopping – visual search can be applied in many areas outside of e-commerce. We’re heavily optimized or commerce which is big enough for any of us to focus on already.

Anil:Let me give you a little background about what we are doing at Absolutdata and what role AI plays in it. So, unlike a lot of other companies that you would probably speak with who would be startups that are building AI products. We are actually a very different company. We were founded about 16 years ago in 2001 with the goal right from the day one was that we wanted to build tools, technologies and analytics to help companies make better decisions. That was our mission and we’re still very focused on that. However, when we started the company at that time there was not a lot of analytics being used as much. So, we actually focused on doing data analysis for companies using traditional statistical methods. That’s what we have done over the last 16 years. We have about 40+ customers now, about 350 to 400 people at this point of time. Headquartered in here, in the San Francisco Bay Area and at a big part of our team sits in India, even though our customer reach is fairly global. Our focus traditionally has been on the marketing and sales function of an organization. We also are funded by the private equity arm of Fidelity from Boston and in fact, we used a lot of that investment into building up our AI based products.

To give you a little history of how we came into AI and how we started using it. About three or three and a half years back I had this epiphany where I realized that the analytics that we had been doing, all the big data and everything that people talked about for quite a while wasn’t having the kind of impact on business results as it should be having. The gap that we realized was in how analytics were getting implemented. Typically analytics tended to be more like as services organization either within the company or working with companies outside. It was very good at providing insights. However, it was not doing a very good job of converting or translating those insights into actions the business teams should be taking. So, we decided to take a step in that direction and said we will integrate technology very deeply into what we do at Absolutdata and build solutions that not only provide insights based on the data but actually recommend actions to take. As we started working on that business problem, very quickly we realized that the traditional statistical methods which typically tend to be more focused on regression kind of methodologies are not the best methods to be used in a situation where you are going to be providing recommendations. So, that is when we started going and using machine learning. We started building with some deep learning techniques available at that time. But the need for using AI into what we’ll building actually came for us from the type of solution that we were building rather than saying AI is a cool technology and let’s use it.

So, what we have done is, we’ve built up a platform which is called Navik which means navigator in Hindi. There is a Navik component which targets the Sales function of a B2B organization and there is a component that targets the marketing function of a B2C organization. We also have a third piece in there which is focused on introduction of new products but let me focus more on the marketing and sales pieces. To give you a sense of what Navik can really do, maybe I can spend a couple of minutes just talking about our SalesAI product, Navik SalesAI as we call it. So, if you think of a salesperson, a salesperson today is in a situation where there is a lot of information that is provided to him/her. However, that information does not guide them to what they should be doing. There are some companies that are certainly providing some guidance but most of that guidance today is focused on prioritizing which particular leads a salesperson should focus on. However, once given a lead, the salesperson is supposed to figure out on his/her own what should they do about that lead. There might be some guidelines that sometimes are available but beyond that everything is left as the responsibility of the salesperson to figure it out. So, what we did was, we said let’s look at what is the day-to-day life of a salesperson and let’s build a platform which works like an assistant to the salesperson. So, what Navik SalesAI does is, it literally creates a weekly schedule for the salesperson and what I mean by weekly schedule is, it goes, looks at all the data that is available in the organization particularly in the CRM system and in the ERP system, it looks at any data that is available about outside information, for example one critical component that we bring in to Navik is the various news information that is coming out, announcements and so on, and based on that it does three things.

First, it identifies new opportunities. Opportunities that might not be included in the leads that they already have. That’s a very critical piece that it can actually come up with leads that are not part of the leads that you’ve been given. For example, It can look at your current set of customers based on their past buying behavior, predict what are they likely to buy now. So, it creates new fresh opportunities in the pipeline. It also then prioritize all the opportunities that are available to you for that particular week. So, it says for this particular week, here are the top ten opportunities that you should focus on. So, that’s the step number one.

The second step that it does is, for each of these opportunities, it creates the optimal path to closure. So, it identifies sales actions that the salesperson should take. So that they can close that opportunity in the fastest possible time. This information is informed by the past history that you might have of that particular account or of similar accounts in the CRM database. For example, you would know what activities actually correlate with a deal going through the pipeline faster at each of the pipeline stages. You might have information available from other accounts that are similar and most importantly you have information available from best sales people in the team, in the company. They’re doing certain sales actions and through AI and through our models we are able to learn from that and recommend what will be the best closure path for this particular account.

Cem: Could you provide a few examples to these actions on the path to closure?

Anil: Yeah! So, actions on the path to closure are things like setup a meeting,  invite somebody to a webinar, meet additional people. So, in this organization you need to connect with four people to close this deal and here are the four people that you should be connecting with. It optimizes and figures out what is the best sequence of connecting with these people. Should you start your conversation with person A or person B and when should you bring other people in the conversation. It looks at all this information and essentially puts together the path that we are able to discern from the data so, that account can be closed by a certain date.

Cem: That’s quite interesting indeed. You are using only that company’s data to estimate this, right?

Anil:That’s correct. We just use any data that they have.

Cem: Because what could also be interesting is, if you had multiple companies that are interacting with a possible customer but about different products then it could be, I don’t know if this would be possible to anonymize and so on but if you could pull together those pieces of data and build an understanding of the company, that would be interesting, right?

Anil: Absolutely. you know, I hope we can do that someday. The challenge that we’re facing right now is about being able to pull out all of that data together and be able to leverage that data for that company but yes, that is absolutely a possibility and it would make the recommendations significantly more better but at the same time the I think the good news is that even within the data that you have, just using the data at typically companies have, you get very significant improvements in closure rates through this. So there’s a lot of learning already sitting in the data. Of course if we were able to combine outside data, other interactions, then that would become that much more stronger and better. So, that was our second piece which is for every account: Tell me how do I close this account.

Then we have third piece which is every contact that you have with the account, we recommend what that engagement should look like. If you are going to be doing an email what should the message say, what product should you be talking about in this email. It also recommends whether you should do an email or an in-person meeting. There’s lot of other information that sometimes companies have, sometimes don’t. For example one of our clients also provides guidance about if they have a contract with that particular customer, how much of the contract has been used. So, there’s a lot of information that is provided to the salesperson so that they can make that contact in the best possible way. The salesperson still has to go and do it. They still have to have that conversation but we will provide, for example a cheat sheet of what to talk about in that conversation. In one of the situations, we also recommend what is the most likely competitor they might be considering for this product and how do you counter that and what information do you provide. So, the whole idea is that we want to make the sales people successful in what they are selling and for that, provide them guidance on through the whole process of what actions to take, what is the sequence of actions that you should be taking and where and which particular account should be focusing on doing this. So that’s essentially at the end of the day, the core of what SalesAI does. It is not just a machine learning system that just tells you what you do. What we have also done is built a lot of interactivity in there which is that what we’ve done is we looked at how a salesperson plans their week and in the tool, we provided them ability to actually plan their week because just providing recommendations is not enough. This is a tool that a salesperson can use to plan their entire week. So, they can add opportunities on their own. They can take some of the recommendations and say I will not act on it this week, I will do it next week. They can invalidate a recommendation and every time you are doing this our algorithms in the back are learning that and so that, next time we are making those recommendations, it will take this learning into consideration. So that’s what our

SalesAI platform is.

The other type of platform is MarketingAI platform and I will do a fairly quick version of that. The MarketingAI platform has a very similar thinking as we have over SalesAI platform, except it recommends what campaigns you should be running in a particular week. So, marketing companies that do a lot of campaigns on Facebook, email and others have a very traditional model of doing campaigns. The traditional model of green campaigns is, I come up with a campaign and I find out what’s the best set of customers I should target with that. What we do is, we conceptually look at each customer and say what is the best campaign for this particular customer to run and aggregate from that to create the campaign’s that you should be running. So, we would be, for example at each customer level, predicting what is the next product they’re going to buy and when are they likely to buy that product and how would that get impacted by a promotion you are running. So, we then find the best combination of promotion that will get the customer to buy that particular product. It uses mostly the data that a company already has on their customers and is able to provide recommendations pretty strongly.

Just to give you a sense of impact on the SalesAI, we have done it really tight A/B test with one of our customers, it’s a  10 billion dollar company out of Chicago in the B2B space. We were able to give them a four percent increase in sales in seven weeks of testing. This is for a company that for the last three years has been flat or negative in terms of sales. So, that’s the impact that we’ve been able to create there and again, a really clean A/B test.

On the MarketingAI side we’ve done a pilot test, A/B test with one of the largest hotel companies in the US. There, our recommendations produced a %51 of the increase in revenue over their traditional methodology. So that’s the power of what you have there in the AI. The goal that we have for our platform is, build a platform to become like your AI assistant for what what you should be doing at your job. That’s what we’re trying to build.

Cem: That’s pretty impressive and it’s also pretty quick to set up I guess, thanks to integrations, right? Because in seven weeks you were able to get pretty solid test results.

Anil: Yes. Again, the both the platforms have been built with a view of quick integration with existing systems. So we’re not trying to replace any of the existing systems the companies have. Because most of the existing systems and existing technology tends to focus on execution. So, on the marketing side, you have a lot of CRM systems which are very good at executing. What we do is, we sit on top of that and make recommendations on which campaigns to run or on the sales side, what the salesperson should be doing. So, we work with for example Salesforce, Microsoft Dynamics, and integrate very easily. So, we built those integrations. The other thing that we’ve also done is, our analytic systems in the back, are also built in a manner that, once you bring in the data, it literally is a three-day setup for us. So within three days everything is set up and you can get up to speed and running. Again, typical setup doesn’t take only three days because there is a time involving getting access to the data, mapping the data to our system. So, it only takes about three to five weeks to do the setup but at the same time the whole goal behind all of this is being efficient in how we do that. In fact, we actually use AI even in our data mapping. This is part of the philosophy that we have at Absolutdata which is that we want to become an AI first company which means that we’re not going to use AI only for our products, we use AI in everything we do. Even, when I’m talking to our HR team. I am looking at opportunities how to make on HR decision-making better and faster through AI. Is there a tool that we can build that can do something for our HR team. We actually have a lot of small things that we’re looking at throughout the entire company and trying to build AI into every little aspect of what we do and make this very core part of our DNA. So, that’s a very exciting journey and I’m actually very much looking forward to just not having AI on the product side but also on how we’re doing things internally.

Cem: Yes. it’s quite exciting indeed. Finally, could you talk a bit briefly about what you can on the pricing aspect?

Anil: Our SalesAI product is priced like as a SaaS product so, it’s based on per seat. The pricing depends on the number of seats you’re assigning, so tends to be significantly less than what clients typically pay for their CRM license. It’s usually half of that. On the marketing side the pricing is based on the number of campaigns that you run or through our platform and again, the pricing for a campaign is between 4000$ a campaign to 8000$ a campaign. But again, all of the pricing we’ve done in a manner where you can test it and start running it without making a big investment and without having to make a big commitment up front. The whole goal is, you use it, you see the value and as you see the value you scale up the usage. Then you start paying more as you’re scaling up.

CEM:

Hi Nikunj, thank you very much for your time. I wanted to learn a bit more about how you view the industry and unique value proposition of Falkonry.

NIKUNJ:

Thanks for asking, Falkonry is pretty much focused on industrial predictive analytics and this is is a field that has gained prominence in the last, I would say, two years, but we’ve been doing this now for five years at Falkonry. The reason we are doing it is because of the strong understanding of the industrial sector and its need to improve on reliability, productivity, safety, as well as efficiency. What we saw was an industrial and analytical state which was predominantly dependent on highly trained people who can complete manual analysis one at a time. Now I won’t blame the people for this it is also a result of the lack of investment in technology in the industrial sector in general. But then companies like GE, Schneider Electric and the German governmental institutions made a lot of noise about the need for better data analytics technologies and greater software investments. I think the whole world has sat up and taken a look at what is about to come. I think this picture that was put together by AGC partners. Actually, does a pretty good job explaining what is the spectrum of analytical technology that should be considered by customers. So, first of all, think about manual analysis as people putting together either MATLAB or Excel spreadsheets to solve individual problems that get reported to them. Most companies that are operating at scale have professional manufacturing or process engineers whose job is analysis and they solve individual problems. They put together calculations based on their knowledge of the area or of their own systems design. They try to solve those very specific problems. Therefore, they are very intricately familiar with the systems that they work with. It works for some problems but typically the cost of solving any such problem can be in the million to two million dollar range. So, naturally you cannot do this for every problem because industrial world will present you hundreds of thousands of problems.

In the context of oil and gas distillation columns and you’re trying to create an equation to determine what temperature you should perform the distillation at or what volume you should use for the given fluid you’re processing. In those kinds of situations, people have developed virtual physical twins where an engineer has constructed a productized version of an equation that can be parametrized through specific physical character criteria that are relevant to a specific application of that physical twin. Then they can operate that physical twin in very regulated conditions so that its results can be relied upon for operation. Now it’s a better model than the manual analysis because it is productized and therefore can be used in more than one place. But, it is limited to those situations where a governing law or a governing set of equations apply and you can control these systems’ behavior in a very tight manner. We’ve seen these in oil and gas pretty extensively especially in the upstream processing. Now this is also quite common in some areas that, for example, are operating turbines for power generation and hence have very tight control parameters. Companies like GE, have made a pretty big deal out of those twins that they are offering because of the decades of experience they’ve accumulated and the number of models that they have developed over the last twenty-thirty years. And it’s a great model because it actually gives operators the degree of confidence to operate on the basis of manufacturer provided models. So it’s great. The problem is that it requires those tight controls of operation. And that is not possible for at least these days sixty to seventy-five percent of the systems that are being operated. The world is changing too fast for that. I’ll give you one simple example; turbines like industrial gas turbines may have been designed for base load and are now being used for peak generation. So therefore their operational state is going to be very different from how these turbines have been designed. And therefore there may be no great physical twins available for industrial gas turbines when they are operated at peak power generation.

So therefore now people have realized that a physical twin is not going to be enough and also that physical twins are also expensive because they have to be created by the manufacturer. So, for these two reasons now people have realized that they need to look at the virtual statistical twin. Now there are some companies, especially most startups that have come about in the last five years or so that have made a business out of creating virtual statistical twins. It requires data scientists to find the right kind of data to collect to create the right kind of data science pipelines and to put together whatever knowledge they can glean about the machines operation to figure out what is the right statistical twin which is basically an empirical model built from data to predict or evaluate the condition of a system. It requires a large staff of data scientists and some companies are able to do that for very specific problems. One example that I’ve seen is that there is a company that put together a patch to go on your heart so that it can continuously measure electrical pulses traveling through your body primarily close to your heart and from that it can analyze the heart behavior into one of twelve possible classes of arrhythmia or perhaps even a breathing congestion problem etc. Now that is a virtual statistical twin. It was put together based on specialized data being collected by data scientists for a very specific problem. Now that company that put together the patch has created a deep learning- based model that performs as well as or better than a cardiologist and so it’s a very valuable company but there is only one signal will being colleccted and that is the heart electrical field. It can only be used to do cardiological analysis of the heart. Therefore you can do these kinds of virtual statistical twins where the potential payoff is very very large and when you have the organizational brand to be able to attract the best people to put together that virtual statistical twin.

That’s a problem because most companies are still industrial companies, they can not attract these kinds of people and their systems are very very complex and so explaining to somebody else what is the design of your system is very hard. People give up and this is happened to many of our customers. That’s where the full loading automation comes in. That’s a world in which software is able to recognize patterns from the sensor data automatically but it does not require that sensors be designed in a particular way. It also does not require that the data be collected in any predefined kind of a way. There is a good amount of responsibility division between the designers and operators of systems and the designers of the software that provides full learning automation. In this category the software can discover patterns from the data without too much guidance and therefore it does not require data scientists to begin with. Secondly, operators are able to put in their knowledge in the form of labels that they are able to fix over time and in a very sort of unmanaged sort of a way because you cannot manage the producing label for every moment in time. Now this kind of software can be deployed to any kind of a problem and therefore the majority of problems that we still have not solved will be immediately solved using the full learning automation software. It’s also a technology that creates incentives for existing staff within organizations to solve problems because that’s their job and ultimately they are the ones who are solving the problem but this approach of full learning automation is the basis for creating a scalable industrial predictive analytics practice within large organizations. We think about this as Six Sigma or lean manufacturing which did not get introduced by bringing in a huge team of consultants from an Accenture or PwC or an IBM Team. There was a small team of people from the Six Sigma coaching firm or the lean coaching firms. People visited their colleagues and peers from other companies and shared their experiences with each other to then become good at it across their entire companies and across their industries. We believe that full learning automation will be the same thing for industrial predictive analytics. Now of course the goal for many of our customers is to get to autonomous control and they know quite well that autonomous control may be possible in some of their business needs but for others they will still require a human in the loop. They are actually fine with it because full learning automation moves the needle way beyond their current state of practices. Therefore it creates huge economic benefits for them. So, this was a long winded answer but it addressed what you’re trying to understand about analytics in the industrial world.

CEM:

Great! If you can also tell us a bit more about the financial and time aspects of rolling out such a system. How long does it take to roll out an implementation and how it works on pricing?

NIKUNJ

Sure. Falkonry’s approach to its pricing is to give customers complete control over how much data, at what rate, and across what amount of computation they want to use. That is because in the early stages of a technology, you want to get the best possible results at the lowest computational and transmission cost. And only once you can prove to yourself that there is value in the analytics does it make sense to take into account what is the cost of the analytics itself. So, in other places people have charged based on the number of streams of data being collected or the volume of data being produced or even the number of users that are using it or the number of CPUs on which work is being done. So, we believe that all of these are incorrect for an analytical product, especially during the early days of that technology wave. On the other hand, all of our customers are aware that they need to solve problems. Problems that in their mind may be causing reduced availability or longer time to resolution, poor quality, whatever it might be. And so, each one of those problems, they have a sense of what impact can be created if the problem can be solved. So, what Falkonry has done is it offers a uniform pricing structure that is based on the number of entities, the number of signals, and the number of predictions that Falkonry is creating on a continuous basis. Remember, Falkonry is designing these models for real-time predictions. Therefore the predictions are not of one-off kind of a thing they are happening continuously. We make it easy for our customers by simply banding the signals into buckets of 25. Thereby, you don’t have to count for every signal. You just have to divide use your users between the low signal count or the high signal count. That gives you freedom to choose how many signals you want to use for solving any problem. Similarly you can use Falkonry to solve a problem on a custom designed system but you can also use Falkonry for a distributed fleet of assets that are all designed similarly. Now this approach is comparable to saying how many data scientists and how many software engineers and how many IT people do I need. On a per problem basis you’re going to spend probably half a million dollars to a million dollars a year for that type of a solution. So Falkonry’s pricing is going to be a fraction of that, plus of course the time that your own process engineers or manufacturing engineers are putting in, therefore, it presents a more attractive value proposition that allows the organization to keep more value for itself without diverging the benefits to Falkonry and having to negotiate a performance criteria on an annual basis etc. So the pricing approach we have developed allows both parties, Falkonry as well as Falkonry’s customers, to feel confident about the future to not be vulnerable to intellectual property or other proprietary trade secrets that we have to manage ourselves. And that is one of the unique things that Falkonry has navigated successfully in the growth of its business.

CEM:

Clear, thank you very much indeed. Can you also give a real-life example? Maybe from an industry that you are most experienced in. What would be the annual costs of working with Falkonry?

NIKUNJ:

So, every problem or need that  Falkonry solves is typically a million dollars or more per year and the costs of operationalizing Falkonry in the context of that kind of a problem or use case is going to be in the low six digits. So, typically within a year you get your payback and you are spending on a recurring basis a very small fraction of the value you are creating. Because of volume pricing because a lot of customers try to solve many problems at once they are able to get an even better leverage.

CEM:

Can you also elaborate a bit on the time to set up? I know it depends a lot on the data integration issues and so on but I also know that you have you have tackled these problems before so if you give an idea about the average case.

NIKUNJ

Sure, first of all, most benefits are visible in the analytics and people have been collecting

data some way or the other. So, what we suggest to people is to validate that the analytics produces value without bothering to invest in data integration. So that way, they are able to develop a good understanding of the benefits within two to three weeks from the existing data collection practices they have in place. In some cases where they are customers of one of our partners, we have a data integration offering available to them and they can avoid any long-term costs of data integration. Now, once they have validated the value in the analytics then they have to start figuring out what data they want to bring in and what the systems of record of that data are and how to integrate it with Falkonry. Falkonry worked, for example with one customer who was bringing together ERP, manufacturing execution system, as well as machine operational and sensor data. There were three different systems that were being used whose data was to be pulled together for pattern analysis and predictive analytics. In that case, the customer worked with a vendor of their manufacturing execution system, this was a company called Vegam Solutions. Vegam Solutions built an extension to their manufacturing execution system that used Falkonry client development kit and within a matter of two months they connected all of those systems to Falkonry, so that a user of the manufacturing execution system could very easily select what problem they wanted to solve and the data related to it and bring it into Falkonry so that all of the pattern recognition could be done on it. They, as a user did not need to know how  complex this integration was or change the way they work from one problem to another.

CEM:

Just to understand what you mean by starting with the analytics without the data federation or without the full data integration. You mean you start with analyzing their available data so your model is available from day one, for predictions or did I misunderstand?

NIKUNJ

That’s correct. So, most of our customers are already collecting data and would like to have a model on day one. Now that does not cover everybody. So, for some customers they do not have historical data and they want to start collecting data and analyze it. To answer your question specifically, our customers want a model on day one and they are happy to use historical data that they have collected for that.

CEM:

This is a great solution for relatively large entities we discussed before in manufacturing but is there also going to be a lighter version for much smaller entities?

NIKUNJ:

So, there are simple solutions that are available for two types of situations. One, where the data patterns are not complex and you’re working with one or two parameters. So, we try to stay away from that space, because there are other technologies like rule-based systems or streaming analytics that are well suited to those kinds of problems and where the costs are very very low because of strong open source offerings. So, we try not to indulge ourselves into working in those markets. Those are also areas that are less dependent on subject matter experts and can be codified directly by software engineers. So, there is no data scientist needed in that context either. There is also another category where there are certain known patterns that occur in one or at most two variables that people are trying to keep an eye on, on a continuous basis. That’s also a problem area at Falkonry could be used for but there are other alternatives that our customers often use and they use it while they are also our customers. So, we see that such complementarity is beneficial to the market. So, we are happy to support our customers who want to use both solutions. Now, Falkonry is by itself not a very expensive solution. You know, the starting cost on an annual basis are not very high, considering the value that somebody will be creating using that kind of technology. And we certainly are open to working with governments as well as with educational institutions to make available to them some of these technologies for public benefit as well as for educational benefit. In those kinds of situations, profit making is not our our objective.

CEM:

Vey clear. This is not that much business-related but I wonder where the name Falkonry comes from.

NIKUNJ:

That’s a very nice question especially because this is also a topic which is very dear to my heart. I know a lot of people all around the world have heard about falconry in their own languages because there has been a relation created between a human and a raptor bird, in many many parts and many cultures around the world. Falkonry derives from that original concept of training, a falcon to hunt for the benefit of a human. And so, that’s where the world Falkonry comes from. The reason I picked that name is because Falkonry in terms of its analytics, is just as powerful as fast and as sharp as the bird falcon is when it is going out and hunting. It’s able to figure things out on its own. You just have to train it to attend to one type of problem and not another type of problem. Falconry is also, if you look at the word origin, a place where you train the falcon, in that same way Falkonry software is where you create these analytical falcons and then you set them free to go hunt problems in the wild meaning on real-time basis. So you know falconer is the person who is doing that type of setup, in our case that’s a subject matter expert and so Falkonry, Falkoner and Falkons are the terminological basis for what we are doing in our business.

I’ll say one more thing: In our world of industrial work, falconry and falcons are a strong message to our customers that it is industrial grade, it is the best of the best and therefore it creates a strong brand perception. Our customers ask all the time, how did you come up with this name, it sounds really cool and we generally receive very favorable feedback for this.

CEM:

This is a really nice name for the business, then. I have also spoken with vendors in robotic process automation landscape. We were talking with the WorkFusion leadership team and I really appreciate their free as in beer offering. I am just thinking five years down the road maybe, but even smaller companies will need to deal with more complex manufacturing processes. Would that be something that you think industry could turn into to or you think this will remain in the domain of large manufacturing entities that are ready to pay six figures digit fees.

NIKUNJ:

We think that Falkonry will build a bridge between both the small entities and the large entities. While our focus is currently on large entities, we have served smaller companies in 2015, as well as in 2016. Our focus right now on large companies is only so we can gain some scale while the industry understands how patterns can benefit them. So our objective is not to only have large customers, especially because the industrial world is going through so much change that many companies that are very small today are poised to grow very rapidly. So, Falkonry will also be meaningful for a lot of smaller customers that have a clear understanding of the problems they are trying to solve and the business value they are trying to create. So we are certainly desirous of working with small companies when they are clear that they can benefit from working with a vendor like us than to try and spend all of their core competencies on coming up with a custom solution for their own pattern recognition needs. We’ve come across such startups ourselves. One is in the agricultural context that is managing irrigation of crops based on soil and weather conditions. Another is doing railway transportation analytics for equipment in the distributed rail system and both of them came to the conclusion that it was just not worth it for them to hire and retain data scientists to solve so many little problems that come in their world because it would take them too much time and would require constant tweaking because their world is always changing. So, those are the kinds of situations where we think startups can benefit from this type of prepackaged technology and they can pay as they go just like they pay as they go for the cloud for example.

Naturally, Falkonry can work out a pricing scheme for start-ups where they take a lot more responsibility for the support they would require from us but also, because they are far more nimble and technically competent group of people, we expect to be doing less hand-holding as well.

CEM:

Thank you very much indeed. If there are any final comments you have, love to hear them.

NIKUNJ:

So, Falkonry’s primary benefit is to the industrial practitioner but the immediate indirect benefit is to the industrial world as a whole. As we know 40 percent of the global economy is industrial activity. So therefore we believe that by making that sector of our economy, our global economy more efficient we are actually reducing the dependence on natural resources. We are also making the industrial engineer a lot more effective and therefore providing opportunities for engineers all around the world who may not have access to the best data scientists and the best manufacturing design know-how to be able to do a good job operationally. So just in the same way that countries in Asia became extremely good at manufacturing even though they did not invent all of the manufacturing equipment themselves, we believe that Falkonry-like technologies will help them become extremely good operationally using data even if they themselves do not come up with the analytical techniques themselves. So both from a developing world perspective but also from a global economy perspective Falkonry is going to have a substantial impact to improve the natural resources and sustainability as well as the livelihood of people all over the world who are involved in industrial activity.

Cem: What are the big challenges you are facing with vendors?  What are the companies that you are finding that are easy to work with? What sets you apart from competition? I guess these are the typical questions you get.

Mario: We are exclusively focused on big industrial transportation, primarily equipment and helping them move unplanned maintenance to planned maintenance. So, obviously, these big industrials have years of experience maintaining their equipment. If they have any predictive analytics it’s usually a rules-based engine. So, the expert knows that if the temperature of the motor goes above 500, something’s wrong, right? There’s a rule built for that and we’re taking a data AI based approach through machine learning algorithms. There’s a lot of companies out there that are messaging that way. You can, you know, buy Microsoft Azure which comes with ML libraries and then other things. What makes us different is that we have automated about 80% of the process of creating that actual machine learning algorithm. So, the process of creating features, we do that automatically in a massive scale. So, in a deployment, we might create 30,000 features. Normal data scientist would create maybe 50 or 100 features right with the data sets. We massively scale that and we automatically score those features and we find the top 100 or 500 features for that particular piece of equipment in that particular environment. We have every known modeling technique out on the planet plus a few we have for them that we’ve created. We’re heavy users of classification algorithms that predict if the failure will or will not happen in the next day to 30 days. So, for a customer that has already hired IBM and failed, has a lot of data and wants to create a lot of algorithms, the manual approach doesn’t scale very easily. So, that is our sweet spot. For us, customers are large industrials like GE is a customer and we are about to start with two manufacturers of aircraft engines, turbines. They have very sophisticated smart teams and scientists but they just can’t scale their analytics with custom approaches and our software is able to do this right. We go from project kickoff to go live about 90 days.

Cem: That includes of course data integration as well.

Mario: There is no magic to that. The data ETL is a challenge. we’ve built our own ETL engine that enables us to go incredibly fast. It’s purposely built for preparing that data for pancaking that and preparing it for machine learning. In a 90-day deployment, the first 30 are the data ETL

consolidation into our environment. We use elastic search as the technology in our platform where that data is stored. The second month is the machine learning automation, the algorithm creation. We call that the max engine, and then the third month of a deployment is the configuration of Predikto Maintain. It’s a front-end GUI application that takes daily predictions and turns them into notifications in a way that a user can actually operationalize the real-life environment. So, that’s usually what a deployment for us looks like.

Cem: Just to understand the exact transportation setting, is it planes or trains or something else?

Mario: Yeah, we started with rail. Working both with freight locomotives diesel and electric, the German DB freight railroad system is a customer with their Bombardier electric locomotives. We then did a pilot with bullet trains as well in Germany. We then expanded to aviation with non-engine

parts like landing gears and hydraulics. Now we’re getting into the engines with two of the largest OMEs in the world. Then we landed Maersk as a customer. So, we’re working with them on cranes that load containers from the port to the ship’s. So, technically our software is not specific to just locomotives or aircrafts. We can predict anything. From a sales perspective we’re focusing on transportation first. We’re about to potentially get into wind turbines as well. Big, heavy equipment, that’s spread across the vast area, that’s kind of our sweet spot.

Cem: Maintenance is critical for optimal performance and reducing downtime. In the crane case, I guess failure is something extremely rare, right? Because then you lose the crane operator and it’s a big disaster. I mean even before predictive technologies, they could achieve operations without accidents, right? So, what is the value add there? Because I don’t know much about that business,

Mario: Great point. So, if you think about a port, ports may have twenty or twenty-five cranes, and the ships come and park next to the port and they get unloaded and loaded. If you have one of those cranes that’s not working, you reduce the speed by which you can unload and load the containers and a backlog of other vessels are waiting to come into the port. So, it creates a very big disruption in the supply chain.

Cem: Yeah, that’s very clear but what’s the frequency?

Mario: Yeah, you know, one port could have probably 800 to a thousand hours a year in downtime. It’s expensive. So, if a gantry motor does not work, the crane can’t move from left to right or the boom hoist might not able to pick up the container. I mean, it’s a complex big piece of equipment and it’s like a locomotive. It’s just performing different functions. Instead of pulling, it’s moving and carrying it has a lot of moving parts and those moving parts fail. They are exposed to the weather, near the ocean, salty air. There’s a lot of things that can go wrong with them.

Cem:Final question. This may be confidential but I wonder the pricing. If you have an idea about what you are worth to the companies and if you can be opaque about your pricing then it is a great opportunity.

Mario: Yeah, we actually tried to do value-based pricing in the beginning. Struggled significantly with that because customers are really smart. They’ve been buying technology for years. Value-based pricing naturally creates opaqueness and lack of transparency and that was not working for us. So, our price is pretty straightforward, we charge 250 K dollars a year for the base platform. Then we charge per prediction endpoint. So, depending on how many parts you want us to monitor and how many locomotives or cranes or engines you have, you multiply the two and that gives you the total number of prediction endpoints. That allows us to start small and grow with the customer as their needs and their use of a software grows.

Cem: You were extremely open. Thank you for your time.