Intelligent Spending: A Q&A with Coupa’s Rob Bernshteyn
In a recent interview for the Early Adopter Research Podcast, Dan Woods spoke with Rob Bernshteyn, CEO of Coupa, about how he has led the adoption of AI and ML in the Coupa product. While some companies will have success with AI and ML by incorporating them into their own systems from scratch, most companies are going to be using AI and ML as part of products in the way Coupa offers them. Regardless of the approach, however, being a sophisticated buyer who can evaluate a vendor’s claim about AI and ML is essential. But AI and ML will also need to be augmented with human context and intelligence. Obviously, mapping and cataloging all the data available in an enterprise and partner ecosystem is important because AI and ML live on data. The more data it has, the better the data it has, the better it is going to work. And in order to do this properly, you are going to have to move data around between repositories and applications to create a complex and agile data supply chain, while also learning how to debug data. With AI and ML, data is not just a descriptive quantity. It is also like the source code. And just like source code can have errors and lead to mistakes, so can data. And you want to be able to get good at understanding the bias in datasets and any errors that create suboptimal results. In this edited Q&A version of their conversation, Woods and Bernshteyn discuss how the Coupa product takes advantage of AI and ML to help companies save money and the lessons Coupa has learned that would be useful for other companies that are also seeking to apply AI and ML.
Dan Woods: What is Coupa, and what is its mission and origin story?
Rob Bernshteyn: We are what we call a value as a service company, and our mission is to help organizations around the world spend smarter. We think that companies are not employing information technology as effectively as they could inside their organizations, empowering employees in virtually every department and every geographic location to have access to very simple, intuitive tools that help them purchase the goods and services the company needs to operate effectively. We have created both a platform as well as a company that is focused on this mission. We think that this mission allows us not only to help companies spend smarter but ultimately allows them to have more operating resources and money, frankly, to reinvest in their own missions so they can pursue them vigorously on a day-to-day basis.
As for our origin story, I spent my entire career in enterprise software. I had spent a lot of time in the financial application software area and was responsible for internal training for Accenture, where we brought on new consultants, fresh from college, and helped them get good at implementing complex SAP ERP systems.
I always felt that there had to be a better way in terms of better alignment with customers. The software was very expensive. The consulting fees that the company I worked for was charging were very expensive, and a lot of times the customer was sort of left out there in the cold. I wanted to come out to Silicon Valley and be part of the solution to creating the type of information technology that drove value for customers and removed a lot of that friction.
Along the way, I had the opportunity to work at Siebel Systems in the CRM space and then helped develop a company where I was head of products in the human capital management space. That’s where I first saw the power of having these applications delivered in an on-demand experience over the Web. I then thought, What is another space where there is this opportunity for not only an on demand as a service offering, but a space where there is a lot of opportunity for greater operational efficiency, greater usability, greater value delivery for customers? Having been in the financial, the customer and the employee side, I though the last entity was really suppliers and the way companies spend.
Coupa does provide enterprise software in the way we think of it, which is usually an application that is internal to a company that is then providing a way to do a function. But you also have relationships with suppliers so you offer something more like an Amazon Marketplace, not just an internal application. How would you describe your product as a platform that has one foot in the enterprise and one foot out of the enterprise?
It is very interesting the way you frame it. Our first question was where is the bigger problem today? Is the bigger problem that suppliers or vendors don’t have access to great tools to help them sell or is the bigger problem that companies themselves don’t really have their arms around the way they spend? We came to the conclusion that the latter is the bigger problem. If you look at the situation today, now 10 years later, the number of tools that we have to help us sell are unimaginable. The people trying to sell to me know my address, my location, and the other types of products I’ve bought. But when you look inside many of the companies we work with, they really don’t know whether or not they are paying the same price points for goods and services across a global enterprise. It often takes a very long time to get something approved. There is still a lot of paper — purchase orders going out, paper invoices being matched — and a lot of inefficiency in the process. So we said, well, why don’t we begin on the buy side? Let’s really begin with the spenders and charge them fairly for being able to get their spending in some level of control, minimize maverick spending, make sure their spending is on contract. We saw very quickly that we were interfacing these buyers with millions of suppliers. At Coupa we have enabled business for our buyers and we have hundreds of customers around the world, with millions of suppliers. So in fact, to your point, we have gotten to a place where there’s this collaboration happening between our buyers and suppliers. They are not only receiving orders, but they are sending their invoices back against those orders. They are collaborating around early payment discounts.
What you have done is said, look, the center of action is buyer and you are going to serve that buyer, provide them the tools they need, and then reach out from there, allowing them to create essentially their own marketplace of what they are interested in.
That is exactly right. The huge advantage we have today is we were able to build this platform from the ground up as a cloud platform, and also as a cloud business, where everything we care about is the recurring revenue that we develop and the ability to keep our customers forever as they add on more applications, more use cases, and as we encode more logic that is highly applicable.
Now you have chosen an unusual platform to build your application on. When we spoke last, you told me that you built all of Coupa on Ruby on Rails. What went into that decision?
That decision was actually made very early on as part of being an open source project to figure out the best tools to use. Our technology is effectively similar to what is called LAMP stack with Ruby at the core and hosted primarily on Amazon. But the logic was that if you are going to help companies spend smarter, you have to have visibility to spend, and the only way you are going to get visibility to spend is if you get very, very wide end user adoption. And the only you are going to get very wide end user adoption is through the intuitiveness, the simplicity of your application. It has to be easier to use than any other alternative for spending money. If you are going to spend the company’s money, you could go to the store to buy something and expense it, you could ask your assistant, you can call your procurement group, you can collaborate with other employees. We want to avoid every other way of company spending other than wanting to go to Coupa. So our concept here is the best user interface is no interface. In other words, how do we avoid having to login and do anything unless you absolutely have to as a human? The letter “u” in Coupa stands for “user centricity.” It is something we are committed to. It also gives us an opportunity to stay up to date. As later and later releases of that platform come available, we stay true to that and our customers don’t even know it, but on the back end we are staying up to date on some of the latest capabilities.
You have built this application yourself, primarily, and only rarely have decided to do acquisitions. Recently you purchased a company called Aquiire. What was the motivation behind that acquisition?
We have had a very clear strategy around what we build and what we buy and the transactional capabilities of our application: those are procurement, procure, expenses, expense management, invoice, invoice processing, and payments. Those are all organically built because we have literally millions of transactions going through our platform on a daily basis in any one of those areas and we could not trust that to some sort of acquisition that we try to snap in and integrate. But as part of our strategy, we decided we may buy either what we call power user applications or distinct technology components. Power user applications are capabilities that are used by maybe a dozen users within a company on an ongoing basis. They live in the application. These are things like spend analysis or strategic sourcing or contract management, as examples.
We may also buy very distinct technology components that we think we could snap in relatively easily into our platform and take advantage of innovations that are out there. Aquiire falls into that category. This company has over a dozen patents and some very interesting technologies. It is allowing us to do is deliver real-time local access to supplier catalogs for goods and services that customers need, at the prices and terms negotiated by them or by Coupa on their behalf. With this technology, we are really the only platform out there that is going to offer this real-time customer value.
What is the value to the user of Coupa of real time?
There are really three ways to offer a good or service to an end user when they are in a platform such as ours. One is catalogs. The challenge of that is someone has to actually load those catalogs into your system. That may be the buyer themselves who gets these catalogs in some format from suppliers and loads them in. That burden may be on the suppliers themselves to keep those catalogs up to date and upload them through a supplier portal, such as ones we offer to keep those up to date. A second option is something called punch out, which is very typical in our space and that is you are working within a platform like Coupa, but you have to punch out to a third-party website. In that website, you see your preferential pricing and then you have to pull it back into Coupa to get it approved, routed and then eventually ordered. It is a little bit tough for the end user.
We bought a technology capability from a company in Switzerland about seven or eight months ago where they have the capability for caching that data. So rather than going out and having to pull it back in, it can be cached inside our system and so when one person orders it, the next time someone else goes in and looks for the same item, it’s there. With Aquiire, we are now capturing the third way to get this, and that is real time. That means the moment that you are searching within Coupa for a given good or service, behind the scenes Coupa is going out, getting that preferred pricing, and pulling it in in real time in front of you.
So customers can change the pricing confidently knowing that that is going to be reflected in anybody that looks at their catalog?
That is correct.
When did you realize that AI and ML were going to be important to Coupa and, how do you think of these concepts?
I have realized the importance of this since my days getting my MBA and taking just one class down the road at MIT where I saw some really interesting things being done in the AI lab there. I always thought that there would be a point where enterprise software would have to adopt these types of technologies because enterprise software captures a lot important contextual data. Clearly there will be opportunities for using AI to make these systems smarter. Traditional enterprise software has been data in, data out. It may be in nicer pie charts and different visualizations, but not necessarily prescriptive or offering real intelligence to the end user or the executives that use them.
I knew there would be a point in time where there would be an opportunity to employ AI, but there have been two limitations. First, there hasn’t been enough data. You obviously need a lot of data so that AI becomes interesting. It’s the same way that we as humans, as we get older, the amount of experience we have arguably should make us more intelligent. Second, you need computing power. You need speed. You need the ability to run through that data at pace, and the CPU has to be able to handle that. Both of these aspects have evolved a lot over the last two decades, which gives us the opportunity to employ some very interesting AI techniques.
We have a whole host of these around our application platform. We began simply. We bought a very small company with a couple of smart folks who were working on AI, and effectively machine learning or rule-based machine learning, for the problem of taking data off of an invoice and dropping it into an enterprise software product. Now the reality is, even today, so much of the invoices around the world are paper, emailed or sent in in letter format or even emailed as a PDF to companies, and some data entry clerk has to pull the data off them and punch it into a system. There are some OCR type tools that have gotten relatively decent at pulling off data, but they are not necessarily smart tools.
Our technology is very smart. It is rule-based, so it learns based on the types of invoices it is seeing. What data is the header data? What data is the line level data? What field should it pull the data off of at the pixel level and then drop it into Coupa. And as more and more invoices come, it starts to learn and creates that closed loop situation where the quality and the number of times people have to get involved becomes less and less and the data store becomes bigger and bigger.
The value proposition is very powerful. There’s a reduction of employee expense and payroll for data entry clerks and others. With very high-quality, we can automate a process that was effectively paper-based through rule-based machine learning.
It often seems that in the adoption of enterprise software, software gets adopted by one department or one division. Then everybody sees it working, and they want it too. Is this true of AI and machine learning in your experience?
Yes. For instance, we are doing some very interesting work in spend analytics. It is also a different approach than typical rule-based machine learning. We are employing what is called a naive Bayes classifier approach which is a more probabilistic classification of data. We go into companies and it may start in a certain division, but very quickly they start looking at the whole organization. That is, we analyze where have they been spending their money, and how much have they been paying and where have they been getting and not getting discounts. We normalize their items tree. We normalize their supplier base. And we come back and tell them, “We’ve identified $200 to $300 million dollars in savings that you can get if you simply route all of your purchases to the best contract you have in the company. The AI capabilities that help classify their data, normalize it, and help them understand it, allow for that opportunity. So what often begins as a divisional conversation quickly spawns into a global multinational level dialogue about how a major Fortune 100 company can save hundreds of millions of dollars.
After you had your first foray into optical character recognition, you then realized that there were a lot of other places that you could apply this. How did you decide which to go after first, second, and third?
The underlying driver was finding the point that is the biggest area of value creation for your customers. We’ve started with the problem. There is a big problem, as I mentioned, in paper-based invoices. Can AI be used to automate processes there quickly? Yes. Is there a big problem in people understanding all the areas where they’re spending and perhaps spending in a way that isn’t too smart? Yes. How can we apply AI there? Then we went to a third problem. Think of a company like a Walmart. Every one of these stores has to have apples in the area of produce and in that area they need all those apples to be fresh, and there need to be enough in inventory and they need to be ordered at the right price and you need to think through the freight and you have to think through how much you want to have in inventory and how much you want to have on delivery. You have to think through every possible constraint to ensure that every store has exactly the right amount. That is a problem that humans can continue to look to solve through spreadsheets or you can apply algorithmic AI-based approaches to try to solve these problems. We started to address that problem. Now we have some of the largest companies in the world using our strategic sourcing capabilities and our mathematical optimization algorithms to optimize their purchasing.
What new skills did you have to acquire inside Coupa to understand and build these extensions to your product?
I would have to put them into two categories of competencies. One is obvious, which is the development capability. You need folks that have some exposure to either rules-based machine learning or approaches to AI that could be applied to problems. But equally importantly – and I can’t underestimate this – is product management capability, the folks who can really identify the value drivers that these technologies can be applied to. Technology for technology’s sake or product ideas that are never really supported with properly created technical capabilities are worthless.
Where are you going next?
What we are doing now and where we are going next for me is by far the most exciting. Because everything I shared with you so far are things that are really one customer at a time, if you will. What we are starting to do now is a concept we call community intelligence. Now, community intelligence takes advantage of the now literally hundreds of billions of dollars in transactional spend running through our platform. We are getting closer and closer every day to a trillion dollars running through our platform. We have, as I mentioned, more than four million suppliers with hundreds of companies all over the world. How can we use the intelligence of our entire community of growing customers to help each individual customer get smarter and smarter about the way they spend? We have now launched products that are doing very well in the marketplace that leverage this community intelligence concept. One is a product called Risk Aware. This product looks at all of the suppliers that our customers are working with and without bothering them in any way, pulls data such as how often are our buyers having disputes with our suppliers, how often are these suppliers sending invoices that are higher than they ought to be, how often are they sending products that are broken, for example, or not on time? And we take all of this data and we clean it, sanitize it, normalize it, and then massage it on the backend to then come up to each individual customer. We give these prescriptive tools to each of our customers so they can get smarter and smarter about the way they spend. And we are just touching the tip of what is possible with community intelligence.
Which of the AI techniques have you found to be most productive for your applications? Have you found that it is better to start with simpler techniques and then gradually move up the sophistication chain?
You have to begin with the problem that you are trying to solve in mind rather than technology for technology’s sake. If we tried to solve looking at invoices from a neural network perspective, I don’t think we would be really successful. We began with a rules-based approach. If you tried to solve very complex strategic sourcing events and algorithms with a rules-based approach, that might not be the right approach either. You have to begin with the problem set. Our experience has been that you begin with the most simplistic approaches first and then evolve from there.
The one thing I will tell you is there has to be some element of closed loop if it is AI. With every interaction that that AI system is having, it needs to pull in the outcome of that and make that part of the data store upon which it applies its next set of intelligence. I think a lot of times people classify these capabilities in ways that are part of AI but really aren’t because they don’t have that closed loop concept.
The closed loop allows you to understand how well you did and then do better next time?
The closed loop incorporates the outcome of what just happened into your experience for the next interaction. Every interaction I have builds upon my dataset and then next time I make a decision, it incorporates my entire set of experiences. Otherwise, that wouldn’t be intelligent. It is the same thing with artificial intelligence. Artificial intelligence at some level has to incorporate that closed loop concept.
In your journey, starting with the invoices and moving through the other areas where you’ve applied AI and ML, what would you have done differently now looking back?
I think we could have gone even earlier in terms of investing in this area. I think we are actually very early relative to anyone around our space, but I think we could have been even earlier. We are a fast moving company in general and we tend not to sit on things, but we could have gone even faster.
What mistakes do you see other vendors or other companies making in adopting AI and ML?
One of the biggest mistakes I see out there is that a lot of vendors put AI and the subcomponent ML all over their marketing literature and they say, “We have that too.” And when buyers go out to decide who they want to work with, they check the box on AI/ML with vendors without really knowing what they are getting. That is a big mistake. It is on partners like ourselves to make it much more explicit. Folks understand there is a difference between the intelligence level of human beings. There’s also a very distinct difference between the level of intelligence a platform can offer. And it is on us to show those value drivers so they understand it in business terms and do not just resort to checking the box on something that may or may not deliver.
What advice do you have for companies attempting to develop their own solutions, either an internal company, like somebody who would use your product, or a vendor?
Begin with the problem set in mind. What are the problems in your company that are not operating as efficiently as they could or are subpar in terms of efficiency? If you could apply intelligence to the problem, how much bang for the buck can you get?
What would you advise a company to do in order to get the best results when buying AI products?
If you are a vendor like us or a smaller company or a larger company, I would say you want to capture as much data in a key area that’s important to companies first. Because data—this is a known metaphor, but data is the new oil. There will be different innovations in terms of the rigs that pull the oil up, just like there will be new algorithms and new AI approaches. But the oil itself is what is so critical. If you are a vendor, develop a business around a core dataset that you believe is going to be very important to customers for the long-term and then think about ways to leverage that dataset to bring intelligence to those companies around that data. Too many companies that are getting funded have interesting algorithms and different AI approaches but they are not necessarily easily connecting to any data nor are they coming with an existing dataset.
What is your advice to companies that are going to be using AI purchased through products? How can they make the best of their situation?
It begins in the selection process. They have to really dig through the value proposition a vendor is offering rather than just checking it off. They should go in and have vendors show them how the use of AI is going to deliver a business value for them that is measured in dollars and cents. And then secondly, the quicker you get transactional data going through a system, the quicker you start to automate underlying business processes, the faster you are going to get to a place where you are actually able to get some intelligence out of that data and improve. A lot of times these projects begin with a bold approach, but they don’t necessarily do the underlying work required to get to that bold outcome. The key is doing the grunt work first and then you can reap the benefits of some of these things that we are producing.
For the average company that is going to be using AI and ML systems, either that they built or that they purchased through products, what is success going to feel like?
If you think about what is happening in the world, with 7.5 billion people, we are in a global marketplace. Yes, there is some protectionism currently, but that is probably a short-term dynamic. If you have a business and a great value proposition, you could take your product or service to the market very quickly. You are going to compete based on, one, your operational efficiency. And you are going to compete on how intelligent you are in that market, how quickly you can adjust, how agile you are, how quickly you can reinvent your business, and what type of insight and outcomes you can create for your customers. Artificial intelligence is at the very tip of that. When we talk about pundits and others talking about software continuing to take on more and more of what is happening in companies and the entire digital transformation underway, once everything is digitized, it is about how smart and thoughtful we are in understanding our customers, our suppliers, our employees, our cash. This a very big thing that is just starting to happen and I would urge anyone who is thinking about exploring this area to explore it deeply and understand it because the future will be based on it.