Joe Fuller: Though often heralded as a game-changing technology, artificial intelligence remains largely invisible as it is applied across the real economy. There are no humanoid robots; there is no “HAL”—the humanlike AI system imagined in the film 2001: A Space Odyssey. Nevertheless, AI is driving results across a wide array of business functions—ranging from better management of supply chains to the allocation of marketing resources. Improvements from AI are numerous, if not glamorous.

Welcome to the Managing the Future of Work podcast. I’m your host, Harvard Business School professor and visiting fellow at the American Enterprise Institute, Joe Fuller. Today I’ll be speaking with Fred Laluyaux, CEO of Aera Technology and an executive with deep expertise in the enterprise software industry. Fred has described the company’s platform as “self-driving technologies for running a business.” He’ll show us how AI offers value throughout the enterprise—from sales to strategy to the supply chain—and discuss the associated organizational changes required to exploit the technology. Fred, welcome to Harvard Business School.

Fred Laluyaux: Thanks for having me. Great to be here.

Fuller: Fred, companies are very intrigued by artificial intelligence. There’s a lot of discussion about it—I’d say a lot of loose talk about the type of impact it’s going to have. What’s your view on artificial intelligence, how it’s going to affect companies, and how is that informing Aera’s strategy?

Laluyaux: AI delivers value with its ability to actually predict and project business performance over time. That’s actually a real insight. You can predict maintenance. You can predict customer churn. You can actually really deliver the insight of a company’s need to optimize their business performance. It’s very driven by the emergence of big data, by access to very large data sets, structured and unstructured, that are coming from the world of Internet of Things—IoT—from the web, but also from the massive amount of data that’s been accumulated over time by large enterprises in their enterprise resource planning [ERP] systems, and their CRM systems—customers relationship management—and so on and so forth. Now everybody has access through cloud computing to the computing power that allows you to make sense of these giant data sets.

Fuller: Fred, one thing that certainly I witnessed in talking to executives, historically, was that the advent of these large ERP systems and CRM systems, there was just an increasing torrent of data, which they couldn’t structure in a way that they found actionable, or it took a long time for them to get their arms around it. How are changes in technology—and how is Aera’s approach—addressing that problem?

Laluyaux: The issue with that is those systems were built with a certain DNA. Think about IT systems 20, 30 years ago, that didn’t really fit—it’s not really a good fit for analytics. You think about large organizations that grow through acquisitions. They have a lot of these systems sitting side by side, but trying to reconcile all this data—this very large amount of data—into a single instance that gives you the insight that you need to actually understand how your business works is actually very difficult to do. I need real-time visibility into my end-to-end business, where I can leverage that harmonized data to actually predict and optimize my performance—two very different paradigms, but the same data. We’re enabling those very large—sometimes old—organizations—50, 100 or more years old—that have implemented those systems over time and are a little bit stuck in that paradigm. If you think about the Amazons of the world, those companies that were born in the last 20 years that didn’t have the burden of those systems, they’re able to actually run analytics. They’re able to get customer insight. They’re able to optimize their supply chain in real time, because they didn’t build their organizations on that foundation. Aera Technology’s goal is to actually enable those very large non–digital-native companies to actually perform and understand their business and get insight and predictions the same way digital natives would do.

Fuller: How do you do that? Because you’ve got this information structure you’ve described in often dozens of different ERP systems …

Laluyaux: Yeah.

Fuller: … and sometimes incompatible across geographies. Plus those systems inform a management decision-making process, which in large companies is very focused, very siloed, doesn’t have visibility across the sweep of data.

Laluyaux: What we’ve done is we’ve actually built a technology that’s fairly similar to—we can take the analogy of Google. When you do a search on Google, Google doesn’t start searching the internet for an answer. What Google has done, it’s created a hot replica of every single website around the world and brought those—that image, if you wish—of your website on their servers, and then process that website by indexing, by ranking, by doing a series of steps that allow that data to be now searchable in a way that Google understands. What we basically do is, we create a hot replica of all the transaction data that we need, and we bring this data into our cloud. Those are billions of records at scale. Those are refreshed multiple thousand times a day. But once this data is in our cloud, we can actually process it—the same way Google processes your website, we’re going to process your transaction data to bring into a single data model that allows them to connect the dots of your business, to harmonize your data, to give you an insight into your business end to end. If you have a time series and a comprehensive data set, you can start predicting and projecting.

Fuller: When you talk about billions of records, this could be anything from a customer record to supply chain to standard costing. It’s cutting across all those different systems.

Laluyaux: Yeah, it absolutely does. It goes exactly as you said—from supply-chain manufacturing, finance, customer relationship management, safety—anything that you need that is structured or not structured. But also, if you start doing projections and optimization, you need things like, if you’re a pharma company, bring the FDA approvals in real time, bring weather patterns, bring consumer data, bring Nielsen data if you’re trying to optimize your promotion.

Fuller: How does this system know what data to query and combine to provide a human being with some insight? How does it know what correlates with what in a way that’s meaningful for decision making?

Laluyaux: We basically rebuilt an entire data model with what we call “subject areas”—order management, inventory management, and so on and so forth. It connects the different functions of an enterprise. Now, the next problem we have, the challenge that we have to resolve, is how do I connect those disparate systems to our data model? Here is a simple problem of translation. It’s a question of hard work and mapping all the fields in those systems, one by one, so that we know that a ship date or an order number is actually that field in that table. Where it gets a little bit more complicated is that, as I mentioned, those systems have been in use for many, many years. In some cases, the intention of the field is different than the reality. We have to do a bit of reconciliation with our clients, because the textbook we understand, by the way, the meaning of the data, can be a bit different. So we use artificial intelligence as well to help us. But sometimes it’s just good ol’ hard work and making sure that we actually have a good understanding of what each field actually means. Because we have a single data layer at the output, it’s very easy to realize when there is a mistake. The numbers will look off right away, and then you can go back, drill back to the transaction, and say, “Oh, we misread, or we misunderstood, the way that field is actually ... the intention of that field,” so to speak.

Fuller: And how is all this actually made comprehensible to a decision maker? Companies have their own internal taxonomies. They have different decision rules, risk profiles, even the effective, the incentives, and measurement systems inside companies.

Laluyaux: We’re a hyper-standardized world. If you’re making and shipping goods, you’re going to be using the same logic, the same lingo. Now, there will be variations by industry. If you look at the CPG companies out there—consumer packaged goods—the companies that deliver the goods that we consume every day pretty much all work the same way. Take OTIF: on-time, in-full. Your job as a vendor is to deliver the right amount of product that you’ve ordered as a customer on time and in full. Now, your OTIF standards for Walmart might be different than for another retail store. You, as this company, will use a definition for OTIF that will be a bit different than from that other company. We have to adjust that. We work with our clients who say, okay, let’s look at all the key formulas that actually drive your ability to understand how your business is performing, or your business operations are performing, and we’ll adjust the rules and the thresholds to make sure that we stick to your standards.

Fuller: One thing we regularly hear about machine learning, artificial intelligence, is that it’s extremely helpful at deriving correlations between data, but not so helpful—in fact, not helpful—on understanding causation of why something works the way it does. Is that just a facile platitude?

Laluyaux: Yeah, I think the crystal-ball aspect of AI has been overhyped. What we’re really focusing on is answering very simple questions that can help improve a company’s performance on a day-to-day basis. If I understand my demand better by looking at data, by correlating it with weather patterns or all the sides of patterns, and I can literally optimize and reduce my error when I project my demand, I can optimize my entire supply chain and save a lot of resources and save a lot of working capital and save a lot of money. So we focus on this pragmatic problem. The problem that we see in large organizations today is that doing your work as a planner is becoming increasingly complicated. The speed at which you need to make decisions is increasing by the minute. The number of data points that you need to assess or analyze or derive sense from is increasing by the minute. A combination of all of that makes the job that you were doing with a combination of instinct, a deep knowledge of the market, experience.

Fuller: Heuristics.

Laluyaux: Yes. This has actually now being rushed by a need to process a vast amount of data in real time. And the human brain is actually not cut out for this. We’re moving into an era of people doing the work supported by a computer system, including some statistical models that have always been there, to an era where the computers are doing the work controlled by people. So the revolution you saw on the shop floor over the last 100 years, and accelerated in the last 30 years, is now coming up into the offices, so to speak. I think that this is where the AI, the big-data modeling, the computing power, is combined with people’s experience and delivers incredible value.

Fuller: It sounds like you’re almost talking about now moving to a new paradigm, where historically, these tools—and ERP tools, generally—were deployed in support of the way workflows were structured in companies. What I’m hearing you’re saying is that the future is to design the work flows around what computing can do to simplify the choices to harness these massive flows of data that are now available and then, in recognition of that, create workflows that support and exploit those outcomes, which is a fundamentally different organizational paradigm.

Laluyaux: I think the fundamental difference here is that we’re starting to digitize the decision-making process in large enterprises. This is not new. If you think about trading, electronic trading, what is it? It is a digitization of the trading process.

Fuller: So how far can a system like the one you’re describing get to replacing the human altogether? And what’s the role in human governing that the system is concluding?

Laluyaux: I think I got this one wrong when we launched our company a couple of years ago—which would have made it 2017—where I thought that our recommendation engine would deliver recommendations to the information workers, let’s say a planner, and say “Hey, Mr. Planner or Ms. Planner, I recommend that you increase your forecast for this product for that period for that region by 1.2 percent.” What we realized was that, in so many cases, due to the acceleration that I’ve described earlier, people don’t really know what to say. The role of the human is really moving toward controlling and understanding what the machine is doing, as opposed to interacting with it in real time and deciding. To summarize: We went from a shift of people doing the work, to computers making recommendations, to computers doing the work and the humans controlling and learning and optimizing how the algorithms are working.

Fuller: Is that to look for anomalies that don’t make sense to them? Is that to ensure it doesn’t violate limits that somehow haven’t been expressed in the system? What’s the level of control, and what kind of visibility does the human being, the knowledge worker, have into the underlying reasoning that the system is applying?

Laluyaux: Some problems are just, imagine like a workflow, right? If this happens, do this. You see the if/then processing. We can actually map that, but run it really, really fast, and run it in real time and combine it with some projections if I need to. Imagine that, as an end user now, you’re in a control room, and you’re seeing those flows, right? Over time, you can see if the system was right or wrong. So your job that was basically making these decisions now has been moving toward analyzing and then re-tuning the algorithms. In some other cases, you actually do need a human interaction in the decision-making process. I’ll give you an example. Fifty percent of the sale of the consumer packaged goods companies come from promotions. You can actually digitize the entire process, but you still need to go in front of the user, the person, and say, “We recommend that calendar. What do you think?” Then you’ll be able to catch, “There is a fair in the region that affects … people for that weekend don’t go to the store.”

Fuller: Don’t shop.

Laluyaux: And that maybe wasn’t captured by the data. We’re building a permanent memory of all the decisions that are being made. Because we’re digitizing the recommendations, we know when the recommendation is made, for what product, for what expected outcome. We also know the business context at the time the recommendation is coming to you. Then we will see over time what was the impact of your decision. It’s very important to keep track of all of this and to build that extra data layer that allows you to optimize how decisions are being made.

Fuller: It sounds like you’re both provoking the decision maker—the executive—to question their own assumptions and look for patterns that are actually more the scale of human cognition. That a decision maker who’s knowledgeable about some business phenomena can think through a matrix of 10 variables by 10 considerations; they just can’t think through a matrix of 250,000 variables vs. 1,000 considerations. And you’re doing that level of thinking but presenting it to the decision maker in a way that they can actually apply their knowledge and discernment with the benefit of getting that massive data stream to power the recommendations that are being presented to them.

Laluyaux: It’s a combination of what we call “cognitive automation,” where we were able to fully automate the decision-making process. And that’s usually when human judgment has a little part. It’s really like looking at the data, and as I described earlier, running through a very complex process workflow. But computers are incredibly good at this. That’s cognitive automation. The other part is cognitive augmentation, where I will use my computing abilities, nothing more than that, my computing ability, my ability as a computer to work 24/7 for you to monitor certain levels across a vast array of data 24/7 to detect things. I will be your augmentation engine, but ultimately I’ll come to you with a very clear and crisp recommendation. The way computers do the work is they’re actually now moving to, “I can make that decision.”

Fuller: Fred, what you’re describing seems very compelling and, in retrospect, kind of obvious that we’d want to integrate these big-data flows, take advantage of this massive amount of data that’s available to people, and try to build this type of cognitive memory that you’ve been describing. Why is it only happening now? Is it a function of technology? Is it just a recognition that we need to get to a new level of efficiency and performance?

Laluyaux: A few years back, I looked at the LinkedIns, the Yahoos, the Googles of the world. And I said, “How can they manage the incredibly vast amount of data that they manage on a day-to-day basis in real time, with that level of concurrency, millions of users every day in that system, and the level of complexity of their algorithms?” And we in the enterprise world, so to speak, are stuck in the old world of pushing data from one spreadsheet to another, to a tool, to another tool, to an ERP. I think the answer to your question is simply there. The internet scale—the ability to compute incredibly large amounts of data in real time at scale—is what’s really enabling this whole new set of cognitive technologies in the enterprise. It’s combined with the fact that data is now available. Companies have been putting sensors and captures and all sorts of devices to understand their manufacturing, their supply chain. You have now sensors on every truck, you have GPS data, you have Internet of Skills data. So the combination of these data being available, plus the computing power, is really enabling the transformation that we’re seeing today. If I take an analogy, people look at self-driving cars, and they say, “Oh well, you need a set of sensors in the car.” No, to actually have self-driving cars running, you need to have satellites in the sky, and that’s not a small ordeal. So here is the same thing. To have a self-driving enterprise, you need to have a lot of data and a lot of infrastructure in place. But at the end of the day, the processing, the ability to compute these vast amounts of data is what really enables these kind of technologies today.

Fuller: A distinction that you often hear, and many of our listeners will have heard, is the difference between structured and unstructured learning by AI. Is that a relevant distinction anymore? And how would you describe your system’s structure for learning?

Laluyaux: We’re talking about building a cognitive operating system, a giant brain that sits on top of your transaction systems, and it can actually augment and automate a lot of the decisions that are repetitive by nature. That allows us to create a memory of the decisions that are made, and then, as I mentioned, project those memory points in time to see how accurate the system was. The key of our technologies is really, first, to build a permanent memory of the decisions that are made in an enterprise. The second—and I think this is always a bit ignored—is the ability to operationalize what you learn. The challenge today in large enterprises is you have your, you’re still working in silos. Every large company now has built a team of data scientists. What do the data scientists do? They spend 80 percent of their time trying to prepare that data set. It needs to go in front of some business users, or whoever they are, for a decision to be made. By the time they try to take an action or take a decision to operationalize the decision, the context is already changed. When I talk about self-driving, it’s really around the automation of that entire thinking process. As soon as the system can understand the impact of that decision and derive from it a learning moment, what we call “a learning moment,” I can operationalize it. So you reduce the time. We’ll still make errors, but you want to make them very, very fast. You want to keep building that loop where you can learn from your errors and improve over time.

Fuller: What we see across industries and have discussed regularly here on our podcast as routine work of different types is under extreme pressure. It’s shrinking as a percentage of the workforce and, specifically, good-paying routine work being one of those knowledge workers who maybe in the past has been asked to take four different types of spreadsheets coming out of four different ERP systems and somehow make sense of them. That’s the type of high-paid routine work, where the variability around the outcome and the task is low but is highly subject to automation. You see that on the shop floor, you see that in white-collar work, you see it in science.

Laluyaux: Because there’s high sense of dependability in those works. You’re basically paying people to do a fairly painful job, 8, 10 hours a day of repeating either a task or processes. You pay people in those large organizations 50 percent of their time just to know how to do things inside that organization. If you’re a planner inside a large organization today, half of your time or more—sometimes it goes to 80 percent from what our customers are saying—is spent around trying to make things work inside that organization, and 20 percent is really doing your work of selling, of understanding the consumer. That painful work is going away for sure.

Fuller: Fred, could you walk our listeners through an example of maybe a company adopting the tool to change a specific process?

Laluyaux: One is so simple. It’s called ATP—Available to Promise—or CTP—Capable to Promise. What does it mean? The expected delivery date. Think about a very large enterprise—some of the largest pharma companies in the world or others—that are simply not able to provide the information to their customers. The orders can be very complex, lots of parts, lots of components. The answer to that simple question sits at the intersection of 10, 15, 20, 30 different transactional systems that are spread across the world, that the components of that order needs to be shipped by train, and so on and so forth. What we’re able to do is look at data, collect all the information from all those different systems, run some projections, and predict with a high level of accuracy when your order will be available. So another example where companies are trying to resolve with what we call “cognitive automation” is demand management, is forecasting—trying to accurately predict the level of demand for a given product in a given market. Not working in aggregate, but working really at the lowest level, where the level that will drive efficiencies in your supply chain and in your manufacturing. So deploying this kind of cognitive-augmentation approach to demand management is absolutely critical, not just to the one company’s performance, but to our society as a whole. Another example that I like to take is promotion planning. We’ll collect in real time point-of-sale data. So we’ll get real insight into what’s moving off the shelves, or not, in a specific area. We have elasticity data, we have a bunch of other data points that can drive consumer demand. And we correlate that to your supply because it makes no sense to build a great promotion if you cannot supply the goods. Right? And if you think about this for every SKU, for every point of sale, around a country or around the world, you have a very big data problem on your hands. And this is where cognitive automation, this is where cognitive augmentation, can play a massive role.

Fuller: What are the workers of the future—the information workers that are living in this type of inverted environment we’ve discussed, where computing is making many more of the routine decisions controlled by people, as opposed to computing, supporting people making routine decisions—what’s going to be required of them in terms of skills and background?

Laluyaux: I don’t think anyone can actually say for sure. But it’s clear that if you do not understand how those systems work, you will be lost. Right? So your job now is to control how the decision-making processes are happening inside those—let’s call them “computer systems”. And your ability to understand those computer systems is absolutely critical empowerment. But that’s not enough, right? If you think about the structure of a large organization to date, some gigantic pyramid with 10, 11, 12 levels of hierarchies, I predict that we’re going to observe, we’re going to see a massive de-layerization of those organizations. We’re going to move from those giant pyramids to cellular networks with the small cells of people coming together to work on a project to create something. But the execution part of the job—which is the foundation of that big pyramid—the decision-making process and the execution processes will be automated. We’ll decide long range at the high level. But the computers will take over a lot of the tactical day-to-day optimization operationalization of the work in a large organization. So I would say, from a skills perspective, you’re going to see a lot more around interpersonal skills, creativity, coming up with ideas, ideation qualities, and less around, “I’m acting like a robot.” Basically, if the thinking that is involved in new work is highly repetitive, that is also going to go away.

Fuller: Fred, it sounds like this type of technology is going to allow companies to really move over time to a new level of efficiency and effectiveness, in terms of their inventories, in terms of their time to market, their responsiveness that things are happening in the market, how they’re incurring costs, whether it’s for promotions or in terms of order cycles. When you see this becoming a stable part of the operating system, what does that look like?

Laluyaux: When you talk about the operating system, you’re exactly right. I think what we’re seeing right now is a rapid push toward the concept of self-driving. How can I bring automation, how can I bring augmentation, to some of our core processes, where the human really doesn’t have a lot of value, and actually where the value of the human is decreasing because the complexity and the speed at which decisions have to be made is such that we cannot follow the pace? And that’s where you see a massive gap now being created between the digital natives and those who are not digital natives. That concept of self-driving capabilities applied to those larger enterprises by enabling them to retrofit their entire infrastructure is absolutely critical. For the first time, I see these conversations are being highly relevant and driven by the boards, by the CEOs, of the largest companies in the world. They know that they have to change right now if they want to continue to thrive, or they have to come up with an alternative solution. And I believe that this cognitive operating system, being able to leverage what they currently have and bring them into the future, is really where they need to go.

Fuller: Fred, thanks for sharing all these insights about how technology is beginning to be put to use in new ways in companies and might provide a lease on life to those storied 100-year-old companies and prevent them from going the way of the dinosaur.

Laluyaux: Thank you for having me.

Fuller: Thank you for listening to this episode of the Managing the Future of Work podcast. To find out more about our project on the future of work, visit our website at hbs.edu/managing-the-future-of-work.