Demystifying AI and machine learning for executives

In this episode of our Inside the Strategy Room podcast, senior partner Tamim Saleh cuts through the hype around artificial intelligence (AI) and offers clear guidance for executives looking to make precise strategic decisions about where and how to employ AI in their businesses. Tamim shares insights on the impact of machine vision on AI, the future of voice recognition, and the latest developments in advanced analytics, virtual assistants, and robotics. He outlines the challenges companies face when adopting AI and the steps CEOs can take to overcome them. You can listen to the episode on Apple Podcasts, Spotify, or Google Podcasts.

Podcast transcript

Sean Brown: From McKinsey’s Strategy and Corporate Finance Practice, I’m Sean Brown, and welcome to Inside the Strategy Room. I’m joined today by Tamim Saleh. Tamim is a senior partner in our London office, and he is with me at our Global CFO Forum, where he’s speaking about AI and machine learning.

Tamim, one of the things you’ve talked about is the notion of five different developments of AI. I’d like to first focus our discussion on the impact of machine vision on AI.

Tamim Saleh: Machine learning and AI are limited by the fact that when we input data as humans, first of all we are slow, and we make mistakes. One of the fastest-growing technologies is capturing data through image analytics and cameras. And the beauty of this is, cameras don’t make the same mistakes we do, because they capture things the way they are, and they don’t see the world the same way that we do. In fact, the spectrum is much wider than what we see. It includes infrared, et cetera.

So there are a lot of business problems [that image technology can help]. Take, for example, mining, where traditionally people—geologists—will go and look at the ore, spend some time, and write a report.

And then you adjust the angle of digging accordingly, and then you do this once a week to optimize the yield. Now you can do this in real time. There are cameras that can actually monitor the geology and, in real time, adjust the angles of digging. For a mining company this could be worth hundreds of millions of dollars.

This also could be applied in safety, for example, in oil and gas, where the cameras monitor people’s movements. And if there are any likely compromises, the algorithms would give warnings and something could be done. In fact, there are hundreds and hundreds of use cases or real-life business problems that will be resolved by image.

Imagine this: the amount or the level of information that you get through a combination of image and sensors is up to a billion times more than the traditional methods. And with machine learning, when you get so much input, you get the most out of machine learning much faster. And this era is just starting.

Sean Brown: What are some of the gating factors? What is keeping companies from adopting this notion of machine vision?

Tamim Saleh: There are probably two things. And neither of them is technology. One is talent—people who know how to use image analytics in a useful way in business; people who could translate this into actionable, tangible outcomes. There is a quiet revolution going on in the world where, gradually, different companies are training and skilling people to translate these types of problems into action, but it’s taking time.

Problem number two is, how do you connect these types of propositions to an existing business model that doesn’t work that way? If you’re an insurance company, for example, to capture a claim, you still have to fill forms and call the call center. Some insurance companies are beginning to say, well, actually, if I take the image, and I run algorithms on it, it’s much more accurate than somebody calling me. They could lie in the first place, and I fill a form through an agent. But it’s going to take time.

Sean Brown: So they take an image of the accident, of where the accident was?

Tamim Saleh: Correct.

Sean Brown: Interesting. You talked a little bit about machine learning. That is one of the other big developments of AI. What are some of the things that are supporting that? And could you take a minute to describe how machine learning works?

Tamim Saleh: There’s a lot of hype about machine learning, and people get intimidated. Actually, it is a really simple concept. Machine learning is pure statistics. When algorithms work to solve a problem, they predict an answer to a question. A question could be, for example, “What is the probability of somebody defaulting on a mortgage?” And machine learning basically is to test what was predicted versus what really happened, and then adjust the algorithms to get a better level of prediction. And that’s what it is.

Machine learning becomes extremely useful as you combine this with the human judgment, because human judgment is not going to disappear, and as you absorb a lot more data, you get more accuracy. The driverless car, for example, works on machine learning. A real driver drives, and there’s a program that observes what the human does, and then you improve all the time.

Sean Brown: And so if the human is correcting, that becomes part of the learning.

Tamim Saleh: Correct. We need to think of machine learning as human and machine, rather than just machine.

Sean Brown: You talked a little bit about virtual assistants as well. I’m not sure if you saw that Google video where the virtual assistant had a conversation with the salon owner and made an appointment for their client. Can you talk a little bit about how those virtual assistants currently work and what the future might hold there?

Tamim Saleh: Despite all the hype, the virtual assistants now are still very primitive. But in six to eight years, as the computational power continues to improve, the software technology continues to improve, and the infrastructure that supports this continues to improve, our relationships with machines will change.

My generation sees machines as something you input information into. You type. The next generation, they will see the machine as a genuine assistant that you can talk to. You can put information there through image and voice and the machine gives you advice and talks back to you and goes into a process, a continuous thinking reaction. We are not there yet, but it will come. And when that happens, it will dramatically change the way companies work—whether it’s a field force out there working and repairing ducts or somebody in a factory checking quality or somebody sitting in an office in the finance function trying to understand data in a report and trying to find more insight. And that revolution is just starting.

Sean Brown: How about robotics and how that ties in?

Tamim Saleh: We need to think about robotics in two different ways. One is software robotics. That already exists, and it works extremely well. Software robotics is like a combination of text mining, image recognition, automated letter writing, et cetera. When you combine these types of assistants, you can dramatically improve repetitive tasks that humans do and let the humans do the judgment part. Because judgment is where we are gradually going to migrate to while repetitive tasks gradually pass to machines in the next decade or so. This software robotics already exists, and it works extremely well, and it’s improving all the time. And in the next three to four years, in most call centers, for example, the agents will be using this stuff. They’ll be talking on the phone. Somebody talks to them. The machine will be giving them advice in real time. They will not type, or if they’re trying to extract data again, they will talk to the machine to get it back, and it will dramatically improve productivity. Some organizations are looking at efficiency improvements of 50 percent or even 60 percent because of that.

Now robotics in the physical world has existed for a long time, but we should not mix that with the robotics using machine learning, because there has been robotics in the car industry for 30 years that does very repetitive tasks. It does not change what it does and is programmed to do this same thing. But then there is future robotics, which is coming along very nicely now—it adjusts what it does, and it learns from what it does. It uses machine learning, visual analytics, the concept of voice analytics, and being able to talk to people and gradually changes what it does and adjusts it and improves it all the time. That will be a major revolution as well.

Sean Brown: And now voice recognition and analysis: Could you talk a little bit more about that, please?

Tamim Saleh: Voice analytics is also one of the technologies that is rising very fast. In many, many call centers in the United States and Europe and so on now, when you have a call, the system will capture what you’ve discussed; it will be converted into text, in real time; and triggers and patterns of that text will be captured. In some cases your sentiment through your voice fingerprint will also be captured, so how you feel as well. For example, if you’re on with a call center and you say, “Look, I’m not happy about this bill, and this is the third time I’m calling you,” “not happy” would be captured, as would “bill” and “third time calling.” Immediately you can work out something. Your sentiment is captured, and if you are a valuable customer, you might be right at the head of the queue in front of everybody else when the call center is connected through telephony to the back office, making the process much faster.

Sean Brown: So you might get a manager much more quickly, for example.

Tamim Saleh: Now voice analytics has thousands of different applications, and it is rapidly growing. Combined with image analytics and sensor technology, it is going to become part of what we do in everyday life. And depending on the organization, if you have that channel or you have interactions directly with people, whether B2B or B2C, almost certainly voice analytics is very relevant and also within the organization.

Sean Brown: How about video analytics? Have you combined the audio and the picture, and how is that coming together?

Tamim Saleh: It’s beginning to. There are lots of applications of video analytics separately from voice analytics, but there is a whole science, for example, which is team performance. Team performance is to optimize the way people work together—team mix—the way they communicate, how they are measured, their physical location. And image analytics plus voice are used to provide input into the many, many variables to get the most out of these teams whether they are R&D teams, for example, or sales teams, working within the constraints of regulatory requirements of course.

Sean Brown: Let’s talk a little bit about data. A number of the examples involve massive amounts of data. How are companies protecting that data and keeping it from being used in a negative way?

Tamim Saleh: Companies are learning how to do this because these developments are happening so fast. To give you an example, in the UK, one of the companies has been capturing people’s images—real-time videos of when they’re looking at advertising boards and walking the streets. They are capturing how people feel when you have, for example, a shocking image, and you look at the product or the image and then use these techniques to adjust advertising images and so on.

Then there were questions: Who has this data? What are they doing with it? And of course the regulation does not yet fully answer some of these questions. It is coming along. So what companies are doing is developing capabilities to respond to regulations like GDPR [the European Union’s General Data Protection Regulation].

And essentially there are four things that companies are trying to be good at. One is to be able to justify doing this if they’re asked the questions of why they have this data and what they’re doing with it. Two, they need to be able to track that data and understand the context where that data came from and whose data it is. And then third, they need to be able to be transparent quickly, should the question be asked, whether from the customer or the regulator. And fourth, they need to be able to intervene even before problems happen. Now, these four things I’ve said, they’re not always straightforward, especially when you have image analytics and sentiment and all of this. But in the next five years a lot of developments are going to happen in this area.

Sean Brown: If you’re CEO of a company that’s not using any of this right now, and you talked about vision, virtual assistants, machine learning—let’s say it’s a B2B, it’s more of an industrial company—what are some of the first steps that you recommend companies take if they’re trying to get smart and then really move the needle on leveraging these tools?

Tamim Saleh: If I’m the CEO, the very first thing I would do is that I would get the management team and I would spend a day to two days and really learn what true advanced analytics is and what these techniques are. It’s very important to understand that, and I personally spend a lot of our time with our clients to make sure that at the top level they understand that.

The second thing I would do would be to make sure that we were very precise about where the value was. If I’m in a B2B business, what is the opportunity from advanced analytics that optimizes the effectiveness of the sales force, for example? How about my demand forecasting and the entire supply chain: What is the value? What is the value on operations? What is the value on pricing? It’s important to understand that. That would be my second step.

The third step I would do would be then to think about the problem in terms of people. This is not a technology problem, and it’s not mathematics either. The main barrier is the mind-set and not having enough talent within the organization to be able to do that. And the answer almost certainly will not be hiring people from outside; there aren’t many, either, and they’re expensive. The answer is to start from the business problem, prioritize and understand the sequence where you’re going to deploy answers to these problems in use cases, and then have a training-program certification and a clear career path for data scientists, engineers, translators, et cetera. And then as you develop these use cases gradually—each one with a business case and return so it pays for itself and more—you build the capability in the organization. Typically, it would take 18 months or so where you really begin to have enough critical mass that you know that the momentum is irreversible. But you start with educating yourself as a CEO.

Sean Brown: Thank you. To flip to consumers, as companies become more and more adept at leveraging these developments toward AI, in some ways do you see that tilting the field back? Right now, digitization creates a lot of consumer surplus. How do you see AI being used to—I won’t call it the opposite of consumer surplus—but how can companies use it to improve their value proposition or to create more consumer loyalty without necessarily having to “give away the store”? So how will advanced analytics help companies be more profitable?

Tamim Saleh: In a number of ways. Over time, companies will be able to know the customer much, much better. Part of the problem today is that we don’t have a single customer view, and the data that we have is not reliable, so we don’t really have a good way to extract external data and match it to the customer. All these barriers can be overcome to a large extent using advanced analytics and the new technologies—open architecture, et cetera—that allow you to run analytics on unstructured data.

What will happen is that instead of, for example, having eight relatively generic customer segments, you can have 15,000 microsegments. You can even go more extreme than that and have the single customer DNA. You can run millions of correlations per customer based on their behaviors, what they’re buying, their socioeconomic groups, where they live, et cetera, and find patterns and correlations with their propensity to buy something, for example, or with their propensity to react to a message or what makes them happy or not and then make sure that you do the right thing for the customer.

The ethical part of analytics is extremely important. You need to do the right thing for the customer, based on that very microsegmentation. You combine that with developments like, for example, being able to see traffic movements within a 300-meter radius in real time at an extremely accurate level, plus you know the individual customer, and you begin to have options that you never had before around pricing and promotions, et cetera. It’s an extremely exciting time, actually, knowing your customer better and being very precise about what you can offer them, using the digital channels.

Sean Brown: At what point does this become a black box to the provider? In other words, if you’ve got 15,000 customer microsegments, how do you know that they’re the right ones, that you’re doing the right thing for each of those segments? You talked about this notion of machine learning and human-involvement segments—how does that work in the microsegment example that you just shared?

Tamim Saleh: The analytics are statistics and correlations based on facts, based on what actually happened.

Sean Brown: Observable facts.

Tamim Saleh: Correct. So if you have 15,000 microsegments, whether it’s relating to propensity to buy something or reaction to a message, it would be based on how these microsegments have reacted in the past and prediction of what might they do going forward. If you have got your algorithms in the right way and you are testing them continuously using machine learning, you should have a reasonable level of confidence in terms of knowing what the customer wants.

I’ve given you a long answer, but it is not guesswork. It is actually statistics based on facts and predictions based on that. You can still have 15,000 microsegments, but the way you reached 15,000 is that you used hundreds of millions of correlations and data to get to that level of granularity, and even at that granular level you can be very precise in terms of what they might want, what messages you should do, and you test that all the time. The important part of this question is that in the existing channels—the call centers, for example, or branches or face-to-face interactions—it’s very difficult to work at the microlevel, but that’s where the digital channels become extremely important, because the messages could be tailored and coming from the algorithms and so on.

Sean Brown: Now, to go into the physical world, what are some of the ways that companies are offering or executing those microsegmentation techniques but in real life, in person?

Tamim Saleh: There are some consumer companies that are testing in places like Japan, South Korea, and the US. When a customer comes in to buy a TV, for example, the customer has two options. They could either talk to a machine or to a person. Interestingly, 55 percent of the customers in South Korea prefer the machine. They ask the machine about the product and even trust the machine more. You’re capturing data all the time, including about the voice and how they feel. It’s the physical world: the customer is actually buying a product, seeing the product and holding it, but using a robot or machine in the interaction. Or if they decide to work with a human, the human also can have all the data and is assisted with the machine. So over time, basically there will be a blurring effect, a conversion of AI, machines, and humans in the real world. You’ll be talking to a human, but he or she will have access to data that could make him or her much more useful to you.

Sean Brown: It’s like having a virtual assistant in the store that’s helping you.

Tamim Saleh: Correct. Connected to their CRM [customer-relationship-management] systems. And all this stuff is being experimented with. Sometimes it’s business as usual, but we are still at the beginning of these things.

Sean Brown: For the five developments of AI, is any one the first step, or do they tend to all happen at once in terms of that interaction? Because you’ve been talking about the interaction between machine learning and vision—they seem to go hand in hand. Are there any that come before the other?

Tamim Saleh: Yes, in some situations. You could think about how machine learning and AI are evolving for business in three waves. One is using existing data that we have for insights. And that’s almost pure analytics. You’re running analytics. And most of the application of advanced analytics today falls into that category.

And then wave two is when you combine vision and image analytics and voice analytics with machine learning, and the machine becomes more like the agent, where we begin to interact with the machine as an agent. You have an ongoing process of interaction as opposed to us being proactive, putting data in, and then running a program. That wave is beginning to happen.

Wave three is when you take this into the physical world, when there are robots that are using this technology. For example, you go to a shop, and it will make the sandwich for you and the coffee and carry the boxes and all that. Each one of those waves exists in today’s world, but we are much more in wave one at the moment.

Sean Brown: And are there any technology constraints right now? Are we hitting any limits in terms of Moore’s law? Or is the technology staying ahead of all of this?

Tamim Saleh: The problems are talents and mind-sets and changing the way we work. That’s 70 percent or 80 percent of the problem. We have done surveys that show that 75 percent of the problems are in these categories that I’ve just mentioned—talent, organization, mind-set.

When it comes to the more sophisticated techniques of, for example, the virtual assistant or physical robots that use image, et cetera, there are areas of those that continually need to develop. Processing power—we still need more of that. Dealing with issues like latency, for example, especially when you’re using real-time data. Security issues. And the actual robotics themselves—how the robots operate, for example, handling and so on, is extremely complex, and the current technologies that we use, they’re not the most efficient way of developing these types of robots. We still need more development, but it is incredible the rate we’re going—every two or three years we are doing orders of magnitude better than where we were before.

Sean Brown: If you’re looking at implementing AI within your organization, you talked about a number of different key elements. The first one that you discussed was vision and strategy. You talked earlier about how if you were coming in as a CEO, the first thing you’d do is get the team together. Is that also when they get smart on advanced analytics and they also agree on where the value is and what the vision will be?

Tamim Saleh: It’s the kickoff of that process. When you understand the truth from hype, so you’re not intimidated as a CEO, you are much more comfortable to begin to discuss strategically what advanced analytics really means for you. In almost every value chain, whatever the industry, there are opportunities to have new business models or new products, and so on, as well as improving the core business. You start with education, and the next stage, which should not take a long time, is to go precise and think, “What does this mean to my business? What are the specific opportunities? What’s the value? What are the use cases?”

Sean Brown: In designing organizational models to better support AI, are there any specific pitfalls that you’ve seen that companies should try to avoid or specific steps in rethinking the organizational model that really help speed things up?

Tamim Saleh: Yes. We humans are very similar, whether we are in Africa or North America or Asia: we repeat the same mistakes. There are probably three or four that are important.

One, do not centralize the analytics capability and expect that the organization will adopt it. It just doesn’t work that way. If you have a central team of data scientists and so on, and then they go to somebody with 30 years’ experience and say, here is a model, guess what? It is very unlikely that he or she will accept that. At the same time, do not go the other extreme. Do not have complete chaos where everybody everywhere can experiment with analytics without having a common language and common methods, protocols, and methodologies for mathematical modeling and so on. That will create new complexity, and you’ll never be able to scale it. You need to get the right balance between uniqueness and freedom within the business units and centralization in terms of the standards and the commonality of methods to apply. And actually that is the trend that is happening. Typically, organizations will have a small layer, centers of excellence that define the approach and methods and tools and so on, and they do the training for the wider organization. But the agile labs that deliver the analytics use cases, et cetera, are inside the businesses. So that’s one learning.

The second learning is that you need to be very precise about the roles. What is a data scientist? What is a data engineer? What is a data architect? What is a translator? And embed them into your existing organization. If you don’t have a very clear definition of a data scientist, they’re unlikely to stay in the organization because they’ll get frustrated and because everybody else will call themselves data scientists. You need them to have a career path and training and certification so that they are motivated. A lot of organizations don’t do that.

And then the third element organizationally is that you need to understand as a CEO that this is not just about mathematics and data science. Actually, most of the value is in your existing people in the business who can translate their knowledge into information that can be modeled and used by a data scientist. You need to pick up 10 to 15 percent of the population and train them while they are doing the job. Train them to have that additional skill of being translators. Without doing that, almost certainly you are not going to be able to scale analytics in the business.

Sean Brown: To switch gears very briefly, if you were in the CFO’s role, what are you seeing as the most common things that CFOs need to think about with the advent of advanced analytics and how it can help the CFO really drive the organization to success?

Tamim Saleh: The CFO will be one of the greatest beneficiaries of advanced analytics in the future. Why is that? Because at the moment for the CFO, the information that he or she gets is not real time most of the time, it’s not always very accurate, and he or she is on the receiving end. Now imagine if you are in an organization that uses image analytics in real time, voice analytics in the call centers, demand forecasting, and inventory management. If an organization is using advanced analytics in what they’re doing, a CFO should make sure—this is action number one—that whenever use cases are being developed by the organization, the KPIs [key performance indicators] and the measurements that would help the finance function in the business are designed as part of this process.

Over time you’re beginning to build an information center where you are much more proactive. When demand for costing is being done using algorithms, the same information needs to go to the CFO, and the context is understood—how it links to the full picture, for example. Now when you add this into hundreds of different points over time, you will find that the CFO as an information-center leader is much more proactive and strategic than being a bookkeeper, for example. It’s completely different.

In addition to that, the CFO needs to have his or her own capabilities inside the finance function to use analytics for costing, accounts receivable, accounts payable, all the things that they do—risk management, all that. But the opportunity is much bigger than the finance function.

Sean Brown: And if you are a CFO interested in getting smart on this, are there any specific books or things that you’d recommend for a CFO who really wants to get quickly up to speed on advanced analytics?

Tamim Saleh: The CFO, like any other executive, needs to really learn the truth from the hype. Typically, a day or two of training is excellent. There are lots of places that you can do this. Reading books, yes. But be careful because a lot of the books also have hype.

Sean Brown: And they can be outdated as well.

Tamim Saleh: Correct, but understanding the basics of statistics and how advanced analytics work and how it is being delivered is important. The concept of agile teams, labs, is really important, because that’s where the failure tends to happen as opposed to the mathematics. Then, beginning to think, “Okay, how do I build that capability in my team, and how could I be part of the wider organization?” This is really, really important. If you intervene early as a CFO and begin to have your fingers in the different parts whenever the organization is developing use cases, then as an information center, you’ll really be ahead of the pack.

Sean Brown: Tamim, thanks so much for taking the time with us today.

Tamim Saleh: It’s a pleasure.

Sean Brown: We appreciate it.