This week we speak with David Carmona, General Manager of AI at Microsoft. Carmona discusses how redefining a business process is a very different kind of AI adoption project than working on something that is horizontal.
He discusses how to attack both of these scenarios, which to handle first, and why.
In addition, Carmona talks about proprietary data and things that are close to your own IP. How do you take advantage of the real strategic data value within your own organization? How should you be thinking about that differently? Carmona poses three different questions to determine where those valuable opportunities are for you.
Subscribe to our AI in Industry Podcast with your favorite podcast service:
Guest: David Carmona, General Manager of AI – Microsoft
Expertise: AI and machine learning
Brief Recognition: Carmona has worked at Microsoft for 18 years in various managerial roles, becoming its General Manager, Artificial Intelligence in 2017.
(02:45) What do you think some of those most common misconceptions about AI adoption are across industries?
David Carmona: That’s a very, very good question and we see many misconceptions. I think that if I have to focus on two, I would say that the first one is actually what are they capabilities, right? So there’s so much noise right now in the market on what AI can do and what AI cannot do. The first question that I always get is, what is truth behind all of that? Right? So what are the real capabilities of AI? How would you define AI? How it can impact my business. And there’s a lot of disclaimers to do in there because of that common misconception of AI because of press and because of they hear everywhere about what are those capabilities on AI.
(03:30) So for you, it sounds like the misconception is not having a firm grasp of what’s hype and what’s not?
DC: Exactly. Yeah and what are the capabilities? So going to the next level, the question that we get all the time is, “Hey, make this real.” So what exactly can I with AI and what exactly can I not do with AI? So that’s usually the first question that I get. The second one, by the way, and we can go in that in more detail, is that misconception on, “Hey, what do I need to get there?” So what are the skills that I need? There’s this huge perception that you need to be a PhD to even get started with AI. That I think I for specific scenarios may be true, but for the most common scenarios in an enterprise is definitely not true.
(04:30) When you think about getting an executive team up to speed on, what our real capabilities? If you were advising executive folks as to what the heck AI can do in finance, that’s not hype, how do they get started?
DC: Yeah, I think that the first thing that I do is to focus on the core capabilities. So I try not to talk about AI as a technology, but just as a concept that can give you new capabilities that you haven’t available before. So it’s like a new tool that you can use for your business instead of technology by itself, right? The three things that I always use to explain that and they are very basic. So it’s the very foundation of it.
But of course, the most important thing to explain in there is the concept of learning, right? So how we’re moving from just developing manually, providing the structures to the machine to do something, and to have the machine learning by itself with either data or experience, right? So the whole concept, of course, of machine learning and that has a huge implication on how you can use AI for the business.
But then, that’s the core aspect of it. But then to make it real, I usually focus on two areas that build on top of that. One is perception: the ability for the machine to understand the war around us. And that’s where you can go in more detail on areas like vision, speech recognition, or text understanding and so on, right?
Which are one of the first steps on any enterprise to get into AI. Then the last one that is usually the most complex one to explain is cognition, right? So it’s being able to use that learning capability to have cognition on top of data for very basic stuff, like regressions, right? Predictions or classifications or anything that can help you either, help you make decisions based on your data or improve processes or your core business because of that.
(06:45) Do you guys use some kind of a visual map to showcase that or is it more of explaining that conceptually? How do you convey that message?
DC: Yeah, so for each of them, we go then in more detail on the scenario. So for us, the way that we’ll look at it is that each of them will provide you a map of the core capabilities. So for example, perception, that’s an easy one. We go in more detail on things like vision, text to speech, speech to text and so on. Right?
So in cognition, same thing. So we go into six patterns that are part of cognition that can help you understand what is the detail behind that. So we use that as the core map and yes, we do have content for customers two or three. I think the important thing is that those are your building blocks, that then you can apply to a real product, right? So that’s the core foundation that we use for the business to understand not the technology behind it, but the capabilities that now they can apply to the business. So it’s a good dictionary for the business user to understand what they can do with AI.
(09:00) What are some expectations businesses should have up front if they haven’t worked with AI a lot in the past?
DC: So the first thing that I always try to explain is how you shouldn’t look at AI like the most complex scenarios directly, how you can get it started easily and this connecting also to the misconception that we mentioned before on the business users and business leaders, thinking that they do require from the get-go a huge data science scaling organization or a scale organization to address AI.
I use another structure for that, that I try to use to explain what are those short term scenarios that they can do with AI, that they don’t require that deaf knowledge on AI or deep learning and so on. Those three things that I always say is first, it’s connecting a little bit with the three key capabilities of AI that I mentioned before.
The first one is thinking that AI…as a better software, a better way of creating software. So don’t think yet about transforming your entire industry with AI. Think of how you are already a software company. So there’s no industry nowadays that is not a software company as well.
So I look at AI as a way to do that software better, right? So you can have many, many new capabilities in your business applications or external applications with your customers by using AI.
In most cases, you don’t require either a huge amount of data or a huge knowledge of the AI techniques to do that. You can use previous services; for example, cognitive services in Azure that you can use out of the box to infuse AI into those applications and make them better. So that’s a very, very easy way of looking at AI with which we see a lot of enterprises having big success in the short term.
(11:30) What level of in-house talent is often needed?
DC: Yeah, so it depends on that level of maturity. If you are infusing AI into your applications…these pre-built AI services, they are targeted at developers. So you can use your developer…if you have one in-house to infuse AI into your applications, right? You can customize those models and you can do a lot of customization and connection to your business without any data science knowledge, right? So that’s powerful.
But if you go beyond that, I think the next step where it is important to go to the next level of skills is to bring AI not only to those applications but to your business processes, and now we’re talking. So now that is a different league. That is not about modernizing your existing applications or creating new applications. For example, I don’t know, a conversational AI application, which you can also do with development skills.
But now you’re talking about changing, improving, or even redefining a business process with AI. That usually requires data science knowledge and skills. Although, I have to say that even in that case, I double click there a little bit and I distinguish between two different processes.
So I think we have to look it as horizontal processes. So think of processes like marketing, sales, customer service that are very, very shared or very, very similar across different industries. So customer service or marketing, they are very similar, no matter if you work in manufacturing or you work in healthcare.
For those, you already have out-of-the-box solutions that you can try to use. So you have this concept of SAS AI, that in many cases with a little bit of customization, without again, requiring deep knowledge of data science, you can customize for your business and those are very powerful.
I think what you need that data science experience for is for vertical processes. So think of processes that are very unique to your differentiation as a company. If you are Uber, you probably don’t want to use an out-of-the-box solution for your core distribution algorithm. So that’s your uniqueness. That’s what makes your company your company. So in those cases, we do see custom development and those usually require, of course, more skills and experience on data science.
(15:30) What you’re saying is those types of data, types of processes, types of systems that are really unique to us, that probably there isn’t a vendor doing it with 300 other companies. It’s kind of tailored to something specific about what we’re up to. Now we’re going to need a lot of that business context and in-house data science knowledge to really be able to mold a solution to fit that. Is that kind of where you’re headed?
DC: Exactly. Yeah, and there’s this rule of thumb that I always use to identify quickly those processes and the three questions that I usually ask is, what are those processes where you have unique experience? So if you’ve been doing this for 20 years, you are probably going to know better than anybody what is required to make it better with AI.
The second question is, do you have specific data? Do you have unique data that nobody else has on that particular process? That is also a sign that you may want to use custom AI to leverage that data. The third one is, do you have IP? So is there anything that you develop that is very unique to you, that you think you can take it to the next level with AI?
(17:00) When you think about an enterprise who’s trying to decide, “Where do we apply this first?” What are your ways of thinking through that problem at an enterprise level?
DC: It is definitely so important to pick that first project correctly because I saw so many projects that are either disappointing because the project was big or not meeting the expectations on the impact because you went too low on the ambition for that first project.
So it’s a really critical decision. I think the most important thing that I say here is always make that decision from the business side.
I know that the first AI project should be small. So we always say start small. Don’t take so much that you are not yet prepared to deliver. But at the same time, you should look at it as the first step in your journey.
So what I mean with that is that, don’t look at that first AI project. That’s just a pilot on some random place that you think you are going to get benefits in the short term.
Look at your long term strategy. Think about where you want to be in three years, how you want to really find your company with AI and pick a first AI project that is a step in that direction. Yes, you’re not going to fulfill that entire strategy in the first project, but at least you are going in the right direction. So that’s something very important.
I heard yesterday, this way of referring to the stage where many enterprises are today. It was referred to as the pilot purgatory. So you are always in pilot mode, right?
I love that expression, and it’s so true. It is so true. So we see so many companies that are in pilot forever, and yeah, it is great; they are learning, but hey, what are the results? When am I going to have this in production and get real results, right?
I close the loop so I can keep learning and that is so, so important. I think there’s something, like common factors in here that I see all the time, when I see this pilot purgatory. But I think the most important one is that disconnection with the business.
So it’s looking at AI as a way for the business to be empowered, supported by the technical organization and not the other way around. So not looking at the technical organization, making those distinctions and then the business trying to get that.
An example that I think here is actually our internal experience in Microsoft. So, of course at Microsoft we also use AI internally for our processes, and we use it across sales, marketing, customer service.
One of them is finance. So in finance, we actually do…Today, we do our forecast as a company with AI. So a process that was primarily done by the finance department on a manual way, now it’s all driven by AI. So all the forecast that you see externally from Microsoft, that is being generated with AI, with human, of course, involvement and integration.
So, we move that process from being 3% budget to budget to now 1.5%. So it’s becoming even better, and what is even better for the finance department is that now, they can focus on things that are less operational, like forecasting, and more impactful, like understanding or identifying what the next businesses that Microsoft should be part of are. So that was huge.
Now how do we get there? That reminded me a lot of the pilot purgatory because at the very beginning, the way that this worked, the way that we tried to transform our finance organization with AI was very top-down.
It was basically, “Hey this is where we want to be, this is how we’re going to drive these AI transformations.” It was very ambitious; it was great. It just didn’t work, and it didn’t work because of the business user.
So that final user in the finance department was not involved from the beginning, and the way that we changed that dramatically was doing it just the other way instead of top-down. It’s getting the business user, the finance person involved in the process. We actually did a training to that business user base, and it’s very similar to… So it’s going more detailed on those three capabilities that I was mentioning before.
We created that training; that is actually external too. So people, any enterprise, if you search for AI business school, is the training that we also provide externally for business users to understand what AI is. We delivered that internally in the organization without any guidance on what projects we should address with AI. So very different from the initial approach, and then we waited for the business user to be the one providing, “Hey, these are the use cases that I want to target.”
By a long shot, the most popular use case that they wanted to target was forecasting because it was a pain for them. It was not something that they enjoy. They preferred to do other things.
That changed it dramatically. So that process, to identify the first AI project, made the difference for them to be successful. Because you have the business user involved, they’re going to be part of it, they’re going to provide feedback, they’re going to use it, they’re going to provide the right data and you are going to get into a loop that is going to cause that project to be successful.
(24:30) What you’re saying is, shoot for a place to land that is not just to play around and build the skills, but is in line with your ultimate goals and can be seen as kind of a bit of learning, a bit of a building block to get towards where you want your company to ultimately be. So it’s not a toy application in a dark corner. It’s something, maybe you won’t have tremendous returns, but you’ll build core skills and you’ll build toward something important. It’s in line with longer-term strategy.
DC: That is exactly it.
Header image credit: BankIslami