Top
Podcasts
5.25.2020

Fabian Merkel – What do weed killer, pharmaceutical drugs and artificial intelligence have in common?

The Cogniverse Podcast: In this episode we talk about what is the right approach and strategy to start working with AI from a business perspective.

Attila Tóth

Digital Strategist

“I believe that curiosity and courage make the difference.” Fabian Merkel

Fabian Merkel is a Digital Innovation & Venture Consultant from Germany. He is dedicated to building great products and ventures which make things easier, faster and better. On the path of achieving this goal, he’s currently exploring the opportunities of AI and ML.

As an imprint of this process he’s also a Podcast Host & Community Manager at ai-zurich, an initiative that aims to bring AI closer to DACH region (Germany, Austria, Switzerland) and to build an AI community through specific events and meetups.

You can connect with Fabian on Linkedin.

Listen to this episode on Apple Podcast, Overcast, Spotify.

Please enjoy:

Podcast transcript:

Attila Tóth: So, hi Fabian, welcome to the Cogniverse Show. I'm glad you're here.

Fabian Merkel: Hello Attila, thank you for having me.

Attila Tóth: You are a true AI enthusiast, who doesn’t like to beat around the bush, right? Could you tell me where your interest related to artificial intelligence comes from?

Fabian Merkel: Sure. Actually, it started like, almost two years ago, I was working at a big consulting project, we did was data mining, in particular, was about process mining. And within this project, everybody was talking about artificial intelligence and how they want to apply it and which artificial intelligence should we buy. So actually, to me, I had the impression that there were a lot of people talking about AI who didn't really understand what it is. And I felt that some way, and then from that I started to educate myself about AI, in particular about machine learning. And somehow that really grabbed my attention in the way of its potential, in the way how it will transform our economies, our society, yeah, our actual life. And yeah, since then, I'm really passionate about it, started my own podcast, which is called "AI for business", so I'm interviewing people, who work with AI, who do projects in some field of AI, and I want to give them the voice so they really tell how it is, they do not use all the kind of buzzwords, you know, and then just tell the same story which the mass media is always telling. So I really want to have an eye on something, how it really happens.

Attila Tóth: Sounds pretty good. This means you probably have a lot of definitions on AI. If I would ask you to give, let’s say a Cambridge dictionary definition, how would you define artificial intelligence from a business perspective?

Fabian Merkel: That's a pretty good question, actually, and having a definition of AI it's pretty complicated itself. A colleague or former colleague of mine, she was writing her master's thesis about a definition of AI. So that's a big topic. But a really easy definition - and I think that's the most applicable - is: AI is a technology that can take over tasks that previously could only be solved with the help of human intelligence. So I would say that's the official definition. But that doesn't really mean much to me. So for me, a better way to explain what AI is, what it can do, is by starting: okay, what is actually not AI? Ai is not a robot AI is not an in-kind of intelligence where you can talk to, so we do not have real artificial intelligence at the moment. What we have is actually machine learning, so that's only actually a bunch of algorithm, it's a bunch of statistical models, you know, all this probability calculation we had in school, that is what we have at the moment. So the question is, what is the cool thing about that, yeah? Why is there hype now around that and what can we do with it? The main difference between, like, traditional software and machine learning or artificial intelligence - however you want to call it - is in traditional software you have to program everything, yeah, it's an if-then condition, you know, if this case happens, then do this. So you can imagine when you do an excel sheet with a lot of formula, that's a lot of work, it's getting complicated, more and more. The cool thing about machine learning is that you do not have to program it, it learns itself. You can train it a bit like a small child, you know, you have a child and you show him a dog, and the child realizes after seeing a dog two or three times, okay, that's the dog now. And you can do pretty much the same with machine learning, of course, you need to show them some more pictures of dogs, that more than average child, but that's the principle how it works. So now you can do things which you couldn't program before actually, and that is from business perspective, the really- the cool thing about.

Attila Tóth: So, you are confirming that what you experienced with AI is related to making  data processing faster, more assisted, more efficient. And you also agree to the fact that we are not looking at Terminator phase robots, Skynet is not really here yet, right?

Fabian Merkel: No, definitely not. Yeah, you're absolutely right. The thing is why... I mean, artificial intelligence has been in place for - I think 60 years or something - the basic concepts were developed in the 50s. But nowadays, we have the data, because in the last 20 years, we were really collecting a lot of data. And also that we have the computing power, you know, we have martial law. So every two years, the prices decrease by half for the computing power, so now you have the ability actually, to put those concepts from the 50s and 60s into practice.

Attila Tóth: It took us quite a lot of years to put it in practice, but I agree. What is interesting that in my field of work with digital transformations and digital marketing, when we see the use of AI, it's totally - as you said - in the direction of statistics, or something that's more like a predictive assistance, rather than what people imagine and what Hollywood is putting in front of us. But, to move to our next topic, what do you think are the key benefits for businesses to think about implementing AI solutions?

Fabian Merkel: Sure, but maybe let me add something to the definition, what we have as narrow AI at the moment, we do not have a general AI so what AI is being able currently is to solve really specific tasks, like, okay, "recognize a dog" or "recognized a cat" or "say it's a cat or it's a dog." It's not a general artificial intelligence where you can ask, like, really complex questions and they will be answered. So maybe that makes it a bit more tangible to understand.

Attila Tóth: Absolutely, yes.

Fabian Merkel: Yeah. Now we're talking about the benefits. Your question was, why should businesses use AI technology?

Attila Tóth: Yes, and basically you met AI, myself met AI, but a business owner - or let's say - a head of marketing, a CMO who hasn’t touched base with AI and is looking to implement new technology in its business to be more efficient, to increase revenue, what would be the key benefits? How do you see it? What are the things why AI is worth thinking of?

Fabian Merkel: Sure, I think that two angles to look on this particular topic, there is this benefit: you can speed up processes, you can maybe save money, you can gain insights you haven't had before and great value upon that. So there are certainly some benefits, you can benefit from that, on the other hand, if you don't use it and your competitor starts using AI and he's really successful with it, then you have a problem because he will operate on a lower cost by the basis, he will be faster, he will have the insights. So I would look from that kind of perspective on the topic of AI, and why you should use it. I do not say rush totally into AI, spend millions on it, and transform your whole company right now. But what I'm actually saying is you at least should start informing yourself now, and you should start experimenting with it. Yeah, make little experiments in your company, see what you can do with it, build a knowledge, build a kind of confidence to work with it, and then see how you can apply it in a really beneficial way for your business.

Attila Tóth: Sounds good. As we are talking about general implementation and piloting with AI, what do you think are the minimum data requirements to really start leveraging the power of machine learning?

Fabian Merkel: It is also a really important and really good question, is the question everybody's asking. And the honest answer is it's hard to say. It depends on the problem you want to solve, it depends how good you want to solve it, and it's not also about... it's not only a question of quantity, it's also a question of quality, yeah? How are your cases distributed over the data? Is the data up to date and format, you know, are they accessible, are they complete? That's also kind of factors you have to look at. I talked to some people who say: "okay, with the data sets of... when the data sets have 10,000 entries, that is a starting point to train a model, but you cannot have high expectations towards this model. So I would say that's the baseline. The question is, what can you do if you do not have enough data? Well you can sometimes buy it, yeah, if it's not a really particular company, like, case - you want to solve it, you might get public data, you can buy it from another company, you can wait and try to produce more of this data, maybe you can, yeah, build a model which is less complicated so you divide your problems in smaller problems and only use those parts to automate or to work with a machine learning. So that is a lot of things you can do, and you can also really start producing it by yourself, there's little nice story about a bunch of students who developed an application for farms, you know, you have... on the field you have a lot of weed, and you have to spray weed killer on them. And what we did until now, you just got like a tractor and it was pulling weed killer all over, and weed killer is expensive and also weed killer is quite harmful to our environment. What they did is actually develop the model, which is able to identify those weeds and only spray the weed killer directly on the weeds. And when they started, of course, there was no big collection of pictures of weed. So what they did, like, they spent a few weekends outside on fields with their iPhone, and they were just taking pictures of weeds. And then they had a couple of few thousands, and then they had a starting point to work with. And then they continued. So yeah, thinking in a creative way, maybe there's a way for you to produce that data to work with.

Attila Tóth: It's an interesting story. So basically, even companies who are not yet owning a big data set or having huge data sets available can apply this strategy to build their own data. On one hand, I think this can be a story of inspiration for those who want to start piloting, and on the other hand, it can be a story of motivation to start building your own data. And for that, there are actually tools, for example, in terms of digital marketing many times what we like to use with our partners are data management platforms, where you can choose different data sets and purchase them, which are relevant of course, to your business, because otherwise, if you just tap into some kind of data pool that is not qualified and you don't know the level of trust you're having on that quality, then you can train your model to do something that is not in scope of your business.

Fabian Merkel: Exactly. Yeah, I think sometimes it helps to think about data it's like oil. A lot of companies think" okay, we have this bunch of data and it's like oil, it's a treasure, we can use it", but what you have is crude oil, yeah? You can do nothing with it unless you have a refinery, and then you can create value out of it. And if it comes to data, yeah, this refinery is a really difficult topic, because you have to take care of the quality after scope and all the things you have mentioned.

Attila Tóth: Exactly. I really like this symbol, the symbol of oil. As lately the internet started discussing that data is the new gold, and what I found and what I can say that maybe it's not the new gold, but as you said, it can be the new oil or it can be the new soil. So data is kind of like soil. If the soil is good, you can plant your seeds, grow your plants. But if the soil is not good enough, meaning the data is not precise or it's not relevant to your business, then whatever you're planting, the results won’t be something you'd like. Staying on this topic, what do you think are the most important AI-driven tools that a digital strategist, digital marketer, or basically any business executive should be aware of?

Fabian Merkel: That's actually pretty difficult to say, and actually, I do not want to recommend a specific one. And the reason is that it really depends on the problem you want to solve, there's no one size fits all solution. So what I recommend is do your homework, look, what kind of problems you want to solve, line the tools you want to implement, along with your data strategy, because the point is if you start training models, you have a kind of lock-in effect, you know, you can’t just download it and place it somewhere else that might be pretty difficult. So think about AI or machine learning tools a bit like an ERP system, once you've implemented an ERP system and you use it, it's really, really hard to change that system, is almost impossible. On the other hand, there are a lot of new companies in this space. And if you're like a big company working with a small company together, maybe they're out of business in two years and then maybe your models are gone, so that is also a problem. So make sure you really select the right vendor, do not rely too much on this vendor and have a kind of exit strategy. So do not make really long contracts or something. However, most companies do not have the problem of selecting a vendor for any kind of machine learning application. They have bigger issues to solve first. So they actually need to implement the data strategy, they need to maybe create a data lake, they need to make their data accessible. So that are the issues most of companies - at least in Germany - have to work on, and then later, they can think about the right tools for machine learning to use and implement.

Attila Tóth: I'm glad you put this on the table because before you can choose a tool, you have to make sure that you are ready to use it. But, before we go into this readiness, I just want to stay a bit more on the tools. From what I experienced, many companies believe that they are using AI because they bought a Software as a Service Cloud application, which states AI-driven X or Z, but actually when somebody who has worked with AI checks it, can really shortly find out that it has nothing to do with AI. What would be your suggestion for companies who are out there trying to pilot with different vendors, what would be a good way to avoid these fake AI software? Do you have any recommendations on that?

Fabian Merkel: Sure. First of all, you shouldn't try to use AI for the sake of it. If you have a problem or any kind of opportunity or challenge, where it makes sense to solve it with AI, fine, then do it. If you have, like, a problem which can be easily solved with a normal piece of software, why should you use AI? So keep that in mind, okay? Just because someone is claiming to use AI, you know, what is the matter? If they are solving your problem and in a good way, fine, yeah, you don't have to care, but I totally agree AI is a hype, everybody claiming he's using it, you know, it's a bit like teenage sex, everybody is talking about and nobody's really doing it. And of course, putting a lot of peer pressure on anybody else. I think we all can remember that teenage days. So don't get stressed about that. It's difficult to tell the difference, because, of course, you can't really look into their system, into their code, and maybe if they're really overusing the term, you know, they're somehow maybe not really right about it, and I also know a lot of companies who actually use ML models heavily, and they do not say anything about it. They don't want anyone to know because they're afraid that they lead competitors to their way of solving a problem in a really efficient way. So, yeah, if somebody is really overusing the term, that's maybe not true, there's also a little joke it's called "machine learning is written in Python, and AI is written in PowerPoint". So if they're also talking about AI, maybe they're not really doing it because AI doesn't exist, what do people use as machine learning, deep learning, etc, etc. So maybe it gives you an indication.

Attila Tóth: Yeah, it's an interesting joke, but yes, I fully agree. And now let's get back a bit on what you said about putting the processes in the right phase before starting to think about AI. What would be the starting point when you want to think of solving a solution, you realize that AI might help you, what are the things you have to make first ready in order to be, let's say "AI friendly"?

Fabian Merkel: Well, it really depends on the problem, really depends on what kind of company you are and what is the current state of your IT, but let's assume you do not know much about how to use artificial intelligence, you just have a glimpse, and you only have the impression that might be something which, you know, can help us as a business and might be really interesting for us, but you do not know much more. So, first of all, I would really start learning about it, really read a few books, do a few online courses, maybe listen to podcasts, go maybe to a conference or something. So get to know the topic. Once you do that, put it on the map in your company. Colleagues talk about it, maybe you find some nice use cases and do not wait for the management because the nice and interesting use cases for AI, actually, they're mostly found by the subject matter experts, because they've said "okay", that's the kind of problem, that's the kind of daily task I can maybe automate with machine learning. Management or the really higher management is usually not the best address to find those really interesting problems. So really integrate your colleagues integrate subject matter experts, also learn from others, you find a lot of interesting use cases online. You can also sometimes transfer that directly into your business because all that kind of businesses have the same issue or sometimes as far away but still you can find some similarities and then experiment with it, do a few proof of concepts. Just try it out, but then make sure you get the right approach. So you should really think about your strategy, you really should put to play things like data governance, all that kind of stuff, so it's not becoming, you know, headless, we just tried to integrate some, you know, AI models here and there. So AI was the strategy, and that's the big topic, and then it's also the level where your really unique problems will arise. Do you have a data lake, you know, are you in the cloud, are you on-premise, do you have any kind of regulations, do you have like all these GDPR issues, all that kind of stuff comes up at this level, but you also have to solve that kind of problems.

Attila Tóth: So you're saying, start getting to know the topic, then bring it to the people closest to you in the company to see what would be the potential of using machine learning and basically, once you have one or two pilot ideas, then start thinking about data governance and, of course, your data strategy.

Fabian Merkel: Exactly. Maybe there's something I would like to add. If your work was machine learning, then you shouldn't confuse it with traditional software development. In traditional software development, you can nearly solve any problems just with time, money, or you put smart people on it, you know, they can program it, they can make it real. If you work with data, that's not the case because, you know, the performance of your model depends on your data and you can not influence the data directly, you know, there's no magic formula behind. If the data is not good enough, your model won't be good enough. So just you decide to maybe, integrate AI doesn't mean that it will work at the end. It's more experimenting, it's more working in the lab was with chemicals, you know, you can't tell from the beginning if it will work or not, you just can find out by doing it. So that's why I would start with a small approach, fail fast, do a lot of proof of concepts, you know, find out what it means for you, how you can work with it, instead of really starting this big, big program with a lot of software in it and maybe like 50 consultants right away. Don't do it. Really go the small scale approach and then if you really know how to handle it, and if it works for you, then go big.

Attila Tóth: It's a good point, and thanks for bringing this up, because I heard stories and I had people reaching out to me saying that "we've tried AI, it's not working for us, it's probably not likely to work us at all." And that's not true because, as you said, "it's an experiment first and people should be ready to fail", and actually from that failure you can learn, okay, what were the things that were missing or why the data was not the right data?

Fabian Merkel: Exactly. And that's actually a pretty dangerous situation, if you tried AI is not working for you, and now you say "okay", we do not pick on this topic anymore, and someone else will solve it and someone else will maybe disrupt the business, and that's a problem because a lot of people have too high expectations towards AI. Maybe there's an example with this: autonomous driving. You know, everybody, at least in Germany, people are frightened that these kinds of self-driving vehicles might kill anyone you know, because there's a child jumping on the street and there's another old lady so, you know, the cars decide which person to kill. And people are totally afraid about it. On the other hand, if you look on the street how many accidents they are, how many people are drunk on a car, or maybe taking the phones out and killing people, that's not a problem at all. So people have higher expectations towards machine learning, towards AI then they have towards other humans. And this kind of high expectation makes it also really difficult in these early days where we are now for AI to really create value. You should really lower your expectations. So if you also do product management in this area, expectation management is key. AI is not a hundred percent, it will make mistakes, it won't be really nice at the beginning. So keep that in mind. And it also means, if your experiments turn out to not, you know, go that well, keep doing it maybe do not waste too much money and too much resources, but you know, keep doing it.

Attila Tóth: Absolutely. And as most of our show listeners are working at big brands, I think you touched on a key point here: is that with AI, you cannot be satisfied just having one or two pilot activities. You have to do more until you find out what is working for you. And the bigger the company is, the more pilots it will need because it's complicated to find from ground zero, right away, the perfect solution, even if you have a lot of data. Sometimes what I see happening is that, people say, we have huge data sets, we are meeting every requirement to work with machine learning, let's do something! And then they fail. And they fail badly at their first attempt and then they put it in a box and they say, okay, we'll wait another 5 years and when somebody else does it right we'll restart the process again. And in order to not to get stuck with AI, I think it's critical, as you said, "keep doing it, keep going onwards." It's not a simple input-output model. It's more complex, and you need to find your way and there aren’t perfect solutions and there can be multiple scenarios.

Fabian Merkel: Exactly. I think you also mentioned another really difficult problem. It's the problem that people look at the data and then they think "oh, okay, what can we do with it", and then they build a really nice, good looking use case for the management, and obviously, in the end, it's not going to work, you know. They're not able to satisfy those experts' expectations or it's not working at all. So always make sure you have a clear business value behind it. Okay. And maybe in this "experimenting and get-to-know" phase, it's good just to look at the data and play with it around, but then take the other approach, really look for problems and then think "okay", might that problem be a problem which can be solved by AI? And if yes, okay, how we can do it and what is the real business value behind so really qualified, okay? Process duration will decrease or we can save money or whatever. So that is the strategy you should go. And maybe there's a little story which I can tell: We built a little machine learning prototype to recognize documents. It was for a pharmaceutical company, and they have a lot of documents about drug development and maybe they sell the license for drug to another company, okay, they don't want to produce it anymore. You know, any kind of reason to sell it. So then they take this 10,000 pages or 100,000 of pages of documentation about this drug development and whatever, just give it to another company on the day when you're selling it. The problem is the health authorities, if they have any question about a drug, you must be able to answer that really, really fast otherwise, you know, they will block your drug from selling. So but imagine you get, like, hundred thousand of pages and then the question and maybe these pages are in different languages and, you know, they are sometimes from the 70s, so, there's actually no human able to work with it. So we develop a model who was able to not read those documents, but to classify it, okay, to say "okay", that is a kind document kind of that type, and it has that kind of keywords in it and it's written in that language and it might be related to that. And the model wasn't working hundred percent, okay, it was maybe correct in 70% of the cases. And many people would say, okay, that's a failure. But look at the other way, I mean, 70%, were right so you were still able to create any kind of value, which wouldn't be possible before because no human was able to do that, and 70% is still better than 0% of the human, and it's still fast and still cheaper. So, even though the 30% were wrong, you still had a major benefit, so that's the kind of calculation you should do when doing that kind of product.

Attila Tóth: It's a really good example and to put this 70% into a bigger perspective for our audience, imagine that seven times out of 10 you have good results and only three times you didn't find in those documents what you were looking for. So 70% in AI, in my opinion, it’s quite a good result, something that's impressive and as you said, people need to lower their expectations. 100% perfect results will not be achievable on the first go, not because you don't want to achieve that, but because maybe there are so many variables, for example on these types of documents that you just can not classify each and every one of them. So that's part of the deal.

Fabian Merkel: Yes, absolutely.

Attila Tóth: Remaining on this topic and also thinking a bit on a bigger scale. As you said, you had experience in Germany, how do you see the openness and responsiveness of the market for accepting and specifically implementing digital trends like machine learning, deep learning? What do you consider as being currently the biggest obstacle of the companies in pursuing strategies that might involve this unconventional thinking, to be disruptive to try out new things?

Fabian Merkel: Well, I can only speak for the German or German-speaking market. Well, to be honest, it's not going pretty well for us. The problem is that a lot of companies here in Germany do not feel the pressure. They're still making too much money, they're still too happy, they underestimate the change who will come over us, will transform our markets will transform our companies, so what they do is they, I would call it, they play this innovation theater, you know, they do design thinking workshop from time to time and you know, maybe they have a nice innovation, that room has a lot of posters in it, but they're not really serious about it, because, you know, going out of the comfort zone to form yourself, it's quite difficult. People have to learn new skills, their job description will change, some people will lose power, some people will gain power, nobody wants that. That's our biggest issue here in Germany and that really blocks us and that really avoids unconventional thinking all that kind of stuff because you only do it if you really want to do it and if you have a need for it, and currently, we, the people, do not see the need to do that.

Attila Tóth: Yes, and on the one hand, I think this is an opportunity for those companies who want to think ahead and build ahead, because it's still being underrated in what impact can have on a business level. What you're seeing for Germany, I can confirm for Europe in general, UK, Austria, Hungary, Eastern Europe, what I experienced, that these countries have the same mindset. The only difference where I see more openness are countries in Northern Europe, like Sweden, Norway, and of course the US, they are really aware that this is happening and it's not only a hype. So there's a huge opportunity here people, and you can of course, stay back and wait but those who work and who are currently working on it, for sure will get ahead and the advantage that they are gathering now it will be a gamechanger for them, and once they are ahead, it won’t be easy to keep up with them, if your current mindset and strategy is to wait 3 to 5 years.

Fabian Merkel: Exactly. There was one guest on my podcast and I want to quote him and his name is Danny Nicholas, he's working, he's implementing AI solutions for banks mainly in Switzerland, he was saying okay, "I can't guarantee you that when you are studying using AI now that you will be in business in 10 years, but what I can guarantee you if you don't, you will be out in 10 years."

Attila Tóth: Very well said. I really like how he put it, because it sums it up so nicely. And as we are getting close to the end of our discussion, I have one final question to you: what would be the one thing you would recommend for companies, for people who are thinking about AI, to do in 2020?

Fabian Merkel: Learn about AI, inform yourself. It's not only that companies should do it, I think everyone should do it. Even if you totally do not work in business, you might be a kindergartner or something, do it, because it's not only a business topic, of course, it creates a lot of value in the business sector, but it also has implications on our society, democracy, and how, yeah, we deal maybe in healthcare, all that kind of stuff. So if you're not informed if, you know, if you do not know what it is and what it does, and what are the opportunities and also what are the dangers, you're not able to make any kind of informed decision and then you just become, you know, someone who are, ya know, ahead of the development somehow only react. So that is my only recommendation and learn about it. And there's a lot of stuff online in any kind of language in any kind of level, there are YouTube videos explaining basic topics in 10 minutes, you can also do online courses for six weeks or eight months or something. So pick something easy, and just go with it.

Attila Tóth: Thank you very much. I think this a great takeaway and a good ending note to be open-minded, learn about it, and be prepared as it's not only coming, it's already here.

Thanks Fabian for joining the Cogniverse Show. I'm glad we had this discussion and looking forward to new discussions with you.

Fabian Merkel: Thank you so much. It was a pleasure.