I mean asked me to present a little bit about you know what can we do about AI here in Europe what are some of the pitfalls what are some of the experiences perhaps that we can learn how can a start-up help big corporations especially in the European sector and that's what I wanted to present a little bit out of sort of the past experience I worked before there was like mentioned now part of BCG so I see a lot of different customers and different clients different situations but these are kind of real world examples not as much on the research side so you know it's not going to be hyper technical on that but I think it will help to see how we can drive AI really at the business scale and therefore push it and pull it into Europe so the thought is really what we want to try and do in the business is we want to start with thinking big not just you know have a small technical solution for something that
seems to be interesting because you know I can now unlock the hat as a chatbot but more of well here is a big vision you know we think with AI we can unlock not you know a couple of thousands not a couple of hundred thousands but perhaps a hundred million dollars for a company and that might be both disruptive in a disruptive way and evolutionary way that has to be seen and some of that is very difficult to apply into a company but the interesting thing is after we laid out the big stage we want to identify where do we start really you know we want to find the key lighthouse project or two or three and build out of that and out of this create really a drive and a pull in the company to pull in this kind of new thinking and a lot of that is unfortunately you know trying to change the mindset and some of that I wanted to present today so you know we've seen these kind of slides
already quite often I'm sure but what is interesting in the mindset of a lot of top executives is to try to say it's the data now that starts to drive the change we have now more data not that we did not create that data before but we have the ability to tap into that data and rather than having a very standard approach of you know engineering thinking we build something we see what what would it produce we now start looking at the data and see what it tells us if it tells us the same story as we thought going for before but now we have also technology and on technologies the storage layer it's the NVIDIA CPU said you know which is absolutely amazing but also some of the new algorithms that can unlock suddenly problems that we couldn't unlock before and the interesting thing is and that's the gold standard of course we want to get to is that where once we have generated these insights we want to have them become actionable so
it's nice to have a dashboard culture not that it gives that us that much but we want to have kind of actionable insights that where the system can do something like a prediction and say I think that and that is going to happen what are you going to do about that and see what the result is ideally that will generate even more data and it will have the system learn and that's you know what we're all after having a system not that she does historical back testing but actually can work in the real time and adjust so in the III part you know a lot of AI talk and I'm not going to change once more the definition but a lot of the work that have been doing is in round machine learning advanced statistics advanced analytics but it's been working mainly on the structured set of data that we have available so this is traditionally in applications and databases locked into different silos into the enterprise and there's huge value that we can drive
out of that now interesting enough though most enterprises if not all of them 80% of the data that is available is unser so this might be sensor data that is we don't know what it actually means till we actually look at it from different perspective or a lot of documents and this part is untapped well most of it at least and that's where the knowledge lies in and for me the interesting part of our AI is not looking at just the data and seeing patterns out of the data but actually compressing the data and turning it into information and now learning about that so I'm going to show two little use cases that I think are a good list relief but one of the things that I often show to executives is this kind of slide which I think really nicely explains where the difference lies between you know traditional software and rule base or engineering based systems and machine learning so you have on the left side you have sort of the kind of
expert build system somebody who has really a lot of knowledge and we thought about how we can do something so you have the data that comes in you have the rule sets and you have then an output out of that we'll see how it reacts in real world often different than we expected there's a lot of chaos going on and then we need to adjust it so the difference to machine learning the thinking of it is different in that we think about the input data in the same way where we now take the output data from the past and out of that we derive the rules seems very very smart has of course a couple of problems but in essence we learn first from the past you know like we all think we the way we learn and out of that we want to sort of create yeah this is this is kind of the adjustment of the rule and the interesting thing is that as it goes along the circle can continue and we
can just adjust the model so new data we get new rules or refined rules now how does it look in reality so we know if we look at some of the applied some of the applied AI a lot of the business in Europe especially is around manufacturing that's where working at a lot of the manufacturing is also a high level of outputs but one thing is how can we attack this how can we change some of those aspects and of course quality you know made in Germany is a big thing and that's why it's some of the core use cases that we see in a lot of the industries here now here's a use case and that we did in the past and we have a customer who has a new product wants to ramp up that product at the beginning before you go really into mass production you have a lot of quality issues now there's a new product so you can really learn from the past can you well depends you might have
had past project past products that somehow resemble and this is something where we can use machine learning for right machine learning is good to compare apples and oranges like I usually give my example on that where if you can you compare apple and oranges where the saying says no but machine learning says yes if you're looking for a fruit with high level of vitamin C both work if you want to have something that's kind of round yeah both work if you wanna have something that you can throw on the head of somebody both work but if you look at a color coding they're not so given the context we can see supposedly different situations suddenly being the same so in this case what we did is we looked at past hundreds of thousands of past issues and we try to extract using natural language processing so handwritten sometimes machine written notes to sort of discern what is the same what is not and present that to the human operator as a point of decision saying
the issue that you're facing right now we think we've saw something similar in a different station different product two years ago do you think it's the same so you know we thought we're doing good we're putting the actual decision back to the humans now the units are actually really smart about that they they are expert and they can say yes that's that's the same issue I'm facing right now so the system great learns and says well if that's the same issue can we use the same kind of solution that we used to apply last time that were good versus the bad one why don't we do the same solution path and with that we can actually show the tremendous amount of effect so you can speed up production time you can speed up root cause analysis and you can have a stoppage of the typical hands up so you have a problem listen I can't decide I don't have all the information that somebody else who was in this station who might have that information
now we're unlocking that and being bringing that to the shop floor sounds all good now there was just one little problem is that we knew that the prediction rate is 87% meaning that in 87% we knew that the algorithm would probably spit out the correct answer the correct hypotheses after a couple of months the prediction rate went up and we thought it was great the system learns but went up to a hundred percent so that's where we thought there might be something wrong so contrary to what we're thinking that the humans actually train the machine the machine actually trained the humans it was so good that humans said this just right yeah next time this is right yeah this look good next time is a key I know this is good so yes yes yes so they answered more often yes to the system than the system actually thought of it himself so the trust issue sort of reverted to what we didn't think and which is a problem because all of you know about
biased data now even worse does it get when we create biased data we know that it's wrong and 13% of the cases but these are being ignored and the system just feeds on its own bad behavior so we actually had to put in guardrails to say the system said yeah I think that's right but I think you should say no which seems a little bit counterintuitive but I think it highlights one of the problems that we have where we want to create a self-learning system that wants to learn out of human output but we're actually controlling that output so I think that's something to be taken into account when we look about the data in real life I brought in another example which I really liked because I thought when I when I started this project it's really cool the AI in there turns out I was wrong so this was a problem where you have a lot of general ledger spend so it's the ins and outs you know of a company that buys
goods or buys elements or you know travel all kinds of expenses and they want to map it usually you have it mapped fairly well most of the time in reality there's a lot of uncertainty a lot of wrong bookings in there so we have on the one side a general ledger on the other side we have around 300 cost categories to which we want to map these kind of expenses sorry I said a good let's do some NLP and are you to try to find out that so you have sometimes invoices fortunate turns out that those invoices of multinationals are actually multi-language so first thing is try to translate that which comes out to be a little bit challenging in certain cases because it's specialized language and doesn't translate all that well could we work with this and then you have this big blob on the right side even though we already extracted what we think is the most relevant aspect of the invoice to categorize it now out of that these are the things
that you need to extract in this case it was actually quite simplistic that it was like keywords how do you find these and how you do that with over a hundred million or a million cases of these invoices and of course being beasty in six weeks right so there other companies that use thousands and tens of thousands of people actually in certain countries to go through these very quickly and try to classify them and we said can can we do that better can we do that like with a couple of orders of magnitude better with eight people interesting enough we could and this is the the approach remember the invoice I showed you so if we have that big blob of tics how do we condense it remember that AI in my opinion is about trying to extract the information out of data this is very loose data so meaning that there's a lot of words but what's the information in there so we go successively through it and what we try to find see
if this comes over we look at different elements of the text and we see oh wait this is a whole phrase phrases typically because it's a sentence it's explanation information does not relate necessary to the information I want get out we have superficial information such as names locations and that's not important one could think but if I have something like I have a flight from London to Munich it doesn't really matter that I have London Munich if I can replace them by cities so that I have motor flights suddenly fall into the same category so we're condensing further the information or we're throwing it out then we get super Fisher's information you know filler words etc and successively reduce the amount of data in there till we extract the number of keywords that are relevant to it so what we wanted to do is really to combine a little bit the this kind of methodology with you know a very high speed working of that and what we were able to do is is this
kind of method so first we have the 800,000 invoice data that is kind of like messy and we cannot go through all of that our goal is to reduce this to something that's more reasonable with a structured approach that I just showed you we condense this data we look at the similarities of that and we try to find you know what can we group together that we think belongs together and suddenly we're down to 40,000 things 40,000 groups we can now treat us now what we do is we see which ones of them have the highest spent and usually in a Pareto analysis you can see that in this case five thousand five thousand of those groups actually mapped to almost 90 percent of all the remaining spent so I actually only have to concentrate on those now five thousand what was interesting his became such a manageable number that we said that's like kind of two weeks for for people to go through so we could actually just do manual stuff but let's do
some machine learnings and let's have some smart AI try to figure this out actually feels harder than we thought because in the first instant you know those kind of invoices are quite esoteric let's say you're a a coffee brand company you buy a lot of milk well is milk is that you know for another company this might be something for the cafeteria just in different spends for that coffee company that actually part of the resources and is capital spend so we cannot treat it the same way we have to generate a context that is that is common between those and what we think is egg after having done 2030 we can build up a kind of context-sensitive interaction so we want to start quick but built up this database of context-aware knowledge that can use that so with with just a simple set of rules we can actually spend most of that now and this we validate with humans and then we train our machine learning algorithms as of now in the first case machine
learning actually just accounted for additional 15% but that can grow significantly the interesting thing is this we can now do just with out super duper AI but just a very data-driven approach we can have very quick results in very short amount of time and therefore collects in the time that it takes others to go through four or five customers we can do a hundred of these and create at the same time this machine learning database in the back end so the I wanted to stop the talk on just a little bit of the the kind of approaches that I think work really well if you're a start-up or if you're a tech company to go and gain traction in some of their businesses especially in Europe where we have a lot of engineering focus what is important a two aspects one they are not data driven thinking I mean we've heard that a lot of times but the whole organization is not necessarily set up like that if you are in a small function and
you succeed in helping that small function the question is how how does that get you reuse later on so what we need to build this well is this this thing is big and I think what I try to tell always to executives is I've been to the same thing you know a is a disruptive technology now what does that mean then can we do things better yeah but the interesting thing is most of the AI disruption means destruction for the incumbent and the customer you are the incumbent so they need to move they need to move very very quickly and not just do things better that they used to do we need to actually try to map the entire process change the value chain see if we can redo it in a completely different way or extend that value chain right is it enough to just sell a product or do we need to think about selling the service around that product if I'm a car producer or standard example but what is the service
that I'm providing is transportation right from A to B well I I never went from A to B I don't know about you but I went from a specific origin to a specific destination now my car doesn't know that my test my house does it asked me it looks in the calendar and sees the kind of location and knows where I want to go and perhaps if I want to go to my mother-in-law perhaps I don't want to go the fastest way but that's the most scenic way that's relaxing whereas when I want to go to work you know what's interesting in after 10 years I have the same car and I go at 8:30 to my car still doesn't know that I want to go to work you know yet I have a smartphone that costs a fraction that does learn about that so there's a disconnect I think it's not the technology that that is not there it's the thinking that's not there so we need to attack at the value chain and
say what does autonomy for example mean for a car you know does it just mean that you have a cool car on running on a battery mean that you're not gonna produce anymore that more than many cars and that suddenly your business model is going to change and I think that's the important part that we need to go into the vision on the top level and then attack with a use case that they can actually see okay I get it now and now you need to find the balance you know do you attack on the core business and see how they can revolutionize it what do we do is that separate part where they just say we built up a new company that's gonna eat our own dog food or that's gonna eat our own cash cow and I think this is this is where we need to find a balance right a lot of the startups in the AI that are successful disrupted the market and a lot of the successful players like Amazon
that have stepwise evolutionized throughout this journey are now eating one business after the other because they have this massive advantage of a platform strategy or an AI strategy that completely revolutionized the way that we do business so here is a great opportunity I think for us because what we have in comparison to the u.s. we have a lot of industry here with us us has an outsourced a lot of that and they try now to bring it back I mean think about it that testa has this main production company in California in the Silicon Valley the most expensive work force there is but it is crucial for the way that they want to produce their cars so they're trying to bring manufacturing back we have manufacturing we have data let's use it thank you