AI that delivers business value
hi everybody and welcome to this seminar today it's organized by RISE Growth analysis and Digital today it's called ai that deliver business value and the agenda is first i'll talk a little bit about the drivers of ai adoption it's a literature review and then we have the circle from rice applied ai center and he'll talk about swedish ai use cases and then we have aleister nolan our international guest that will talk about ai diffusion and a new project that OECD is having and then we're having the Q&A session and that will also be assisted by andres yet another colleague from OECD and then we go to next steps and for housekeeping you can ask your questions using the Q&A functions and we are recording this seminar and we will comply with the gdpr so if you have any questions please just send me an email and the recording and slides will be disseminated to registered participants after the seminar and yeah here we go to the first slice it's a research on the business value of ai and what we can see in the literature is that a lot of firms are seeing that they can add business value with ai however it's not a lot of firms that have managed to deliver business value so far 40 percent of firms say that they have done significant investments in ai but they have not reported any business gains yet so there is a lot of things that we need to know that we do not know at this moment in order to get ai to contribute to farm performance and then we have the current adoption level that's important we need a a level of ai use that we are currently having and you can see the figures from sweden its official statistic you see that four percent of small firms use ai whereas 30 percent of large firms use ai and that is something to bear in mind that the aius when we accelerated these are the levels that we are accelerating it from so that is also important going forward and here we have the drivers of ai adoption and this is what firm need to combine in order to adapt ai successfully so it's about data that i use to train the air models it's ai technology it's algorithms and ai system but it's also iot and cloud and you also need to be able to compute all the data and there's ai skills so it's both technical skills and management skills and we also have intra-firm coordination with in-firm collaboration between business units and also you need ai business model to generate ai business you need those important business model and there's also the ai innovation innovation ecosystem so collaboration between the government finance years and then ai research and universities and firms and there's also connected coordination across organizational boundaries so firms um always not always but many times they buy ais an external from an external provider so this need to be collaboration between the purchaser and the external ai provider and that is new for ai we haven't really seen that to that extent when it comes to digital transformation and within this literature review i also added a checklist it's a an evidence-based checklist so it's based on practice based research and these are the bullet points to pick a problem that is small but yet strategic to make sure that ai is the right tool for the job so don't use ai if you don't have to but use it when it's right and set air specific objectives and metrics for success so you know when you're delivering business value and list the data requirements as we know data is crucial and identify the resources and competences needed and then also pinpoint the potential challenges uh which could be regulatory compliance for example with the gdpr and the artificial intelligence act and then we have the report this is what it looks like it will be available online from october 20th it's of course free so i will also send you a note on that when it's available and it's all in english with a swedish summary and then we go to the next speaker we have the serger johnson he's director of rice applied center for ai go ahead sparky thank you then so i will talk about their cases that deliver value and from my rights perspective next please so i'll talk about this perhaps from a slightly technological uh point of view and along two themes of of the solutions so one strand is industrial data analysis which is a very powerful toolbox developed over a long time and now increasingly applicable and used in the digitalized era and secondly the more state-of-the-art leveraging the power of state-of-the-art ai models next please so a few words about rice so we are the swedish national research institute 2 800 employees in 30 40 sites from north to south and covering many industrial domains most industrial domains and also some public sectors next please and this also is reflected in how we work with ai so this image each box here represents people projects partner contacts activities in different areas within rice in over 120 projects with major ai content so we have several research groups working on the driving technologies we have research groups again working on what we here called the synergistic areas which are actually absolutely essential for successful uh application of ai innovation management ai aspects of security and security for ai ethics of ai and so on and not least policy recently with the eu ai act proposal and then lots of excitement really really i'm i'm so happy to work as sort of in the center uh for ai within rice because there are so many exciting application cases from ranging from materials next generation energy systems every aspect of the vehicles and transportation systems etc etc drug development and you name it climate change solutions and so on next please so the first of the themes is the industrial data analysis and just to give you some feeling for this this is based on a large toolbox of sort of statistical machine learning tools that provide situational awareness and monitoring based on data you can detect anomalies in data sets and data streams you can diagnose the faults that you may be discovered using anomaly detection you can predict the next steps or assist the development of systems and more generally find structure causal relationships explanations and so on for for various phenomena next please and just to give two examples uh of this uh so statistical anomaly detection is very widely used we have done dozens of projects this is just one example from condition monitoring of cranes uh with abb and uh as you can see in the figure to the right these are various sensor readings in these systems and some of these readings are very obviously or not necessarily but in this case obviously anomalous but some are more subtle and the point is that the anomaly is derived from the combination of values which is unusual and strange and needs to be sort of modeled in very carefully to find the analysis and this is very important to always monitor i think this is a technology that should be applied everywhere where you have data you should apply next please and then for doing diagnosis so diagnosing is a sort of open-ended thing there can be almost anything can go wrong and you need to start somewhere so we have developed a rather clever system uh called incremental diagnosis where you can incrementally add what's called fault mode prototypes where you don't need so much knowledge you just enter what you think are a few related things a few sensor readings and some type of fault and the system will then start to incrementally learn based on the data how these are related and help you do efficient fault detection and the figure here is based on diagnosis of vehicles but again this is something we've worked with many companies on so these were two examples and they sort of are almost no not ai but they are ai i mean there is no sharp border between ai and uh sort of statistics it's a it's a continuum and increasingly what we call ai methods are used also for this data analysis tasks next please so just to take one step back and sort of talk about models so the machine learning part of ai is about moles of data obviously so this is just an illustration we have these data points of areas and prices of houses and we have the very very very simple model which is the straight line which uses only two parameters and we can do prediction and we can find anomalies and so on based on this next please and then we come to uh sort of what we start thinking about as ai and which appeared uh maybe 10 years ago and now we take for granted these capabilities it's really funny how rapidly it has happened so that we can see what's in images uh translate sound to text translate between languages and sort of do captioning of images and these are still just not the same but similar enough mathematical models it's just that they are much larger they have tens and hundreds maybe they vary uh a lot the million parameters that sort of embed the knowledge in each of these items and so these are as you know in production in your mobile phones and everywhere next please and what i want to convey here is now a sort of state of the art and how ridiculously expressive and sort of powerful machine learning models it can be these days so this is a case from open ai dolly and it's a model where the data points are images and descriptions of images from i guess all over the net and the train model can be used to predict an image from a text a new text and create entirely new images and the point here is that the model has sort of built such deep domain knowledge about objects and their visual properties and so on so as to be able to synthesize of images where something looks like an armchair and like an avocado at the same time so of course you can't learn anything very specific from this but just get sort of an intuition that oh wow we have huge representational capabilities in these models so this brings us to the deep domain of moles and i'll give two examples next please so first language models mods of english and swedish etc and these have appeared in the last couple of years and they are incredibly powerful even larger than the model we saw and revolutionize how we handle documents streams of documents streams of text of many kinds we work a lot with swedish authorities on this where they already produce value they are easy to just put into production and and also with some companies on every report just automating document streams next please and uh my last example is my favorite in in many ways it's sort of how we can now start modeling or we already do model chemistry we can model the properties of molecules of chemical reactions and then use these uh to automate in in silico uh various tasks that were until now very slow and cumbersome and costly to perform so we can predict the toxicity of compounds we can predict the outcomes of chemical reactions we can plan uh the most efficient synthesis of molecules and then apply these in for example drug development and the development of new battery chemistries and this uh for for drag development companies such as sarsenica are investing heavily and already are seeing payoff from this and this is definitely the future of all chemistry so with that i have the next just a summary slide so i gave a number of examples from two sort of streams of the more basic in some way industrial data analysis which very well established methods which are already a staple of digitalized industry i think maybe even more companies than the ones you list use products which use these methods and sort of implicitly use ai without even thinking about and then the deep domain models which really have the potential to revolutionize and uh several industries and i think we'll see just more and more examples of this in the future so that ends my presentation thank you great thank you sverkin and then we'll move on to the next speaker we have alistair noland our international guest go ahead alistair thank you very much iren and it's a pleasure to be here so i'm going to be presenting some new ongoing work at the oecd and this is being led by myself and by my colleague andres balanechi who is also joining us for the q a next please so i'm going to be talking about uh ai and production and why this matters and how this is the foundation for and the ratio it has to do with the rationale for our new survey work next please so why does ai in production matter well a key reason is because of the problem of sluggish labor productivity really across the world so i'm referring to the value of output produced per hour worked um and as i say this has been uh stagnant for a number of decades it has been particularly hard hit since the financial crisis of 2008 next please and this productivity challenge also click again please this productivity challenge is all the more important given that the old age dependency ratio across oecd countries will double almost double the next 35 years so those who are in work the those who are um those who are of an age where they are working will have to become ever more productive more productive than they are now in order for society to pay for the age-related burden of a much larger non-working cohort next please so why is productivity growth stagnant next please when we have the advent of these productivity enhancing um technologies such as those that sverka has just very interestingly presented well one reason has to do with the slow pace of technology diffusion in this case a.i next please this slow page was already alluded to by uh irene in her uh opening slides now how do we what do we know about the um the pace of uh the penetration of ai in the uh across production well there's a there's a number of different types of evidence coming from survey work done by national statistics offices and other national agencies and these are survey modules on ai which have been embedded in already established large-scale surveys of ict use or of expenditure on r d and so on a second source of evidence comes from survey work done by some of the blue chip management consultancies bcg mckinsey pwc and so on and then this work done by academia and by a number of ngos um all of these types of evidence have distinct characteristics and have different strengths and weaknesses so the work done by the nsos the national statistics agencies these have a strength of being conducted with very large sample sizes 300 000 firms in the case of the of the us but some of the modules themselves are rather shallow just a few questions on ai and they're not necessarily very revealing about some of the policy uh constraints on adoption and possible enablers the work done by the management consultant is often very small and unrepresentative sample sizes some of the methodologies not always transparent and dealing with constituencies with respondents who are already intensive in their use of ict so next please so we um and this as iran had already said that the uh the large-scale surveys of ai in firms conducted by the nsos these are tend to show very low rates of uptake really across the board to some variation in the types of in the survey responses in the numbers um across eurostat encompassing areas these uh these numbers are fairly constant but we see big differences for example with respect to japan and some in the united states but the overall story is that adoption is much lower than one would um the more would expect if one's reading the sort of ubiquitous news about the take-up of ai and particularly low in smaller firms next please but this is a challenge not just for a.i i think there are particular obstacles particular difficulties in the adoption of ai which differ as iren said from other types of digital technology but even in a mature digital technology such as cloud computing here we see big differences in rates of take up in the corporate sector from say from finland on the far left something like 70 percent or so of the entire enterprise population are using cloud right down compared to a country like germany for instance where it's only something like 15 percent of firms are using cloud so next please so diffusion is a challenge in uh in ai but in not just in ai so in response to the some of the gaps in the availability of evidence we're beginning new work in collaboration with the boston consulting group with insead business school new survey work and the aim is to generate a variety of new cross-country data and policy analyses on the diffusion use of ai and to do so in ways which are helpful to policymakers which are novel which are new which complement existing survey evidence and which if possible lead to um these are conclusions findings which are actionable on the part of the policy community next please actionable in the sense for example that all oecd countries operate a wide range of institutions which have technology diffusion the acceleration of the spread of these technologies as part of their remit from well-known applied technology centers such as fraunhofer to innovation vouchers used for example in um in the uk and many other countries to dedicated field services like technical extension programs and so we want to incorporate some questions in our survey work which try to better understand what kinds of services ai users or potential ai users would best require what do they most want what do they most need on the part of institutions of this sort next please so for example in our um maybe just one click back if you don't mind thank you um in our survey we are asking questions which really are not been incorporated um in survey work so far we don't think it's necessary just to and we don't have the resources to replicate the work done by the national statistics offices but that's not so important to do because we already have those gross or aggregate numbers we know something about the broad pattern of ai diffusion across the corporate sector so we want to try and hone in as i mentioned on some novel issues for example asking questions about whether firms that use ai have engaged in public tenders around ai do firms use any public services for example for advice information finance capacity building in the ai adoption process is the potential of ai in firms well understood including in novel more recent uses such as ai being deployed for training um do firms consider that public agencies could assist ai adoption and other priorities amongst the types of services that public agencies could seek to reply we all know that ai skills are scarce everywhere but can we get a more nuanced picture of skill scarcity does quality qualifications frameworks for example play a role next please and do firms have high no or low awareness of their national ai strategies at all um and various other questions of this sort that we intend to put so next please so what's our sampling strategy where are we going to focus next please so far we plan to look primarily at manufacturing and to look initially with a concentration on smes in manufacturing why focus on manufacturing in part because um the whole issue of industry 4.0 of advanced technology in manufacturing is of strategic and policy priority in most oecd countries so over 30 countries now have industry 4.0 strategies next please as flicker was also uh was pointing out ai is now being used in almost every stage of manufacturing every stage of production from product design to fabrication for example in the manufacture of silicon chips where ai feature detectors look for some of the irregularities and they're all in the otherwise very ordered grid pattern on um on a on a semiconductor and you see those little red rectangles picking up where the floors exist in process control and training next please and giving like cognitive uh and other kind of uh support for workers so here's the uh factory for the a350 the um produced by airbus in toulouse and this ai system monitors every stage of process of the production process and also every shift um and when a worker now encounters a problem the ai control through uh examples where similar problems may have arisen and suggest possible solutions and this system actually has managed to reduce the time required to address disruptions by by one-third which is a large return on investment um and my last two slides please one more click again please and another reason for focusing on smes in manufacturing is that the national ai strategies maybe are very strong in looking at what's going on at verticals in health education cyber security often with the collaboration of of very large service providers in these areas and the national strategy is also at the lower end of the spectrum putting in money into venture funding for example but in the in the middle there may be a an under attention to the needs of smes particularly manufacturing next please and again thank you this is my last slide and we will complement the actual survey work we'll marry this survey work with a series of case studies particularly of ai adoption in larger firms and we intend to undertake interviews with the chief technology officers the chief information officers chief digital officers etc to try and get at a range of issues which are maybe too novel or too diffused too complex to really encapsulate in a survey how firms are dealing with a range of uh regulatory questions for example managing novel sources of ai related risk uh dealing with um a number of of data issues so uh and even how they are responding to a number of high-level principles which have been enunciated for example by the g20 and by the oecd um intergovernmental agreement uh orchestrated by our organization um to 18 months ago so with that i will leave the presentation and thank you all great thank you speaking and thank you alistair so we'll uh go ahead to the q a so if you have any questions on this presentation or if you have any reflections go ahead and put it in the q a and in the meanwhile i'll ask the participants on something on the link between ai and digital transformation you mentioned that alistair so i was wondering have anybody of you thought a little bit deeper on that what is the connection between ai and digital transformation well maybe two two points one as uh sverka very interestingly pointed to i mean this really is the sort of natural evolution of the sort of data data revolution it's where we will go next but the point i would make is that there are in the process of transformation there are certain difficulties with ai which maybe distinguish it from previous generations of digital technology one might be but there's a number one might be that it's really not a plug and play technology it requires complementary teams with uh maybe not just knowledge of data cleaning and the algorithms but also as you pointed to here and the understanding of the hardware so that the engineers so that in a manufacturing context where you may have many machines of different vintages with different operating systems so that these systems can all talk together and it's in a seamless way and the data in silos that you know the data relative relevant to um supply chains might be held in a place other than the data held on customers that these need to be these need to be linked brought together but perhaps most fundamentally it is maybe a question that the there's a degree of uncertainty in the in ai projects the it's more difficult to calculate the return on investment there's a degree of experimentation which distinguishes i think from other digital technologies there's no a priori guarantee of success and i think that also requires different approaches on the part of managers to deal with um that higher level of uncertainty in in their investment decisions great thank you and now we have uh andres who has comments on that yeah just to add to this uh crucial point that uh alistar was raising that uh not only firms need a mix of skills and uh capable it systems that uh enable the development of ai solutions but also like these these sorts of uh investments are very ad hoc they're tailored to a specific situation in the firm and then those solutions cannot be easily exported to other firms as well so there's no out of the box solution and this is yeah one one of the main reasons why the diffusion of ai is not is not progressing as one would expect great thank you we now have uh three uh four questions so i'll start with the first one it's from uh lagoon she's wondering when focusing on manufacturing in the service would you consider the service link in the manufacturing supply chain arguably three of the functions in the chart on manufacturing are services often provided by external supplies i think that question is for for our ocd guests would anyone you'd like to ask yeah i mean there's a number of points that come to mind um this um you know one area we're looking at is r d and sort of r d services and the way that uh the well as smith already alluded to with respect to drug research but in other areas of research a corporate r d where for example is using new materials in the aerospace sector are extremely important maybe services around supply chain management um so uh bmw i think has the ambition at the moment of knowing in real time the status of every machine involved in producing important components in this entire supply chain network so using ai to to to help manage that that process is also important um great and i see we have a follow-up questions on the aik studies that will perform will be performed by the oecd it's from patrick he says great presentation and it's a question to alistair are you also planning to capture the degree to which firms are adopting responsible practices of ai yeah well you know in the i think there's something like 60 overarching principles around a.i which have been developed internationally and by various national bodies and um ai associations
and so forth um i think there's a question and their time and again the whole issue of responsible and ethical and human centered ai these are the sort of buzz words which um which are ever present but in an industrial context it's very hard to know what a human-centered ai system an industrial production process actually means or responsible ai the primary concern is accurate a i am reliable ai effective ai um and so i think what what we'll be interested in knowing maybe is where issues of um quote responsible ai actually how they link to success in an industrial context if at all are um are ceos and ctos actually concerned with this concept do they see that there would be um business value in somehow what do they understand by the concept of responsible or ethical ai what relevance is it to their business and so i think we'd want to try and get that in the in the case studies get to that in case that is great thank you and we also have a question now on uh from linear holland on whether we have a strategy for industry 4.0 in sweden and it's included in our smart industry strategy and i see that andres is also typing and answer that that will come shortly we have yet another question from sophia do we know that the production is increasingly done in we do know that production is increasingly done in global value chains is there a role for international trade policy in supporting the diffusion of ai and and this is that may be something that you would like to take yes that's that's right yeah so um for for the um yeah for for trade policy and and diffusion of ai i think it's the the issue is is um the roles that that the big players have in in uh so we see that the the aai um solutions or or the um firms that are uh most capable in ai are are the the big international players and and um yeah countries can can attract such uh players um so we see in in ireland the presence of facebook and other other big tech firms there with with so there are these types of uh measures that allow uh attracting uh um large firms that would eventually connect with the with the country's internal um uh value chains and and this is a way of indeed of uh helping a division um but but um this is this is only one i mean one part of the of the picture here uh and um uh yeah for it it requires as i said was saying um more efforts on on the side of the um diffusion agencies that countries have to to uh really give examples for um um uh small firms uh and smes to um see what what what is the potential um for for ai investments that's that's one of the also like one of the problems that small firms have that they don't know they don't have a clear visibility of what the possibilities are uh so having these types of uh proof of concepts that are available uh by by the public actors made my uh showcasing them uh having demonstrators are uh other ways that that um could help diffusion um yeah great thank you maybe just one point there yeah i think it's maybe not so widely appreciated just how much data is generated in industrial processes industry actually generates more data than any other part of the economy and so a big issue around relevant to supply chains and relevant to the efficiency of the of a are using firms is that uh you know global trading data the transmission of data across borders is far the rate of growth of this is far exceeded the rate of growth of trade and that if countries are going to adopt practices which restrict the transmission of data across borders then the evidence is beginning to show that this will raise the cost of doing business for firms and may even actually slow down the rate of fdi foreign direct investment into those countries great thank you for that and we have the last question it's from thomas in the north of sweden and he was wondering i'm interested in the views of sme their views as me has on on data do they accept data-driven decisions do they see value in data and they're looking into the status of data and sme and i was wondering this this might be a question for you that are working actively with the different types of organization is this something that you've seen maybe thank you i think someone with a more uh sort of economy outlook should answer as well but i can just give you some anecdotal evidence that the small companies we come into contact with typically are very proficient as very advanced in leveraging data for creating value and within the context of their products but i don't think they are anymore if if data-driven decisions refers to sort of their business practices and how they understand their market and so of course some some market data i think they can analyze but uh on the whole i would say they are uh do business in the same uh sort of a human oriented way as as most companies still do you know but maybe someone of the other panel members has data on on this yes yes i mean i think an issue is in addition to size it depends what sector they're working in as well so firms for example which in some sectors like insurance like in finance like in online retail etc these firms may have quite a long history of working with with data there's more of a culture of data there whereas for firms in the manufacturing sector which would be more product-led than data oriented i think that these firms maybe will face more of a challenge is more of a cultural transition to engage in especially in the country like germany for example where many of the firms are like family held for a long period of time this for the same reason that as i pointed out in the graphic that there's a great deal of hesitancy in using cloud computing in germany there's a whole issue about data security and um letting data leave the company boundaries is something about which there's a great deal of hesitancy um so yeah i think it's more to do with the sector uh than with them with size although as i say and as you've pointed out um the smaller firms are adopting a whole slew of digital technologies at a much lower rate than our larger firms right andres would you like to add something or do you think that that no just just to say that uh yeah indeed these smaller firms they they struggle more in making the necessary investments to uh have like a data management uh that is required to to work with ai so that's just adding to that yes and uh now i see that we have come to the end of the seminar so i would like to thank our excellent speakers very alistair and also andres and i would like to thank also the participants for excellent questions and we'll keep on producing reports in this area and i will make sure to let you know when we have the literature available and there's also a forthcoming study on aiu's cases so unfortunately that will be in swedish so but we'll make sure to have a thorough english summary of that so i would like to thank you all for making this a really great seminar thank you thank you
2021-10-26 07:05
