Rise of AI Conference 2023 · Berlin

Generative AI for Industries

Feiyu Xu · SAP

At Rise of AI 2023 in Berlin, Feiyu Xu, Senior VP and Global Head of AI at SAP, presents SAP's view of generative AI for enterprise and industry. The talk opens with remarks from Thomas, an SAP board member, who says SAP identified more than 80 generative AI use cases within a few weeks, warns that almost three quarters of AI models come from the US and 15 percent from China, and argues Europe must not miss this shift, pointing to SAP's role in the LEAM initiative for large European AI models. Xu then explains SAP's two-pillar strategy of embedding AI into business applications and enabling low-touch, no-touch innovation across four core business processes, describes advantages of generative over narrow AI such as reduced annotation needs, and outlines a three-step adoption path for companies: use existing models as-is, fine-tune on proprietary data, or invest in further research for data types like tabular, time-series, chemical and biological data. She closes on SAP's AI ethics governance, including a monthly cross-functional steering committee.

Key takeaways

  • SAP identified more than 80 generative AI use cases within a few weeks of the technology's emergence, according to board member Thomas.
  • Almost three quarters of AI models come from the US and 15 percent from China, which Thomas says makes Europe's LEAM initiative urgent.
  • Xu organizes SAP's AI strategy around four business processes: lead-to-cash, design-to-operate, recruit-to-retire, and source-to-pay.
  • Xu argues self-supervised learning's reduced annotation burden is what makes low-touch, no-touch enterprise AI applications realistic.
  • Xu proposes a three-step adoption path for companies: use models as-is, fine-tune on proprietary data, or invest in further research for underserved data types.

Detailed notes

SAP's early generative AI momentum and Europe's position

Thomas, an SAP board member responsible for product success, opens by describing rapid movement inside SAP: within a few weeks the company identified more than 80 generative AI use cases, including creating test scripts and code documentation, supporting employee HR queries, defining business process models in natural language, and customizing product information for customers. He warns that almost three quarters of AI models come from the US and 15 percent from China, and that Europe must not miss this paradigm shift, citing analyst estimates that the global AI software market grew 22 percent from 2021 to 2022, before accounting for ChatGPT's impact. He points to SAP's participation in LEAM, the large European AI models initiative, and notes SAP set AI ethics guidelines as early as 2018 and introduced a global AI ethics policy in 2022. As an example of AI's impact, he cites the Hamburg Port Authority, which handles millions of cargo containers a year with no room to expand and uses AI to run its infrastructure more efficiently.

SAP's strategy: build AI in, make innovation adaptable

Xu describes SAP's AI approach around two pillars. The first is building AI directly into business applications so customers' own processes become more efficient, since technology alone has no value without application. The second is adaptable innovation: because SAP serves more than 400,000 enterprises across more than 20 to 25 industry verticals, each with different workflows, applications must be flexible, low-touch and no-touch, which she argues was nearly impossible with narrow AI but becomes achievable with foundational, generative models. She organizes SAP's business processes into four end-to-end flows: lead-to-cash (from first customer contact through contract and payment), design-to-operate (product design, customer feedback, supply chain resilience), recruit-to-retire (writing job descriptions, finding candidates, supporting employee development), and source-to-pay (finding suppliers, managing procurement risk and payments), with AI needed throughout all four.

Why generative AI matters for enterprise

Xu frames generative AI's advantages against narrow AI: where narrow AI mainly classifies, for example recognizing a face to unlock a phone, generative AI can create content, raising new authentication and provenance questions about where content originates. Self-supervised learning sharply reduces the annotation effort needed, which she says is central to making low-touch, no-touch applications possible. She also highlights emerging capabilities: a single AI system that can perform multiple tasks, unlike earlier narrow models built for one task each, potentially replacing many micro-services with one industry foundational model. She cites ChatGPT reaching 100 million active users within two months as evidence of how fast adoption can move, expressing hope generative AI will similarly raise adoption of SAP's own AI applications. She expects customer interaction to shift toward chat-based natural language, for example requesting vacation in a sentence instead of filling in start and end dates, and predicts developers will increasingly become debuggers and evaluators of AI-generated code rather than writers of it, illustrated by using GitHub Copilot to build a search engine in a couple of days.

A three-step path for companies

Xu lays out three adoption stages for companies working with foundational models. The first, and lowest-hanging fruit, is using models as they are, which an audience member describes using for coding, including debugging assistance and building a web scraper in about an hour with GPT-4. The second is fine-tuning models on a company's own data, such as SAP's proprietary programming language, which requires working through legal conditions with vendors. The third is continued research into data types large language models still do not handle well, including business process, tabular and time-series data, and chemical and biological data, which she argues requires sustained investment rather than pausing AI development. She illustrates simple applications already possible: retrieval-augmented generation combining embeddings and a large language model to answer business process questions, and a document-processing pipeline, demonstrated by an audience member named Wolf and his team, that uses OCR to convert an invoice image into structured data and then extracts key business information with a language model.

AI ethics governance at SAP

Xu closes by describing SAP's approach to AI ethics in practice, not only on paper. SAP set its AI guiding principles in 2018 as an early European mover and holds a monthly steering committee with legal, human resources, engineering and marketing experts who review use cases and identify data-related challenges, inviting external AI ethics experts when needed. A dedicated trustworthy AI workstream accompanies use cases from ideation through development, deployment and continuous improvement.

Q&A highlights

Asked about the biggest barrier to adoption, Xu says identifying the truth and originality of AI-generated content is a central challenge, less concerning for text but more so for video, making content authentication important for building trust. Asked whether enterprises mostly buy "wrapped" AI products rather than building their own large language model capability, as Deloitte AI reports suggest, Xu agrees this reflects SAP's own approach: SAP mostly sells business AI applications built on top of large language models rather than general-purpose models themselves, and customers typically ask either how SAP can use large language models to improve their business processes, or how to apply AI successfully in their own specialized domains, such as the chemical industry.

Quotes

The question whether AI will revolutionize the way we live and work is no longer relevant. The only relevant question now is how we will shape this revolution together.

Thomas (SAP board member)

Europe must not miss this paradigm shift.

Thomas (SAP board member)

In the future we will not have programmers, we mainly have debuggers.

Feiyu Xu

Questions answered in this talk

What generative AI use cases has SAP identified?

According to Thomas, an SAP board member, at Rise of AI 2023, SAP identified more than 80 generative AI use cases within a few weeks, including test script and code documentation generation, HR query support, and customer support automation.

What is SAP's AI ethics governance process?

According to Feiyu Xu, SAP's Global Head of AI, at Rise of AI 2023, SAP set AI guiding principles in 2018 and a global AI ethics policy in 2022, with a monthly steering committee of legal, HR, engineering and marketing experts, plus external AI ethics advisors, reviewing use cases from ideation through deployment.

How will generative AI change how people interact with enterprise software?

According to Xu, customers will increasingly interact with business software through natural language chat instead of clicking through forms, for example requesting vacation in a single sentence rather than manually filling in start and end dates.

Mentioned

generative AI, foundational models, low-touch no-touch automation, lead-to-cash, design-to-operate, recruit-to-retire, source-to-pay, retrieval-augmented generation, prompt engineering, AI ethics governance

Keep exploring

More from Feiyu Xu

More from Rise of AI Conference 2023

More on Enterprise AI & Adoption

Transcript

Auto-generated transcript, not edited. · 4,443 words

Show the full transcript

foreign it would be impossible for me to stand here today in Germany talking about AI with not having one of our most important players in the ecosystem Our Nation Beloved uh sap and we're lucky to have fee join us here today who's the head of AI and we'll be talking about specifically the million dollar question on all of our minds on how do we use genitive AI for industry something I can't wait to hear your talk so thank you today we have listened to many talks about generative Ai and the foundational model I'm going to share with you about how generative AI can contribute to the success of Industry of course we are still in the very beginning before I start with my path I want to invite Thomas our board member of sap and he's responsible for the product success and the floor is Thomas let's show His Radio foreign thank you for inviting me to talk to you today as leading Minds in AI research as policy makers and those who recognized the

Imports of AI it's my pleasure to talk about Sap's AI perspective and I wish I could have joined you in person today the question whether AI will revolutionize the way we live and work is no longer relevant the only relevant question now is how we will shape this revolution together the title of today's conference is absolutely fitting the rise of AI it's exactly what we've been experienced at warp speed over the past few months generative AI has completely disrupted the AI conversation and there's a huge excitement here at sap in just a few weeks we have identified over 80 use cases like creating test scripts and code documentation as well as supporting employee HR queries and of course we are also developing customer use cases like defining business process models using natural language in our sap sign up your process manager or quickly addressing customer support requests or customizing the product information we share with our customers and we are also engaging with multiple partners to embed foundational models into business applications but what is

needed today is high quality training data for the European markets that our customers operate in and access to a powerful digital infrastructure this will take collaboration across many sectors including industry policy makers investors and innovators speaking on behalf of our customers they need applications that are developed on top of generic large language models adopted to their specific needs currently almost three quarters of all AI models come from the US and 15 from China Europe must not miss this paradigm shift analysts estimate that the global market for artificial intelligence in software grew by 22 from 2021 to 2022 not even considering the ChatGPT impact so we are paying attention to this Mega Trend at sap we are perfectly positioned to influence the future use of AI foundational models we help run the world's most business critical processes integrate across the whole value chain with the broadest portfolio of business applications carrying more than 25 Industries the Hamburg Port Authority is a great example of these it handles millions of cargo containers every year with no space

to expand death need of trucks to keep up with the growth number of containers we help them use their infrastructure efficiently and intelligently this was only possible through AI we are committed to our vision of supporting every customer to become a networked intelligent and sustainable Enterprise AI made in Europe is essential to that vision that's why we are participating in initiatives like leam the large European AI models to advance models based on our shed European leads but we want to run this race basically we were the first company in Europe to establish our own sequi guidelines in 2018. in 2022 we also introduced a global AI Essex policy which applies to all saps globally the policy is essential to build trust in AI by ensuring transparency oversight and address bias and discrimination because we believe in our responsibility to build and use AI basically while we all are on the verge of many unknowns here one thing is for sure sap is in the race thank you and enjoy the conference and now back to

you faith and you can see right the sap we are middle in the game of this Innovation and also sap is supporting Liam Hans be happy that you have a very strong partner who supported our European large language model initiative when we talk about AI for Enterprise we are not talking about only AI Technologies because we want to infuse Air Technology into applications because this application is in particular applications developed by sap will contribute to the efficiency of our customer spins processes therefore for saps AI the first pillar is build AI into our business applications make our customers business processes efficient optimized they can do optimized planning right technology is only for application for our customer success however if we look at sap we can be very proud of our ecosystem we have more than 400 000 Enterprises companies the success of sapcai will be the success of our customers however since we cover more than 20 Industries verticals and each company company and each vertical they have different workflow therefore our Technologies our applications

must be very flexible easily adaptable and low touch no touch is very very important with narrow Air Technologies it's almost impossible to make this low touch no touch happen therefore with foundational model generative AI this support for customer will be possible therefore I'm really happy with this General TV AI development therefore the second pillars adaptable Innovation how can we make our customers successful they are Innovative even the if they don't have a data sense team if even if they don't have a team of AI Developers we have talked a lot today I will not repeat so much and if you run a company even run a very big company in our saps term it's simple we can simplify into full length of business the first one called lead tooth cache what does little Cache mean right we don't use this term in our everyday life it means one company you start with the first contact with your customer and then cite the contract and then make them pay which you deliver and also serve your

customer customer experience in all these end-to-end business process AI is needed money through the markets find the trends and also support with natural language with chatbot with digital assistant and also designed to operate what does designed to operate such a simple term but for a manufacturing company is so important design design the product and also get the user feedback improve the product create new product and also control the supply chain make sure the supply chain be resilient optimize the supply chain and all the needs AI in order to make it successful we know people people are the most important sources for a company recruit retire is which sap we call it starting to find the talents job description we leverage generative AI to write job descriptions automatically and also find the best candidates and also how can we help our employees to develop and match the best positions and help them also retire smoothly and the source will pay very important right how can we find the best supplies how can make sure the procurement

successful and avoid the risks and the get the payment down and all this business processes needs Ai and the sap we have been working on this all these years when we talk about AI when we talk about the AI for intelligent Enterprise we should think about intelligent ecosystem because it is not only about Technologies it must start with the data for intelligent ecosystem or Enterprise the data is not only the data within one company where even within one company we have a lot of data we have emails we have business transactions we have data for supply chain we have a finance we have human resource all these data there are structured data their unstructured data their texture data their semi-structure data however just imagine right if we want to monitor the trends of our customers where can you find this data inside the company or outside company Tina and if you want to monitor the supply chain resilience where can you find the data also inside company and outside company right and also not only

perhaps one company perhaps the chain right the dependence you have to monitor the defendants of the company right therefore very important in order to make successful we have to identify which datas within the companies are relevant which data also Beyond offset company and also the dependence among the company and Link them how can we make them linked semantically right we need AI in this case again even in order to make data available we need AI right entity linking natural language processing and also all these Technologies and if we have linked data the first step of value and we can build machine learning models on the one hand implicit knowledge Hans also mentioned implicit knowledge and also explicit knowledge if we talk about sap right we have 50 years of business process experience business Logics this information we should not ignore not these are not data driven the combination of the explicit knowledge also regulations right and also policies also policies for customers they are not data driven they are defined by humans the combination of

the implicit and explicit knowledge help us to develop applications they are explainable they are transparent right for automation for human machine interaction today I think every customer expect the access to software should be natural language chat driven right we have to make it happen and also optimization planning forecasts as Thomas mentioned actually he gave half of my talk already and all applications have to contribute to the intelligent transformation of our customers right our kpi is the kpl of our customer's success and furthermore supply chain resilience right inspiration resilience is so important because of the challenges the challenges we have face have been faced in the past three years and now right and for us it's also very important as we mentioned also the quantum Computing this morning right we cannot do a supply chain demand forecasts for weeks actually we have to be able to react almost every day and real time right where companies have to react very fast this is the reason the AI is needed heavily and also sustainability sap we have

been providing AI Solutions in order to help companies to identify monitor their sustainability kpis and also do the optimization and of course as a global head of AI I'm so impressed by the success of jgbt right within two months 100 million active users this is my dream in the past three years I hope with generative AI foundational model the adoption can really generative Ai and the foundational model can help mine team sap have high adoption of the AI applications because of the low touch no touch less annotation needed this can really make happen and also my pressure will be lower and as I also talked to many investors and also the market right the attention of investment Market is really high now and also I think I am really looking forward also to the success of Industry AI and I want to also summarize what I think is it's really the most exciting advantages of foundational model and the generative AI also from the perspective of uh from industry right for example before and we

have the the narrow AI mainly do classification right given a phase if we use face recognition in the morning unlock our mobile phone three technology behind it right fees recognize the fees recognize which phase behind it whether it's a face of a real person right but now with generative AI they can create faces right therefore it's so important the authentication what and also know where the content come from is very important the second one from the supervisors to self-supervised learning this means the load of annotations and very lower right and for the low touch no touch application is very important and the emerging capabilities this part I think also will be important for Enterprise AI because now at sap we provide a lot of stand-up applications standard business processes but if the our ecosystem has developed also new business models with emerging capabilities we can perhaps or cope with the long tail of the business processes and the human the AI also make us humans surprise right before we have the narrow AI one year

technology only do machine translation another do speech recognition now when AI system can do different tasks that's like our humans this means also how we will involve our software development perhaps we will have one industry foundational model which can cover different tasks we don't need so many uh micro software Services anymore perhaps the Microsoft service derived from the industry Foundation model these are all things which have to think about this will change also the architecture of the Enterprise Solutions and if I want to share with you what will be the impact of foundational model in industry for our customers right the expectation of our customers they will change they expect they will use our softwares with very easy chat right text they don't want to click anymore like in order to apply a vacation they want to don't want to click and fill the starting date and the date and they want to see I want to have a vacation from when to when and for uh one sentence or two sentence that's it right

and also business transactions right because the many business transactions can easily be done also like travel reimbursement and a lot of boring work for my assistant this will be done by chats and easily right and also for the product development as I mentioned the rifle manufacturing companies creative industry perhaps the AI will make the first suggestions also afeps will make the suggestion based on the customer feedback and do the proactive inspiration for the designer design of for the for the product designer for the Innovation teams right and because of the less annotation effort also our customers because before they have to do a lot of implementation themselves right in the extension the customers of sap they are also business process experts right if you look at their codes after 10 15 years they have built their business process themselves right and with less annotation perhaps this kind of feedback of their input will be part of this solutions tool and for developers I give many Keynotes I often make jokes where I see in the

future we will not have programmers we mainly have debuggers right the generative AI will provide us the codes we just to have to do the evaluation I think two two or three days so I let codex and write a search machine even it's very simple keywords simple keyword search machine but they they are able to define the indexing part the retrieval part it's amazing right I think it's also for for University for Education very very important and what does mean for for the ecosystem for sap for sap ourselves right because most of the companies we are not companies who provide general purpose large language model right we don't have the whole stack and if we don't have the whole stack we leverage some large language models what will be our strategy of course the most companies start with the low hanging fruits start to use the model as is I think many companies use it already right and also as individuals we use jgbt anyway and perhaps can you share with me who are you

use the model as is build some applications which kind of pumps are they submitting several problems team that you that he was calling avoid now to say everything yeah but but yeah next one shall I repeat that did you hear me no okay good perfect yeah we used a lot for coding exactly so we use stockholder for example just came out like a few days ago and the co-pilot from GitHub and it's it's amazing so the debugging part I told the I told the Android that it's working like a charm so you don't have to code for three hours you have to check for half an hour maybe and they're like the the gdp4 can if you tell them that there's a problem with the web scraper for example in time where it can REO cue that and then do it by itself and then you have a web scraper and I know and I would say an hour and it's done and it's really good so we use it all the time yeah great

right I think if the first bubble use models as it is no brainer right I think every company will go for it and we have to leverage this opportunity right the second one is tune the models um your own data because for example as ssap we have our own programming language and I don't know whether you know it about right I I I don't think in the in the Codex there are so many other codes because we don't disclose to uh outside the world and in all but uh um but we if we want to divide new above codes or the rewrite and we have to infuse about code into the language model and this means also under legal conditions right and we have been also working with different vendors in order to make this possible and the third one third ones even Elon Musk they wrote a letter C we have to stop General Tai for six months however if I look at the third bubble we have so many so much work to

do right given the larger language model how can we deal with other important industries data right like business process data tabular data time series data also chemical industry chemical formulations and also biological data these are datas the language model still did not understand there's still no connection because also the application will be different right for the time series data transaction data we need a prediction we did recommendation perhaps also we need a composition of the business processes right these are still missing and also for the chemical formulation how many chemical companies they ask for help how can large language model plus their data in order to contribute to their Innovation therefore I think we we cannot stop research and development we have to continue even invest more in order to make it happen make our industry successful and I will not go deep because this morning these slides in different style has been presented several times and I want to only share with you some very simple example right and which is you can start

very easily we can leverage embedding in order to combine two steps right Retriever with and also leverage large language model in order to provide very natural response like how can I adapt my business process for what calculator and we can embed the business knowledge business processes and and also leverage the questions in order to find the right results and the results can give the large language model can help us formulate a good answer this is a straightforward application the second one is also oof is in the audience wolf has oof and his team has been working on dealing with business document with generative AI we are able to easily adapt to new languages and new business document types for example given a invoice image with OCR we are able to convert into 2D structure in GSM format and then we can put this as pumped under the the large language model can extract major business information the structured information right it's a very easy application however we cannot stop with simple prompt engineering of course

Trump engineering can be also complex right I like this uh there's a lecture given by open Ai and deep learning Andrew Inc there's a very nice uh prompt engineering lectures you guys perhaps some of you have visited already I think it's very useful even prompt engineering is not a easy thing and I think for most companies there are two or three most urging things one is up skill ourselves and up skill our teams let them to become excellent prompt engineer because I think generative AI foundational model is not a elite group like Central EA digital sense team I think every every employees every product developer should have this knowledge in order to build great applications the second one is also identify the needs of your company right whether you are happy with leave the model as is or you want to infuse your own data I want you want to build your own industry foundational model if we go beyond prompt engineering I think it's very important with the continued training and fine-tuning this will

make your wish to model your results more successful and I close with Essex you cannot do AI without talk about it as Thomas mentioned already right sap we are one of the Pioneers in European Union we defined AI guiding prints as a guiding principles in 2018 and also I I want to share also with you our practice and not only we have this guiding principle on the paper we also do the implementation in our everyday work and life we have a governance in every month we have our steering committee within the company they are experts from legal experts from Human Resource experts from in technology engineering marketing we come together and study the use cases and identify the challenges with data and use cases furthermore we also invite International eisic experts who can help us if we Face some challenges money to our use cases and that our trustworthy AI work stream they work with the different teams together starting from the ideation of a use case and accompany them in development deployment and continuous

Improvement and also the marketing I think and this together will make the ear Essex not only at a very conceptual level and Good Will right in the practice in the in the real world life and work as of here I have still 30 seconds left thank you very much Faye for a very insightful presentation and so you do have 30 seconds and I'm going to give the floor the opportunity to ask any questions and whilst you're thinking through you've had a very positive largely positive and even in the crowd view on generative AI I'd like to maybe look at the other side of that and what is it from your perspective the biggest challenge that it faces in actually being adopted and where is where is that focal point and still need to be jumped through as a hurdle if there are any in fact yeah I think there are a lot of challenges right I think one of the biggest challenges on the one hand is generative is so exciting they can also take

over so much work right writers essay for us create a PPT for us however I think one of the biggest challenges to identify the truth the truth even for one company right the also identify whether the information really uh the origin originality of the information right I think from the text point of view it's less chilling if you create videos right I think this is the authentication of the content it's very important and ultimately help us build more trust yes thank you are there any other questions on the floor yes please and it's microphone lady in blue blue blouse thank you so yes when you the chart that you showed regarding the evolving of basically people using all alarms now as is and then going Enterprise AI of all my yeah this one uh different consumption and tuning patterns thank you so um definitely currently we're at the used model as is and and partly the tune models on data I'm wondering um when you say Enterprise AI um so at least when you look

at the Deloitte AI reports it always shows that Enterprise AI mostly uses wrapped products now so they they use wrapped AI products and they take this and try to adopt it and I see this also with llms I see that currently most Enterprises would try to get an alalam product and an AI product that works well for legal or that works well for chemistry or so do you see that as well or do you really think that enterprises will adapt the knowledge how to build llms and then integrate that into their I.T systems I think two since right our customers when they approach me they have two questions one questions how can sap leverage large language model improve our portfolio in order to make their business process their ID successful right this is the sap we we have to work on Infuse structured data transaction data into language model in order to do the prediction and also planning right this is one thing another thing is also our customers also ask how can we help

them to deal with like chemical industry and in order to help them to build their successful products I think this is two points right when I talk about Enterprise AI I mainly talk about from the sap point of view right the business process part and as sap you would mostly sell wrapped AI products right I don't I will not call this air product I would call them business AI applications right AI is part of the product yeah fantastic thank you very much for your questions thank you Faye it was a delight thank you thank you a truly insightful talk so before you all stand up to leave I'm have a pleasure to introduce our next esteemed guest who

Rise of AI 2027 · Berlin · May 2027

Be in the room for the next one.

Ten years of talks started as conversations in one curated room in Berlin. The next edition brings 300 selected leaders together in May 2027.