In this blog adapted from the podcast “Artificial Intelligence (AI) for SAP S/4HANA and SAP BTP: A deep dive”, Terry Penner and Jürgen Butsmann discuss the AI priorities, practical applications, and key use cases for SAP S/4HANA and SAP BTP.

Topics discussed:

Part 1: The value of AI at SAP
⦁            AI models explained simply
⦁            The different levels of AI
⦁            Use cases where AI adds the most value
⦁            Generative AI explained under the hood
⦁            Why GenAI now?

Part 2: AI applications in SAP S/4HANA and SAP BTP
⦁            Using AI in SAP business processes
⦁            Ensuring AI at SAP is relevant, reliable, and responsible
⦁            What AI use cases are available in SAP S/4HANA now
⦁            The role of SAP BTP in AI at SAP
⦁            Future plans

About the speakers:
Jürgen Butsmann is with the SAP S/4HANA Cloud Solution Management team responsible for AI in the context of S/4HANA, finance and digital supply chain. Jürgen has been with SAP for more than 26 years.  He is a recognized expert on SAP S/4HANA and SAP BTP.

Terry Penner is part of the SAP BTP Marketing and Solutions team, focused on BTP for SAP S/4HANA. He has more than 20 years of technical and business experience in the SAP platform and analytics, including working directly on hundreds of customer implementations.

Part 1: The value of AI at SAP

Terry: Today, we’re discussing artificial intelligence or AI. Jurgen, could you explain why AI is important for SAP, and what is its value?

Jürgen: AI is more than just a buzzword; it’s a technology that we can use in many areas, particularly in business software. Fundamentally, AI is about replicating or enhancing processes that we humans could use help with from computers.

AI models explained simply

We as humans build our world models based on our experiences and senses. These models grow more complex as we learn. However, some decision-making based on these models can be simplified and implemented into computer science. We’ve done this before with rules and models in simple programming.

But with AI, we’re dealing with software learning from experiences, or in our case, from data. The models we develop through these algorithms, created by data scientists, will likely only capture a limited complexity of the models that we have in mind.

The models we use in business processes may not be as complex as human thought, but they can process vast amounts of data that we can’t retain in our heads. They can effectively analyze and correlate this data, using algorithms for tasks like clustering or projection. This allows us to understand situations, such as a customer’s status or a business process, based on a large volume of diverse data.

The models quickly process this information and provide valuable insights, helping us identify problems or decide on the appropriate action at a given time. They function like a brain that pools all past experiences, applying them to a model that, while not overly simple, is less complex than human cognition.

The different levels of AI

Terry: AI can handle simplified problems due to its capacity to process an immense amount of data almost instantly. However, AI isn’t always necessary. Can you elaborate on the different levels of AI and when its use becomes advantageous?

Jürgen: That’s a good question. Various levels of intelligence are supported with decision support. Understanding what constitutes intelligence can be tricky. My view is that if something is perceived as intelligent, then it is intelligent. For instance, if we have a key performance indicator (KPI) or a report that provides a timely figure useful for customer interaction or decision-making in a sales process, it could be considered intelligent. Even though this is static and doesn’t show any changes or indications, its relevancy at that specific time makes it perfect.

While a static figure or graph can give you a visual representation of trends and help you make decisions, it’s limited to two dimensions due to our perceptual limits. AI, on the other hand, can analyze huge amounts of data and derive conclusions that are beyond human capabilities. However, this comes with expenses such as programming and data collection.

Use cases where AI adds the most value

Terry:  You mentioned the expense of processing large data quantities. Where do you see AI adding the most value?

Jürgen: AI is most valuable where it can accelerate processes. However, it’s not always cost-effective. It largely depends on the data, which is the most crucial part of any algorithm. The data’s type, correctness, and heterogeneity are vital for properly training the algorithms. Despite this, you still need to determine if AI adds sufficient value to a specific business process or decision, considering the costs associated with processing the data.

Terry: Let’s delve deeper into generative AI, which has rapidly developed in the last couple of years. What makes Gen AI unique and interesting to you?

Jürgen: What struck me was the sudden increase in awareness and understanding of generative AI. The examples are tangible; for instance, you ask a question and get an immediate summary from a vast amount of data. It’s clear how valuable this can be. Discussing its value is much easier compared to, say, an algorithm created in the backend that suggests recommendations based on certain data. While those are useful, they’re more abstract. With Gen AI, everyone can easily grasp its practicality.

Considering use cases such as automatic summaries, language conversions, and sentiment analysis, AI touches our lives daily. However, not all business tasks can be anticipated from a private perspective. What’s intriguing is the generation of new information, not just processing data and obtaining a result. This creation process can be fascinating, and even a bit daunting. But if we keep it within the right parameters inside our software, there’s no need for fear.

Generative AI explained under the hood

Jürgen: From a process perspective, you have a request, or what we call a prompt. Based on this, the system generates results from existing data. It doesn’t just summarize by cutting away but creates more distinct, crisper information. The large language models we use can also modify the requests, translating or rephrasing the prompt into words that the system better understands. This results in more precise questions that help the algorithm work and generate more accurate outputs.

Terry: I’ve found generative AI very effective for tasks like summarizing articles or cleaning up podcast transcripts. It simplifies many aspects of my day-to-day work. I also find its potential for translation and democratizing language understanding fascinating.

Jürgen: It’s amazing that we can now have a podcast in one language automatically recreated in another using the original speaker’s voice. This not only globalizes information but also enhances accessibility. It can help refine questions based on user interaction and understand different dialects or language types, allowing more people to participate in this field.

Terry: Can you elaborate more on how generative AI models work?

Jürgen: The first step involves gathering a broad and diverse set of data. This could be from public or business sources and could even include customer experiences or proprietary data. The models are trained on this data, allowing them to process and extract new insights from it.

For example, if you feed it a text, it compares this with other similar data, like podcast structures. By comparing and combining these different models, it can create new content. In our example, a new podcast could be created from the content of your question and experiences from other podcasts. The system’s capabilities improve the more diverse information it receives, as it can generate, compare, and create the best output from this data.

Why GenAI now?

Terry: It’s crucial that the data fed into this model is relevant and good. Why do you think generative AI is gaining traction now?

Jürgen: I believe it’s due to the impact of the first version of ChatGPT, which sparked a fascination with its broad-scale applicability beyond just scientific domains. Since then, many technologies have emerged. SAP has always been a company that embraces and adapts technology, but always in the context of business processes, as that’s our purpose.

We aim to enhance accessibility to systems like our ERP. The transition from large computers to more compact ones, and now to client-server, has significantly increased this accessibility. It’s not that SAP invented all these technologies, but we’ve successfully integrated them into our business processes. While SAP has shown the value of these technologies, we are now focusing on the implementation of large language models and generative AI capabilities. They have proven use cases and are offering real value, making it the right time to utilize them.

However, the implementation needs to carefully consider the associated costs. We need to ensure that the processes we’re supporting are worth the investment. It’s also crucial to understand that not all these technologies perform the same or offer the same quality. This is why we’re integrating the large language models and generative AI capabilities into our technical architecture, SAP BTP.

Part 2: AI applications in SAP S/4HANA and SAP BTP

Using AI in SAP business processes

Jürgen: Our goal is to provide solutions to as many business problems as possible. These large language models will be available to our customers through our architecture and will be commercialized through our software.

To use the correct model for the right purpose, it’s not a given that we have this capability. Many of our customers have been seeking AI, especially generative AI, recently. However, the unique aspect of our approach is its versatility, the capability to leverage many features, and its integration within our technical and business architecture. This is the direction we’re heading, and it necessitates significant investment in this area.

Terry: To summarize, we aim to integrate generative AI as closely as possible into the business processes that SAP understands best, instead of trying to make it general purpose.

Ensuring AI at SAP is relevant, reliable and responsible

Terry: We’ve talked about ensuring our AI is responsible, reliable, and relevant within SAP. Can you expand on what it means for our AI to be relevant?

Jürgen: Relevance means that the AI should be necessary in the context of our business processes and should provide value to our customers. The value of the AI is determined by its impact on the business process. Each customer assesses whether the AI is relevant based on factors such as process duration, quality, potential repetitions, user skill shortages that could be improved by AI, and how information provided through AI affects their process execution.

We understand that, for example, in a process where we need to scan a certain number of documents, understanding the manual process to process those documents is crucial for a customer. They need to evaluate whether the cost they would incur is worthwhile. In this way, every customer must make this decision. We need to anticipate what this means on a large scale, as it’s probably most valuable to most of our customers or to a larger group. We’re not just focusing on smaller cases, but also those with a big impact.

Terry: Understandably, reliability is crucial for AI at SAP.  There have been instances where AI models provide inaccurate or outdated information. How does our approach to generative AI at SAP ensure reliable results and training models?

Jürgen: Indeed, reliability is critical. The quality of results largely depends on the data input. If your data is very heterogeneous, it may not be suitable for some algorithms, like account matching. This could lead to an inability to capture outliers. Therefore, it’s essential to understand how the data should look, how much data you need, and where the quality signals lie. Strive to select the best data for your specific organization or even for a part of your organization. This will best reflect the occurrence of certain analyses in the system.

It’s crucial to explain how and why the results are moving in a certain direction. If some expected values are missing, remember it’s a process. The more you use it, the more confident and trusting you become in the system. We’re all learning to embrace this. It’s simple to test something in a familiar context. If the results don’t meet expectations, you might question its reliability. However, consider an occasional user trying to answer a complex question. They might accept an incorrect answer because they can’t judge its validity. Therefore, we need mechanisms that explain results to our customers, building their trust and understanding in how to handle the data.

Terry: Absolutely – if you’re going to make decisions based on what the model tells you, you need to trust it. This applies whether the data is from people or AI. Now, onto the third part of the question around responsibility. I believe this is a big differentiator for SAP, ensuring that we are responsible with our AI and building confidence with our customers and partners. Can you elaborate on what being responsible means to you?

Jürgen: Responsibility to us means ensuring data security. We don’t want to share our intellectual property externally, and many large language models are situated in an external cloud. We need to ensure that our data is converted, encrypted, and cleansed in a way that it can be used in these models without providing any access to the source data. It’s a big task, but we’re committed to it. Sometimes, we have to deploy some of these models in our own environment to ensure security. Additionally, responsibility also means that the use cases we provide are ethically sound.

What AI use cases are available in SAP S/4HANA now

Terry: Let’s discuss a practical aspect. What AI capabilities are available now in SAP S/4HANA Cloud?

Jürgen: Currently, SAP S/4HANA Cloud offers over 25 use cases based on AI technology. Many of these are built on our SAP BTP platform. Our strategy is to focus on solutions within this architecture. For example, we have automatic matching of incoming payments with open receivables, fraud detection, and automatic derivation of sales order information from unstructured data. We’re enhancing these capabilities with generative AI, which can automatically identify the structural elements of an unstructured document. We’re constantly improving our services by getting more out of these algorithms.

The role of SAP BTP in AI at SAP

Terry: Could you elaborate on the role of SAP BTP with AI in S/4HANA Cloud?

Jürgen: Sure. SAP BTP is our standard environment for technological extensions, including the foundation for all of AI and generative AI in SAP S/4HANA Cloud. It’s where we process our algorithms and integrate both the AI models we own and those we don’t. This connection between business applications and SAP BTP allows us to manage intelligent scenarios where we combine business data and lifecycle management with the technical structures in SAP BTP. This is also the route through which we handle all other models.

Future plans

Terry:  Jürgen, could you discuss some of the key topics Gen AI is focusing on or the development team is currently working on?

Jürgen:  We’re considering and developing a variety of use cases, with some already in production for the first half of 2024. Let me highlight one of our major projects: SAP Joule, an automatic co-pilot and digital assistant. This human-to-machine interface can be used in many ways, including conversational, navigational, and information gathering use cases. It’s based on SAP proprietary data, process flows, and applications. It can help users gather new data about functionalities they want to onboard with or guide them to execute tasks like master data extensions or new sales orders.

Joule helps with navigation and enables a conversational interface for data retrieval. It’s one of the earliest applications and has broad uses in various settings where communication is needed. It often finds use in interfaces where hands-free or verbal communication is beneficial. There’s a lot to explore in terms of making work and system interactions more efficient. As such, it serves as a helpful tool in many scenarios, playing a crucial role in process enhancements and optimizations.

SAP Joule

Jürgen: Communication Intelligence is a key area in our roadmap. It’s a generic use case where decision-making processes like managing a potential dunning case can be executed, incorporating sentiment analysis, task prioritization, customer interaction, call script creation, and other guidance. It allows for an automated execution of transactional tasks, meaning you gather customer information, decide what to do with it, make a decision, create a process, and it helps complete the task with recommended ways to navigate and execute functions.

Enterprise Service Management supports a full-service process, including intelligent ticket handling. It covers multiple areas, offering an intelligent way to deal with requests. For instance, in a shared service center, it aids in ticket handling.

Our feature, currently named ‘Just Ask’, is a natural language interaction tool that aids in finding and executing the right reports and KPIs for end users, regardless of their language or lingo. It enables them to access the data and information they need.

We are aiming to have Signavio not only run process analysis and process mining but also submit system recommendations for process enhancements. We’re also working on code generation. Having written millions of lines of ABAP code, we understand the rules and methods for creating code. So, if there’s new code to be created, it would be beneficial to have a template, or even an almost complete ABAP program, based on your requirements. Of course, this is not about creating a PowerPoint-ABAP converter as jokingly mentioned in an old SAP saying. Rather, it’s about reducing some of the time-consuming and monotonous work involved in scripting. Refinements will definitely be necessary, though.

Terry: Yes, code generation is an exciting area we’re exploring, especially for building on the BTP side. This could greatly benefit our developer community. Jürgen, thank you for sharing your insights about the progress with AI at SAP. We’re eager to see how this evolves and check in again to discuss the new offerings for our customers.

Listen to the full podcast at “Artificial Intelligence (AI) for SAP S/4HANA and SAP BTP: A deep dive”

[ Note:  This blog was adapted from the podcast transcript by an SAP system using Azure OpenAI–gpt-4 ]

Key links:

• SAP BTP for S/4HANA
• SAP S/4HANA
• SAP AI
• Podcast: Artificial Intelligence (AI) for SAP S/4HANA and SAP BTP: A deep dive