- SAP Managed Tags
I have been actively involved in the SAP ERP domain for nearly 30 years and have witnessed numerous pivotal innovations across SAP’s evolving technology landscape. My ABAP journey began in 1996, when I developed custom code for SAP R/2 HR. As a consultant, I’ve experienced the complete evolution of ABAP—from its early days to the most recent releases—closely following every significant change, enhancement, and breakthrough. Over the past decade, these innovations have had a profound impact on the developer community. However, in my view, no shift has been as revolutionary as the introduction of Joule for Developers and the rise of AI capabilities within the ABAP environment. In this blog, I will share my thoughts with you.
In October 2024, I had the privilege of participating in SAP’s internal workshop series, Joule’s ABAP Developer Capabilities, through my role as an SAP Mentor. This engagement provided early access to Joule for Developers within ABAP in Eclipse, enabling me to conduct hands-on testing, provide structured evaluations, and contribute feedback ahead of its official launch in February 2025. I use ChatGPT regularly both at work and in my personal life. However, during these workshops, I realized my lack of in-depth knowledge about AI.
I realized that to be future-proof, I must invest in myself, and I was happy that my company entrusted me with developing an AI introduction training program tailored for SAP developers. Additionally, in one of my client engagements, my responsibilities as a solution architect were expanded to include the role of AI advisor. In this role, I had the opportunity to join a dedicated two-week hackathon, where I explored how existing AI capabilities could be leveraged effectively and identified key success factors for deploying AI within complex SAP landscapes. With these opportunities, I'm on my way to closing the AI knowledge gap necessary for success as an SAP Architect and SAP developer, and I want to share some of my experiences.
Joule for Developers
As an SAP enthusiast, I began my journey at the SAP Learning Hub to expand my knowledge and skills. And even though Joule for Developers is relatively new, I found detailed information in the delta course to be certified as an ABAP developer. In this course, it is explained that Joule for Developers supports SAP ABAP developers with AI-based capabilities throughout the development process. This includes scenarios like
- Predictive code completion: generating ABAP code in the source code editor
- Joule chat: natural language interaction with an AI-based chat to get support for any development-related questions
- Explain: explaining development objects or selected lines of code in natural language
- ABAP Unit: generating and improving unit tests
- CDS test generation: creating ABAP test classes for CDS entities
The Code Assistant is integrated into the ABAP development tools in Eclipse. It functions as a chatbot that understands complex development objects within ABAP Cloud, allowing developers to ask interactive, natural-language questions and provides real-time answers.
And with the ABAP AI SDK in ABAP Cloud, the developer has even the possible to develop custom generative AI features directly within your business applications. But this capability is currently (June 2025) only possible when your application is running on an ABAP Cloud environment (such as SAP BTP ABAP Environment or ABAP Cloud in SAP S/4HANA Cloud). You need also access to SAP AI Core and Generative AI Hub, which you only get when your company has an SAP BTP subscription for SAP AI Core, SAP AI Launchpad, and Generative AI Hub.
For many developers, the journey with Joule for Developers has remained out of reach due to high costs of the BTP services and limited availability within their company SAP environments—but that’s about to change. With the September 2025 release of S/4HANA, AI capabilities will become available for both on-premise and private cloud deployments. SAP also plans to extend compatibility to the 2023 S/4HANA releases. This means we’re just a few months away from a major opportunity with Joule and the broader AI potential SAP is unlocking for developers.
However, as Joule becomes accessible to a wider group of developers, this shift also raises important questions:
Are ABAP developers prepared to leverage the AI capabilities of SAP Joule for Developers entirely?
Are we, as a developer community, truly ready for this transformation?
If not, should we not seek additional guidance and support to navigate this new chapter successfully?
These are critical questions because the answers will ultimately determine the success of Joule for Developers. To explore the answers to these questions, I decided to dive into hands-on research and practical testing.
Research and practical testing
During the development of the AI training, I frequently used ChatGPT to find answers to my questions. Because I also had access to Joule for Developers on an embedded cloud environment of SAP S/4HANA Cloud, public edition, I could check my thoughts and findings directly in Joule for Developers.
I began by asking some general questions about Joule, directing them to both ChatGPT and the Joule platform. Interestingly, both AI tools provided similar responses. SAP Joule is specifically tailored to understand ABAP, SAP data models, and business context, making it more suitable for SAP-centric development. In contrast, tools like OpenAI's ChatGPT and GitHub Copilot offer more general-purpose coding assistance across a wide range of programming languages and platforms.
All the AI tools recommended using SAP Joule as an assistant for ABAP development. However, as a seasoned ABAP developer—with a few gray hairs but an ABAP Cloud certificate—I need more than just a recommendation. I want to be convinced. So, I decided to put both Joule and ChatGPT to the test by asking them to create a function module using the following prompt:
Both tools provided similar ABAP code for the function module. However, Joule went a step further by also offering a step-by-step guide on how to create the function module within an SAP ERP system.
However, here’s the catch: neither response enables me to run the function module in my S/4HANA Cloud, public edition system. The provided code doesn’t align with clean core principles and utilizes unreleased artifacts. Moreover, Joule’s step-by-step guide is essentially useless in this context, as I’m not allowed to use SAP GUI or traditional transaction codes in a public cloud environment.
As an ABAP developer—especially one without a background in AI—I found this experience frustrating. From an SAP-specific AI assistant, I expect valid ABAP code tailored to my system landscape, along with a step-by-step guide that reflects the actual constraints and best practices of S/4HANA Cloud.
Naturally, I tried to solve the issue by asking follow-up questions, but I continued to struggle with incomplete or incorrect answers. It wasn’t until I deepened my understanding of generative AI, large language models (LLMs), and prompt engineering that I began to see why the AI tools responded the way they did. In the world of AI, these issues are referred to as bias and hallucinations. In other words, the problem wasn’t an error in the AI tools themselves—it stemmed from my initial lack of understanding.
It also reminds me of a key insight I came across while researching for the training I developed. In the paper 'On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?' by Emily M. Bender, Timnit Gebru, Angelina McMillan-Major, and Margaret Mitchell, the authors emphasize — in my own words
GPTs are trained to generate human-like text, not necessarily accurate text.
I found this to be especially true, notably when someone lacks a solid understanding of how generative AI and large language models work. And this so true for the question I asked.
Both models were trained on vast amounts of ABAP code and SAP documentation, learning to identify and reproduce common patterns. So, while Joule was designed to generate ABAP code and step-by-step guides, it wasn't necessarily generating correct or system-appropriate solutions. It was doing exactly what it was trained to do: generating content that looks plausible but is not guaranteed to be valid in a specific context, such as S/4HANA Cloud, public edition.
Prompt Engineering
Several techniques—collectively known as AI grounding—can be applied to overcome incorrect answers caused by bias and hallucinations. Some of these, such as supervised fine-tuning (SFT) and reinforcement learning from human feedback (RLHF), are the responsibility of the AI vendor—in this case, SAP, the provider of the Joule assistant. SAP has already made considerable progress in applying these methods to improve the quality and reliability of Joule's responses.
However, the user of AI tools often has the most significant influence on the accuracy of AI-generated answers—in our case, the ABAP developer. By using effective prompt techniques, selecting the proper AI tool settings, and adhering to established best practices for prompting, developers can significantly enhance the quality and relevance of the answers they receive. Prompting well has become a dedicated discipline, leading to the creation of a new job: the Prompt Engineer. This means that the ABAP developer should learn a new skill, prompt engineering, at least to become a successful developer with SAP Joule.
If I try the same prompt question as before, add context, and follow the zero-shot chain-of-thought (COT) task step-by-step, a widely used reasoning technique in prompt engineering, the Joule result improves significantly.
In the response, Joule provides a step-by-step guide that aligns well with my ABAP development environment in Eclipse. It clearly outlines what I need to do and suggests creating a class rather than using a function module—an approach more typical in an ABAP Cloud environment. It also highlights that direct access to SAP tables and unreleased APIs is not permitted, which aligns with ABAP's core principles of cleanliness.
The provided code skeleton matches what I would expect in an ABAP Cloud setup. However, it falls short in one crucial aspect: it doesn't specify the correct ABAP CDS view to use.
As an ABAP developer, I’d naturally ask follow-up questions like:
What is the CDS view for customer information?
Which released CDS views are available for customers?
But the answers are often unhelpful, forcing me to search and experiment on my own, which hurts productivity.
A skilled prompt engineer, however, would recognize this approach as weak. Instead, they would refine the prompt using additional context and prompt engineering techniques, iterating multiple times until the correct result is achieved. He will likely develop a prompt like this. In many cases, these tailored prompts can also serve as reusable templates for similar scenarios in the future.
This response reflects the level of support I expect from a SAP AI assistant—providing clear guidance and ready-to-run code.
Conclusion
Based on my research, I can conclude that ABAP developers need at least basic prompt engineering skills—and ideally, a broader understanding of general AI concepts—to effectively utilize SAP Joule for Developers and fully leverage its AI capabilities. Currently, most ABAP developers lack these skills, which limits their ability to derive the maximum value from the tool.
In my opinion, the developer community is not yet fully prepared for the AI transformation, primarily due to a lack of foundational knowledge in AI. I was fortunate to be in a position that allowed me to conduct this research and enhance my understanding of AI while developing AI training for our company. However, this is not the case for most ABAP developers.
To ensure success, SAP needs to provide more structured guidance and support. At the same time, companies must invest in upskilling their developers to capitalize on the opportunities offered by SAP Joule fully.
This article was initially posted on the linked-in profile of Robert Eijpe
Disclaimer:
SAP Joule is an AI tool, and the results of prompts may vary over time or when repeated. Thus, the responses Joule provides can differ from those shown in this blog.
| User | Count |
|---|---|
| 7 | |
| 6 | |
| 6 | |
| 4 | |
| 4 | |
| 4 | |
| 3 | |
| 3 | |
| 3 | |
| 3 |
