A Secure AI Environment for SAP Customers

Artificial intelligence supports developers in SAP programming by helping with code analysis, troubleshooting and optimization. It can automate routine tasks, giving development teams more time for more complex tasks. By using AI, for example, we at T-Systems can develop customized solutions more efficiently and adapt existing codes more quickly. This is particularly advantageous for SAP customers who operate in a dynamic business environment.

An example: companies have unique internal processes that often go beyond standard SAP functionality. ABAP enables them to adapt their SAP systems to individual or industry-specific business requirements. Over time, SAP systems become larger and more complex. In addition, the ABAP developers involved may change companies in the meantime or retire. Due to the long service life of the systems and changing development teams, companies have often not documented their ABAP programming consistently. It is therefore unclear for which functions ABAP code is used. This is where AI comes into play: with its help, it is possible to determine what “old” ABAP code is actually used for, as AI tools can help to understand and evaluate individual pieces of the coding puzzle.

The linchpin of our AI environment are Large Language Models (LLM), which are hosted by T-Systems in a protected, data protection and compliance-conformant IT environment. The advantage for SAP customers: your ABAP code is neither used for learning nor for training LLMs. In addition, the data they enter into the LLM or the requests they make via it do not end up in a public cloud. The situation is different with the LLMs operated by large, mostly US providers: there, the data is freely accessible on the web, so that unauthorized persons can relatively easily reproduce code snippets entered by means of clever queries.

Our AI Engineer uses the LLMs of major providers, such as Llama from Meta at the moment. The challenge: common LLMs are trained with huge amounts of data from the Internet. As ABAP is mainly used within companies for software development, coding or comprehensive information on ABAP is usually not available on the Internet. This is why LLMs such as ChatGPT 4.0 fail to generate compilable ABAP code in around 50 percent of all cases. In addition, the ABAP code pieces generated by such LLMs regularly fail testing because they do not meet the required quality standards.

T-Systems relies on Retrieval Augmented Generation (RAG) to make optimum use of AI Engineer for the development of ABAP code. RAG extends the capabilities of LLMs by supplementing the knowledge from the training data with additional sources of information to generate more accurate answers. An example: if SAP customers have been programming with ABAP for some time, they have their own coding guidelines. Using RAG, we load such additional information into a secure vector database, which provides relevant information with every LLM query and improves the AI’s responses.

The big game changer will be SAP’s own LLM for ABAP code, which was presented at TechEd in October 2024 and is scheduled for official release in 2025. SAP has been training this LLM for some time now with more than 250 million lines of ABAP code. For us, it will then be a matter of connecting the SAP ABAP LLM with the data protection-compliant environment of T-Systems. In addition, our ABAP developers will test the LLM extensively.

Prompt engineering: The key to high-quality ABAP code

Another key element for the code quality of our solution is prompt engineering. AI Engineer offers workflows for multi-level prompts that support developers with routine tasks. For example, the existing code can first be analyzed before new code is generated in the second step. This structured approach further improves the quality of the generated ABAP code. The plan is to use the AI solution in SAP S/4HANA transformations, in particular to facilitate the analysis and documentation of existing code bases. This will be helpful in understanding old systems more quickly and transferring them to modern architectures.

Quick help in day-to-day development thanks to AI

The following example shows how AI can be used successfully in ABAP programming: at T-Systems, we use our AI tool to create code segments or perform simple tasks such as string operations and database queries. The AI tool also provides us with lightning-fast answers to standard questions that ABAP rookies in particular would otherwise have to spend a lot of time researching in SAP forums themselves. However, there is still room for improvement in the use of AI: when it comes to writing entire programs or generating lots of code tailored to customer-specific requirements, AI is currently reaching its limits.

The future of ABAP development as a symbiosis of AI and humans

We are certain that AI will change both ABAP programming and the SAP world as a whole in the future. At T-Systems, we always want to stay on the ball with all developments, try out new things – and always involve our customers in the process. This allows us to determine together where the opportunities and risks of AI lie. But it is also clear that in the foreseeable future, it will hardly be possible to feed the AI with “old” SAP program code and believe that you will get a new SAP S/4HANA application at the touch of a button. This is good news for ABAP developers, who will therefore remain indispensable in the future.