How the AI chatbot works in the SAP environment
- The SAP user submits a request for a document via a simple web interface.
- The AI analyzes the context of the request in the background, for example that it is a production order and the number of the element.
- The request is technically processed and transferred to the SAP system via the interfaces (APIs).
- In response, the SAP system returns a technically structured file that is sent to a language model.
- The language model interprets the response and prepares it in an appealing way, for example as continuous text or in structured sections with important elements such as tables or diagrams highlighted.
- The finished answer is presented to the user clearly and comprehensibly - without the need for long waits or research in the SAP system.
The user no longer has to memorize transactions, search for the right Fiori app or search applications for the required field. Efficiency also rises to a new level when preparing reports or interpreting data!
The good thing about the CONSILIO prototype is that it can be implemented in both a cloud and an on-premise version. In the on-premise version, all data remains within the company network. Both self-hosted language models and the interfaces of external providers such as OpenAI can be integrated. This increases data security and reduces data protection concerns, which is particularly important when processing sensitive data.
Developing an AI chatbot: how to proceed
Before developing an AI chatbot, a proof of concept (PoC) should be carried out: the first phase of development in which the concept is tested for feasibility. Depending on the problem, it may initially be advisable to search the internet for existing models on platforms such as HuggingFace.
Freely available models can be adapted to your own data through fine-tuning, which is a special form of model training. In this case, it is not necessary to start from scratch. The pre-trained model is then further trained over time with company-specific data and thus specialized to improve its performance in specific use cases.
If no ready-made models can be used, a separate model is trained from scratch. Depending on the complexity and scope of the use case, this can be time-consuming and cost-intensive, which also applies to fine-tuning. An iterative approach helps to gain initial impressions of the quality of the results, the expected effort and the scalability of the training. A deep understanding of the available data is crucial for training and fine-tuning the model. It is therefore necessary to clarify in advance which data sources exist in the company, whether this data can be retrieved and how it is to be interpreted.
Experience has shown that combining different data sources often leads to more meaningful results. The data must be prepared for model training. If no suitable data is available, a separate data set is developed. This includes the collection, cleansing and formatting of the data. The model is then trained with the prepared data and subsequently evaluated. This makes it possible to check how well the model fulfills the tasks set. Depending on the results, it may be necessary to adapt the data set or add additional data. The model can be further optimized based on the evaluation results. As soon as the model has been satisfactorily trained and evaluated, it is saved and made available for integration into the chatbot.
With its prototype, CONSILIO shows interested companies how they can increase their efficiency and successfully integrate innovative technologies relatively easily by using AI. Get in touch with us!
Summary
The CONSILIO prototype shows that it is possible to develop your own AI solutions within a reasonable framework if strategic planning, investment in technology and employee training are taken seriously during the project. This also includes first implementing a small application area/functional scope before scaling the solution. Anyone who also attaches great importance to data security and data protection because they process sensitive data should implement the on-premise version of the solution. In this case, all data remains within the company network. This protects company data from access by external providers.
