To ensure that the article does not come across as clickbait: In this article, we will cover how the underlying algorithm behind ChatGPT, rather than ChatGPT itself, can be leveraged in the SAP ecosystem. ChatGPT, as you may know, is an AI model developed by Open AI, which leverages a specific type of algorithm, reinforcement learning, to engage and interact in a conversational dialogue. You can learn more about ChatPGT here and access the chat interface using this link. Like me, you may have derived your interpretations if you have played with ChatGPT interface for a while. I have seen many people deride it because of stupid answers to simple problems, whereas others applaud it as it provides accurate and easy-to-understand answers to complex questions. As mentioned at the beginning of the article, we will not analyze ChatGPT specifically. But my verdict on ChatGPT is what will be the premises of this article so I will start with that. ChatGPT is an excellent outcome of the research work that has gone into it. You may ask at this point that if the outcome is so good, why does it answer seemingly simple questions wrong? You will find many such examples on professional forums where it answers simple elementary-grade math questions wrong, and my interpretation is what I always highlight- a model is as good as its training.  So, was the training flawed? Not in my opinion. Since ChatGPT is being trained as an “all-encompassing” chat interface that can interact with users on topics spanning almost everything, there will be “holes.” Now, the initial training was probably done by smart people who were not interested in answer to simple problems. Hence, the algorithm does a better job of debugging codes than answering 2nd-grade math problems. But that does not diminish the utility of the algorithm. In this article, we will explore how the specific type of reinforcement algorithm, Reinforcement Learning from Human Feedback (RLHF), can be used to significantly transform the SAP end-user ecosystem. Before we get into this SAP-specific use case, let us understand RLHF in simple terms. And I will use an example from the real world. My son, Sahaj, had speech delays as a toddler, so a therapist used to visit us to help him with his speech. The goal of these therapy sessions was to present language in various contexts to my son so that he could understand how to communicate in specific scenarios. An example would be when the therapist visited Sahaj on Mondays, she would encourage him to ask her what she did on the weekend. When Sahaj would use language like “What did you do on the weekend?” the therapist would encourage him to use more natural language, like “How was your weekend”? The goal, she insisted, was to provide feedback during my son’s interaction with her so that his language becomes more and more natural. Sahaj used that feedback as a form of reinforcement to update his language and responses (Like when he got a response to his “weekend questions”, he would say- “That sounds like fun”). Over time, based on the feedback he received, my son could synthesize conversations that came across as more natural. But this language naturalization also needs to be married with context. If one of his friends says they attended a basketball game and start talking about it, Sahaj needs to understand baseball to carry the conversation forward. In the applied RLHF context, that means integrating the capability of naturalized conversations with subject matter expertise. So if you decide to discuss your Python code with ChatGPT, it needs to have the subject matter expertise AND be able to leverage the “learning” on how to present that expertise to the interacting human in a way that seems like human conversation. And now let us start thinking about this from the SAP technology ecosystem perspective. But before we get into that, let us quickly touch upon a question- Can we consider ChatGPT a decent outcome, considering that it cannot answer some very simple questions correctly? My answer is -Yes. The notion that there can or should be an AI algorithm that is all-knowing and omnipresent is very flawed in my opinion, as suggested in my other writings. There is no need to have one. The world can use AI algorithms that excel at more difficult problems, and it is ok if they suck at solving elementary-grade problems. It does not mean that the algorithm is flawed. And what this means is that while GPT is an exciting project, the key learning is that you can, indeed, build a tool that not only presents information when asked on intricate subjects (if it has been trained on that subject) but can also present that in a “natural” way. And this takes embedded and augmented analytics to a whole new level. With this context, let us jump back into the SAP ecosystem. Complexity is a word that has been associated with SAP portfolio of technologies for a while. Organizations believe that complexity is embedded deep within SAP systems from implementation to configuration to maintenance. An ecosystem of partners has thrived for decades, helping SAP customers deal with this complexity. However, this complexity can quickly become a strategic disadvantage in the age of more agile, easy-to-implement, and intuitive systems. And this is where a solution like ChatGPT can come to the rescue. Imagine a conversational solution that can help perform configurations of SAP modules, suggest data quality approaches, and possible enhancement and enrichment opportunities. With its own conversational AI solutions, NLP algorithms, and embedded analytics features, SAP has been pushing to reduce complexity. But as some end-users complain, some of these efforts in some cases have added to the complexity. SAP needs to develop an all-encompassing solution, like ChatGPT, with expertise in its systems. It needs to be one single solution that can have features beyond conversational AI. The future of enterprise systems is not just low-code, but systems that interact with end-users who are not techies but those working "in the process." And the primary driver behind that is to reduce the complexity that is associated with current state systems. While piecemeal approaches of reducing complexities are good progress, a more comprehensive approach, in the form of one solution, is needed. This also means essentially re-envisioning the underlying thesis of enterprise platforms. The thesis needs to evolve from systems that users interact with in the form of primarily one-sided interactions into systems where the system provides business process aid, context and information to the end-users as they initiate interaction. This is eventually going to be the future state of ERP and approaches like ChatGPT can play their part in building such enterprise soluttions.