Can we use ChatGPT to enhance Supplier data collection for ESG?
Well, precisely because of that!
AI is as smart as we build it. An “Artificial” Intelligence is (still), in its essence, build by humans, and run on some electronic “circuits”. Some would even argue – there is no such thing as an “Artificial Intelligence” – there is only “Intelligence” or “no Intelligence”… But this article is not the debate of what AI is – this is not the topic I want to cover. In fact, I do find AI to be very useful for many applications, and in this article, I am exploring one possible use-case of AI – using AI to support ESG data acquisition and thus accuracy of the overall ESG reporting.
Problem to be addressed
In the previous article Using Ariba for getting Supplier inputs for the calculation of the Greenhouse Gas emissions | SAP Blogs I have elaborated few concepts of getting Supplier data inputs relevant for ESG reporting.
If our Suppliers will provide input data in the electronic exchange formats – which we can integrate with our systems, then we have no problems. But in general, this is still not so common. ESG relevant Supplier inputs (e.g. certificates), whether we are connected via Business Network or not, very often would come as unstructured electronic (or even non-electronic) documents. Assuming, majority of Supplier would be able to provide inputs as an electronic document (e.g. PDF), and considering that even non-electronic document (e.g. paper) can be scanned – I will focus only on unstructured electronic input documents.
What kind of Supplier inputs we are talking about?
It could be various ESG relevant inputs, either for GHG reporting, or EPRs, or various ESG frameworks – like GRI or similar:
- % or volume of the delivered renewable energy;
- % recycled connected with delivered raw materials;
- Costing part of the recycled connected within delivered raw materials;
- CO2e footprint of the raw materials;
- GAP or similar certificates;
How does ChatGPT fits-in?
Standard ML for content recognition needs to be trained. If Supplier is changing the layout/format of the document, or if we are bringing different suppliers with different document formats – each time ML model would have to be re-trained.
Large Language Model (LLM) are proven to be very good in providing answers with no or very little learning (basically, unsupervised learning). ChatGPT on the other hand is based on Generative Pre-trained Transformer (GPT) version 3.5 (at the moment this article has been written). It is “more” than LLM. Optimization is achieved using Reinforcement Learning with Human Feedback (RLHF) – and this “Human Feedback” means it has been pre-trained on vast amounts of data written by humans. This is what makes it so compelling for humans during the interaction – as it sounds quite human-like.
For the particular problem of recognizing content form the Suppliers’ inputs – the ability to create human-like interaction is irrelevant. However, what is relevant is the ability to recognize the content without the need to be trained on the any Suppliers’ input data forms.
I am presenting a Solution Concept – an idea of potential use of ChatGPT (or similar AI) model to support the collection of Suppliers’ inputs for the purpose of the ESG reporting. This does not imply ChatGPT is “ready” product for this purpose and can be used out-of-the-box. ChatGPT is primarily designed to “chat”, but the model behind is promising for other uses as well.
The Process steps:
- The process would start by collecting Suppliers’ inputs – e.g. via Business Network.
- The next steps will be text recognition (OCR) as ChatGPT works with text.
- Then, we would have to pass the text and questions to ChatGPT (e.g. this could be a certificate of the overall quantity of the renewable energy delivered by the electrical energy Supplier)
- ChatGPT would (ideally) recognize the context, and extract relevant answers from the text (e.g. what was the quantity, in which period etc.)
- Finally, ChatGPT provides answers, but those answers would have to be mapped with the specific business content (e.g. by mapping “questions” with specific Business Objects like Quantity, Period, Date; we can map “answers” and assign values to the respective Business Objects).
Creating human-like responses by ChatGPT in fact would not even be beneficial for this use-case. We are only looking for the ability to recognize the “text” and pull-out the answer from the text – where it was written.
For testing this Solution Concept, I have used “rich dummy” PDF document – example of Supplier certificate stating, among others, quantity of the renewable energy delivered in the specific period of time.
Then, I have asked several questions about quantity of the renewable energy and the period when it was delivered.
The answer was correct, and it was encapsulated in the first sentence.
The answer was also correct, although the dates are formatted for the human interaction.
The approach with Ariba SLP
If we would like to put ChatGPT (or some performant & capable LLM) in motion for this use-case, some configuration work would still be needed:
- Acquisition of the electronic documents from the Suppliers – APIs should be able to fetch the Suppliers provided input documents – e.g. certificates. This however is not a problem if using Ariba SLP Modular Questionnaires – as fetching attached certificates is supported by Supplier Data API with Pagination.
- Classification of the document might be needed (e.g. is this certificate for the renewable energy, or GAP certificate) – this is needed in order to ask “right” questions. Either the classification of the documents would happen during the data acquisition, or some training might be needed for AI to understand “what is this document about” and classify it accordingly – and only then “ask” appropriate questions. Again, if using Ariba SLP Modular Questionnaires – either different questionnaires or different questions are used for different certificates, so each incoming document text is already labeled/ classified and thus appropriate questions would be asked.
- We do need answers are correctly extracted from the “text”, but we do not need human-like answers. Here, recommendation would be to pre-train model to do what we need.
- Business Mapping of the answers with specific Business Objects would need rule-based engine to be built on top of the AI. However, building rule-based business mapping should not present a major configuration task in the SAP Integration Suite (CPI).
High level business process:
*) for additional details, please visit my previous article Using Ariba for getting Supplier inputs for the calculation of the Greenhouse Gas emissions | SAP Blogs
In all examples tested (5 test samples in total), ChatGPT was correctly understanding the questions and extracting the answers – although answers tent to be too human-like (which is expected from ChatGPT). We may ask ourselves, if we do not need this human-like interaction, why don’t’ we just use any other LLM? The thing is, ChatGPT does demonstrate superior quality in understanding the context of the questions – thus easily extracting the correct answers from the large text document. Other models based on the LLM, of course, might also provide similar capability – however, this small showcase is limited to only one AI model which was tested – and for this showcase I have selected “most” mentioned AI model these days – ChatGPT…
In general, this test does demonstrate generative AI can be of use in the enhancing (or overall simplifying?) document Business Content recognition. This is noting unexpected – we do expect AI to be “smart”. However, the main benefit of using “smart” AI like ChatGPT would come from following facts:
- superb understanding of the semantics within the “text” – being able to understand the question and extract right answer;
- no real learning is needed to recognize specific “text” forms.
Another interesting feature, worth mentioning – I’ve loaded some “certificates” in different languages (e.g. German) – and ChatGPT was processing them without any problem.
Now, the question is – do we want to use ChatGPT for this use-case, or we would rather look to “build” more specific AI models (e.g. LLM or so)?
And of course, share your thought and comments on my article, in the comments section.
*) the original Blog has been posted on March 21st, 2023 – republishing due to the technical reasons.