SAP and ChatGPT - Project Chatbots

Imagine your SAP Project has ChatGPT running as a private Chatbot on all the project's documentation...

How much easier would it be?

Here are some examples (there are loads more)...

• You're a harassed PM needing updates. You get your phone and ask the project bot (by voice) ?"What WRICEFs are overdue, who is working on them and give me a summary of what they do"
• You're a user who needs help - "Explain like Im 5, what are Inbound Deliveries are and how does the interface work with them?"
• You're a Functional Consultant – "I think this cose is causing a short dump - Please explain what it does and what errors it might cause.

Can ChatGPT actually do this?

Yes!

..well

...sort of.

Right now you could take ChatGPT, lock it down to your project documents, plans etc and it WILL answer all the above questions - very convincingly.

Just not always correctly. Which is a pity.

The Problem

It goes to the heart of what these AIs are. ?What we want is an “answer-fetching” bot, but LLMs like ChatGPT are “I-make-convincing-conversations” bots. It might look like its trying to answer your question, but its not.

Its trying to make a convincing reply to your question. It doesnt understand accuracy. This means that not only can its answers be wrong, but they can be subtly, dangerously wrong.

But its fixable - you need a few more steps.

You cannot get it to want to answer correctly. What you?can do is make it want to give you convincing replies on SAP questions. If you set this up rightly, amounts to the same thing.

So how do we do that - well here are a few simple steps that can make a huge difference.

Easy level - Improve the users - Train them to use clever prompting to ask specific questions - Ask questions specifically - so the only 'convincing answer' is the right one. Ask for references back to the docs - so you can double check if you have doubts.

Quite Easy - Improve the interface - Use coding behind the scenes to improve the prompts and improve context. Consider 'splitting' bots to cover specific areas then combining them back together in the reply. This can help prevent confusion over similar terminology and reduces hallucinations.

Harder - Improve the data - Think about the sort of questions you will ask and the answer you want to get – then use this to structure your data into smaller chunks and create contextual docs into your private library. You then need to index these chunks and store them in a vector database (its not has complex as it sounds and gives you a tool you can deploy on other projects).

All of this can be very effective, especially if combined effectively and creatively – which is where the art is. Small changes can combine to give you surprisingly powerful solutions - particularly if you use the AI to help you solve the AI problems :).

The exact mix will depend on the project, the industry, the sophistication of the users and the AI you are using.

So yes you can get ChatGPT as a chatbot on your project, but you have to do it right and not just plonk it on top of your library.

Get it right and a small bit of effort can give you a chatbot that saves huge amounts of time and hassle and doesnt lie.