Hey! What you need is complex.
Making the AI respond to questions is simple. We connect it to a RAG (retrieval-augmented generation) so it can search documents to answer properly. We can also fine-tune the model to adjust the tone. Easy!
The tricky part is getting the AI to take actions. Connecting it to an API for order info is simple, but trusting it with sensitive tasks, like handling a refund, is much more complex.
I would suggest that we start with a read-only security layer. This means the AI can properly respond to the client in all the media you mentioned (email, social media, website, etc.), but it won't perform any actual actions, like changing a shipping address, issuing a refund, etc. This would be version 1. Those hard actions would be added as pending to a pipeline to be reviewed by the human operators once they're online.
As time goes on, if confidence in the system grows, we can increase the AI's permissions and start using it in any operation.
We'll also need a dashboard to manage all these operations, including monitoring the model, fine-tuning it, and modifying the data or APIs it has access to. It will give you full control over the AI + the AI's pending actions to be reviewed by a human.
For the connection to APIs like Facebook, Gmail, etc, we'll centralize and abstract it in such a way that we won't need to do any changes in the model to add new APIs.
I have a bunch of ideas for what we can do. Send me a message, and let's discuss this!