I am working on a task where I am required to automate the customer service request channel. The process is quite typical. A customer queries about a product via email, the person on the front channel checks emails, forwards it to the relevant department and then answer is provided.
The problem is that customer query can be about one of hundreds of devices listed. Each device has its own pdf documentation which is quite extensive. Finding the right pdf and then finding the right section where information could be listed is really a tedious process and wastes a lot of time. Sometimes the information is not even listed and answer has to be improvised by product specialist (the last part hints me about reinforcement learning, what do you guys think).
What I want to achieve is that this whole tedious and repetitive process is automated and may be if possible, the model learns over time as well. The task output is quite open ended as well. Different approaches and models can be tried out (like chatbots and etc). Rapid failure is highly appreciated here.
Below mentioned are some more details:
- I have customers queries about devices in the form of emails.
- PDF documentation of devices. The documentations are quite extensive.
- I also happen to have some excel files where some sample queries and sample answers are listed. But since queries can be of very dynamic nature, it doesn't seem like a classification problem (to me at least).
I have googled quite a lot about the topic but mostly what I get are topics like 'How AI will transform the customer service' and then something more specific to NLP and a lot of company ads etc. So far what I have understood from online surfing is that possible approaches need to use NLP library (Nltk) in Python and do some topic modelling for documentation and for email. Still how I approach the whole task is not clear to me.
What I want from you guys is that maybe guide me how this task can be achieved step by step. I am not looking for any code! Just which methods can be used and how the problem can be approached. Right now, I don't know where to start and how to approach it.