Business Case
A telecom company wants a support chatbot
The story: A phone company's call center is overloaded, and the online ticket queue keeps growing. The manager wants a chat assistant on the website that can handle most questions, with two clear targets: far fewer tickets, and customers who leave happy.
In AI/AWS terms: AnyCompany, a telecom provider, wants a generative AI support chatbot to cut the cost of phone support and the load of online tickets.
- Goals: reduce online tickets by 70% and reach a customer satisfaction score of at least 4 out of 5, measured by a survey after each issue.
- Model: an LLM, because the chatbot needs to understand natural language.
For the exam: Goals are measurable: 70% fewer tickets, satisfaction of at least 4 out of 5.
The problem, and two fixes
The story: The new assistant is smart and chatty, but it has never worked at this company. It doesn't know the plans, prices, or common problems. So the manager does two things. First, they give the assistant a binder of past chats, tickets, and call transcripts, with customers' names and numbers blacked out and the mess cleaned up. Second, they give the assistant a login to the customer system so it can actually do things, like order a new phone or upgrade a plan, instead of only talking about them.
In AI/AWS terms:
- Problem: LLMs are trained on public data and don't know AnyCompany's services.
- Fix, part 1: give the model company knowledge from chat logs, past tickets, and call recordings. That data must be collected, anonymized, and cleansed into a knowledge base. This is RAG (the binder).
- Fix, part 2: let the chatbot take actions in customer accounts, such as ordering a phone or upgrading a plan. This is agents (the login).
For the exam: Company knowledge the model lacks → RAG. Taking actions in systems → agents.