_This_ is exact the place where a LLM could step in and could really help:
1. Customer describes the problem
2. You get the problem, parsed by the model, including a summary and a suggestion what could help.
3. You write "do that 3 things"
4. The model tells the customer to do the 1st, then the 2nd and so on
5. The model reminds the customer in intervals that things are not completed
6. the model notifies you after the 3 things are done and the issue is still there.
You don't need an expensive, opaque, antisocial system to run interference against your users. What you need is literally just a checklist.
Give the user a checklist instead of just freeform text -- where they can check off the three individual things to try. When they've checked them all off, they can go back to the well for another response from support.
1. Customer describes the problem 2. You get the problem, parsed by the model, including a summary and a suggestion what could help. 3. You write "do that 3 things" 4. The model tells the customer to do the 1st, then the 2nd and so on 5. The model reminds the customer in intervals that things are not completed 6. the model notifies you after the 3 things are done and the issue is still there.