This matches what I see with small support desks: the win is retrieval and timing, not clever replies. One practical way to start this week: export your last 200 tickets and ask ChatGPT or Claude to cluster them into the top 10 question types, with a count for each. Usually 3 or 4 types (where is my order, return window, sizing, address change) cover more than half the volume.

Then for each of those, write one approved answer template with placeholders, and give the model this prompt: "Using only the policy text below and the order data provided, draft a reply. If anything is missing, list what to ask the customer instead of guessing." That last line cuts invented policies a lot.

Keep a human click before sending for refunds and anything angry, and review ten drafts a week to tweak the templates. Thanks for keeping it grounded for three-person shops.