Customer messages are the tax you pay for selling online. Ten a day feels fine; sixty a day across Amazon, eBay, Shopify and TikTok Shop quietly eats every hour you were going to spend growing the business. The instinct is to hire someone. The cheaper move is to build a system that answers the boring 80% automatically and only escalates the messages that genuinely need you — and to do it in a way that doesn't read like a chatbot wrote it.
First, Audit What People Actually Ask
Before automating anything, export the last 200 customer messages from your busiest channel. Almost every seller discovers the same thing: five or six question types account for the overwhelming majority of volume. Where is my order. Can I return this. Does it fit. Is it compatible with X. When will it be back in stock. Paste the whole export into Claude or ChatGPT and ask it to cluster them. You'll get a ranked list of what to automate first, based on your real data rather than a guess.
Write a Reply Bank in Your Own Voice
This is the step most people skip, and it's why automated support sounds robotic. Don't ask AI to invent replies from nothing — feed it ten of your own past responses, the ones you were happy with, and tell it to learn the voice before drafting anything. Short sentences, no corporate padding, an actual apology when something went wrong. Then have it write one reply per question category in that voice, plus two variations of each so the same customer never sees identical wording twice.
Automate the Routing, Not Just the Writing
A reply bank saves you typing. A routing layer saves you reading. Connect your inbox to Make.com or Zapier and put an AI step in the middle: every incoming message gets classified into one of your categories plus an urgency score, then either gets an auto-reply, gets drafted for one-click approval, or gets flagged straight to you. Anything mentioning a refund dispute, a safety issue, or an A-to-Z claim should skip automation entirely and land at the top of your list. That's the whole design — machines handle volume, you handle risk.
Use Draft Mode Before You Use Send Mode
Run the system in draft-only mode for the first two weeks. Every reply gets written automatically, but nothing sends until you approve it. You'll catch the failure modes fast: the tone that's slightly off, the policy it invented, the sarcastic message it read as sincere. Fix those in the prompt, not in each reply. Once you're approving 90% of drafts without edits, switch the safest one or two categories — usually shipping status and basic product questions — to fully automatic and leave the rest in draft.
Feed It Your Policies, Not Just Your Products
Most bad AI support replies come from missing context, not bad writing. Build one plain-text document containing your returns window, delivery times per region, warranty terms, restocking fees, and the exact wording of anything legally sensitive. Attach it to every request. And add one hard rule to your system prompt: if the answer isn't in the document, say you'll check and escalate — never guess. A confident wrong answer about a refund policy costs far more than a slow one.
Track Two Numbers and Nothing Else
Vanity metrics will waste your time here. Track deflection rate — the share of messages resolved without you touching them — and escalation accuracy, meaning how often the system correctly flagged something that really did need a human. If deflection is climbing and escalation accuracy stays high, the system is working. If deflection climbs while accuracy falls, you've automated your way into angry customers, and you should pull a category back into draft mode immediately.
The Bottom Line
You don't need an AI agent that handles everything. You need one that handles the five questions you answer forty times a week, in your voice, with your policies, and knows when to get out of the way. Start with the audit, build the reply bank, run it in draft mode for a fortnight, then let the safe categories fly. Most sellers get to 60–70% deflection within a month — and that's an afternoon of setup buying back several hours every single week.