Set up an AI agent that answers your support email — in 7 steps

Written by Daan · Commercial & Creative

Builds and runs these systems at Shopgentic. Updated July 2026.

Go through your sent folder and copy the 20 support replies you are most proud of, across refund requests, delivery questions, product questions and complaints. Paste them into one document. This is the raw material for your tone, so pick the ones that sound like you on a good day. Check: the document covers at least 4 different ticket types.

Make a project called "Support replies". Add three things as project knowledge: the reply document from step 1, your returns and shipping policy, and a short list of facts agents get wrong about your store (delivery times per region, what is and isn't in stock permanently, warranty terms). Check: ask it "what is our return window?" and it answers from your policy, not a generic one.

Paste this into the project's custom instructions, then adjust the bracketed parts: You draft support replies for [store name]. Match the tone of the example replies in your knowledge: [two words, e.g. warm and direct]. Rules: - Answer from our policy documents. If the policy doesn't cover it, say "NEEDS HUMAN" and explain why. - For refund requests, always propose the best alternative first (exchange, credit) with the refund as fallback. - Never promise delivery dates we can't verify. - Keep replies under 120 words unless the customer's problem needs more. - Write in the customer's language. Check: a made-up edge case ("my dog ate the invoice") comes back flagged NEEDS HUMAN.

Paste in 10 tickets you already answered, one at a time, and compare its draft to what you actually sent. Where it misses, don't fix the draft, fix the instructions or the knowledge. Every correction here saves you a hundred later. Check: 7 of 10 drafts are sendable with light edits or none.

The single biggest quality jump: paste the customer's order info with the ticket. Use this shape every time: TICKET: [paste the customer email] ORDER: [order number, items, order date, shipping status, total] HISTORY: [first-time customer / repeat, any past issues] Draft the reply. Check: drafts now reference the actual order instead of asking the customer for it.

Decide now, not per ticket: anything involving money, a legal question or an angry repeat customer gets read word for word before sending. Everything else gets a skim. Put the rule in the same document as your policy so everyone on the team plays by it. Check: the rule fits in two sentences and is written where the team works.

Run support as one 15-minute batch each morning: paste, review, send. Track two numbers for two weeks, tickets per day and minutes spent. That's your before-and-after, and if you ever talk to us or anyone else about automating further, it makes you a much sharper buyer. Check: after two weeks you know your minutes-per-ticket, before and after.