Safe, helpful customer replies with AI and a human in charge

A short library of approved reply templates, a list of topics a person must always handle, and a tested draft assistant.

Best for: Small businesses answering the same questions by DM, email or chat who want faster, consistent replies without risky automation.

• Around 30 real messages you have received, with names and contact details removed.
• The true answers: prices, hours, policies, delivery and refund rules.
• Access to the tools where you reply (inbox, Instagram, Telegram, your site chat).

1. 1

   Collect and anonymise real questions

   Remove names, phone numbers, addresses and any private detail before pasting messages into an assistant. Check what your tool and your privacy policy allow before you share customer data with any service.

2. 2

   Group them into question types

   Most inboxes contain six to ten recurring types. Count how often each appears, so you build templates in order of real value, not guesswork.

3. 3

   Write the approved answer for each

   One short answer per type, built from facts you have confirmed, ending with one clear next step. Put [CONFIRM] in any slot where a fact may change, such as stock or opening hours.

4. 4

   Draw the human line

   List the topics a person must always handle: complaints, refunds, safety, medical or legal questions, anything about money owed, and anything emotional. The assistant drafts at most; a person sends.

5. 5

   Test with difficult messages

   Have an assistant play an annoyed, vague or out-of-scope customer, then check each draft reply. Look for invented promises, wrong facts and a tone that would make a bad day worse.

6. 6

   Start with drafts, not auto-send

   Begin by having the assistant suggest replies that you approve. Automate only the narrow, low-risk cases (opening hours, link requests) once they have worked reliably for weeks, and keep a visible way to reach a person.

7. 7

   Review and update monthly

   Read a sample of sent replies every month. Fix templates when policies change, add new question types, and remove anything that caused confusion.

Replace anything in [square brackets]. Prompts are starting points: results vary between tools and versions, so review every output.

Prompts are written in English because image, video and most text models follow English instructions best. To get answers in your language, add “Reply in English.” to a prompt for a text assistant.

Here are 30 real customer messages I received: [paste, with names and contact details removed]. Group them into 6 to 10 recurring question types. For each type give: how often it appears, the best short answer using only facts I will confirm, and whether a person must always handle it (refunds, complaints, medical, legal, anything about money owed).

Write a reply template for this question type: [type]. Facts you may use: [facts]. Tone: [voice]. Max 60 words. End with one clear next step. If any needed fact is missing, put [CONFIRM: what] instead of guessing. Never promise a date, refund or outcome I did not state.

Act as an unhappy customer and send me 8 difficult messages for this business: [describe]. Include a vague one, an angry one, one that asks for a discount, one outside our scope, and one with a sensitive personal detail. For each, say what a safe automated reply would be and when it must be passed to a human.

• ✕Pasting private customer details into a tool without checking the rules.
• ✕Letting a bot answer complaints, refunds or sensitive questions.
• ✕Giving the assistant outdated prices or policies.
• ✕Automating everything on day one and never reading the replies.