Answer Customers Faster with AI · Lesson 4 of 7

Draft, check, approve: the daily routine

25 minutes. The AI drafts. A person checks the facts and sends. Every time.

---

This is the lesson where you actually start answering faster. The rule that makes it safe is simple, and it doesn't bend: the AI writes the first draft, and a person reads, fixes and sends every reply. Nothing goes to a customer on its own.

That one rule protects you from the two ways AI replies go wrong: a wrong fact, and the wrong tone at the wrong moment. A person catches both in about thirty seconds.

The routine, in four moves

Rosa at our fictional Maple Street Plumbing runs this three times a day: at 8:00, after lunch, and at 4:30. Pick times that suit your day. Batching beats jumping at every ping.

1. Sort (2 minutes)

Look at what came in since last time. Put each message in one of three piles:

  • Answer bank questions (the A and D questions from Lesson 1). These go to the AI.
  • Hand-offs (the P list below). These go to the right person, today, with no AI draft.
  • Everything else. Spam, vendors, internal. Handle as usual.

2. Draft (seconds each)

Open your Customer Replies project, Gem or agent. Paste the customer's message, with names, addresses and account numbers removed. Add one line if it helps: "This is a text, keep it short" or "They've emailed twice already."

Get the draft. If you have several similar ones, you can paste three or four at once and ask for a separate draft for each.

> If your AI is connected to your inbox (see Connect Your AI to Your Business Tools), it can read the message and create a draft reply directly in your email. Same rule: it drafts, you send. Never set it to send on its own.

3. Check (30 seconds each)

This is the step that matters. Read every draft against three questions:

  • Is every fact right? Each price, time, day, area and policy. Compare to the answer bank, not to your memory. If the draft has a fact that isn't in the answer bank, treat it as wrong until you've checked it.
  • Does it promise anything it shouldn't? Arrival times, coverage, refunds, prices for unseen work. Look for the words "will," "guarantee," "covered," "free" and "today."
  • Does it answer what they actually asked? AI drafts sometimes answer a neighbor of the question. A customer who asked "can you come Saturday" doesn't want your weekday hours.

Also look for any [CHECK: ...] marks. Each one is a fact the AI didn't have. Fill it in or take that sentence out. Never send a draft with a bracket still in it.

4. Send, then fix the source (seconds)

Edit the draft until you'd sign your name to it. Then copy it into your email, text or messaging app and send it yourself.

Here's the habit that makes this better every week: if you had to fix a fact, fix the answer bank too. If the draft said $79 and the price is $89, the answer bank is where the $79 came from, or it's missing the fact. Update it, update the file in your tool, and that mistake doesn't come back.

The hand-off list: these always go to a person

Some messages should never get an AI draft. Not because the AI can't write one, but because the reply needs judgment, authority or care that belongs to a person. Your voice instruction from Lesson 3 already tells the AI to write "HAND OFF:" for these. Your team should spot them first.

  • Angry or upset customers. Read it, then call them if you can. A fast, human "I'm sorry, I'm looking into this now" is worth more than a polished paragraph.
  • Refunds, credits, discounts and billing disputes. Only someone with authority decides these.
  • Damage, injury or safety. A leak that damaged a ceiling. A customer who slipped. A gas smell (that one gets "leave the house and call the gas company's emergency line" from a person, right now).
  • Anything legal. Mentions of a lawyer, a lawsuit, a contract dispute, a complaint to an agency.
  • Anything medical or health-related. For clinics, any question about symptoms, results, medications or treatment goes to clinical staff. For everyone else, a customer's health details don't go into the chat at all (Lesson 3).
  • Privacy requests. "Delete my information." "Who has my records?"
  • Anything you're not sure about. When in doubt, a person answers.

Write your own list on the Reply Voice and Hand-Off Card, with a name next to each kind. "Refunds: Dana. Clinical questions: the nurse on duty. Everything else unclear: Rosa decides."

A note on "D" questions

For "depends" questions, the best first reply is often a question back. "Happy to price that. Is your water heater gas or electric, and about how old is it? A photo of the label helps." The answer bank's What to ask first column tells the AI what to ask. A fast, useful question is still a fast first reply, and it gets you to the real answer sooner.

Do this now

Run the routine for real, once, today:

  • Sort everything that came in since this morning.
  • Draft replies for at least five answer bank questions.
  • Check every draft with the three questions. Mark each one: sent as is, small edit, or big fix.
  • Send them yourself.
  • For every big fix, update the answer bank or the voice instruction so it doesn't happen again.

Then fill in the Hand-Off section of your card, with names.

Check it worked

  • Five real replies went out today, each read and sent by a person.
  • None of them contained an unchecked fact or a leftover [CHECK] bracket.
  • At least one fix made it back into the answer bank or the instruction.
  • Everyone who answers customers knows the hand-off list and who handles what.

---

Next: the most public replies you write. Reviews.

→ Lesson 5: Reply to online reviews

Previous: Teach it your voice, and what stays out of the chat · Next: Reply to online reviews