Use AI to answer common, low-risk questions (order status, hours, booking, rescheduling) and to draft replies for everything else. Send upset customers, refunds, legal matters and anything new to a person, with the AI’s summary attached. Start with a person approving every reply, then loosen it for proven question types.
- AI handles the repeat questions; people handle feelings and exceptions.
- Draft first, send later: start with approval on every reply.
- Route by sentiment, value and novelty.
- A good handoff includes a summary, so customers never repeat themselves.
What customers actually want
Most customers don’t care whether a person or an AI answers “what time do you open?”. They care that the answer is fast and right. Where they do care is when something has gone wrong. That’s the split to design around: speed for the routine, people for the moments that matter.
What AI can handle, and what it can’t
What a good AI draft looks like
Rules for when a person steps in
Write simple “if this, then a person” rules. The five most useful triggers are:
Email, chat and phone
- Email: the easiest place to start. AI drafts, a person approves in batches twice a day.
- Chat: answers instantly from your FAQs and offers a handoff button that reaches a real person.
- Phone: voice AI can answer, book and take messages after hours, then send a written summary to your team.
A handoff should never make the customer repeat themselves. The AI passes the person a short summary: who it is, what they asked, what was already said, and what they want.
How to roll it out
- List your twenty most common questions and the right answer to each.
- Turn on AI drafting for email only. A person approves everything.
- After a few weeks, let proven, low-risk question types send automatically.
- Add chat, then phone, using the same rules and answers.
- Read a sample of AI replies every week and fix the instructions behind any repeat mistakes.
This is human-in-the-loop review applied to customers. For the instructions behind the drafts, use the 5-part prompt recipe.
A day in the inbox, with AI
Picture a small home-supplies store getting forty customer emails a day. By 8am, AI has sorted them: twenty-two order-status questions with drafted replies, nine booking requests matched to open slots, five product questions answered from the catalog, three complaints and one refund request flagged for a person. The customer service lead approves the thirty-six drafts in about twenty minutes, then spends the rest of the morning on the four that need real care, with summaries already written.
Nobody waits overnight for a simple answer, and the upset customer gets a thoughtful reply from a real person before lunch.
Common mistakes to avoid
Be open where it matters, and always offer an easy route to a person.
If customers have to repeat themselves after a handoff, the AI made things worse, not better.
When a price, policy or opening time changes, update the AI’s information the same day.
Acknowledging a complaint quickly is fine. Resolving it is a person’s job.
One of our builds answers calls and emails for a service business using a voice AI and a writing AI working together. The part clients notice isn’t the AI. It’s that the angry customer, the refund and the unusual request all reach a person quickly, with the context already written up. That routing is the real product.
Where to go from here
The fastest way to learn this is to try it on one real task this week. If you want a guided start, take the free course. If you’d rather talk it through, book a call.
Questions, answered
Can AI handle customer service for a small business?
It can handle common, low-risk questions and draft replies for the rest. Keep people on complaints, refunds, legal matters and anything unusual.
Will customers know they are talking to AI?
Be open about it where it matters, and always offer an easy way to reach a person. Customers mainly care that answers are fast and correct.
Should AI send customer emails automatically?
Start with AI drafting and a person approving. Move proven, low-risk question types to automatic sending only once they rarely need edits.
What is the best first channel for AI customer service?
Email. It’s easy to review drafts in batches before anything is sent.