How-To: Use AI for Customer Service in a Small Business
You don't need a chatbot platform to start benefiting from AI in customer service. A few well-chosen prompts can save hours of drafting, summarising, and sorting — while keeping a human in the loop the whole way.
Why this matters
Most small-business customer service work is repetitive in shape — a complaint about a late order, a question about a refund policy, a request for product help — but different in detail. AI is excellent at this: drafting responses, summarising complaints, categorising feedback, and suggesting solutions. Used as an assistant (not a replacement), it frees your team to handle the genuinely hard conversations that need human judgement.
Step 1. — Draft customer responses
Paste a customer message into ChatGPT or Claude and ask for a draft. A reliable starting prompt:
Draft a professional, empathetic response to this customer complaint. Offer a [refund / replacement / apology]. Keep it under 150 words. Do not promise anything beyond what I've listed. Sign off as "Customer Care Team". [Paste customer message here]
Edit the result before sending — AI drafts are starting points, not finished replies. The prompt constrains tone, length, and promised remedy, which keeps drafts consistent and on-brand.
Step 2. — Build a library of response templates
Most enquiries fall into a small number of recurring scenarios. Build a library of AI prompts for each, so anyone on your team gets a high-quality draft without re-inventing the prompt every time:
- •Refund request — empathetic, confirms policy, advises on timeline.
- •Shipping delay — apologises, gives the current status, sets expectations.
- •Product question — answers from a knowledge base passage you paste in, no invented specs.
- •Complaint escalation — defuses, routes to a manager, promises follow-up by a deadline.
Keep these in a shared document so the whole team uses the same wording and the same AI behaviour.
Step 3. — Summarise and categorise feedback
Paste 50 customer emails (with PII removed — see Step 6) and ask the AI to sort them:
Categorise the following customer messages by issue type (shipping, product quality, billing, feature request, other) and by sentiment (positive, neutral, negative). Output as a table with columns: Message ID, Issue Type, Sentiment, One-line summary. Do not quote the originals verbatim.
In minutes you get a structured view of what's actually going wrong and how customers feel about it — the kind of analysis that would otherwise take an afternoon. Use the output to spot patterns and prioritise fixes.
Step 4. — Generate FAQ entries
Your past customer emails are a goldmine of real questions. Feed batches of past enquiries to the AI and ask it to draft FAQ entries:
Here are 30 customer questions about [product]. Identify the most common question themes and write a FAQ entry for each. Format: "Q: [question] / A: [answer, plain language, under 60 words]". Flag any question you don't have enough information to answer; do not invent.
Review, correct, and publish to your help centre. Refresh quarterly — new questions surface as your product and customer base change.
Step 5. — Consider an AI chatbot (when you're ready)
Once your prompts and FAQ are stable, the next step is putting them in front of customers automatically. Options span a range of budgets and control:
- •Intercom — full-featured, AI-assisted helpdesk with chatbot builder.
- •Chatwoot — open-source, self-hostable, lower cost if you have technical capacity.
- •Self-hosted options — for those who want full control of customer data and the AI model.
Whatever you choose, don't replace humans entirely. AI chatbots are most useful for deflecting the easy 60-70% of questions and routing the rest, with context, to a person.
Step 6. — Privacy and best practices
- •Don't paste customer PII (names, emails, order numbers, addresses) into public AI tools if you can avoid it. Anonymise before sending.
- •Always review AI drafts before sending. AI can sound confident and still be wrong about your own policy.
- •Maintain human tone. Edit stiff or generic phrasing; customers can usually tell when a reply is auto-generated.
- •Don't over-automate. Upset customers, complex cases, and anything involving money or safety should reach a human fast.
Quick tips
- •Keep response templates in a shared doc that the whole customer-care team can find and edit.
- •Review AI output for tone before sending — too chirpy on a complaint reads as dismissive.
- •Train the team on the prompt patterns so everyone produces similar-quality drafts.
- •Escalate sensitive cases — legal threats, safety, vulnerable customers — to humans immediately, every time.
Need More Help?
StarCaller Academy offers 1-to-1 sessions to help you with any of these topics and more.