Tuesday, 9:40 a.m., at an online clothing store doing about US$ 90,000 a month. The shared inbox has 64 unread emails, 23 live chat transcripts from the night before, 11 Instagram DMs and one new BBB complaint titled "They took my money and disappeared." Two people handle all of it. Most of those 98 messages ask the same five things. Three of them could cost the store a chargeback. That mix is the whole case for using AI for customer service, and also the whole warning.
After watching a lot of small businesses switch on a chatbot and switch it off two weeks later, I stopped asking "should we use AI?" and started asking a different question: which messages can the AI answer alone, which ones does it draft for a person to approve, and which ones should it never touch? This guide is built around that split. You get the three boxes, the tools with US$ prices, a mini agent flow, the US rules that apply to returns, late orders, cancellations and texting, and the prompts I copy every week.
The split I use: three boxes
| Box | What goes in it | Who sends the reply |
|---|---|---|
| 1. AI answers alone | Order status with a tracking link, shipping times, sizing chart, payment methods, store hours, "do you ship to Canada?", password resets | The AI, using only your written knowledge base |
| 2. AI drafts, you review | Complaints, late orders, damaged items, public reviews, BBB replies, returns that fall outside the policy, anything with a dollar amount | A person, after editing the AI draft |
| 3. Never leaves your hands | Refunds above a limit you set, chargebacks and disputes, legal threats, injury or safety reports, a customer in visible distress, anything involving a minor | A person, from scratch |
The boxes are about what happens if the AI gets one wrong. A wrong answer about store hours costs nothing. A wrong promise about a refund is a promise your business made, and a customer who screenshots it will hold you to it.
In practice, box 1 covers 50 to 70% of the volume at a typical online store. Box 3 is small, but that is where the money sits.
What AI does well in customer service
First response and repeated questions
In e-commerce, six or seven questions make up most of the volume: where is my order, how long does shipping take, how do I return this, can I change my address, do you take PayPal, can I cancel. Once those answers are written down, the AI can reply at 2 a.m. on a Sunday in the tone of your brand. First response drops from hours to seconds, and that alone cools the second message.
Drafting replies to complaints
This is the biggest time saver I see. Answering complaints is tiring because it takes patience, and by 5 p.m. nobody has much left. The AI hands you a polite draft that names the problem and states the next step. You fix the facts, add the human sentence it could not know, and send.
Summaries and tagging
A 60-message thread that went through three agents becomes a five-line summary for whoever picks it up next. The AI can also tag tickets by reason (late, damaged, wrong size, billing, cancellation) and urgency. At the end of the month, those tags show you where the operation is bleeding. One store found 31% of its tickets came from a single carrier: a shipping problem, not a support problem.
The knowledge base
Paste 200 real conversations (with names, emails and phone numbers removed) and ask the AI to list the most frequent questions with the answer that actually solved them. In one afternoon you have the FAQ your chatbot will use. Without a written knowledge base, no chatbot works. It will improvise, and that is how a bot promises free returns you never offered.
Answering a BBB complaint or a 1-star review with AI
A real-world case, with details changed. A customer bought a jacket on Black Friday. The order shipped 12 days late because the warehouse mislabeled a batch. She filed a BBB complaint and left a 1-star Google review: "They took my money and disappeared." The team had answered her by email on day 5, but the message went to spam. The jacket was delivered on day 14.
A public reply is not written for the person who complained. It is written for the next 500 people who read it before buying. The prompt I used:
You are the customer care lead at [STORE], an online clothing store. Write a public reply to the complaint below, for [BBB / Google / Yelp / Trustpilot]. Facts (use only these): order placed [11/28], promised ship date [12/2], actually shipped [12/10] because of a labeling error at our warehouse, we emailed on [12/3] and the message likely went to spam, delivered [12/12]. Rules: apologize for the delay without excuses, state what happened in one sentence, say what we are offering [free return shipping or 15% store credit], invite her to reply privately to [EMAIL]. No order number, no full name, no address. Under 110 words. Do not promise anything not listed in the facts. Complaint: [PASTE]
Real output:
Hi Dana, you are right to be upset, and we are sorry. Your order should have shipped on December 2 and left our warehouse on December 10 because of a labeling error on our side. We emailed you on December 3, but it looks like that message never reached you, which made a late order feel like silence. Your jacket was delivered on December 12. As an apology, we would like to offer you free return shipping or a 15% store credit, whichever you prefer. Please reply to care@[store].com so we can set it up today. Thank you for telling us, it helps us fix this for everyone.
Two things I always change before posting: the first name stays only if the platform already shows it, and I add one line a person would write ("I checked the tracking myself this morning"). On BBB, keep the dollar offer in the private part of the complaint process when you can.
A legal point worth knowing for this exact case: the FTC's Mail, Internet, or Telephone Order Merchandise Rule says that if you do not give a shipping time, you need to ship within 30 days, and if you cannot ship on time you must tell the customer and give her the option to cancel for a full refund. The AI should know that option exists. It should not be the one deciding it.
Returns, refunds and cancellations: where AI slips the most
Returns are where AI invents the most, because it mixes up three different things:
- Changed my mind. There is no general federal right to return an online purchase in the US. Your posted return policy rules. Some states require that the policy be posted clearly, so the AI has to quote yours, not "the standard 30 days" it saw on another site.
- Defective or damaged. Different path entirely: warranties, carrier claims, photos, replacement. This goes to box 2.
- Wrong size or "not what I expected." Usually an exchange under your policy, often with a fee for return shipping. The AI should explain it, not waive it.
Two more traps. Chargebacks: a customer who disputes a charge with the card issuer has rights under the Fair Credit Billing Act, and the dispute has its own process and deadlines. The moment a customer writes "I'm calling my bank" or "chargeback," the conversation moves to box 3. Subscriptions: many states have automatic renewal laws (California's is the best known) that require easy online cancellation. If you sell subscriptions, the bot must give the cancel path, never a retention maze. A bot that makes cancelling harder is what state attorneys general look for.
My fix is simple. I write the policy in three labeled blocks (CHANGED MIND, DEFECTIVE, WRONG ITEM) and the system prompt says: "Quote the block that applies. If none applies, hand off."
AI chatbot for a small business: is it worth it?
Yes, with three conditions. Without them, I would not turn one on.
- A written knowledge base with real prices, real shipping times and the real return policy. "Ask us" is not an answer.
- A visible way to reach a person. Typing "agent" or "human" has to work, every time.
- Honest disclosure. California's bot disclosure law covers bots used to sell or influence purchases, and disclosure is simply good practice everywhere. "I'm [Store]'s virtual assistant" in the first message solves it.
Below about 30 conversations a day, I would skip the bot and use AI only for drafts.
A mini agent flow for chat, email and SMS
The smallest setup I would build for 50 to 300 conversations a day, wired with n8n or Make on top of your helpdesk:
- Intake. A message arrives (site chat, email, Instagram DM or SMS). A keyword filter runs first: chargeback, lawyer, sue, injury, allergic reaction, fire, "my child." Any match skips the AI and alerts a person.
- Classify. The AI tags the reason and picks a box (1, 2 or 3). It returns only the tag, not a reply.
- Box 1. The AI looks up the order in Shopify (or your platform) by email or order number, answers using the knowledge base, and closes with "Anything else?"
- Box 2. The AI writes a draft and parks it in the helpdesk as an internal note. A person edits and sends.
- Box 3. No draft. The conversation goes to the lead with a five-line summary.
On SMS, one extra rule: the TCPA requires prior consent for marketing texts, and carriers expect business numbers to be registered (A2P 10DLC). Replying to a customer who texted you first is fine. Sending them a coupon afterwards needs an opt-in, and every text needs a working STOP.
Personal data and privacy in support
California's CCPA (as amended by the CPRA) and a growing list of state privacy laws give customers rights over the data support touches all day, and your AI vendor counts as someone you share it with. At the desk:
- Never paste card numbers or full addresses into a consumer chat app. Use the order number.
- Use the business plan or API of your AI vendor, where data is not used for training by default, and check the vendor's data terms.
- Strip names and contact details before pasting conversations to build the knowledge base.
AI customer service tools and prices
| Tool | Best for | Price (check the site) | My take |
|---|---|---|---|
| ChatGPT or Claude | Drafts, knowledge base, monthly tagging | Free, or about US$ 20/month | Start here before buying anything |
| Shopify Inbox | Chat for Shopify stores, AI suggested replies | Free | Good first step if you are on Shopify |
| Gorgias | E-commerce helpdesk with order data and AI agent | From about US$ 10/month, AI priced per resolution | The one I would pick for a Shopify store above 50 tickets a day |
| Zendesk | Larger teams, many channels | From about US$ 19 per agent/month, AI add-ons extra | Excellent, but the bill grows with every seat |
| Intercom with Fin | AI agent that answers from your help center | Per seat plus about US$ 0.99 per AI resolution | Very good, watch the per-resolution cost at high volume |
| Freshdesk | Ticketing on a budget | Free plan, paid from about US$ 15 per agent/month | Solid choice for service businesses |
| Tidio | Chat plus the Lyro AI bot for small stores | Free plan, paid from about US$ 29/month | Fastest to set up, less control over the rules |
My pick for a small store: ChatGPT or Claude for drafts and the knowledge base for the first month, then Gorgias (on Shopify) or Freshdesk (everyone else) once the volume justifies it. Per-seat plus per-resolution pricing hurts below US$ 50,000 a month in sales.
Ready-to-use prompts for customer service
Besides the BBB reply above, these are the ones I use most.
1. Build the knowledge base from real conversations
Below are [200] real support conversations from [STORE], with names and contact details removed. List the 15 most frequent questions. For each one: the question in the customer's own words, how many times it appeared, the answer that actually solved it, and whether the answer depends on information I need to confirm (mark [CONFIRM]). Do not invent policies. Conversations: [PASTE]
2. Reply to an angry customer without sounding like a bot
Draft an email reply to this customer. Facts: [WHAT HAPPENED]. What we can offer: [OPTIONS]. Rules: first sentence acknowledges the specific problem, no "we apologize for any inconvenience," no "valued customer," one clear next step with a date, under 120 words, plain American English. Message: [PASTE]
Real output (late order, customer wrote "this is the worst store I've ever bought from"):
Hi Marcus, your order was supposed to arrive Friday and it still hasn't, so I get why you're frustrated. I checked: the package has been sitting at the carrier's Memphis hub since Wednesday. I've opened a trace with them today. If it hasn't moved by Thursday at noon, I'll ship a replacement by express at no cost, and you can keep or refuse the first one when it shows up. I'll email you Thursday either way. Sorry this one went sideways. Jess, [Store] Customer Care
3. Tag the month's tickets
Tag each ticket below with one reason (late, damaged, wrong item, wrong size, billing, cancellation, product question, other) and urgency (low, medium, high). Then give me a table: reason, count, percentage, and one sentence on the likely root cause. Tickets: [PASTE CSV]
4. Say no without losing the customer
A customer is asking for [a refund 45 days after delivery; our policy is 30 days]. Write a reply that says no clearly in the first two sentences, explains the policy without legal jargon, and offers [store credit or an exchange] as an alternative. Warm, not apologetic for the policy itself. Under 100 words.
Real output:
Hi Priya, thanks for reaching out. I can't issue a refund on this order because it was delivered 45 days ago and our return window is 30 days from delivery. What I can do is give you an exchange for any size or color, or store credit for the full US$ 68, good for a year. Just reply with which one you'd like and I'll set it up today.
If you also run a restaurant or a food business, the review and ordering side is covered in AI for restaurants. And if most of your support lives in Instagram DMs, AI for social media covers keyword automations in the inbox.
Common mistakes when putting AI in customer service
- Launching the bot before the knowledge base. It will make things up within the first hour.
- Hiding the human option. Customers who cannot reach a person go to the BBB, the state attorney general's office or their bank. All three cost more than an agent.
- Letting the AI quote refund amounts. Dollar figures belong in box 2 at least.
- Measuring only first response time. Track reopen rate too. A fast wrong answer comes back twice.
Where to start this week
Monday: export 100 to 200 conversations from your inbox or helpdesk and build the knowledge base with prompt 1. Tuesday: rewrite the return policy in the three blocks (changed mind, defective, wrong item). Wednesday on: use prompts 2 and 5 for every complaint and track how long each reply takes. After two weeks, check the share of box 1 tickets. Above half, with 30+ conversations a day, it is time for an agent.

Want to skip this part?
Ready-made AI agents for customer service: nothing to build, no prompts to write, they already work for you.
- Answers order, shipping, return and payment questions around the clock, using your policy
- Leaves a draft ready for every complaint and public review, waiting for your approval
- Hands the case to your team with a summary when the customer needs a person
Step by step
- Export 100 to 200 real conversations and remove names and contact details
- Build the knowledge base with the 15 most frequent questions and the answers that solved them
- Write the return policy in three blocks: changed mind, defective, wrong item
- Sort message types into the three boxes and set a refund limit for box 3
- Use AI drafts for complaints for two weeks and measure reply time
- Turn on a chatbot only for box 1, with a visible way to reach a person
Frequently asked questions
What is the best AI for customer service?
For a small team, ChatGPT or Claude for drafts plus your helpdesk's built-in AI. Gorgias fits Shopify stores, Freshdesk fits service businesses, and Zendesk or Intercom Fin fit larger teams with a budget per seat and per resolution.
Can AI answer customers on its own?
Yes, for repeated questions with a written answer: order status, shipping times, sizing, payment methods. Complaints, refunds and anything involving money or legal risk should be drafted by AI and sent by a person.
Do I have to tell customers they are talking to a bot?
California's bot disclosure law requires it for bots used to sell or influence a purchase, and disclosing is good practice everywhere. One line in the first message is enough.
Can I use AI to reply to BBB complaints and Google reviews?
Yes, as a draft. Give it only the facts, forbid promises not on the list, remove personal details, and read it before posting, because the reply is public and speaks for your business.
Can my AI chatbot send marketing texts?
Only to customers who gave prior consent, under the TCPA, from a registered business number, and with a working STOP option. Replying to a customer who texted you first is a different case.
