ChatGPT for Customer Service: 3 Ways a Small Shop Can Use It, the Limits and Privacy Notes (2026)

The answer up front: what each of the three levels can do

"Using ChatGPT for customer service" means three quite different things.

What all three lack is really the same list; there is a section on it below.

Which level do you need right now?

Level 1: ChatGPT as a drafting assistant

You can do this today. A customer sends a message, you paste it into ChatGPT, it writes a reply from the shop details you supply, you read and edit it, and you paste it back into LINE (the messaging app most Taiwan shops use) or Instagram yourself. You are still the one replying; it is a colleague who types faster.

Plans and cost

This is what the official ChatGPT pricing page showed for individual plans when opened from Taiwan today. Amounts are the monthly prices displayed that day; whether tax is included and what you are actually charged is settled at checkout.

There are also Business and Enterprise plans; the page's FAQ states Business starts at 2 users. If you only want drafts, try Free for a week before deciding to pay. For each plan's usage limits, read the official page directly; we do not restate them.

A drafting prompt you can paste

You are the customer service drafting assistant for "(shop name)". Write replies using only the "Shop details" below.

Requirements:
1. Reply in Traditional Chinese with Taiwan wording, friendly, three sentences at most.
2. If something is not in the shop details, do not guess. Write 「這部分我確認後回覆您」 instead.
3. Never promise discounts, refunds, delivery dates or any exception.
4. If the customer left out a field needed for a booking, ask only for that one field.
5. After the reply, on a new line starting with [Note], tell me whether anything in it needs me to check personally.

Shop details:
(paste your opening hours, price list, booking and return policies)

Customer's message:
(paste the question, with name, phone and address removed first)

English gloss of the Chinese line: "I will confirm this part and get back to you."

The "Shop details" section is your knowledge base. Paste the same version every time and the drafts stay consistent. How to organise it into a form an AI can use, with fill-⁠in templates for eight blocks, is in How to build an AI customer service knowledge base. Ready-⁠made reply patterns are in the customer service message templates hub.

Three habits that keep drafts steady

First, keep the shop details in one text file and paste that same file every time rather than retyping from memory. When a price changes, change the file first, then paste.

Second, paste one question from one customer at a time. Drop five messages in together and it can easily write customer A's date into customer B's reply.

Third, note what you had to change. The same mistake three times means the shop details are missing a line or the prompt is missing a requirement. Add it and you will not have to fix it again. That list is also a ready first draft of your rules, whichever level you move to later.

Where this level stops

It does not know who the customer is or what you discussed last time, and it will not answer while you sleep. Once volume grows, the copying and pasting becomes a job in itself. And fluent is not the same as correct: read every reply before sending. A wrong answer delivered instantly is just a more efficient way to cause trouble.

Level 2: a project or custom GPT for your team

Level two keeps your shop's documents in one fixed place so colleagues do not paste them every time.

What OpenAI says about Projects

According to OpenAI's help article, a project keeps related chats, files and instructions together, and ChatGPT uses that context when it answers. You can upload PDFs, spreadsheets, documents and images, or paste text, and write instructions that apply only to that project.

Projects can be shared. The article states that sharing is available on Free, Go, Plus, Pro, Business, Enterprise and Edu, and that invitees get either chat or edit access. The limits it lists:

So a three-⁠person shop can create one project, add the price list and return policy, and invite two colleagues.

Keep just three things in the project: one plain-⁠text file of shop knowledge, one list of "do not answer, ask the manager" topics, and the project instructions. For the instructions you can reuse the drafting prompt above without the "Customer's message" line. File limits are low, so rather than uploading ten scattered old documents, tidy them into one clean file first. When a price changes, replace the file in the project too, or colleagues will keep getting the old answer.

Custom GPTs: check what changed in 2026 before deciding

Many tutorials online still tell you to "build a customer service GPT". Here is what OpenAI's help article says now:

• Personal accounts (Free, Go, Plus, Pro) can no longer create or publish new GPTs; existing GPTs remain usable.
• Business, Enterprise and Edu workspaces can still create, edit and publish them when workspace settings and permissions allow.
• OpenAI says it plans to retire custom GPTs and recommends moving workflows to Plugins. Retirement for affected Enterprise workspaces is planned for December 11, 2026; other plans are expected to follow the same timeline, subject to the notices your account receives.

The design of a GPT matches how customer service knowledge should be organised anyway: instructions decide how it behaves, knowledge files provide reference material. The article states a GPT can have up to 20 files attached, each up to 512 MB, and recommends text-⁠forward knowledge files. But with a retirement plan already announced, we would not advise a small shop to start building its customer service on custom GPTs now. The pricing page still lists custom GPTs among plan features; where the two pages differ, go by what your own account shows.

Why this level is for internal use only

A project is for collaborators who are signed in to ChatGPT. The help article states that when a share link is set to "Anyone with a link", that means any logged-⁠in ChatGPT user, and that people who join can see the project's chats, files and instructions. It is not a support widget you put on your website for customers, and it will not appear in your LINE account.

So level two is an internal cheat sheet: a new colleague asks it when unsure of a policy, then answers the customer in their own words.

Level 3: connecting to LINE yourself through the API

To have a model actually reply to customers inside a channel, you do not use the ChatGPT product. You use OpenAI's API plus your own code or an automation tool such as Make or n8n. The API and a ChatGPT subscription are separate things; for cost, see OpenAI's official API pricing page.

On the LINE side, we re-⁠checked the developer documentation today. The key points:

• Verify the signature before processing webhook events, to confirm the request came from the LINE Platform.
• LINE recommends processing events asynchronously.
• The same event may be sent more than once; deduplicate with webhookEventId.
• Events may arrive in a different order from the one they happened in; compare timestamp.
• Content sent as a reply message does not count toward the monthly message quota; push messages, broadcasts and similar do.
• Make has a LINE app that Make maintains. n8n's documentation marks its built-⁠in LINE node as deprecated from version 1.64.0, because the LINE Notify service it relied on ended on April 1, 2025; you receive events with a Webhook node instead.

Instagram and Facebook messaging go through Meta's own official interfaces. See Meta's developer documentation for the requirements; we did not verify them for this article, so we say no more.

The API only does "text in, text out". Who the customer is, what was said before, and when to stop so a human can take over are all parts you build. The full checklist and a requirements template you can hand to an engineer are in LINE AI auto-⁠reply: native features, tools or DIY; we do not repeat them here.

Five things all three levels lack

This table is the difference between a general chat tool and a customer service system. The first is very good at writing sentences. The second does the things that are not sentences.

Human handover deserves a special mention. Customer service rarely goes wrong because one sentence was poorly written. It goes wrong because nothing stopped when it should have: the customer is already angry and the auto-⁠reply is still saying thank you, or the manager has started replying and the bot cuts in. Levels one and two do not carry this risk because a person handles every message. At level three, once the model replies to customers directly, do not go live until this part works.

Logs are similar. While you are drafting, the record sits in your personal ChatGPT account, separate from the conversation the customer sees in LINE. If you ever need to check what was actually said, the message sent in the channel is the one that counts.

Before you paste a customer's message: privacy notes

What OpenAI says

The points below summarise what OpenAI's help and API documentation showed today. Setting names are given as they appear on screen.

• Services for individuals, such as ChatGPT: OpenAI states it may use your content to train its models, and you can opt out.
• How to opt out: in ChatGPT, go to Settings, then Data controls, and turn off Improve the model for everyone; or make the request in OpenAI's Privacy Portal. After you opt out, new conversations are not used to improve the models.
• Even after opting out, if you give a thumbs up or thumbs down on a response, OpenAI says the entire conversation tied to that feedback may be used for training.
• Temporary chats are not used to improve models while they remain temporary, and they do not appear in chat history.
• ChatGPT Business, Enterprise, Edu and the API: OpenAI states that by default it does not use inputs or outputs to improve its models.
• The API also generates abuse monitoring logs, which the documentation says are retained for up to 30 days by default.

These are OpenAI's statements, and they change. Before you go live, open the settings page yourself rather than relying on anyone's screenshot, this article included.

Remove these before pasting

Before pasting into any chat tool, remove:
- Name, nickname, account ID
- Phone, email, address
- Order number, member number
- National ID number, date of birth, payment details
- Medical records, health conditions, treatment history
Rewrite as: "A customer asks: ..."

Customers give their data to your shop, not to the tools you use, and the responsibility stays with you. The articles of Taiwan's Personal Data Protection Act (個資法) most relevant to customer service, and seven questions you can put to any vendor, are in 9 questions to ask before adopting AI customer service.

What changed in 2026

• Custom GPTs are being retired: personal accounts can no longer create new ones, OpenAI recommends Plugins instead, and affected Enterprise workspaces are scheduled for retirement on December 11, 2026. Before following an old "build a customer service GPT" tutorial, check whether your account can still do it.
• n8n's built-⁠in LINE node is deprecated; old tutorials for connecting LINE will get stuck.
• ChatGPT plan names and contents change quickly. This article records today's page only; the official pricing page is the reference.

When to move to a dedicated tool, and when not to

FRIDAI is one such dedicated tool. Its website says it connects direct messages on channels including LINE, Instagram and Facebook, replies from your knowledge base, notifies a human when needed, and keeps the AI out of the way once a human takes over. It has a monthly fee and no free plan. If your volume cannot yet justify that, start with Free AI customer service in 2026; to compare vendors, see AI customer service platforms compared.

FAQ