Your agents can use ChatGPT for customer support work that is only about words: drafting and polishing replies, summarising threads, translating, and writing macros and help content. It can't do work that needs your order data or your rules, such as finding a parcel, approving a return or cancelling an order; for those requests, an AI agent connected to your order systems is the better tool.
The short version: drafts, not actions. Below is where that line falls for a consumer brand's orders, deliveries, returns and refunds, what customer data is safe to paste into ChatGPT (from OpenAI's own documentation, checked on 30 September 2026), and prompts your agents can copy.
In short
- "ChatGPT for customer service" means three different set-ups: ChatGPT in an agent's browser tab, a bot built on OpenAI's models that answers customers, and an AI agent connected to your order systems. They do different jobs.
- In an agent's tab, ChatGPT is good at words: replies, tone, summaries, translation, macros, help content and training role-plays.
- It doesn't know your orders or your rules. Agents should paste the facts in and never ask ChatGPT for them. Checking or changing an order needs your systems.
- On personal ChatGPT plans, OpenAI may use conversations to train its models unless the user turns that off; business plans aren't used for training by default. Give agents a company workspace and a rule about what never to paste.
Three things "ChatGPT for customer service" can mean
The phrase covers three set-ups. They differ in who talks to the customer and whether anything happens in your systems.
| ChatGPT in an agent's tab | A bot built on OpenAI's models | An AI agent connected to your systems | |
|---|---|---|---|
| Who replies to the customer | Your agent; ChatGPT helps them write | The bot | The AI agent, handing over to your agents when needed |
| What it knows | What the agent pastes in, plus any files or apps they add | Your FAQs and policy documents | Your policies, plus live order, delivery, payment and customer data |
| What it can do | Draft, rewrite, summarise, translate | Answer questions from your content | Answer, check the order, apply your rules and make the change: cancel, refund, exchange, update an address |
| What it takes | A ChatGPT plan and a policy for agents | A developer or a chatbot platform using OpenAI's API | A platform connected to your store, courier and payment systems |
This guide starts with the first column, the one your agents can use tomorrow. The second is covered in our guide to AI chatbots for customer service, the third in our guide to choosing an AI agent for customer service, and our comparison of agentic AI vs generative AI explains the difference between AI that writes and AI that acts. For every other way AI shows up in a support team, see our guide to AI customer service.
What agents can safely use ChatGPT for
ChatGPT is safe to use where the agent supplies the facts and checks the result. In those tasks, the worst outcome is a weak sentence the agent fixes before sending. For a consumer brand's team, the list is long:
- Drafting replies from checked facts. The agent looks up the order, then asks ChatGPT to turn "shipped Tuesday, held at the courier's hub, new date Friday" into a warm two-line WhatsApp message.
- Fixing tone. Softening a blunt reply, apologising without promising what the policy doesn't allow, or answering an angry review. Give it your own list of customer service phrases so the result sounds like your brand.
- Summarising. A 30-message email thread becomes five lines for the next agent: what the customer wants, what was promised, what's pending.
- Translating. Replies into Hindi, Tamil or Spanish, and customer messages back into English, with order numbers and dates left unchanged: a start on multilingual customer support before you hire for every language.
- Writing macros from your policies. Paste your returns policy and ask for canned responses for customer service for each case: window passed, item used, tags missing, refund to the original payment method.
- Drafting help content. FAQs, internal knowledge base articles and support email templates, written from your policy documents.
- Training new agents. Role-plays in which ChatGPT plays a customer whose cash-on-delivery (COD) order was marked delivered but never arrived, so new agents practise before their first shift.
- Writing procedures. For the head of customer service, turning rough notes into a customer service standard operating procedure, or a checklist agents follow before approving a replacement.
Each of these produces words that an agent reads before anything reaches a customer. OpenAI gives the same advice in its help centre article Does ChatGPT tell the truth?: use ChatGPT "as a first draft, not a final source", because it can sound confident when it's wrong.
Flowcall does the macro and help-content job inside the product. Admin Copilot drafts FAQs and canned responses from a plain-English request and shows each one as an editable draft before it applies, and the knowledge base turns the policies and guides you upload into answers the AI Agent uses.
The cost of ChatGPT in another tab
A ChatGPT reply isn't free, even when it's right. Follow one ticket: the agent copies the customer's message from the helpdesk into ChatGPT, checks the order in Shopify and the parcel on the courier's site, pastes those facts in too, copies the draft back, corrects the date it guessed, and then issues the refund in the payment gateway. ChatGPT sped up the typing and added one more tab.
To answer and act on one ticket, agents switch between tabs and systems and copy details from one to another, and every switch adds to average handle time. In our conversations with support teams, we've seen agents dictate their replies to ChatGPT and paste the tidied text into the helpdesk. The request behind it is fair, since agents want help writing; the fix is to put that help where the facts already are. Copying also risks pasting the wrong customer's order, and it sends customer data to another service.
In Flowcall's AI Helpdesk, everything an agent needs to answer and act on a case is on the ticket. Customer 360 shows the customer's orders, delivery, payments and refunds; Agent Assist drafts the reply, summarises the ticket and fills the disposition; and a ticket action makes the refund, replacement or order change, with the form already filled from the conversation and your connected systems. The agent checks it and clicks once, without opening Shopify or the courier's site.
What to paste into ChatGPT, and what to keep out
Customer messages are full of personal data: names, phone numbers, addresses, order details, sometimes a photo of an invoice or an ID. Before agents paste any of it into ChatGPT, check which plan they're on. OpenAI's help centre articles on how your data is used to improve model performance and data controls, and its enterprise privacy page, checked on 30 September 2026, say this:
| ChatGPT plan or mode | Can it be used to train OpenAI's models? | Detail from OpenAI's documentation |
|---|---|---|
| Personal plans: Free, Go, Plus, Pro | Yes, unless the user opts out | Content may be used for training unless the user turns off "Improve the model for everyone" in Settings > Data controls. After opting out, rating a reply with a thumbs up or down means the whole conversation may still be used for training. |
| Temporary chat | No | Not used for training while it stays temporary; may be kept for up to 30 days for safety purposes. |
| ChatGPT Business, Enterprise, Edu | No, by default | OpenAI says it doesn't train on business data by default and that you own your inputs and outputs where the law allows; Enterprise and Edu also control how long data is kept. |
| OpenAI's API (for bots you build) | No, by default | The same default as the business plans. |
Training isn't the only question: the data still leaves your helpdesk, so your own privacy policy and local data protection law apply. Three rules cover the main risks:
- Company workspace only. Agents work in a ChatGPT Business or Enterprise workspace your company controls, not in personal accounts, so you manage access and conversations aren't used for training by default.
- Strip identifiers before pasting. Replace the customer's name, phone number, email, address and order number with placeholders such as [customer] and [order]. ChatGPT doesn't need them to write a good reply.
- Never paste payment details, ID documents or passwords. No reply needs them.
Write these into your SOP before agents get access. For the wider risks, see our guide to the disadvantages of AI in customer service.
What ChatGPT can't do without your order systems and rules
A consumer brand's queue is also full of requests that turn on facts and decisions: where is my order, can I return this, please cancel it, where's my refund. ChatGPT in a tab has none of those facts and none of your rules.
| Request | What ChatGPT can do on its own | What the request actually needs |
|---|---|---|
| "Where is my order?" | Write a polite holding reply; any date it gives is a guess | Live tracking from the courier and the order's status in your store |
| "Can I return this?" | Explain the policy you paste in | The delivery date, the product's return rule, its condition and the customer's past claims, checked against your policy |
| "Please cancel my order" | Draft a reply saying it's cancelled | A check that it hasn't shipped, the cancellation in Shopify or your order system, and a refund if it was prepaid |
| "Where's my refund?" | Apologise and explain the refund policy you paste in | The return's status at the warehouse and the refund's status with the payment provider |
| "It arrived damaged" | Write an empathetic reply | The photos checked, your replacement rule applied and the replacement created |
| A complaint that needs an exception | Draft an apology | An agent with the authority to decide |
Three things are missing, and better prompts don't supply them.
- The facts. ChatGPT sees only what someone gives it. OpenAI's connected apps link it to services such as Google Drive and Slack, and some can take supported actions, but reaching your store, courier and payment provider, and matching each message to the right customer, means building those connections yourself.
- The rules. A prompt guides what ChatGPT writes; it enforces nothing. A return window or a refund limit has to be a check that runs the same way every time against the order data.
- The action. The order cancelled, the refund started, the ticket updated: with ChatGPT in a tab, all of that stays with the agent.
The same gap explains why a bot built only on your documents answers "what is your return window?" but not "can I return this?" for a particular order. Answering and completing are different jobs, the subject of our guide to customer service automation. Our pages on "where is my order" questions and returns management cover two of these request families in depth.
One WhatsApp message, two ways
At 10:40 pm, a customer messages a fashion brand on WhatsApp customer support: "Please cancel my order, I ordered the dress in the wrong colour." She placed the order two hours earlier and paid online. The warehouse's first pickup is at 8 am.
With ChatGPT in a tab. Nobody is on shift, so the message waits. At 9 am an agent pastes it into ChatGPT, which drafts: "I've cancelled your order, and your refund will reach you within 5–7 business days." None of that is true. In Shopify, the order shipped at 8 am, so it can no longer be cancelled; she'll have to refuse it or return it. The agent rewrites the reply. Had a rushed agent sent the draft as written, she would also have been promised a refund nobody had started.
With an AI agent connected to Shopify and your rules. The AI agent answers that night. It finds her order from her WhatsApp number, checks that it hasn't been dispatched, and applies your rule: cancellations are allowed until dispatch. It cancels the order in Shopify, starts the refund and tells her when to expect it, from your policy rather than a guess. (Had she chosen cash on delivery, there would be nothing to refund.) Because she wanted a different colour, it shows her the same dress in navy in her size, with a link to buy. Had the order shipped, the rule would have sent her down your return path; had anything fallen outside the rules, the conversation would have gone to an agent with the order and the reason attached.
The first version fails for two reasons, and neither is the writing: ChatGPT in a tab only helps when an agent is on shift, and the reply needed three facts (dispatch status, payment type, refund timeline) and one action, which a writing tool doesn't have.
This is how Flowcall's AI Agent handles it. It runs on the SOP Engine: the AI follows your cancellation SOP step by step while talking naturally, your rules make each decision, and the order, delivery and payment details come from Shopify and your other systems. Once a workflow is set up correctly, it gives the same result every time and doesn't make up a refund date. The same AI Agent works as an AI shopping assistant, finding the right product in your catalogue in the same conversation, and when an agent is needed it creates a ticket with the conversation, a summary, the order and what's left to do.
Prompts your agents can copy
These prompts are for the words-only work above. Square brackets hold the facts the agent has checked, and each prompt stops ChatGPT adding facts of its own. Remove names, phone numbers and addresses before pasting a customer's message.
Reply to a delayed delivery (WhatsApp)
Write a WhatsApp reply to a customer asking where her order is. Facts: ordered [date]; shipped [date] with [courier]; tracking shows [status]; new expected delivery [date]. Warm and direct, no emojis, under 60 words. Apologise once. Use only these facts. Don't add dates, reasons or offers that aren't listed.
Answer a damaged-item complaint (email)
Write an email reply to a customer whose [product] arrived damaged. What we're doing: [replacement or refund], by [date], and what she needs to do: [nothing / keep the item / send a photo]. Acknowledge the problem in the first sentence. Don't blame the courier. Under 120 words. Make no promises beyond the facts above. Customer's message: [paste, with name and contact details removed]
Say no to a return outside the policy
A customer wants to return [product] [number] days after delivery. Our policy: [paste the rule]. What we can offer: [exchange / store credit / nothing]. Write a kind, firm reply for [WhatsApp / email] that explains the policy in one sentence and makes the offer. Don't suggest any exception that isn't listed.
Summarise a long thread for the next agent
Summarise this conversation for the agent taking over, in five bullets: what the customer wants; what we've already promised, quoted word for word; what's pending and with whom; any deadline; how many times she has contacted us. Don't add anything that isn't in the thread. [paste thread, with personal details removed]
Translate a reply
Translate this reply into [Hindi / Tamil / Spanish] the way a customer service agent would write it on WhatsApp: natural, not formal. Keep order numbers, dates, amounts and product names exactly as they are. [paste reply]
Turn a policy into macros
Here is our returns and refunds policy: [paste]. Write a saved reply for each case: return window passed; item used or tags removed; wrong item received; refund to the original payment method; refund for a cash-on-delivery order by bank transfer. Under 80 words each, with [placeholders] for order details. Quote the policy's numbers exactly.
Turn notes into an SOP (for team leads)
Turn these notes into a step-by-step SOP for [damaged-item claims]: [paste notes]. For each step, give the check the agent makes, the system they check it in, and when they must escalate. Where the notes are unclear, flag the gap instead of filling it.
Role-play for a new agent (for team leads)
Act as a customer whose cash-on-delivery order was marked delivered but never arrived. You're frustrated and this is your second message. I'm a new agent. Stay in character and react to my replies. After eight messages, tell me what I did well and what I missed against this SOP: [paste SOP].
Keep the prompts that work in one place. OpenAI's Projects keep chats, files and instructions together and can be shared with your team, depending on your plan and workspace settings. Think twice before building a custom GPT for support: OpenAI's help centre says it plans to retire custom GPTs in favour of Plugins, starting with affected Enterprise workspaces on 11 December 2026, with other plans expected to follow.
ChatGPT, drafting in your helpdesk, or an AI agent: how to choose
| Choose… | If… |
|---|---|
| ChatGPT in a company workspace | Your agents spend much of their time writing, orders are quick to look up, and your team keeps up with the volume |
| AI drafting inside your helpdesk | Agents lose time copying between the helpdesk, the store, the courier portal and ChatGPT, and you want drafts written from the ticket's own order data |
| An AI agent connected to your systems | A high volume of order, delivery, return, refund and COD requests arrives on WhatsApp, Instagram, email and chat, including after hours, and you can write your rules down |
| Human agents, with AI assisting | The request is a complaint that needs an exception, a safety or health issue, or a high-value customer who needs reassurance |
These aren't exclusive. A brand with a busy WhatsApp queue can let an AI agent settle cancellations and refund-status questions, give its agents drafting inside the helpdesk for the exceptions, and keep ChatGPT for SOPs and training material. Complaints that need an exception stay with a human agent, under your escalation management rules.
How to roll ChatGPT out to your agents
- Pick the tasks. Start with the safe list above, and write down what ChatGPT isn't for: quoting delivery dates, approving refunds, deciding exceptions.
- Give agents a company workspace. A Business or Enterprise workspace keeps conversations out of model training by default and lets you control access. No personal accounts for work.
- Write the data rule into your SOP. Say what gets replaced with placeholders and what never gets pasted.
- Share one prompt library. Put your policies, brand voice notes and these prompts in one shared project, so every agent starts from the same material.
- Make the agent own the send. Every reply is the agent's, whoever drafted it. Before sending, they check each date, amount and promise against the order and the policy.
- Review a sample every week. Look for invented dates, promises outside policy, and replies that don't sound like your brand. Our guide to customer service quality assurance covers how to run the review.
- Measure the change. Compare first response time, CSAT and repeat contacts about the same order before and after. If repeat contacts rise, drafts are promising things that don't happen.
A weekly sample catches only what someone reads. Flowcall's AI Auto QA reviews your agents' replies across every conversation, or the ones you choose, such as those with low CSAT. It flags inconsistent information, questions the customer had already answered and unprofessional language, and checks replies against your own standards for tone, accuracy and policy. Each finding quotes the conversation, and because it works on Flowcall or on top of the helpdesk you already use, you can check ChatGPT-drafted replies without changing tools.
How Flowcall fits
Flowcall combines an AI Agent with an AI Helpdesk for the work ChatGPT can't do from a tab. The AI Agent follows your SOPs on WhatsApp, Instagram, Messenger, email and Live Chat, checks orders in Shopify and your other systems, completes cancellations, refunds, exchanges and replacements when your rules allow, and recommends products from your catalogue. Requests that need judgement reach your agents as tickets with the order, the conversation and what's left to do, where Agent Assist drafts the reply and a one-click action, with the details already filled in, makes the change. AI Auto QA then checks the replies, whoever wrote them.
If you'd like to see an AI Agent cancel an order before dispatch on WhatsApp, and an agent settle the exceptions from the ticket, book a demo.
Frequently Asked Questions
Can I use ChatGPT for customer service?
Yes, as a writing assistant for your agents: drafting replies from facts they've checked, fixing tone, summarising, translating and writing macros. Answering customers directly needs a bot built on OpenAI's models or a platform that provides one, and acting on orders needs a connection to your store and payment systems under your rules. Use a company workspace and keep personal data out of prompts.
How do I use AI for customer support?
Start from the requests, not the tool. Words-only work suits a general assistant such as ChatGPT; questions about your policies suit an AI chatbot; checking and changing orders needs an AI agent connected to your systems, with human agents taking the exceptions.
What should you not tell ChatGPT?
Never paste customers' payment details, ID documents or passwords. On personal plans, also keep out names, phone numbers, addresses and order details: OpenAI may use personal-plan conversations for training unless the user opts out, while business plans aren't used for training by default. Placeholders work just as well for writing a reply.
Can AI replace customer support agents?
ChatGPT can't: it writes, but it doesn't check orders, apply your rules or own an exception. An AI agent connected to your systems can settle routine order requests, which changes what your agents spend their day on rather than removing the need for them. More in will AI replace customer service.
How do I contact ChatGPT's own customer support?
That's OpenAI's own support, a different topic. OpenAI's help centre (help.openai.com) has an article, "How can I contact support?", that lists the options.
