By SupportHQ Team · October 6, 2026 · Getting started
ChatGPT for Customer Service: What Works and What Breaks
ChatGPT for customer service can mean two different jobs. In the first job, ChatGPT helps your team write, rewrite, and summarize, and a person still sends every reply. In the second job, ChatGPT talks to your customers directly. The first job works well today. The second job breaks in five specific ways.
This post gives you six prompts to copy, one ticket worked through, the cost of an API bot, and a rule for choosing.
OpenAI details below come from chatgpt.com/pricing, the OpenAI help center, and the OpenAI API pricing page, checked on October 1, 2026.
Using ChatGPT for customer service inside your team
Set up a project first
The safest use of ChatGPT for customer support is inside your team. Start with one ChatGPT project. A project keeps files, chats, and instructions together, and its instructions apply to every chat inside it. Projects are on every personal plan, but file caps differ: 5 files per project on Free, 25 on Go and Plus, 40 on Pro and Business (source). Add three things:
- Your top 10 to 20 help articles and your refund, billing, and privacy policies. On Free, paste your top articles into one or two documents.
- A short style guide: tone, words you avoid, how you sign off.
- One instruction: “Use only the files in this project for product facts. If a fact is missing, say so.”
Decide what data you paste
On personal plans (Free, Go, Plus, Pro), OpenAI may use your conversations to train its models unless you turn off Improve the model for everyone under Settings > Data controls. On ChatGPT Business, Enterprise, and the API, OpenAI does not train on your data by default.
- Personal plan: turn off training, and remove names, emails, phone numbers, and order numbers before you paste a ticket. Temporary chats are not used for training (source). Plus costs $20 per month per person.
- ChatGPT Business: $25 per user per month, or $20 billed annually, with a minimum of 2 seats (source).
Six prompts that work
Each prompt tells ChatGPT what it may not do. That part matters more than the tone instructions. Replace the text in angle brackets.
1. Draft a reply from your policy
You draft replies for a human support agent to review. You do not send anything.
Rules:
- Use only the facts in POLICY and TICKET. If the reply needs a fact that is not there, write [CHECK: what is missing]. Do not guess.
- Do not promise a date, amount, or outcome that POLICY does not state.
- Start with the answer. One sentence of empathy at most.
- Short sentences. Under 120 words.
- End with one clear next step for the customer.
POLICY:
<paste the relevant article or policy>
TICKET:
<paste the customer message, with personal data removed>2. Rewrite the tone, keep the facts
Rewrite this reply for a customer who is upset. Keep every fact, number, date, and step exactly as written. Do not add promises. Acknowledge the problem in one sentence. Remove blame phrases such as "as stated", "you should have", and "per our policy".
Return the rewrite. Then list every change you made, one per line.
REPLY:
<paste>The change list makes review fast. You check the list, not the whole reply.
3. Summarize a thread for a teammate
Summarize this support thread for a teammate who has not read it. Use exactly these headings:
- Customer goal (one sentence)
- What we already tried
- What the customer already told us (so we do not ask again)
- Open question or blocker
- Suggested next step
Quote the customer's exact words for any error message, amount, or date. If something is unclear, write "unclear". Do not guess.
THREAD:
<paste>4. Turn a solved ticket into a help article
Turn this solved ticket into a draft help article for customers.
- Title: the question as a customer would type it.
- Summary: one sentence.
- Steps: numbered, one action per step.
- Common mistakes: only the ones that appear in this ticket.
- Remove every detail about this customer.
- Mark any step you are not sure about with [VERIFY].
TICKET AND RESOLUTION:
<paste>For a full article template, see our post on knowledge base structure.
5. Triage a batch of tickets
Classify each ticket below. Return a table with these columns:
ticket number | category | urgency | missing fact
Category: billing, bug, how-to, account access, feature request, or other.
Urgency: "high" only if the customer cannot use the product or reports money taken in error. Otherwise "normal".
Missing fact: the one thing we must ask the customer before we can help, or "none".
TICKETS:
1. <paste>
2. <paste>6. Find the gaps in your help center
Below are last week's tickets and a list of our help article titles.
Group the tickets by the underlying question, not by the words used. For each group, return: the question in one sentence, the ticket numbers in the group, the count, and whether an article on the list answers it (yes, partly, no). Sort by count, highest first.
ARTICLE TITLES:
<paste>
TICKETS:
<paste, numbered>Ask for ticket numbers, not only counts. Language models miscount long lists, and the numbers let you check each group fast. For POLICY input to prompt 1, our 25 canned responses are a good start.
Worked example: one ticket through prompt 1
This example uses a made-up company: an invoicing app with a $49 Team plan.
The policy we paste:
Duplicate charges: billing confirms that two successful charges exist for the same period. Billing approves the refund within 1 business day. The refund goes to the original payment method and shows on the statement in 5 to 10 business days. Support cannot issue refunds directly.
The ticket we paste (name removed):
Hi, I got charged $49 twice on Sept 3 for the Team plan. I only have one workspace. Refund the extra one today or I am cancelling.
The kind of draft to expect:
If you were charged twice for the same period, we refund the extra $49 in full.
Our billing team checks the two charges and approves the refund within 1 business day. The money goes back to the card you paid with. It shows on your statement in 5 to 10 business days.
[CHECK: confirm two successful $49 charges on Sept 3]
Your next step: check your card statement. If the refund is not there 10 business days after billing approves it, reply here.
What a person must still check before sending:
- The charge itself. ChatGPT cannot see your billing system. If the second charge is a failed attempt or a pending hold, the whole reply is wrong. The [CHECK] marker stops you from skipping this.
- The word “today”. The policy does not allow that promise. This draft does not make it, but some drafts will. Delete it.
- The cancellation threat. The draft ignores it. A person decides whether to make a retention offer.
What it saves. Example assumptions: 40 tickets a day, 6 minutes to write a reply from scratch, 2 minutes to check and edit a draft. Each ticket saves 4 minutes. 40 tickets × 4 minutes = 160 minutes a day, about 2.7 hours. Time your own team for one day before you trust a number like this.
Where ChatGPT breaks as a customer service agent
Now the second job: ChatGPT answers your customers with no person in between. The obvious way to try it inside ChatGPT is a custom GPT, or the plugins that replace them. Both stay inside ChatGPT, so the problems below apply to either.
1. Customers cannot reach it where they are
A GPT or plugin lives inside ChatGPT, and customers must sign in to ChatGPT to use it. OpenAI’s GPT FAQ says GPTs “are not a way to embed ChatGPT in an external website or application” and points you to the API. So there is no widget on your site and no bot in Telegram or Discord.
Custom GPTs are also on the way out. OpenAI plans to retire them on December 11, 2026 and move them to plugins. New GPT creation is already closed on personal accounts.
2. Grounding is optional
A GPT can read files you upload. Nothing forces it to answer only from those files or to say “I do not know” when they are silent. It fills gaps from general knowledge, and that is how a wrong refund policy reaches a customer. Our guide on preventing AI hallucinations in customer support covers the controls you need.
3. There is no handoff
When the bot cannot help, nothing alerts your team. The customer must find another way to reach you and repeat the story. Our post on human handoff workflows explains what good escalation looks like.
4. There is no inbox
OpenAI states that GPT builders “cannot view” the conversations users have with their GPTs. You cannot spot a bad answer, fix the source article, or answer a follow-up in the same thread.
5. Data handling is your problem
When a customer chats with a GPT, their plan’s training setting applies, not yours.
On the API, OpenAI does not train on your data by default and keeps abuse monitoring logs for up to 30 days. Zero data retention needs approval from OpenAI. If you build your own bot, you also own retention, deletion requests, and access control for the stored chats.
The API route: what it costs and what you build
The API fixes the channel problem: the model works anywhere you write code. Here is an example model bill, with stated assumptions:
- 2,000 conversations a month, 4 AI replies each: 8,000 replies.
- Each reply sends about 4,000 input tokens (instructions, retrieved doc passages, the conversation so far) and returns about 250 output tokens.
- One file search call per reply. Your docs fit in the free 1 GB of storage.
| Cost line | gpt-6.1-sol | gpt-6-luna |
|---|---|---|
| Input: 32M tokens | 32 × $2.00 = $64.00 | 32 × $0.10 = $3.20 |
| Output: 2M tokens | 2 × $10.00 = $20.00 | 2 × $0.50 = $1.00 |
| File search: 8,000 calls at $2.50 per 1,000 | $20.00 | $20.00 |
| Monthly API bill | $104.00 | $24.20 |
Rates are OpenAI’s standard prices per 1M tokens, before caching discounts. The model bill is small. The build is the real cost: retrieval and refusal rules, a chat widget, a conversation store, an inbox, escalation logic, team alerts, and a test set of real questions. Each extra channel, such as a Telegram or Discord bot, is another integration to maintain.
OpenAI’s ChatKit gives you an embeddable chat UI, but you still run the server behind it, the conversation store, the inbox, escalation, alerts, and every non-web channel.
The API is the right choice in two cases: you need the bot inside your own product UI, or the bot must look up orders or accounts and act in your other systems. Check whether a packaged tool can do that before you buy. SupportHQ cannot. To price the packaged route, see how much AI customer support costs.
Which one to use: a decision rule
Ask one question: who sends the reply?
- A person sends every reply. Use the ChatGPT app with a project and the prompts above.
- The bot replies, and you have an engineer to spare. Use the API if you need in-product chat, order or account lookups, or actions in your own systems.
- The bot replies on your website or community channels, and you need handoff and an inbox without a build. Use a purpose-built support agent.
| ChatGPT app (project, GPT, or plugin) | OpenAI API, your own build | Purpose-built support agent | |
|---|---|---|---|
| Grounding | Reads files you add. No setting limits answers to them. | What you build: retrieval, refusal rules, tests | Built in. Check that it answers only from approved content. |
| Handoff | None. A person copies text in and out. | You build it. | Built in. Check which plan includes it. |
| Channels | Inside ChatGPT only. Customers must sign in. | Any channel you write code for | The vendor’s list |
| Cost | Plus $20/month per person. Business $25/user/month, or $20 billed annually. | Per token. About $24 to $104/month in the example above, plus engineering time. | Varies by vendor. SupportHQ Starter: $29/month for 2,000 conversations. |
| Setup effort | Minutes | Days to weeks of engineering (our estimate), then upkeep | Same day for SupportHQ. Varies by vendor. |
Start with prompt 1 on 20 tickets
Put your top articles and policies in a project this week. Run prompt 1 on your next 20 tickets. Mark each draft as one of three results: sent with light edits, rewritten, or wrong.
The “light edits” pile is the set of questions your docs already answer. A customer-facing agent can answer those with no person in the loop. If that pile is most of your volume, the third column of the table is worth a trial.
SupportHQ is one option there. It answers only from your approved docs, website, and past conversations, in a website widget, Telegram, and Discord. It escalates on its own with a reason, urgency, and summary into one inbox. It does not do order lookups or email tickets. Starter is $29 a month with unlimited team members. Check pricing, see how human handoff works, or start a 7-day trial (card required) and run the same 20 tickets through it.