ChatGPT for customer service:
what it can do, and where it stops.

What ChatGPT does well for a support team

ChatGPT is a writing and sorting tool. Support is mostly writing and sorting. That is why it works so well for the part of the job that happens before the customer sees anything.

Here is what it does well, and a prompt for each. Paste the prompt, replace the brackets, read the output before you send anything.

1. 1

   Rewrite a reply for tone

   Your agent knows the answer but the draft sounds cold or defensive. ChatGPT fixes that in one pass.

   Rewrite the reply below so it is calm, short and clear. Keep every fact. Do not add promises or apologies.
   
   Reply: [paste]

2. 2

   Summarise a long thread for a handover

   A twelve-message thread becomes five bullets for the colleague who takes it over.

   Summarise this conversation in five bullets for a colleague taking it over: what the customer wants, what has been tried, what is still open.
   
   Conversation: [paste]

3. 3

   Translate without changing the facts

   Good enough for a first reply in a language nobody on the team speaks.

   Translate this reply into [language]. Keep the tone polite and neutral. Do not change any numbers, dates or product names.
   
   Reply: [paste]

4. 4

   Draft a canned response

   Give it the situation and it gives you a template with placeholders. Our canned responses library has 26 ready ones if you would rather start from those.

   Write a canned response for [situation] in under 80 words. Plain words, no jargon. Leave placeholders in square brackets for the order number and the date.

5. 5

   Sort a batch of tickets

   Paste fifty subject lines and get them back in categories. Useful on a Monday morning, and for spotting a spike.

   Sort these ticket subjects into: billing, shipping, account access, bug, feature request, other. Return a table with the subject and the category.
   
   Subjects: [paste]

6. 6

   De-escalate

   The reply an angry customer needs has a fixed shape. ChatGPT knows the shape.

   A customer is angry because [reason]. Draft a reply that names the problem in the first sentence, says exactly what happens next, and gives one date. Do not apologise twice.

7. 7

   Turn engineer notes into a help article

   The fix is in a Slack thread. The article is not written. This is the fastest way to close that gap.

   Turn these notes into a help centre article with a title, a two-sentence summary, numbered steps and a "still not working" section.
   
   Notes: [paste]

8. 8

   Find what the FAQ does not answer

   Paste last week's questions and the FAQ, and get the list of gaps.

   Here are the questions customers asked this week and our FAQ. List the questions the FAQ does not answer, grouped by topic.
   
   Questions: [paste]
   
   FAQ: [paste]

Each of these is a job for one person, on one ticket, before it goes out. That's what ChatGPT does well. The next section is what it doesn't.

What it cannot do as the thing your customers talk to

Four things. None of them is a model problem. They are all product problems, and OpenAI's own pages say so.

It doesn't know your product. ChatGPT answers based on what it was trained on and what you paste in. The Free plan handles around 12 pages of text at once. Plus and Business take about 40, and Pro and Enterprise about 250. Your help centre is bigger than that, and it changes every week. Business and Enterprise plans have a "Company knowledge" plugin that connects to Google Workspace, Slack, GitHub and Microsoft 365, but it works for your staff inside their workspace. Nothing on the pricing page is a channel your customers can talk to.

It cannot do anything. A customer who wants a refund doesn't want a paragraph about refunds. ChatGPT can't look up an order, change an address or issue a credit, because it isn't connected to the systems where those things live.

It cannot hand off. When a chatbot runs out of what it knows, the customer needs a person, and the person needs the transcript. ChatGPT has no queue, no team, no escalation path, so there's nowhere for that conversation to go.

It reports nothing. You won't know how many questions it answered, how many it got wrong, or which article it should have read. OpenAI's own instruction, on the Company knowledge page, is to "review the answer and any source links provided before relying on the information." That is sound advice for an employee. It is not a support process.

What each plan can read, and what it does with your data

Read from the ChatGPT pricing page and OpenAI's help centre on 13 September 2026. Prices are listed because the plan decides two of the four limits above.

Two things stand out. The Free, Go, Plus and Pro plans are, in OpenAI's words, "designed to be used by individuals," and their chats "may be reviewed and used to improve our AI models" unless you turn that off. If anyone on your team is pasting tickets into a personal account, that is the first thing to fix. And the plan that can read your company's tools is a per-seat workspace for employees, which is the right tool for the eight jobs above and the wrong one for a customer-facing bot.

ChatGPT vs an agent trained on your docs

The four gaps above are exactly what an AI support agent is built to close. This is a capabilities table, not a price comparison. Outlearn is our product; the column describes what it does today.

The agent isn't a helpdesk. There's no shared inbox, no ticket queue. It sits in front of the one you already have, reads the help centre in place with no export, and files what it can't resolve as a ticket with the transcript attached. If you want a helpdesk, keep yours.

Use ChatGPT when, use an agent when

Use ChatGPT when the work is one person's writing: a reply that needs a better tone, a thread that needs a summary, a macro that needs drafting, a batch that needs sorting. The input has no customer data in it, or you are on a Business or Enterprise plan. The answer needs no access to your systems.

Use an agent when the same forty questions come in every day. The answers live in docs that change. The customer needs something done, not described. You need to know the number: how many were resolved without a person, and how many were handed over.

Use both if you can. The agent takes the repeat questions off the queue and hands the rest to your team with the transcript. Your team uses ChatGPT on what is left. That is the setup most of our customers run, and it is what the AI customer service agent page describes.

What a trained agent actually resolves

45.9%

Resolved with no human, of 29,003 conversations

5.7%

Handed to a person, all time

2.7

Knowledge sources per workspace, average

148

Conversations a day in July 2026

Vendors say "up to 80%" and don't say of what. Here's our number, with the denominator. Across 29,003 customer conversations handled from October 2025 to August 2026, 45.9% were resolved automatically with no human involved. 5.7% were handed to a person. The rest ended without either, such as a visitor who left or a quick lookup that needed no close.

The last 90 days run higher: 50.8% of 10,694 conversations resolved automatically, 2.4% handed off. The average workspace connected 2.7 knowledge sources, which is the single biggest reason the number moves. An agent that reads three systems answers more than one that reads one.

Those are the resolution rate numbers ChatGPT can't give you, since it's not talking to your customers. The full snapshot, with the definitions, is on the AI support benchmark page.

How to start

1. 1

   Put ChatGPT on the team's side today

   A Business workspace if tickets will go into it, so nothing trains on your data. Share the eight prompts. Ban customer details on personal accounts.

2. 2

   Count the repeats

   Prompt 8 does it in a minute. If the same twenty questions are a third of your volume, that third is the agent's job.

3. 3

   Connect the docs where they live

   Help centre, Confluence, Drive, Notion, the lot. The agent should read them in place, not a copy you uploaded once. The free plan has 100 credits a month and needs no card, which is enough to test it on your own docs.

4. 4

   Set the handoff before you go live

   Decide what reaches a person, and make sure the transcript travels with it. Our handoff page shows what that looks like.

5. 5

   Watch one number

   Resolved automatically, with the denominator. If it is not published for you, it is not measured.