ChatGPT for customer service: what works, what fails, and the safer option

If your team has tried ChatGPT for support, you know the appeal. It writes well, it is quick to start with, and it can explain almost anything in plain words. The question is not whether it is useful in support. It is. The question is which support jobs it is safe for, and which jobs need a tool that knows your business and admits when it does not.

What ChatGPT is, for a support team

ChatGPT is a general-purpose AI assistant. Its model learned from a very large amount of public text, so it knows how refunds, shipping and subscriptions usually work. It does not know how yours work. Your return window, your plan limits, the product you discontinued last spring: none of that is in its training data unless it was public and widely repeated, and even then it may be out of date.

That one fact explains almost everything below. A general assistant is strong where the job is language, and weak where the job is a fact about your business.

What works: language jobs your team checks

These jobs share one property: a person reads the output before a customer does, and the person knows the right answer.

• Drafting replies. An agent pastes a customer message and the policy that applies, and asks for a short, kind reply. The agent checks the facts and sends it.
• Changing tone. Turn a blunt internal note into a polite customer email, or shorten a long answer for chat.
• Summarizing a long thread before a handover to another agent.
• First drafts of help articles and saved replies. A writer gives the facts, the assistant gives structure, and the writer edits.
• Translating a reply that a fluent speaker can then check.

If this is the job you need, a focused tool can be simpler than a blank chat box. Our free AI reply generator drafts chat, email, social and review replies and uses only the facts in the policy text you paste. The canned responses library has 100 saved replies to start from.

What fails: facts about your business, with nobody checking

Problems start when the assistant answers customers directly, or when an agent sends its output without checking it. These are the common failures:

• Plausible policy that is not yours. Ask a general assistant "Can I return an opened item?" and it gives a reasonable answer based on common practice. If your policy is different, the customer now has a confident, wrong answer in writing.
• No "I don't know". A general assistant is built to be helpful, so it tends to answer. For support, a wrong answer costs more than a refusal: the customer acts on it, then your team has to undo it.
• No record of where an answer came from. When an answer is wrong, you cannot tell whether the assistant misread a page, used an old version or made the fact up.
• Stale facts. Your prices and policies change. Text that you pasted last month does not.
• Data you should not share. Agents paste customer messages with names, emails and order details. Check the data terms of any AI tool before your team pastes customer data into it, and write a rule for what may be pasted.

Three ways to use AI for customer questions

There are three common setups. Each one fixes a problem of the one before it.

1. A general assistant on its own

Agents ask it questions and copy the answers. It costs little to start, and it is fine for the language jobs above. It is not safe for facts about your business, and it should never face customers in this form.

2. A general assistant with your documents pasted in

An agent pastes the return policy into the chat, then asks the question. Answers get much better, because the assistant now has your facts. Three problems remain:

• The answer is only as good as what was pasted. Paste the wrong version and you get the wrong answer.
• Nothing forces the assistant to stay inside the pasted text. When the text does not cover the question, it can fill the gap from general knowledge, and the answer reads just as confidently.
• It does not scale to customers. You cannot paste your whole help center into every visitor's question, and you have no shared view of what customers asked or where the answers came from.

3. A support chatbot limited to your own content

This is the setup built for customers. The chatbot reads your documents and website pages ahead of time. For each question it searches them, answers only from the passages it finds, and refuses when nothing fits. Your team sees the source behind every answer, and the questions it could not answer become a list of pages to write. This is called retrieval-augmented generation (RAG), and our post on how to stop an AI chatbot from making up answers explains the controls in detail.

How to choose for your team

1. List the jobs. Split them into language jobs (drafting, tone, summaries) and fact jobs (policies, prices, how-to answers).
2. Keep a person in the loop for language jobs. A general assistant or a focused reply tool is fine here, with a rule about what data agents may paste.
3. Use a grounded chatbot for fact jobs that face customers. Test it with your own documents before visitors see it.
4. Test refusals, not only answers. Ask 20 questions your documents answer and 10 that they do not. Any confident answer to one of the 10 is a made-up answer. Our checklist for choosing an AI chatbot lists the other questions to ask.

Where Paperlink AI fits

Paperlink AI is the third setup. It answers visitors only from the documents you upload and the website pages you import. When those do not hold the answer, it says so and shows your support email or support page, and visitors can send the conversation to your team when you turn that on. Your team sees the document and page behind every answer in the dashboard, and the Unanswered report lists the questions it could not answer. Replies the bot can't answer are free.

In our internal evaluation (September 2026), it answered 93.8% of the answerable test questions correctly and refused 100% of the questions it should have refused. It starts with a 7-day trial, card required, cancel anytime; see pricing.