Table of Contents
- Key takeaways
- Can you use ChatGPT for customer service?
- How the three setup routes compare
- Route 1: Set up a customer service chatbot with SiteGPT
- Steps
- Cost
- Limits
- Who it fits
- Route 2: Build a customer service chatbot on the OpenAI API
- Steps
- Cost
- Limits
- Who it fits
- Route 3: Make a custom GPT for customer service
- Steps
- Cost
- Limits
- Who it fits
- How to use ChatGPT to help customer service agents
- How much does a ChatGPT customer service chatbot cost?
- How to choose a ChatGPT setup for customer service
- Common mistakes when using ChatGPT for customer service
- Frequently asked questions
- Basics
- Can ChatGPT be used for customer service?
- Can you put ChatGPT on your website?
- Is a custom GPT a customer service chatbot?
- Setup
- How long does it take to set up a ChatGPT customer service chatbot?
- Can a ChatGPT chatbot hand a chat to a person?
- Can the chatbot answer only from your own content?
- Cost
- How much does ChatGPT for customer service cost?
- Is there a free way to use ChatGPT for customer service?
- Data and privacy
- Does OpenAI train on customer service conversations?
- Where should the OpenAI API key go?
- Sources

- The ChatGPT app cannot run as a chat widget on your website. A custom GPT also lives inside ChatGPT, and OpenAI plans to retire custom GPTs.
- A no-code tool such as SiteGPT is the fastest route for a customer-facing chatbot. It trains on your site, installs with one script, and starts at $59 a month billed monthly, or $39 a month billed yearly.
- The OpenAI API is the cheapest per answer. Model and file search fees come to about $29.50 for 10,000 answers a month on gpt-6-luna, before developer time and hosting.
- ChatGPT Business seats help your agents draft replies. A Standard seat costs $25 a user a month billed monthly, or $20 billed annually.
- Whatever route you pick, give customers a way to reach a person.
Key takeaways
Route | Best for | On your website | Price to start | Handoff to a person |
1. No-code tool on OpenAI models (SiteGPT) | A customer-facing chatbot with no developer | Yes, one script tag | $59 a month billed monthly, or $39 a month billed yearly | Built in, on every plan |
2. OpenAI API built in-house | Developers who want full control | Yes, you build the chat window | Usage-based, about $2.95 per 1,000 answers in model and file search fees | You build it |
3. Custom GPT in ChatGPT | Internal team use inside ChatGPT | No | A ChatGPT Business seat, $25 a user a month billed monthly or $20 billed annually | No |
ChatGPT for agents | Drafting and summarizing replies | No, agents use it | Same Business seat price | Not needed, a person sends every reply |

Can you use ChatGPT for customer service?
- A customer-facing chatbot. Customers ask questions on your website and get answers from your own help content. Routes 1 and 2 below do this job.
- An assistant for your agents. Your team uses ChatGPT to draft replies, summarize long threads, and translate messages. A person reviews and sends every reply. ChatGPT Business and custom GPTs do this job.
How the three setup routes compare
- Who builds it. Anyone who can paste a script, or a developer.
- Where answers come from. Your own content, or the model's general knowledge.
- Handoff to a person. Built in, or something you build.
- Cost. A flat monthly plan, or usage fees plus developer time.
Criterion | 1. SiteGPT (no-code) | 2. OpenAI API | 3. Custom GPT |
Who builds it | Anyone who can paste a script | A developer | Anyone in a Business, Enterprise, or Edu workspace |
Where answers come from | Your website, files, and connected sources | Files you upload to a vector store | Up to 20 uploaded files |
Chat on your website | Yes | Yes | No |
Handoff to a person | Built in, on every plan | You build it | No |
OpenAI models | GPT-4.1-mini and GPT-4.1 | Your choice, including GPT-6 Luna | The models in your ChatGPT plan |
Cost model | Flat monthly plan with a message quota | Per token and per tool call, plus developer time | Per ChatGPT seat |
Route 1: Set up a customer service chatbot with SiteGPT
Steps
- Sign in at the SiteGPT login page with your email address. There is no password. You need an active plan to train a chatbot. The 7-day free trial requires a credit card.
- On the Chatbots page, select Create New Chatbot and give it a name.
- Select Website Links, then Scrape Website. Enter your home page address. The quickstart suggests a recursion depth of 2 and up to 50 pages to start. A sitemap works too.
- Wait for training. The Links List shows how many pages are trained, pending, and failed.
- Open Chat with your chatbot and ask three or four questions your customers often ask. Each answer comes from your website. For a question your site does not cover, the chatbot usually replies that it does not have enough information.
- Turn on human handoff. In Leads, then Human Support, turn on Enable Human Support and save. By default the AI hands the chat to your team when the visitor asks for a person, when the situation is urgent, or when the visitor is frustrated after an answer that did not help.
- Select Installation, copy the JavaScript embed code, and paste it into the head section of every page. WordPress, Shopify, Wix, Squarespace, and Webflow have their own steps in the SiteGPT docs.
Cost
Plan | Billed monthly | Billed yearly | Messages a month | Chatbots | Content refresh |
Starter | $59 a month | $39 a month ($468 a year) | 4,000 GPT-4.1-mini or 400 GPT-4.1 | 1 | Manual |
Growth | $129 a month | $79 a month ($948 a year) | 10,000 GPT-4.1-mini or 1,000 GPT-4.1 | 2 | Monthly auto refresh |
Scale | $429 a month | $259 a month ($3,108 a year) | 40,000 GPT-4.1-mini or 4,000 GPT-4.1 | 3 | Weekly auto refresh |
Enterprise | Custom | Custom | Custom | Custom | Daily auto refresh |
Limits
- There is no free plan. The trial needs a credit card.
- Starter has no chat integrations. Slack, Messenger, Crisp, Freshdesk Messaging, Google Chat, and Zoho SalesIQ start on Growth. Zendesk starts on Scale.
- There is no Intercom or HubSpot integration today.
- The plan cards list GPT-4.1-mini and GPT-4.1 only. Newer OpenAI models such as GPT-6 Luna are not on them.
- Starter refreshes content by hand. Auto refresh starts on Growth.
Who it fits
Route 2: Build a customer service chatbot on the OpenAI API
Steps
- Create an API key in the OpenAI dashboard and keep it on your server. OpenAI's key safety guide says never to deploy a key in browsers or mobile apps.
- Upload your help articles and policies to a vector store, so the file search tool can find them before the model answers.
- Write a server endpoint that receives each customer question and calls the Responses API with file search.
- Write instructions that tell the model to answer only from your files and to say when it does not know.
- Build the chat window for your site.
- Build what the API does not include: chat history, rate limits, analytics, and a way to pass the chat to a person.
Cost

Limits
- It needs a developer and a server. Calling OpenAI straight from the browser exposes your key.
- There is no chat widget, dashboard, or handoff until you build them.
- Keeping the files in the vector store current is your job.
- Data sent to the OpenAI API is not used to train OpenAI models unless you opt in. OpenAI may still keep abuse monitoring logs.
Who it fits
Route 3: Make a custom GPT for customer service
Steps
- Use a ChatGPT Business, Enterprise, or Edu workspace. Personal accounts on Free, Go, Plus, and Pro cannot create or publish new GPTs.
- Open Explore GPTs in the ChatGPT sidebar and select Create.
- Write the instructions: the role, the tone, and the rule to answer only from the uploaded files.
- Upload your help articles and policies under Knowledge.
- Test the GPT in Preview, then select Create.
- Share it with your team. The sharing options depend on your workspace settings.
Cost
Limits
- A GPT cannot run as a chat widget on your website.
- A GPT holds up to 20 files, each up to 512 MB.
- OpenAI plans to retire custom GPTs. Retirement for affected Enterprise workspaces is planned for December 11, 2026, and OpenAI expects other plans to follow the same timeline. OpenAI recommends moving GPT workflows to Plugins.
- A GPT has no handoff, no analytics on customer questions, and no lead capture.
Who it fits
How to use ChatGPT to help customer service agents
- Draft a first reply from the customer's message and the relevant policy.
- Summarize a long email thread before a handover.
- Translate a reply into the customer's language.
- Rewrite a reply in a calmer tone after a complaint.
How much does a ChatGPT customer service chatbot cost?
Questions a month | OpenAI API fees (gpt-6-luna with file search) | SiteGPT plan with that quota | SiteGPT price |
4,000 | About $11.80 | Starter, 4,000 GPT-4.1-mini messages | $59 a month billed monthly, or $39 billed yearly |
10,000 | About $29.50 | Growth, 10,000 GPT-4.1-mini messages | $129 a month billed monthly, or $79 billed yearly |
40,000 | About $118 | Scale, 40,000 GPT-4.1-mini messages | $429 a month billed monthly, or $259 billed yearly |
How to choose a ChatGPT setup for customer service
If you need | Pick | Why |
A chatbot on your website this week, with no developer | Trains on your site and installs with one script | |
Customers to reach a person from the chat | SiteGPT, or the API with handoff you build | SiteGPT includes handoff on every plan. The API includes none |
The chat inside your own app or design system | OpenAI API | You control the window, the model, and the data flow |
The lowest usage fee at high volume, with developers on staff | OpenAI API | About $118 a month at 40,000 answers before build and upkeep |
A shared policy assistant for agents | Custom GPT or ChatGPT Business seats | Agents already work in ChatGPT. Plan for the GPT retirement |
Help drafting and translating replies | ChatGPT Business seats | A person checks each reply, and OpenAI will not train on workspace data |

Common mistakes when using ChatGPT for customer service
- Expecting the ChatGPT app to work as a website widget. It does not. Customers would leave your site.
- Letting the bot answer from general knowledge. Tell it to answer from your content and to say when it does not know. The guide on chatbot hallucinations covers the fixes.
- Leaving out a route to a person. Every customer-facing chatbot needs one. The list of AI chatbots with human handoff compares how tools handle it.
- Putting the API key in the browser. Route every call through your server.
- Building a new customer service workflow on a custom GPT. OpenAI plans to retire them.
- Training once and never refreshing. Re-train after every price, policy, or product change. The chatbot best practices guide lists the checks to run.





























