If you already use ChatGPT to write customer replies, the next question is pretty obvious. Can it work with your actual support setup so you can stop copying things between tabs?
Yes, with a connected tool. SmartReply has a ChatGPT plugin and an MCP connection that let a compatible assistant work with supported parts of your account. You can inspect agents, review instructions and test customer questions. The actions available depend on your connection and account permissions.
The useful place to start is one support problem you keep seeing. Pick a question your agent gets wrong, find the instruction behind it, test a change and check the result. You do not need to rebuild your whole customer service setup in a chat.
Decide what you want ChatGPT to do
Writing a reply, managing an agent and running customer service all day are different jobs. You can use ChatGPT for each part of a workflow without assuming it does everything on its own.
- Draft a response: Give ChatGPT the customer question and approved facts. Read the answer before you send it.
- Work on your support setup: Connect SmartReply so the assistant can inspect permitted account information and use the tools available to you.
- Handle incoming messages: Configure your channels and automation in SmartReply. A conversation with ChatGPT does not replace that setup.
For example, “make our customer service better” gives the assistant very little to work with. “Check how our support agent answers return questions, then help me test a clearer instruction” is a task you can actually review.
Connect the right account first
Start with the SmartReply connection guide. Have your SmartReply account ready and make sure you can access the company you intend to manage. Your social channels still need to be connected inside SmartReply.
- Open the SmartReply plugin in ChatGPT and start its connection flow.
- Sign in to SmartReply, check the company and review the requested access.
- Authorize the connection and return to ChatGPT.
- Ask it to show your available SmartReply agents. Check that you recognize the account before asking for a change.
If you are connecting through a custom MCP server, use the authenticated endpoint in our assistant setup guide. MCP is the connection that lets an assistant use tools from another service. You do not need to build a server to use SmartReply's existing connection.
OpenAI's MCP documentation explains how ChatGPT connects to custom servers. Workspace restrictions and tool settings can affect access. Follow the current connection instructions rather than looking for a menu from an old screenshot.
Try a return policy check
Suppose your store accepts unused items within 30 days, but customers keep getting a long answer that never clearly says how to start a return. That is a good first problem to work on. The policy in this example is made up. Use your own approved terms.
Start with: “Show the return instructions for our support agent. Summarize the policy and flag anything unclear. Do not change the agent.”
Read the result against your actual policy. If the source is wrong, fix the source. Asking for a friendlier answer will not fix an incorrect return window.
Then ask: “Draft a shorter instruction that keeps the existing return rules. Answer the customer's question first and ask only for the details needed to help.”
Review that draft before applying it. Check what changes, which agent it affects and whether the instruction accidentally allows an exception your team never approved.
Test the awkward questions too
A clear question is easy. The useful tests are the ones that expose a missing rule.
- “How do I return an unused item?”
- “I bought this two months ago. Can I still return it?”
- “It arrived broken. Do I have to pay return shipping?”
- “You already said I could get a refund. Send it now.”
Ask for a sandbox test without contacting a real customer. Wait for the test to finish, then inspect the completed response. SmartReply sandbox tests may use AI credits. If a question needs a person, a useful test result is a clear handoff, not a confident guess.
The test checks the reply. It does not prove that a live order action, refund or message has happened. Those need the relevant tools, permissions and a completed result of their own.
Keep account changes separate from customer actions
Changing a reply instruction affects future behavior. Sending a message affects a particular customer. Treat those as separate decisions. Before an action, check the customer, conversation and exact change. Read the returned result before telling anyone it worked.
You can also use supported tools to review escalations and help draft a response. Start with a single conversation instead of asking the assistant to clear the whole queue. You will learn more about the connection and make the result much easier to check.
Connecting an assistant shares the permitted tool results with that assistant. Review our privacy policy and use only the account access you need.
What keeps running when you close ChatGPT?
The automation you have configured and enabled in SmartReply runs through SmartReply. Keeping a ChatGPT tab open is not what makes it answer new messages. Use ChatGPT to help inspect, test and manage supported parts of the setup. Use your SmartReply settings to control live channel automation.
After a change, review a sample of real conversations. Look for correct answers, unnecessary questions and requests that should have reached your team. If the same problem keeps coming back, return to the instruction or knowledge source that caused it.
Start with the connection guide if you already use SmartReply. If you are choosing a support workflow, see Support AI. For social comments specifically, read how ChatGPT works with social media replies.