Customer Support in ChatGPT and Claude: From Question to Tracked Ticket

A customer asks, “My workspace export failed. Has anyone picked up my support request?” A useful support app can retrieve the customer's existing ticket, show its current status, and let them add the missing details without starting another email thread.

This is a practical workflow for a business with an existing support system. The example below describes how an MCP-connected app could work; it is not a Kernel Apps customer case study or an announcement of a ready-made helpdesk integration.

Start with the ticket the customer already has

After the customer connects their account, the app retrieves tickets that account is allowed to see. It shows the subject, ticket number, status, last update, and the latest customer-visible reply. Internal notes stay inside the support system.

If two tickets mention exports, show both and ask which one the customer means. Guessing from a similar subject can attach information to the wrong conversation. If there is no matching ticket, offer to create one rather than claiming that a support agent is already investigating.

Turn a conversation into a useful handoff

The customer adds, “It happens when I export more than a month of records.” The app prepares a short update containing the affected feature, what happened, and when it last occurred. The customer reviews it before submission.

The support API saves that update on the selected ticket. Only then should the app say it was added. Show the returned ticket number and updated status so the customer can find the same request later. A generated summary that has not been saved is still a draft.

Keep the support system authoritative

The app needs account-scoped ticket lookup, customer-visible messages, a way to create a ticket or append a reply, and clear failure responses. It should not expose every helpdesk operation just because the API makes them available.

• Let customers read their own requests and add relevant information.
• Keep staff notes, other customers' tickets, and administrative actions out of the customer toolset.
• Treat response-time targets as policy information, not a promise that a particular agent will reply by a generated deadline.
• If submission times out, check the ticket before retrying so the same message is not posted twice.

Make the result easy to inspect

Where the host supports interactive UI, a compact ticket view can show status, the last reply, and an Add details form. A text response should still identify the ticket and explain the next step. ChatGPT and Claude connections need their own verification against the same support service.

Useful measures include duplicate-ticket frequency, the share of updates saved successfully, and whether customers can reopen the same ticket. These are measurements to collect, not assumed improvements.

Read how to prepare your API for an MCP app and how to handle private customer data. Join the Kernel Apps waitlist to discuss a workflow like this.