From Scheduled Checks to Event-Driven Action: Testing ChatGPT’s New Automation for EC Customer Inquiries

Introduction

Hello, I’m Mia Sato, AI Researcher at GDX.

Until now, ChatGPT automation has made it possible to do things such as “summarize information every morning at 9” or “check emails every hour and organize new messages if there are any.”

This is useful, but the basic mechanism is that ChatGPT goes to check information when the scheduled time arrives.

With the update on August 25, 2026, event-triggered tasks were added to ChatGPT Work.

ChatGPT can now start working based on an event itself, such as a new Gmail message, a new Slack post, or activity on a GitHub Pull Request.

In other words, the workflow changes from:

“Check every hour to see whether a new email has arrived” to “Start when a new email arrives.”

This time, I tested how useful that difference can be in real work, using EC customer inquiries as an example.

ChatGPT automation is gradually changing

On GDX’s note, we have previously introduced several forms of ChatGPT automation.

Scheduled Tasks, introduced in 2025, made it possible to run ChatGPT at fixed times, such as every morning at 8 or every Friday.

Later, Workspace agents made it possible to create Agents for team workflows, call them when needed, or run them repeatedly at scheduled times.

The biggest difference this time is what starts the work.

Until now, a person had to call the Agent, or the team had to set a time in advance.

With this new function, the event itself, such as “an email arrived,” can become the starting point.

Personally, this difference felt bigger than I expected.

A comparison of Scheduled Tasks, Workspace agents, and Event-triggered Tasks.

How is it actually configured?

For this demo, I created an event-triggered task in ChatGPT Work and connected Gmail.

The setup had three main parts.

• Trigger: When a new email arrives in Gmail
• Condition: Emails whose subject contains “GDX_WEBHOOK_TEST”
• Prompt: Read the inquiry and organize the priority, customer request, items to check, next action, and draft reply

In the prompt, I also added rules such as “organize internal notes in Japanese,” “write the draft reply in the customer’s language,” and “do not present unconfirmed refunds or cancellations as confirmed facts.”

In other words, instead of pasting each email into ChatGPT every time, you first decide what kind of email should trigger the task and how it should be processed.

After the setup, when an email matching the condition arrives in Gmail, ChatGPT starts processing it.

A Gmail new-email event was set as the trigger, with target conditions and inquiry analysis rules defined in the prompt.

Automatically organizing EC inquiries

Using this setup, I sent several test inquiry emails.

The first test was an inquiry about a possible duplicate charge.

I purchased one item priced at 12,800 yen, but the same amount appears twice on my card statement.

The result was:

High | Payment | Possible duplicate charge

ChatGPT classified it as High because the inquiry could have a financial impact. It also organized the items the person in charge needed to check, such as order and payment records.

At the same time, I did not allow AI to decide whether a refund should be issued. That part is left for a person to confirm, so the draft reply stayed within wording such as “We are currently checking the details.”

Figure 1: Test result for a duplicate charge inquiry. Because it had financial impact, it was classified as High, and the customer request and items to check were automatically organized.

Testing an overseas inquiry

Next, I sent an English email saying that the delivered item was damaged and that the customer wanted to return it and receive a refund if possible.

ChatGPT classified the content as a return request and organized the customer request, items to check, and next action in Japanese for internal use.

The draft reply for the customer was created in English.

Figure 2: Test result for an English inquiry. Internal notes were organized in Japanese, while the customer reply draft was created in English.

For global EC, this flow is quite practical:

Being able to combine this into one task feels very useful.

Benefits I felt while testing it

The biggest thing I felt was that the barrier to trying it was low.

In this demo, I only needed to connect Gmail and set the target email condition and judgment rules.

Another interesting point was that thinking from the event naturally forced me to reorganize workflows that people usually handle somewhat intuitively.

When trying to decide “what should start this work” and “how far AI should handle it,” we need to put into words judgment criteria that are usually not made explicit.

Also, because this function is available inside ChatGPT, I did not need to prepare a separate OpenAI API key to create this demo.

For companies already using an eligible ChatGPT plan, this can be a practical way to try small automations without separately preparing the OpenAI API, such as organizing inquiries automatically, prioritizing important cases, or preparing draft replies.

Does this make Make or Zapier unnecessary?

After testing it, I do not think so.

For a flow like this:

ChatGPT alone can make it quite simple.

On the other hand, if you want to add more detailed branching, Make or Zapier may be better suited. For example, notifying Teams only for High-priority inquiries, sending return-related cases to the CS team and payment-related cases to Finance, or automatically recording the analysis results in Sheets.

I also tried adjusting the Task Update email sent from ChatGPT, but I could not freely customize the appearance of the notification screen.

So my current view is:

Use ChatGPT for reading, judgment, and organization.

Use Make or Zapier when you want to move the result across multiple tools.

Not “when should it run?” but “what should make it run?”

What impressed me most while testing this update was that it is not only about making AI run automatically.

When we think from the event, we first need to decide what event should start the work. Then we also need to think about what kind of judgment is required after that, how far AI should handle the task, under what conditions it should hand off to a person, and what counts as completion.

For the EC inquiry example, the flow can be organized like this:

After making the demo, I realized that I was not only configuring ChatGPT. I was also breaking down the usual workflow into Trigger → Rule → Action → Exception.

To define rules such as “duplicate charge inquiries are High priority” or “refund decisions must be confirmed by a person,” we need to verbalize standards that people usually apply naturally.

We also need to clarify where AI’s task ends, such as having AI prepare the draft reply and then leaving final confirmation and response handling to the person in charge.

From that perspective, event-triggered tasks are not only a simpler setting method. They can also become a trigger for making vague workflows and judgment criteria explicit, and for reviewing the business process itself.

Instead of asking “What time should AI run?”, the more important question for the next stage of automation may be:

What event should start the work, under what rules, and how far should AI handle it?

If your company already uses ChatGPT, it may be worth choosing one task that people repeatedly check and starting by asking, “What event actually starts this work?”

References

• Reference: Full guide to ChatGPT’s Scheduled Tasks feature: setup and use cases / GDX Inc. / GDX note
• Reference: The age when ChatGPT helps team routine work: what changes with Workspace agents / GDX Inc. / GDX note
• Official: Introducing workspace agents in ChatGPT / OpenAI / OpenAI
• Official: ChatGPT Enterprise / Edu release notes / OpenAI Help Center / OpenAI Help Center
• Official: Scheduled tasks in ChatGPT / OpenAI Help Center / OpenAI Help Center
• Reference: Scheduled tasks opened to the free ChatGPT plan, Plus and Pro support Webhook / Mado no Mori / Mado no Mori

※ Part of this article was created with the support of AI and edited by the author. The content reflects the author’s personal views and does not represent the official views or statements of GDX Inc. The information is provided for reference purposes only. Please refer to each company’s official announcements and primary sources for the latest details.