Dear Suzannah

If an AI task runs on a schedule, what still needs human approval?

Anything that changes money, customer commitments, permissions, records, public content, or another person’s work should have a defined approval boundary. Monitoring and drafting can usually run farther on their own than irreversible action.

Here’s the deal

OpenAI’s current Scheduled Tasks guidance makes recurring and condition-based tasks easier to create and manage. OpenAI also describes ChatGPT Work as able to run recurring work across connected apps while keeping approvals available for important actions. The operating lesson is simple: scheduling is not the same thing as delegating unlimited authority.

What this actually helps you do

For an owner using AI to monitor leads, summarize inbox activity, prepare reports, or update recurring work, a stop rule keeps a useful automation from continuing after the business context changes.

Exactly three benefits

• Prevents stale instructions from running forever after priorities change.
• Keeps high-impact actions behind a clear human approval point.
• Makes ownership visible when an automated task produces an exception.

Step-by-step

1. Write the recurring task in one sentence, including what it may access.
2. Separate read-only work from actions that change a system or send something externally.
3. Name the exact conditions that should pause the task, such as missing data, unusual volume, contradictory instructions, or a failed dependency.
4. Name the person who reviews exceptions.
5. Set a review date for the automation itself so old instructions do not become permanent policy.
6. Document how to disable the task quickly if output quality drops or business priorities change.

Tips and tricks

• Give every recurring task an owner and a review date.
• Use a small set of explicit pause conditions instead of vague instructions like “use judgment.”
• Keep exception reports short enough that a human will actually read them.

Common mistakes

• Assuming a scheduled task is harmless because it worked last week.
• Letting one automation both detect a problem and make a high-impact decision without review.
• Failing to remove or update tasks after a project, offer, or policy changes.

Human review checklist

• Is the task owner still correct?
• Are connected apps and permissions still necessary?
• Are pause conditions clear?
• Are irreversible actions still reviewed?
• Is the next automation review date recorded?

How to measure success

• Number of exceptions caught before an unwanted action
• Percent of scheduled runs completed without manual repair
• Time between a business change and task update

FAQ

Should every scheduled task require approval?

No. A daily summary or monitoring task may not need it. The approval point should match the consequence of the action.

What is a good pause condition?

Anything that makes the original instructions unreliable, such as missing records, conflicting inputs, a large change in volume, or a request outside the task’s scope.

How often should recurring tasks be reviewed?

Use a cadence that matches how quickly the underlying business process changes. Fast-moving workflows should be reviewed more often than stable reporting tasks.

Glossary

Approval boundary: The point where automation stops and a person decides whether the next action should happen.

Pause condition: A specific situation that stops a recurring task from continuing automatically.

Exception: A result or input that falls outside the task’s normal operating rules.

Sources and further reading

• OpenAI: Scheduled Tasks in ChatGPT
• OpenAI: ChatGPT for your most ambitious work
• OpenAI: ChatGPT Work and Codex

Related: Scheduled AI Tasks Need an Approval Boundary.

Practical closing note

Good automation is not “set it and forget it.” It is “set it, define the stop rule, assign the owner, and know how to turn it off.”