Quick answer
A scheduled task is a prompt ChatGPT runs later or on a recurring cadence, with settings and run history. Schedule only after the underlying workflow is stable enough to run with limited steering.
Begin with read, compare, summarize, and draft. Add external writes, sharing, publishing, deletion, code execution, or system changes only when the exact permission, approval, validation, rollback, and owner are defined.
Turn a workflow into a contract
| Field | Define before scheduling |
|---|---|
| Objective | The decision or artifact the run must produce |
| Context | Exact chat, project, files, connected sources, and time window |
| Cadence | Time zone, frequency, start, end, and missed-run behavior |
| Method | Steps, tools, filters, calculations, and source priority |
| Output | Format, destination, evidence links, and freshness timestamp |
| Guardrails | Read/write scope, approvals, cost cap, stop conditions |
| Review | Owner, acceptance check, escalation path, retention |
OpenAI’s documentation distinguishes the method from the schedule: a repeatable skill or prompt defines how the work is done; the scheduled task defines when it runs.
Use progressive authority
- Run manually with fixed inputs and compare the result to an expert baseline.
- Schedule a read-only or draft-only version and inspect several run histories.
- Add bounded notifications or saved artifacts with duplicate prevention.
- Add a reviewed write only if the source, target, fields, identity, and rollback are explicit.
- Stop automatically on missing sources, changed permissions, stale data, ambiguous identity, schema drift, budget breach, or repeated failure.
A successful earlier run does not prove the next run is safe. Connected data, credentials, policies, prices, interfaces, and recipient lists can change between executions.
Control cost and repetition
OpenAI notes that scheduled or recurring work can be a high-variance usage pattern. Cost grows with cadence, context size, retrieval breadth, connector calls, model choice, retries, and artifact size.
Set a narrow window, incremental cursor, maximum records, maximum retries, runtime limit, output-size limit, and monthly cap. Alert on empty results, sudden volume changes, partial sources, repeated identical output, permission failures, and spend anomalies.
Review every destination
Keep output in the task or a private draft until it passes checks. Sending email, posting to chat, updating records, deploying Sites, merging code, or changing access affects other people and systems. Require a human or supported approval at the point of consequence, not only when the schedule is created.
The ChatGPT Work overview covers sources and plans. The surface comparison separates Work automation from coding workflows.
Frequently asked questions
What is a scheduled task in ChatGPT?
OpenAI defines a scheduled task as a prompt ChatGPT runs at a future time or on a recurring schedule, with its own settings and run history. Availability and controls depend on the current surface, account, workspace, and policy.
What should I automate first?
Start with a stable, bounded, reversible read-and-draft workflow such as a summary, monitoring check, or report. Keep source scope, cadence, output, acceptance criteria, budget, and escalation explicit before allowing writes or external actions.
Does scheduling make an action pre-approved?
No. A schedule defines when work runs; it should not silently widen data or action authority. Use least-privilege credentials, supported approvals, review gates, and stop conditions for writes, sharing, deletion, publishing, or other consequential execution.
Bottom line
Schedule a proven contract, not an underspecified intention. Preserve least privilege, reviewability, bounded cost, failure visibility, and an immediate stop path on every run.
Official sources
- OpenAI glossary: scheduled task definitions
- OpenAI best practices: scheduled tasks for repeated work
- OpenAI: ChatGPT Work admin FAQ
Source check: August 2, 2026. Verify current availability, settings, execution environment, permissions, credits, and workspace policy.