Scheduled automation

Scheduled automation is the pattern of moving a recurring workflow from "I run it whenever I remember" to "it runs on its own and shows me the result." The AI takes over both the trigger and the execution; the human's job shifts from doing the work to reviewing the output. The shape is broadly the same across AI tools — describe the workflow as a prompt, set a recurrence, get notified when it runs — but the design considerations are tool-agnostic and worth getting right before automating anything.

Recurring work is where AI compounds the most. A workflow you only run when you remember stays at the speed of your memory; a scheduled version runs even when you're on holiday, asleep, or buried in something else. But scheduled automation also magnifies any flaw in the underlying workflow — silent failures pile up, drift accumulates, and a workflow that was acceptable when you ran it manually becomes unreliable when nobody's watching.

When to reach for a scheduled run

A workflow is ready for scheduling when four conditions hold together:

• Recurring trigger. The workflow runs on a predictable cadence — weekly, daily, monthly — or on a clear event signal. Ad-hoc triggers usually mean human judgment is involved in deciding when to run, which is itself worth keeping.
• Bounded sources. The inputs come from a known list of places: specific URLs, a folder, an inbox, a database. Open-ended "look around the internet" runs are harder to make reliable and harder to debug when they go wrong.
• Reviewable output. The result is something a human can scan in minutes and judge — a digest, a draft, a report. Workflows whose output is hard to verify shouldn't run unattended, regardless of how mechanical they look.
• Low single-run risk. A single bad run is recoverable. If a misfire could send the wrong email to a client, post incorrectly to a public channel, or modify shared data, the workflow needs human-in-the-loop confirmation before action — not just review afterwards.

Workflows that fail any of these tests usually become better candidates for an on-demand assistant than a scheduled run. The scheduling itself is not the goal; the goal is reliable output, and the schedule is one mechanism for getting it.

Design considerations

The design choices that matter most are about what happens when things go wrong rather than when they go right.

Source allowlisting. Constrain what the run can read. Most AI tools that support scheduled runs let you define an allowlist of domains or systems the run can reach. A tighter allowlist is better than a broader one — it prevents drift and makes failures obvious because something not in the list will throw a clear error rather than fetch the wrong thing silently.

Output destination. Decide before scheduling where the output goes and who reads it. Output that lands in a place nobody checks is the most common failure mode of scheduled automation — the run succeeds, the file is written, but the human review never happens. Routing the output to a place where you'll see it (a Slack channel, an email, a folder you visit anyway) is often more important than the output's content.

Notification. Coupling the run to a notification you actually receive matters more than it sounds. A scheduled run that produces output silently is fragile because there's no way to tell when it stops working. A run that pings you on completion (and on failure) gives you a passive monitor for free.

Failure visibility. A scheduled workflow that silently misses a run is worse than one that loudly fails. Build in some signal that the run happened — even a "no new items this week" message — so the absence of that signal becomes diagnostic.

Human review on output, not on action. The scheduled run should produce a draft or a digest that the human reviews before any external action takes place. Automated runs that send messages, post publicly, or modify shared resources directly skip the review step and inherit much higher single-run risk. The pattern that scales is "AI prepares, human approves" rather than "AI acts."

How the pattern shows up across tools

The scheduled automation pattern is now native in all three major consumer AI tools, with broadly similar mechanics and a few differences worth knowing.

Anthropic Claude (routines). Cloud-scheduled runs configured at claude.ai/code/routines. Each routine has a prompt, a GitHub repository as its workspace, a cloud environment that controls allowed domains and secrets, and one or more triggers. Output typically gets committed to the repository; notifications can be sent through any service in the environment's allowlist. Routines are available on Pro, Max, Team, and Enterprise plans.

OpenAI ChatGPT (scheduled tasks). Recurring or one-off prompts created from inside a ChatGPT conversation. Up to ten active tasks at a time. Notifications go through push or email. Output stays inside ChatGPT unless the prompt is written to send it elsewhere. Available on Free, Plus, and Pro across web, iOS, Android, and macOS.

Google Gemini (scheduled actions). Recurring prompts with native integration into Google Workspace — Gmail, Calendar, Docs. Up to ten active actions at a time. The Workspace integration is the differentiator: a scheduled action can read your calendar, summarise unread mail, or draft into Docs without separate connector setup. Requires Google AI Pro/Ultra or qualifying Workspace business and education plans.

The user-level pattern is identical across all three: write a prompt that describes the workflow as if no one is watching, set a recurrence, and the tool handles the rest. The differences come down to where output lands by default, what integrations are available without additional setup, and how many active runs the tool will hold for you.

Related pages

• agent-use-case-evaluation — Five-point filter for deciding which workflows are good candidates for any kind of automation
• ai-output-verification — Strategies for reviewing scheduled output efficiently, including sampling and anomaly detection
• compound-engineering — How scheduled automations get better over time through accumulated source refinement and prompt corrections
• ai-productivity-traps — Why scheduled runs that produce output nobody reads are a common form of decorative automation