Aiodic

Back to the table

110 Advanced Sh Scheduled Tasks Agents & Automation

Scheduled Tasks

AI jobs that run on a timetable.

What it is

A scheduled task is an AI job that runs on a timetable instead of waiting for you to ask. You write the instructions once, pick a time ("every Monday at 8 a.m.", "on the first of each month", "every weekday at 7"), and from then on the job runs by itself. At the chosen moment, the AI reads whatever it's been told to read, does the work, and delivers the result: a message in your chat app, an email, a notification on your phone, or a new document in a folder.

Think of it as the difference between a newspaper you go out and buy when you remember and one that lands on the doormat every morning. The content might be the same. What changes is that you no longer have to remember to fetch it, and that it arrives whether or not you have time to read it that day. Both of those facts matter, and the second one is the one people forget.

Here are the kinds of jobs that suit a schedule well for a creator or a small business:

In each case, the AI step is usually some form of summarization: taking a pile of material and turning it into something short you can act on. The schedule is what makes it happen without you.

Schedules and triggers

A schedule is one kind of trigger, the event that starts an automation. The Triggers page covers the other kind: automations that start because something happened, such as a new email or a new form entry. A scheduled task starts because the clock says so, regardless of whether anything happened. That difference shapes everything about it. An event trigger runs once per event, possibly hundreds of times on a busy day. A scheduled task runs a fixed number of times, and usually works on a batch: everything since the last run. That makes scheduled tasks easier to predict and budget for, and easier to review, because you get one tidy report instead of a stream of little ones.

Where you'll find them

Scheduled tasks come in two broad flavours. The first is built into AI assistants. At the time of writing, ChatGPT, Google's Gemini app and Anthropic's Claude (in its Cowork mode) all offer some way to tell the assistant "do this every morning" in plain language. You describe the job in a chat, say when, and the assistant creates a recurring task you can pause, edit or delete from a settings page. The second flavour lives in workflow automation tools such as Zapier, Make and n8n, which offer a "schedule" or "every day" trigger that you can put in front of any chain of steps, including AI steps. Behind both flavours sits an old idea from the 1970s, which the examples section describes.

Why it matters

It turns a good prompt into a habit. Most people who use AI have written at least one prompt that worked brilliantly once and never got used again, because they forgot, or were busy, or couldn't find it. A schedule removes the remembering. If a weekly review of your shop's numbers is genuinely useful, a scheduled task means it happens every week, not just in the weeks you were organised.

It's the gentlest step towards an agent. An agent is an AI that takes steps on its own. Scheduled tasks are one of the safest places to start with that, because the job is usually read-only: it reads and reports, and a person decides what to do. You get the benefit of work happening while you sleep without handing over the power to send, spend or delete. That's why it makes sense to start with scheduled reports and only later consider scheduled actions.

Nobody is watching when it runs. This is the flip side. A task you run by hand has you sitting there to notice when the answer is odd. A scheduled task at 6 a.m. has nobody. If it fails, it may fail silently. If it produces nonsense, the nonsense arrives looking exactly like last week's sensible report. If it quietly stops running, you may not notice for a month. Scheduled tasks need a small amount of care built in from the start: a way to know it ran, a way to know it failed, and a date to decide whether it's still worth having.

Time is harder than it looks. "Every Monday at 9" sounds unambiguous until you ask: 9 in which time zone? What happens on the Monday the clocks change? What if the service was down at 9? What if your laptop was closed? Professional programmers have been caught out by these questions for decades, and the examples below include some expensive ones. You don't need to be a programmer to avoid them, but you do need to know they exist.

Small costs add up over time. One AI brief costs very little. The same brief every weekday is about 260 runs a year. Ten such tasks, some of them doing web searches or reading long documents, is a real line in the budget if you pay by use, or a real share of your plan's limits if you don't. A scheduled task never gets tired, so it also never decides by itself that it's no longer worth the money.

Real-world examples

Cron: the clock daemon (Unix, 1970s)

The idea of a computer running jobs on a timetable is almost as old as the modern operating system. The classic tool is called cron, and it appears in Version 7 Unix, the influential 1979 release of the Unix operating system, whose descendants and imitators run a large share of the world's servers today. Its manual page there describes cron simply as the "clock daemon" (a daemon is a program that runs quietly in the background). It read a file called /usr/lib/crontab, the cron table, in which each line had five number fields for minute, hour, day of the month, month and day of the week, followed by the command to run. According to that manual, it examined the table every minute and ran whatever was due.

That five-field format is still in use. Paul Vixie wrote a widely adopted rewrite in 1987, and the descendants of cron still run backups, reports and clean-up jobs on a huge number of servers. When an automation tool asks you for a "cron expression", or shows you something like 0 9 1 (minute 0, hour 9, any day of the month, any month, Monday), it's speaking the language of that 1970s file. The AI assistants' "every Monday at 9" is the same idea with a friendlier face.

Cron also shows the weaknesses that every scheduler since has had to deal with. It uses the computer's own clock and time zone, so daylight saving changes can make a job run twice or not at all. And it only runs if the computer is on: a job due at 3 a.m. on a machine that was switched off at 3 a.m. simply doesn't happen.

When the calendar itself breaks the job: iPhone alarms (2010) and Azure (2012)

Two well-documented non-AI cases show how scheduled jobs go wrong in ways their builders didn't foresee.

In late 2010, owners of iPhones running iOS 4 found that recurring alarms went off an hour wrong after the clocks changed. The Register reported on 1 November 2010 that alarms in the UK had gone off an hour late after the switch back to Greenwich Mean Time that weekend, and that users in Australia and New Zealand had hit the same problem when their clocks changed earlier in the autumn. Only repeating alarms were affected, the ones people rely on without thinking. Macworld reported that Apple's advice, ahead of the US clock change on 7 November, was to set alarms to repeat "Never" and set them individually until after that date. The phone's clock was right; the schedule built on top of it was not.

On 29 February 2012, Microsoft's cloud service, then called Windows Azure, suffered a major outage. In a post-mortem published on 9 March 2012, Microsoft explained that a piece of software inside its virtual machines created security certificates meant to last a year, and calculated the expiry date by taking the current date and adding one to the year. On a leap day, that produces 29 February 2013, a date that doesn't exist, so certificate creation failed. The bug started firing at midnight UTC on 29 February, which was still 4 p.m. on 28 February in Microsoft's Pacific time zone, and Microsoft's own timeline records services being declared healthy again at 2:15 a.m. Pacific on 1 March. Microsoft gave affected customers a 33% service credit for the month.

Neither case involved AI, but both carry straight over. A scheduled AI task depends on dates in two ways: when it runs, and what it thinks "last week" or "this month" means. If your prompt says "summarise the last 30 days" and it runs on 1 March, or "compare with the same day last year" on 29 February, you're doing date arithmetic, and date arithmetic has edge cases. The simple defence is to prefer plain calendar language ("since the previous run", "the previous calendar month") and to glance at the output after any clock change or unusual date.

AI assistants learn to keep time (2025 and 2026)

On 14 January 2025, TechCrunch reported that OpenAI was rolling out a beta feature called Tasks to ChatGPT Plus, Team and Pro subscribers. Users picked a model option called "4o with scheduled tasks" and asked for things like reminders or a daily news briefing at a set time. TechCrunch noted that ChatGPT could browse the web on a schedule but couldn't monitor continuously. The feature changed a good deal afterwards. On 17 June 2026, 9to5Mac reported that OpenAI had launched a new version of scheduled tasks, replacing a proactive daily-briefing feature called Pulse, with a Scheduled page in the sidebar for pausing, editing and deleting tasks, broad time windows such as "morning", and monitoring tasks that notify you only when there's something worth reporting. According to that report and a second one in IT Brief, tasks couldn't run more than once an hour, and tasks left unattended could be paused automatically. Active task limits differed by plan. All of these are details at the time of those reports; check OpenAI's current help pages before relying on them.

Google announced "scheduled actions" for the Gemini app on 6 June 2025, in a blog post by Dave Citron, a senior director of product management. The examples it gave will sound familiar: a morning summary of your calendar and unread emails, five blog ideas every Monday, updates on a sports team. At launch it was for Google AI Pro and Ultra subscribers and some Workspace plans. At the time of writing, Google's help page says you can have up to 10 active scheduled actions, and makes a point worth knowing about any scheduled AI: responses are prepared in advance so they arrive on time, which means fast-changing information may not be fully current when you read it.

Anthropic's Claude account announced scheduled tasks for Cowork, Claude's mode for longer multi-step work, in February 2026, with a morning brief and weekly spreadsheet updates among the examples. At the time of writing, Anthropic's help page says these tasks can run hourly, daily, weekly, on weekdays or on demand, on its paid plans, and that because they run remotely they carry on "even when your computer is asleep or the Claude Desktop app is closed". It also says they work with connected services and cloud files rather than a folder on your own computer. That trade-off, running in the cloud so it doesn't matter whether your laptop is open, but only seeing what you've connected, is one you'll meet in most tools.

An everyday scenario: the brief nobody read

Here's an illustrative scenario, not a real business. Tom sells handmade leather notebooks online. In January he sets up three scheduled AI tasks: a daily news brief on "the stationery industry", a Monday sales summary, and a Friday review check. For two weeks he reads all three with enthusiasm.

By March, the daily brief has become background noise. It reads the same four trade websites every morning and, because not much happens in stationery on most days, it usually repeats a story from earlier in the week with a new headline. Tom stops opening it. The Friday review check is useful but, when the review site changes its layout, it starts reporting "no new reviews" every week, and Tom takes that as good news. Only the Monday summary stays genuinely useful, because he acts on it every week.

In June, a customer mentions a string of complaints about a broken clasp, all posted in the weeks the review check was reporting nothing. Meanwhile the daily brief has run around a hundred times, each one using searches and AI time, for nobody.

Nothing dramatic went wrong, which is exactly the point. Tom's tasks needed three things they didn't have: a way to say "I found nothing" differently from "I couldn't check", a rule to stay silent on days with nothing new, and a date on which he'd decide whether each task was earning its place.

Try it yourself

This takes about 30 minutes to set up, and a few minutes a week for four weeks. You'll create one weekly AI brief with a written specification, and put a review date in your calendar. At the time of writing, ChatGPT, Gemini and Claude all offer scheduled tasks on at least some plans, and automation tools such as Zapier and Make have schedule triggers; which plans include them, how many you can have, and how often they can run all change, so check the current help page for your tool first. If your tool has no scheduling, you can still follow along by running the same prompt by hand at the same time each week.

  1. Pick one job worth doing weekly. Choose something you would genuinely act on, such as a Monday summary of last week's sales, a Friday check of new reviews, or a weekly brief on one narrow topic. Weekly is a good first frequency: often enough to be useful, rare enough that you'll actually read it.
  2. Choose the frequency by asking how fast you'd act. If you wouldn't do anything about the news until Friday, a daily brief is five times the reading and cost for the same decisions. Match the schedule to how often you make decisions about the topic, not to how often things happen.
  3. Pick a time you'll be there to read it. A brief that arrives at 6 a.m. on a Monday you spend travelling will be skimmed. Put it just before a time you normally sit down to plan. Avoid the small hours between 1 and 3 a.m., when daylight saving changes happen in many countries, and write your time zone down in the spec.
  4. Write the specification. Use this template, fill in the brackets, and keep it in a notes file so you can find it again:
WEEKLY BRIEF: [name]
Runs: every [day] at [time], [time zone, e.g. Europe/London].
Covers: the period since the previous run (the last 7 days).

What to check:
- [source 1, e.g. my shop's orders, the new reviews on my product page]
- [source 2]
- [topic to search for, with the exact words to use]

What to report:
- At most [5] items, most important first.
- For each: one sentence on what happened, one on why it matters to me,
  and a link or reference so I can check it myself.
- End with one suggested action, or "No action needed".

What to ignore:
- [things you never want, e.g. press releases, stories older than 7 days,
  items already reported in a previous brief, general AI hype]

When to stay silent:
- If nothing meets the bar, reply with one line: "Nothing new this week."
- Never pad the brief to fill space.

If you can't check a source (error, login needed, page changed), say so
clearly at the top, naming the source. Do not report "nothing new" for a
source you could not read.
  1. Create the task. In your assistant, paste the specification with a request like the one below. Then open the tool's list of scheduled tasks and confirm the day, the time and the time zone it actually saved.
Please set up a recurring scheduled task using the specification below.
Before creating it, tell me: the exact day, time and time zone you will
use; what you'll be able to read and what you won't (for example, whether
you can see my shop data or only public web pages); and how I'll be
notified. Then run it once now as a test so I can see the format.

[paste your specification here]
  1. Read the test run critically. Did it follow the "at most five" rule? Did every item have a link you could check? Did it admit anything it couldn't see? If not, change the spec and test again. It's much easier to fix the instructions now than after four weeks of mediocre briefs.
  2. Put a review date in your calendar, four weeks from today. Call it "Keep or kill: [name]". Each week until then, give the brief a mark out of three in your notes: 0 for didn't read it, 1 for read it and did nothing, 2 for learned something, 3 for acted on it.
  3. On the review date, decide. Paste your four weeks of marks and the last brief into a chat with the prompt below.
I've been running a weekly AI brief for four weeks. Help me decide
whether to keep it, change it, or switch it off. Be honest; switching it
off is a perfectly good outcome.

The specification: [paste]
My weekly marks (0 = didn't read, 1 = read, no action, 2 = learned
something, 3 = acted on it): [e.g. 3, 2, 0, 1]
The most recent brief: [paste]

Please tell me:
1. Whether it's earning its place, based on the marks.
2. Signs of staleness: repeated items, the same sources every week,
   "nothing new" when something probably happened.
3. Whether the frequency is right, or whether fortnightly or monthly
   would do.
4. Three specific changes to the spec that would make it more useful,
   or a clear recommendation to switch it off.
  1. Act on the answer. Keep it, change it, move it to monthly, or pause it. If you keep it, set the next review for three months away. Only then consider a second scheduled task.

Common mistakes and misunderstandings

Scheduling what you won't read. The most common problem isn't a technical failure; it's a working task that nobody reads. It costs money or plan limits every run, it trains you to ignore notifications from your AI tool, and it can hide an important item in a pile you've stopped opening. Every scheduled task should have an owner and a review date, and switching one off should feel like tidying, not failure.

Choosing a frequency because it's available. "Daily" is the default in many tools, and hourly is often possible. More frequent isn't more useful if you only make decisions weekly. It's also more expensive, and more likely to produce near-duplicate reports.

Forgetting about time zones and daylight saving. A task set in one time zone and read in another will arrive at an odd hour. A task set for a time that is skipped or repeated when the clocks change may run late, early, twice or not at all, depending on the tool. The iPhone alarm bug is the famous case. After each clock change, check that your tasks arrived when you expected, and avoid scheduling between 1 and 3 a.m.

Assuming it ran. If a scheduled task fails, many tools tell you only in a log you never look at. If your computer was off, a task that runs on your own machine may simply not happen, and whether it catches up later depends on the tool. Decide for each task whether a missed run should be caught up or skipped: running a late sales summary is harmless, but sending the same customer email twice is not. Google's engineers, describing their own large cron service in the book Site Reliability Engineering, explain that they chose to skip a missed run rather than risk running a job twice, because a skipped run is usually easier to put right. Cloud-based tasks avoid the laptop problem but can still fail, pause or be switched off. Turn on failure notifications if your tool has them, and notice when a regular brief doesn't turn up. That habit is the heart of Observability.

Treating "nothing to report" as proof. "No new reviews" might mean no new reviews, or it might mean the page changed, a login expired or a search found nothing because the AI searched badly. A good spec makes the AI say which. Occasionally check one source yourself against what the brief says.

Reading the same sources until they go stale. A news brief that checks the same handful of pages every day will keep finding the same stories, rephrased as if new. Ask the AI to skip anything it reported before, check that items have real dates in the period, and refresh your source list at each review. AI can also hallucinate, confidently describing events that didn't happen, so every item needs a link you can open.

Believing a scheduled report is live. Some tools prepare the answer shortly before delivery, and all of them report on the moment they ran, not the moment you read it. For fast-moving topics, the time stamp matters.

Letting a schedule do things, not just report them. A task that reads and reports is low risk. A task that sends emails to customers, posts to social media, changes prices or deletes files on a timetable is a different animal, because a mistake repeats every run with nobody there. Keep scheduled actions behind a human in the loop until the reports have proved trustworthy for a good while.

Doing risky date maths in the prompt. "The last 30 days", "same day last year" and "the end of the month" all have edge cases, as Microsoft's leap-day outage showed on a much bigger stage. Prefer "since the previous run" or "the previous calendar week", and check the brief that covers a month end, a leap day or a clock change.

Forgetting the cost adds up. A fraction of a cent per run is nothing once. It's something else multiplied by every weekday, every task and every web search inside it. Look at your usage or plan limits after the first month, and set a spending cap with your AI provider if it offers one.

Related elements