How do I put my social media on autopilot with AI? Give an AI assistant a recurring instruction and access to your content calendar. Every morning or every Monday it runs on its own: it checks what is waiting for approval, drafts a post from fresh news or prepares a client report. Everything it creates lands as a draft, so a person still approves before anything goes live. This guide covers the whole setup: what your plan allows, how to write a task that works, the permissions that stop it from getting stuck, how to know it really ran, how to keep ten tasks organized and how to run it all as a team.

{{form-component}} 

How scheduled AI tasks work

A scheduled task is a prompt with a timer. Three pieces make it useful for social media teams:

1. The schedule. Weekdays at 8:00, every Monday at 7:00, the first day of each month.
2. The instructions. A prompt you write once, with the rules that never change: which workspace, which status, what not to touch.
3. The connection to your calendar. Through the ZoomSphere MCP server, the agent reads posts, statuses and analytics and creates drafts in the right client workspace. The server checks formats for each network and sends errors back to the agent, so it can fix a caption before you ever see it.

When the task runs, it does the work and reports back: a list, a reminder ready to send, a link to the new draft. The AI runs on the assistant you already pay for, such as Claude or ChatGPT. ZoomSphere supplies the calendar, the workspaces, the approval rules and the data.

Check what your plan allows first

A task that only reads (reports, approval checks) works almost everywhere. A task that creates drafts needs a connector with write access, and that is where plans differ.

Both vendors change plans often, so check yours before you build a task that writes.

Write the task yourself, or let the AI write it?

Let the AI write it, and you check it. Claude has this built in: in Cowork choose "New task" and "Create with Claude", or describe the job in any chat. Claude asks follow-up questions, proposes a task name, a schedule and the instructions, and waits for your confirmation. The AI is good at wording and structure. It does not know your client, your statuses or what must never happen, so give it a short brief:

Before you switch the schedule on, read the proposed instructions line by line (workspace name, status names and hard limits, word for word), run the task once on demand, and save the final version where your team can see it.

Five scheduled tasks every agency can copy

Each prompt works as it is. Replace the words in brackets, paste it into a scheduled task and set the time. The two tasks that create drafts include a duplicate check, because a task that runs twice (a test plus the scheduled run) would otherwise create two identical drafts.

More prompts, including one-off tasks, are in the ZoomSphere MCP prompt library. For the monthly client deck in PowerPoint, see our guide to automating client reporting with an AI agent.

Permissions: why a task gets stuck at 7:30

A scheduled task runs with nobody watching. If it needs your approval to use a tool, it waits for a click that never comes. This is the most common reason a task "did not run". In Claude, every task has an approval mode and every connector tool has a permission:

• Always allow the ZoomSphere tools the task needs: reading posts and analytics, creating a draft.
• Block anything that publishes or deletes, and tell the task never to set a post to Approved, because in ZoomSphere an approved post goes out on schedule. The prompt is a rule, the permission is a lock.
• On Claude Team and Enterprise plans, an admin may need to allow "Always allow" for connector tools in the organization settings. It is off by default there.

A read-only task is the easiest place to start: it never changes anything, so it can run with the tightest permissions.

How to know a task actually ran

A task that fails quietly is worse than no task, because your team stops checking. Use four layers of control:

1. Run history. In Claude, "Scheduled" in the sidebar lists upcoming and past runs, each with its full result. In ChatGPT, the Scheduled page does the same and can notify you by push or email.
2. A heartbeat line. End every task with one line that is always there, even on a quiet day. If the line is missing, the task did not run, and you notice the same morning.
3. Links, not claims. Ask for the link to every draft. One click shows it in the ZoomSphere calendar.
4. A weekly audit. One more task on Mondays: list all Private drafts created in the past 7 days per client and compare them with what the daily tasks reported.

Keeping ten tasks organized

Two tasks are easy. Ten tasks across six clients turn into chaos unless you set a structure on day one.

• One Project per client. A Claude Project has its own instructions, context, memory and scheduled tasks. Put the brand voice, workspace name, status names and house rules into its instructions once. The memory stays with that client, so nothing leaks from one brand into another.
• One Project for the agency. Cross-client tasks such as the Monday analyst or the weekly audit live there.
• One naming pattern: Client · Job · When, for example "Demo Coffee · Approval nudge · Weekdays 8:00". The Scheduled list then reads like a timetable.
• Spread the times. Ten tasks at 8:00 compete for the same usage limit. Stagger them and run heavy reports before the working day starts.

Go through the registry once a month. A task nobody has read for two weeks gets paused, not left running.

Scheduled tasks in a team

A scheduled task belongs to one person. It runs with that person's AI account, connectors and ZoomSphere permissions.

• Every task has one owner who reads the heartbeat and fixes the task when it breaks.
• Share the Project, not the login. On Claude Team and Enterprise plans, a Project can be shared as "Can view" or "Can edit". Sessions cannot be shared, so send results where the team already works: Slack, email or a shared document.
• Give the agent only the access it needs. The task sees everything its owner sees in ZoomSphere.
• Keep the human steps. Account managers still review drafts in the calendar and clients still approve through the usual approval workflow. Every change stays in the post's activity log.
• Plan the handover. When someone leaves, their tasks stop with their account. Keep the final instructions of every task in the registry so a colleague can recreate it in ten minutes.

Your client and their data

Two questions come up before an agency switches anything on. First, where client data goes: business plans of both Claude and ChatGPT come with commercial terms, and on personal plans you should check the privacy settings for model training. Use the business plan for client work and connect only the workspaces a task needs. Second, whether to tell the client. We recommend yes: one line in the contract or onboarding that AI prepares drafts and people approve everything that is published. AI-generated images and video can also carry labelling rules on some networks, so check them when a task creates visuals.

When not to automate

• During a crisis. Pause every task for that client in one move, then use the kill switch prompt from the prompt library to move the next 48 hours back to Private draft.
• Replies to your community. Answering people is relationship work, keep it human.
• Sensitive topics and dates. Holidays, tragedies, elections: switch news-based tasks off.
• The first week with a new client. Learn the brand by hand before you automate it.

Set it up in 10 minutes

1. Connect ZoomSphere to your assistant: the help centre guides for Claude and ChatGPT take about five minutes. It is included in every ZoomSphere account.
2. Create a Project for the client and put its rules into the instructions.
3. Let the AI draft the task from your six-point brief or one of the five prompts.
4. Set the permissions: always allow what the task needs, block publishing and deleting.
5. Run it once on demand and check the heartbeat line and the draft links.
6. Switch on the schedule and add the task to your registry.
7. Read every run for the first week, then trust the heartbeat and the weekly audit.

{{cta-component}} 

Frequently asked questions

Can an AI agent post to social media automatically?

It can, but for a brand or a client it should not. Let scheduled tasks create drafts and let a person approve. You keep the speed and remove the risk.

Should I write a scheduled task myself or let the AI write it?

Let the AI draft it from a short brief, then check the workspace name, status names and hard limits yourself and run it once on demand before switching the schedule on.

How do I know my scheduled task actually ran?

Check the run history, end every task with a heartbeat line that appears even on quiet days, ask for links to every draft, and run a weekly audit that compares drafts with what the tasks reported.

Should I use Claude Projects for scheduled tasks?

Yes. Use one Project per client so instructions, memory and tasks stay with that client, plus one agency Project for cross-client tasks.

How do scheduled tasks work in a team?

Each task runs under one owner's account and permissions. Share the Project with colleagues, send results to Slack or email, and keep every task in a shared registry for handovers.

Which AI assistants can run scheduled tasks with ZoomSphere?

Any assistant that supports MCP connectors and scheduled runs. Claude supports tasks that create drafts on all paid plans. In ChatGPT, writing through a custom connector needs a Business, Enterprise or Edu plan as of September 2026.

Do scheduled tasks use up my AI plan?

Yes. Every run uses your plan's usage like any other task. Keep each task focused on one job and spread tasks over the morning.

Does it cost extra in ZoomSphere?

No. The ZoomSphere MCP server is included in every ZoomSphere account. The AI runs on your own assistant subscription.

Sources: Claude Help Center articles on scheduled tasks, approval modes, Projects and Team and Enterprise settings; OpenAI Help Center on scheduled tasks and MCP apps. Checked on 28 September 2026.