ChatGPT Scheduled Tasks Are Unattended GTM Jobs

A ChatGPT scheduled task is an unattended GTM job. It runs while you are in another meeting, with the plugins, apps, and (on desktop) computer-use access you already granted. OpenAI's scheduled-task docs say these runs happen in the background with your default sandbox settings, and desktop scheduled tasks use approval_policy = "never" when organization policy allows it (Scheduled tasks). Treat the saved prompt like a runbook. Test it in a live chat first. Keep CRM writes off. Review Scheduled after the first runs.

If you treat the feature as a nicer reminder, you will ship a weekday 7:30 "pipeline recap" that still has HubSpot, Slack, and Gmail attached because the live chat needed them.

What is a ChatGPT scheduled task?

A scheduled task is an instruction ChatGPT carries out later, once, on a repeat, or when a supported app event fires. You manage active, paused, and completed work from Scheduled. OpenAI's July 9, 2026 ChatGPT Work launch framed this as work that continues when you are away from the computer, including turning new Slack or Teams messages into updated docs, checking dashboards each morning, and refreshing a recurring meeting agenda (ChatGPT Work).

That is a different product from the ChatGPT workbench described in how to use ChatGPT for GTM operations. A live chat has you in the loop. A scheduled task does not, unless an action pauses for approval.

OpenAI now reports more than 5 million people using Codex every week, and more than 1 million using it for work outside software development. GTM operators are in that second group. The scheduled-task surface is how those non-coding jobs start running overnight.

There are a few shapes, and they are not interchangeable.

Codex still has its own automations. Do not mix those docs with the ChatGPT Scheduled page. If the leftover is a GTM brief, stay in ChatGPT Work. If the leftover is a repo change, use Codex the way GTM operators already should, with a worktree and a review step.

What can a scheduled task reach in GTM work?

More than the reminder UI suggests.

On web, a scheduled task can use uploaded files, connected tools, skills, and plugins available to that chat. It does not keep a local folder between runs. Durable instructions belong in the task prompt or a skill (Scheduled tasks).

On desktop, the same feature can work in a Git worktree or in your live checkout. Running in the local project can change files you are still editing. Non-versioned folders run in place. OpenAI's sandbox note is the operating constraint. Read-only blocks file writes, network, and computer apps. Workspace-write stays inside the project. Full access can change files, run commands, and hit the network without asking.

Computer Use is the extra blast radius. On desktop, ChatGPT can see and operate macOS or Windows apps. OpenAI says you can attach that to a scheduled task when recurring work includes steps on your computer (ChatGPT Work, Computer Use). If HubSpot is open in Chrome and you chose Always allow, the overnight job can click the same signed-in session you use during the day.

Do not use Computer Use for HubSpot writes. Prefer a structured plugin or no CRM connection at all. If the job needs a screenshot of a dashboard nobody has exported, keep the task read-only, name the app, and stay present for the first runs. OpenAI's own safety note is to review website actions as if you were taking them yourself, because the site will treat approved clicks as coming from your account.

Connected-app permissions are a separate lever. Always ask, Allow read actions, Allow low-risk actions, and Allow all actions change when ChatGPT prompts. They do not add new access. Allow all actions is documented as elevated risk because supported actions may run without another confirmation (Connected apps). For GTM, keep HubSpot, Salesforce, and outbound email on ask or off.

MCP connectors still need GTM governance before they sit on a clock. A scheduled task inherits whatever you already connected. It does not invent a narrower HubSpot server for you.

How should GTM operators write a scheduled-task prompt?

Write a runbook. Do not write a vibe.

OpenAI's own test instruction is the whole method. Before you schedule, run the prompt in a regular chat. Confirm the scope, the model, the tools, and that the output is reviewable. Then watch the first scheduled outputs and tighten the prompt (Scheduled tasks).

A GTM prompt that survives 8am without you looks like this.

• Outcome. "Produce a markdown stalled-deal brief for deals with close dates in the next 14 days."
• Sources. Name the export, the Drive folder, the Slack channel, or the dashboard. Do not say "use whatever you can find."
• Method. Point at a skill with $skill-name on desktop, or paste the steps. OpenAI's line is that skills define the action and scheduled tasks define the schedule.
• Output. Sections for risk, evidence, missing context, and recommended human question. No CRM field updates.
• Stop rules. "If owner, amount, stage, or close date would change, stop and ask. If the source is missing, say so."
• Cadence. Weekday 8am is enough for a pipeline packet. Event triggers are for inbound mail and channel spikes, not for a daily recap.

Put durable GTM rules in context files or a skill, not only in the chat that created the task. A task scheduled inside a chat reuses that chat's context. That is useful for a review loop. It is a trap when the chat also contains a one-off "update these three deals" instruction.

Team tasks are stricter. They run under the team's service account. They do not inherit the creator's personal memories, Custom Instructions, or chat history. OpenAI tells you to put sources, work, expected output, and review criteria in the saved instructions (Team tasks). An unattended run cannot complete a new app sign-in for you. If the HubSpot connection is your personal login, the shared 8am job is still you.

"A scheduled ChatGPT task is a job that runs as you. If you would not let a contractor write HubSpot while you sleep, do not give the task a CRM write plugin." Sebastian Silva, Founder, HigherOps

You cannot automate an undefined RevOps workflow. Scheduling does not fix a missing definition of stalled. It only repeats the missing definition five days a week.

When should you use event-triggered tasks?

Use them when the work should start because something happened, not because the clock said 8am.

OpenAI's connected-apps page lists three supported events for Work tasks. New Gmail messages. New Slack channel messages. GitHub pull-request activity in authorized github.com repositories. The task can use only the connected account and the permissions already granted. If an action needs approval, the task pauses until you review it.

Slack has extra setup. Add @ChatGPT to each monitored channel. Direct messages, reactions, edits, and deletions do not trigger tasks (Using Slack in ChatGPT).

Good GTM fits.

• A new inbound Gmail from a known domain produces a draft evidence packet, not a lifecycle change.
• A #deal-desk Slack message produces a recap with source quotes, not a stage move.
• A pricing-page GitHub PR produces a changelog note for enablement, not a HubSpot product update.

Bad fits.

• "When any email arrives, update the contact."
• "When sales posts in Slack, mark the deal commit."
• Any trigger paired with Allow all actions on a CRM plugin.

Event-triggered volume is the part operators skip. OpenAI's help article caps those tasks at 30 runs per hour and 720 per day across all of them, and it may group events (Scheduled tasks in ChatGPT). That is enough to flood Scheduled and enough to send a lot of drafts if the prompt is sloppy. Start with a tight sender or channel filter. Review the first afternoon of runs before you widen it.

Enterprise, Edu, and Healthcare admins have to enable event-triggered tasks. Healthcare keeps them off by default and outside the BAA. Do not point a GTM webhook task at patient or health data.

What should stay off a scheduled GTM task?

Keep the first version in the same bucket as what to automate first with AI in RevOps. Repetitive, high-context, easy to verify. No customer-state change without a human.

Leave these off the overnight job.

• HubSpot or Salesforce writes for owner, stage, amount, close date, lifecycle, or forecast category.
• Outbound send. Sequence enrollment. Customer-facing email from the agent's voice.
• Computer Use against a signed-in CRM tab.
• Full-access sandbox on a laptop that also holds production exports.
• Personal CRM logins on a team task.
• A prompt that says "fix whatever looks wrong."

Keep an audit trail even when the agent only drafts. The Scheduled run is the trigger. The packet is the decision record. The human apply step is the write. If you cannot reconstruct those three later, the job is not ready to repeat.

Route models by RevOps risk still applies. A scheduled recap can use a cheaper model. A task that can send Slack or touch a CRM plugin should not. OpenAI lets you pin model and reasoning effort on the task. Pin them. Do not leave a retired model on a job that still fires. GPT-5.5 left ChatGPT, ChatGPT Work, and Codex on October 14, 2026. That retirement does not affect API access. Review the model field on every saved task that still names it (Model release notes).

Frequently asked questions

Are ChatGPT scheduled tasks safe for HubSpot updates?

No. Use them to draft. A human applies the change, or a scoped integration path does, after review. Connected-app permissions and sandbox settings are not a CRM write policy.

Do I need to leave ChatGPT open?

For team tasks, no. They run in the cloud. For desktop tasks that need local files or Computer Use, yes. The machine stays on, the app stays running, and Windows Computer Use needs the target app visible.

Should the prompt live in the task or in a skill?

Both. The task holds the cadence, the sources for this run, and the stop rules. The skill holds the method the team will reuse. If the workflow still needs a lot of steering, do not schedule it yet.

How is this different from Codex automations?

Codex automations belong to coding work in a repo. ChatGPT scheduled tasks belong to briefs, recaps, and monitoring across connected apps. Pick the leftover first. Then pick the surface.

Key takeaways

• A ChatGPT scheduled task is an unattended job with your plugins attached, not a calendar reminder.
• Test the prompt in a live chat, then schedule. OpenAI tells you to do that for a reason.
• Keep CRM writes, outbound send, and Computer Use on a signed-in HubSpot tab off the first version.
• Put the method in a skill and the cadence in the task. Durable rules do not belong only in the chat that created the schedule.
• Event-triggered Gmail, Slack, and GitHub tasks can fire all day. Filter hard. Review Scheduled like an inbox.
• Team tasks run as the team's connection, not as your personal memory. Name the sources in the saved instructions.

If you want this installed as an operating rule instead of a ChatGPT setting, start with AI enablement for GTM teams.