Automate Work With ChatGPT: Start With a Reviewed Process

To have ChatGPT automate work usefully, choose one repeatable process with clear inputs, a defined result, and a named reviewer. Run it manually first, test the awkward cases, and only then consider a schedule or supported trigger. Automation is easier to trust when you can tell what happened and stop it without losing the work.

fenbs can keep the procedure, human-approved rules, task state, and lessons from each attempt. It holds the work record; it does not run the automation. A good first example is preparing an internal weekly update from approved project notes, with a person reviewing the draft before it reaches colleagues.

How do you choose the first process?

1. Pick work you can describe in one sentence, such as turning approved weekly notes into a draft status update.
2. List the allowed inputs and the exact output. State the project, reporting period, audience, and destination.
3. Name the reviewer and the action that needs their approval. Keep preparing a draft separate from publishing it.
4. Write what happens if information is missing, contradictory, or inaccessible.
5. Run the process by hand with representative inputs, compare the result with the sources, and record the outcome.

Choose a task whose errors you can recognize. If you cannot check whether the result is complete, improving the prompt will not solve the underlying problem. Narrow the scope until a reviewer can compare every important statement with its source without recreating an entire investigation.

What can ChatGPT run, and what needs another tool?

OpenAI describes ChatGPT Work↗ (opens in a new tab) as its experience for longer, multi-step work. Features available in your account and workspace determine the tools a process can actually use. A written instruction to check a source does not establish a working connection to that source.

For recurring shared work, OpenAI documents team tasks↗ (opens in a new tab) that use saved instructions and approved workspace connections. Check its current availability and access requirements rather than assuming a task can run under a colleague’s personal account or reach every company document.

Choose the execution method after the process is clear. You may only need to run the same reviewed request when new notes are ready. If you need a schedule, event, or outside action, confirm that the selected tool supports it and that somebody owns its maintenance. The beginner usage guide covers getting started with ordinary requests.

What should the procedure say?

Write the procedure as instructions a colleague could follow. Define which source wins when two notes disagree, what counts as completed, and how to show missing evidence. Avoid vague goals such as “keep everyone informed.” State the result you want to review: a dated draft with completed work, open blockers, next steps, and source references.

Keep stable instructions separate from run inputs. The reporting period changes each time; the requirement to avoid invented completion claims should remain. Store that human-approved rule in fenbs Decisions and rules, so a later assistant can distinguish the standing instruction from one run’s temporary details.

How do you test more than the easy case?

Start with a normal input, then try cases that should cause the process to pause or qualify its answer. For the weekly update, remove an owner, provide two conflicting dates, include an old completed task, and make one source unavailable. Decide the expected response before you read the generated output.

• Normal case: the draft covers the right project and period with accurate source references.
• Missing source: the result names the gap instead of quietly presenting a complete report.
• Conflicting facts: the draft flags the disagreement and leaves the decision to the reviewer.
• Repeated run: the same period does not accidentally become a second published update.
• Interrupted action: the person checks the destination before retrying anything that may already have happened.

Keep the test result with the procedure: what you supplied, what happened, and what still needs checking. “Works” is too broad if you only tried clean notes. If the first revision causes a new omission, rerun the affected cases before treating the process as ready for unattended preparation.

Where should human review stay?

Put review where a mistaken result creates a real commitment: sending a message, publishing a claim, changing a shared record, or spending money. Describe what the reviewer must inspect and what approval covers. Approval of one draft should not silently become permission for every future version.

Tool permissions also matter. OpenAI’s connected-app documentation↗ (opens in a new tab) explains that available actions depend on the connection and workspace controls. Check the actual settings before enabling a recurring process. Instructions in a prompt should reinforce the intended boundary, while the account’s access limits which actions are available.

For general workflow choices, see how to use AI at work; use the work prompt examples when you need help expressing a specific deliverable. The automation decision comes after you can produce and review that deliverable reliably.

How to do this with fenbs

1. Create a free account and add a project for the process you want to improve.
2. Write the inputs, expected output, exception handling, and reviewer in Routine processes. Record the person’s approval rules in Decisions and rules.
3. Create tasks for setup gaps and failed test cases. Keep the result of each check with the relevant task.
4. Connect a supported assistant through Integrations, and ask it to read the project context before working on the procedure.
5. Save the current state and next step in Where we left off. Add a lesson when a run exposes a mistake that should change future work.

fenbs is not the right fit when you need a workflow runner, a schedule engine, or a service that executes background jobs. Use the appropriate execution tool and keep its reviewed instructions and outstanding work in fenbs. It is free for one person, no credit card.

Start with one process and one reviewer. After several reviewed runs, compare the effort saved with the effort spent checking and correcting. Keep the process manual if exceptions dominate. The AI at work library can help you improve the underlying tasks before extending automation to more of the job.