Decide What an AI Automation May Do and Where It Must Stop
On this page
An automation can draft useful work and still need a person to approve the next action. Clear boundaries say what the system may read, what it may produce and where it must stop. This lesson helps you write those boundaries before connecting any tools.
Practice time: About 15 minutes. Use any text-based AI assistant you already have access to. Interface names and available tools vary. The examples below are fictional, and you do not need a paid plan or private information to practise.
Start with a small, clear example
A fictional club wants an assistant to prepare reminder messages for overdue library books. The first version should use a made-up list and create drafts only. Sending messages, changing records and looking up personal contact details are outside the practice task.
Try this prompt
Copy the following into your assistant. Keep the sample input attached to the request so the task can be understood without guessing. Treat the response as a draft to inspect, not as a finished answer simply because it sounds confident.
Design a draft-only workflow for fictional overdue-book reminders. List allowed inputs, draft output, required human checks and stop conditions. Do not send messages, modify records or retrieve contact details. If a due date or book title is missing, flag the row for review instead of inventing it. Include a short test plan.
Work through the result
Separate creation from execution. Producing a draft reminder is not the same as sending it. The workflow should have a visible approval checkpoint after checking recipient, title, date and wording.
Define what a missing field does. A safe draft-only design can place incomplete rows in a review list while processing complete fictional rows. It should not quietly guess a date to make the batch appear finished.
Test at least three cases: a complete record, a missing due date and a duplicate record. Decide whether duplicates should be combined or held for review. Keep the original input and the output log so you can trace a draft back to its source.
Catch the common mistakes
“Use your judgement” is too vague for consequential external actions. Do not assume that access to a tool means permission to use every function it offers. Start with a narrow workflow, verify it and expand only when the owner explicitly agrees on the next capability.
Check before you reuse it
Compare the response with your original input. Look for a missing condition, a changed number and any information the assistant added without support. Read names, dates and instructions particularly carefully. If the task involves an action, distinguish a suggested action from one you have actually completed.
Revise one specific weakness at a time. A follow-up such as “keep the uncertain date marked unconfirmed” gives a clearer correction than “make it better.” Keep the useful parts and check the revised passage again. An improved format does not guarantee improved accuracy.
Try it yourself
A generated workflow says “If the user does not respond within five minutes, send the reminder automatically.” Does silence satisfy the draft-only boundary? Rewrite the step.
Check your answer
No. Silence is not approval, and automatic sending exceeds the stated scope. Rewrite it as “Keep the reminder as a draft until an authorised person explicitly approves sending. If no approval arrives, leave it pending and record that status.” The system should preserve the boundary even when completing the batch would be more convenient.
Keep exploring: AI learning path.