Know When to Stop

Work through the explanation, apply it to the example, and complete the quick check before continuing.

Not using AI can be the correct design

Do not continue merely because a tool is available or work has already begun.

Stop or use another method when:

• The task is prohibited by policy or outside your authority.
• The tool is not approved for the required information.
• Necessary evidence or specialist review is unavailable.
• Harm could be serious and cannot be controlled.
• A high-impact decision would be delegated without meaningful human authority.
• Prompt-injection or tool-action risk exceeds tolerance.
• Required accessibility cannot be provided.
• Rights, licence, or consent questions remain material and unresolved.
• There is no owner, escalation route, monitoring, recovery, or safe shutdown.

Use the control ladder

1. Remove unnecessary data or action.
2. Reduce scope and permissions.
3. Use a safer tool or fixed process.
4. Add evidence checks, testing, approval, monitoring, and recovery.
5. Keep the work human-led.
6. Stop the task.

Controls should reduce a named risk. Adding many controls that do not address the real failure creates false confidence.

Write an explicit stop condition

Weak:

Be careful.

Strong:

Stop before upload if any application contains personal data not approved for this system. Send the case to the data owner. Do not continue with a redacted copy until the owner confirms the method.

Remember

• Stopping is an active risk decision, not a failure of imagination.
• Reduce data, scope, permissions, and action first.
• Use controls tied to named risks.
• State who receives the escalation and what remains prohibited.

Now you will apply the full method to a synthetic case.