Learning objective
By the end of this lesson, you can replace a reminder-based AI safeguard with a prevention, detection and recovery design.
Why it matters
A fluent draft can blur the boundary between recommendation and action. If one click can send, pay or change a record, a tired reviewer and an overconfident model share the same error opportunity.
Core concept: enforce the condition around the model
Critical controls should not depend on a prompt being obeyed. Build them into permissions and workflow:
- prevent invalid data or tools from being selected;
- validate required conditions independently;
- hold consequential action for explicit approval;
- expose evidence and uncertainty;
- make reversal and audit practical.
The correct action should be easier than bypassing the control.
Visual explanation
Text alternative: an AI draft cannot act until data and rules are validated. Invalid cases are held; valid cases require explicit approval, bounded execution and audit-supported recovery.
Worked example
A payment assistant reads supplier bank details from email attachments. An on-screen warning asks reviewers to check them.
Replace the warning with an allowlisted source: only the verified supplier master can provide a destination. A requested change triggers a separate verification route. Drafting and sending use different permissions, duplicate submission is blocked and every override requires evidence.
Common mistake
Mistake: designing a hard stop with no legitimate exception route.
Users facing valid unusual cases will work outside the system. Provide a controlled override with authority, evidence and monitoring.
Quick check
Which is the strongest guardrail for an AI-generated payment instruction?
A. A longer prompt.
B. A general warning.
C. A verified destination allowlist plus separate approval and execution permissions.
D. Monthly retraining.
Answer: C. It changes the action conditions and can be tested independently of model behaviour.
Practical prompt
Take one AI action in your workflow. Identify the point where a draft becomes an external change. Add one prevention rule, one detection or hold rule and one reversible recovery mechanism.
Summary
- Intervene at the exact error opportunity.
- Prefer enforced constraints to warnings.
- Separate drafting, approval and execution.
- Design accountable exceptions.
- Monitor escaped errors, false blocks and overrides.
Next lesson
Apply the method in Prevent the Wrong-Account Payment, a case about supplier data, permissions and safe override design.