All interview questions
FDEAdvanced
Design the approval policy for an AI agent that can read and send email on behalf of a sales team.
Think first
Risk tiers in code, a model may escalate but never downgrade, measurable approval quality.
Strong answer
Tier 0 reading and summarising: allowed and logged. Tier 1 drafts, labels and CRM notes: allowed and shown in a daily digest. Tier 2 sending any external email: requires the account owner's approval, with the exact text shown. Tier 3 attachments with pricing, contracts, or anything to a new domain: requires a manager. Rules live in code, evaluated before any model judgement; a model reviewer can escalate a suspicious action but never downgrade a tier. Guard against prompt injection from inbound email content. Measure approval rate, time to decide and post-hoc reversals; move actions between tiers based on that evidence.
Weak-answer indicators
- Lets the model decide what needs approval
- No consideration of inbound email as untrusted input
Likely follow-up questions
- Approval rate is 99.8% with a median decision time of one second. What does that tell you?
Related learning
📖 Related tutorial