Overview
How might we design escalation thresholds so people can trust and act on AI output?
When to use
- Essential for agentic commerce, finance, admin automation, and enterprise agents where unattended execution is acceptable only within a bounded envelope of risk.
When to skip
- Products that always require approval and have no autonomy mode to demote from.
- When thresholds cannot be measured reliably and false triggers would freeze work.
- Fully manual tools with no agent autonomy.
Rules
Invisible thresholds the user cannot see or configure.
Escalating without showing which threshold fired.
Thresholds so low that every action escalates and autonomy is fake.
No way for admins to set different thresholds per environment (dev vs prod).
Evidence
| Product | Implementation |
|---|---|
| Enterprise agent platforms | Policy engines that require approval above spend or data-class tiers. |
| Banking / fintech AI | Hard dollar and fraud thresholds before execution. |
| Customer support agents | Auto-escalate to humans on toxicity, VIPs, or refund limits. |
| Cloud ops agents | Prod-write thresholds that demote to plan-only mode. |