Back to Glossary
Escalation Design
Synonyms: escalation paths, escalation policy, tiered support design, human-in-the-loop approval, escalation matrix
Do not index
Definition
Escalation design defines when a problem or decision needs to move to a higher level of support or authority, such as transferring a chatbot conversation to a human or requiring approval before an AI agent takes action. It sets the triggers, approval levels, response time limits, and what the user sees during the process. Good escalation design ensures that complex or high-risk situations are handled appropriately and users understand what happens next.
Use cases
Escalate too rarely and risky actions run unchecked. Escalate too often and people stop reading the prompts.
- No approval step: An AI agent cleans up "unused" database tables and drops one production still needs. Nobody was asked.
- Approval fatigue: An agent asks permission for every file it reads. The user clicks "Allow" 40 times, then approves the one action that deletes their work.
- Silent stalls: A ticket waits in the tier 1 queue for 3 days, and nothing alerts anyone. The customer posts about it publicly first.
How it's used in practice
- Sort actions by risk: Reading runs freely, editing runs with undo, and deleting or paying asks first. Claude Code's permission modes work this way.
- Set a timeout for every level: If nobody responds in 30 minutes (PagerDuty's default), notify the next person.
- Write specific approval prompts: Name the action and what it will change. Use buttons such as "Delete 3 files" and "Keep files" in place of Yes/No.
- Let users escalate themselves: Keep "Talk to a person" visible at every step.
Split escalations into "expected" and "unexpected" in your reports.
A user asking for a person is fine. An escalation caused by a bot error is a bug, and mixing the two hides which topics need fixing.
Challenges & limitations
- Threshold tuning: Every product needs its own approval threshold, and you only find it by watching real usage.
- Stale ownership: On-call lists go out of date when people change roles.
- Users can't see progress: After escalating, most products show nothing until a human replies.
Free resources
- Human-in-the-Loop (OpenAI Agents SDK) — how agent runs pause for approval and resume after a decision
- Choose a Permission Mode (Claude Code Docs) — a real example of tiered approval, from "ask for everything" to fully automatic
- Ticket Escalation: 8 Best Practices (Intercom) — 6 escalation types, from hierarchical to SLA-driven
- Feedback + Control (Google PAIR) — balancing automation with user control in AI products
