Escalation Rule
A configured condition that hands an agent task off to a human when the agent is uncertain, out of scope, or facing high-stakes input.
Also known as: escalation policy · handoff rule · fallback rule
What is Escalation Rule?
An escalation rule is a configured condition that hands an AI agent’s task to a human when the agent is uncertain, out of scope or facing high-stakes input. Common triggers include a confidence score below threshold, an angry customer, an amount above a policy limit, a legal or medical topic, or no matching knowledge-base article.
Christos Papadimitriou, theagency47 · Updated May 15, 2026Common escalation triggers
Production agents typically have 5–10 escalation rules. Common categories:
- Confidence-based. Below threshold X on retrieval similarity → escalate.
- Sentiment-based. Angry / urgent / frustrated language detected → escalate to senior human, top of digest.
- Topic-based. Legal threats, medical questions, complaints about specific staff → always human.
- Value-based. Refund > €50, contract above standard value → senior approval.
- Identity-based. First-time sender, no prior history → flag for review (anti-impersonation).
- Format-based. Input that does not parse cleanly → escalate rather than guess.
- Eval-based. Input pattern that matches a known-fail eval case → bypass agent, route to human.
What good escalation looks like
A useful escalation includes:
- The original input (full context)
- The agent’s reasoning trace (why it escalated)
- Suggested response (if any) for the human to accept or modify
- Priority level (urgent / standard / FYI)
Bad escalation is “I cannot help with this” with no context, that creates work, not handoffs.
FAQ
How are escalation rules tuned over time?
By production data. If a rule triggers too often (false positives), the threshold loosens. If real escalation-worthy cases slip through, the rule tightens or new rules are added. This is part of the post-launch retainer work.
Can the agent override an escalation rule?
No. Escalation rules are hard gates. If triggered, the action does not happen, full stop.
Related terms
Bounded Autonomy
The engineering pattern of giving an AI agent specific permissioned actions and explicit escalation rules, letting it act independently within tight, defined boundaries.
Human-in-the-Loop
A deployment pattern where every AI agent action is reviewable by a human before downstream effects, used during the first weeks of production.
System Prompt
The instructional text that defines an AI agent's role, voice, boundaries, and decision rules. The single most important customization layer of an agent.
