Humans on the right decisions, agents on everything else
Every agent has an escalation policy. When confidence is low, the stakes are high, or a guardrail trips, the right human gets routed a complete context package — not a bare question.
Fully autonomous is rarely the goal. The goal is autonomous where it's safe and assisted where it's not. Escalation is how you tune that balance — per workflow, per action, per customer tier — and how the system gets smarter from every correction.
The escalation path
- 01
Decide when to escalate
Confidence thresholds, action cost, customer segment, and policy rules are checked on every step. Low signal? Escalate.
- 02
Package the context
Reviewers get the question, the agent's plan, the evidence, and one-click approve/revise buttons — no hunting through logs.
- 03
Learn from the outcome
Human corrections feed back into the agent's evaluation harness so the same class of mistake isn't escalated next time.
Capabilities
Policy-driven routing
Escalate a $10 refund auto-approve but a $10,000 refund to a manager — same agent, different routes, transparent rules.
In-tool approvals
Reviewers approve from Slack, Teams, email, or the platform — whichever surface they already live in.
Correction capture
When a human edits an agent's draft, the diff is captured as a labeled training signal — not lost in a tool.
SLA visibility
Escalation queues expose age, volume, and resolution time so ops sees when thresholds need tuning.