← Platform
Human in the loop
Agents escalate by design, not by accident.
Define exception rules once. Low confidence or policy flags pause the run and route to a reviewer with full context.
Pending approvals
· 4
Assigned to you
INV-2847
⚠ POLICY FLAG
Acme Textiles · $147,200
Amount exceeds standard policy threshold ($100,000). Vendor in good standing;
prior invoices cleared in 2–4 days.
Agent confidence 62%
· Waiting 18m INV-2849 Metro Wholesale · Missing PO reference
71%
INV-2851 Alpha Industries · Sanctions match (partial)
58%
INV-2853 Acme Textiles · Possible duplicate (INV-2811)
64%
Routed via policy rules Full audit trail
4 approvals pending · policy threshold, sanctions, duplicate invoice
How governance works
Humans on exceptions. Agents on routine.
When to ask a human
Confidence thresholds and policy flags — declarative, testable, versioned.
Details
Reviewer sees the reasoning
Chain-of-reasoning, inputs, rationale — review in minutes.
Details
Every decision becomes signal
Overrides feed evaluation. Policy rules evolve without prompt maintenance.
Details
Audit & compliance
Every decision, reconstructible on demand.
Who approved what, when
User, timestamp, decision, rationale — on every action.
Details
Full agent context
Prompt, tool calls, outputs — captured on review.
Details
Policy version tied to run
Recreate exact policy state for any past decision.
Details
Export to SIEM
OpenTelemetry spans export to SIEM and compliance pipelines.
Details
Governance the reviewer and auditor both trust.
Map your exception rules in a scoping session.