61%
Active work where some bounded AI execution is allowed
Execution With Boundaries
Assign work to AI only when the scope, controls, and write-back rules are explicit.
The AI execution handoff pattern is now route-native in `apps/web`, making bounded AI delegation part of the real frontend instead of only a demo pattern.
27%
Handoffs requiring named human approval before use
12%
Work items not eligible due to severity, data, or control limits
Required
Every AI hand-off records scope, result, and owner context
The Enterprise Action Pattern
The action should read like governed operations, not casual chatbot usage.
| Execution Class | Meaning | Output | Guardrail |
|---|---|---|---|
| Evidence Gathering | Collect telemetry, prior tickets, known-good baselines, and comparative incident context. | Evidence pack attached to ticket or runbook step. | No unsupported inference when sources are incomplete or contradictory. |
| Drafting And Summaries | Prepare status notes, decision briefs, approval summaries, or vendor escalation packets. | Human-reviewed draft, never autonomous send. | Only factual, source-grounded content may be drafted. |
| Checklist And Validation Support | Prepare standard checklists, pre-check steps, rollback notes, and validation packets. | Structured work package attached for operator use. | Checklist assistance may not replace the human validation decision itself. |
| Ticket Write-Back Support | Update notes, action history, recommended next steps, and linked evidence. | Drafted or governed write-back with operator attribution preserved. | AI may not close tickets, mark approvals complete, or change service state. |
Assign To AI Operation Flow
The hand-off should be easy to understand and hard to misuse.
Select Scope
Operator chooses the exact bounded task: gather evidence, draft update, prepare packet, or create checklist.
Check Eligibility
Platform verifies severity, data class, AI tier, and whether the task is allowed under current policy.
Confirm Controls
Operator sees the scope, blocked actions, expected output, and whether approval is required before use.
Review Output
AI returns the bounded result, which is reviewed, accepted, rejected, or escalated by the human owner.