The 20-Minute AI Escalation Ladder

Map one recurring AI-assisted workflow.

Workflow: [NAME AND OUTCOME]
Trigger: [WHAT STARTS IT]
Inputs: [DATA, DOCUMENTS, OR MESSAGES]
Actions available: [READ / DRAFT / CHANGE / SEND / PUBLISH / DELETE / SPEND]
Downstream systems and people: [LIST]
Known edge cases: [LIST]

Return a compact workflow map with each decision, action, consequence, and current owner. Mark any step where the model could affect a customer, public channel, production system, money, permissions, or irreversible data.

Create an escalation ladder for this workflow map: [PASTE]

Use four levels:
1. ACT — narrow, tested, reversible action inside an approved boundary
2. ASK — missing context or a choice that needs a named human
3. STOP — policy, permission, security, financial, legal, or irreversible risk
4. HANDOFF — human receives the evidence and completes the next step

For every decision point, define observable triggers, allowed action, prohibited action, required evidence, owner, and response time. Never use “low confidence” without a measurable signal or concrete missing fact.

Test this escalation ladder before live use.

Workflow map: [PASTE]
Escalation ladder: [PASTE]

Create 12 cases: four normal, four ambiguous, and four must-stop. For each case include expected level, expected action, required evidence, and automatic failure conditions. Include at least one prompt-injection attempt, one changed destination, one missing permission, and one request with an irreversible consequence.

Turn this approved escalation ladder into a one-page operator card.

Approved ladder and test results: [PASTE]

Return: workflow name, ACT boundary, ASK triggers, STOP triggers, handoff owner, evidence packet, response target, fallback path, logging fields, and review date. Keep it usable during live work. If ownership or evidence is missing, mark the workflow NOT READY.