Escalate
When should AI ask for help, and what should it bring?
Learn to choose a task-specific action mode, test normal/unclear/outside-authority cases and escalate a decision with enough context for the next person to act.
Required: your AI Work Delegation Blueprint and either your own workflow or the fictional starter case. AI/GPT is not required for the core simulation.
Run the three supplied test cases as a simulation, identify the release blocker and prepare an escalation package that preserves useful work.
A three-case test, release blocker and escalation package showing what stays human-owned.
This session uses supplied records, partner review and simulation evidence. Do not label a designed or role-play result as live execution.
What you will leave with
A tested escalation package and human-work map. Add this to your AI Work Delegation Blueprint.
Use your own workflow or the fictional quotation case. No coding or live customer action is needed.
Start with a decision · 10 minutes
Show three test labels, two pass and one fails. Ask for a release decision before revealing that the failure sends an unauthorised discount.
Bite A · Choose a task-specific mode
Question: What may this task do now?
Input: Shared modes; Q-104 preparation and Q-106 discount. Open the case cards.
Try · 12 minutes: Choose Observe, Recommend, Prepare, Act after approval or Act within limits for each step. State which evidence and authority would change your choice.
Compare after you have decided
Preparing a quote and approving a discount have different authority. Modes describe permission, not learner proficiency or a maturity score.
Keep: Mode, scope and evidence needed.
Bite B · Test normal, unclear and forbidden requests
Question: Which failure blocks more authority?
Input: S7 T1 normal, T2 missing facts, T3 unauthorised 15% send. Open the case cards.
Try · 12 minutes: Divide cases across the table. Predict expected behaviour, then read the results. Record pass/fail and consequence; compare as a group.
Compare after you have decided
T1/T2 pass, but T3 is critical. Two passes do not compensate for an unauthorised send.
Keep: Three-case test with a release blocker.
Bite C · Escalate a decision, not a vague problem
Question: What can the next person act on?
Input: Change one condition: missing quantity, rejected approval or absent owner. Open the case cards.
Try · 12 minutes: Prepare a handover with request, evidence, blocked decision, next owner and permitted work while waiting. A partner decides the next action from the card.
Compare after you have decided
“Ask a human” lacks a decision and an owner. Preserve useful preparation while the specific blocked action waits.
Keep: Escalation package and retained human work.
Put it together · 20 minutes
Use the final 20 minutes to revise and retest one failed case. Separate the revised design from an observed successful retest. Revise the weakest item and have a partner use it. Keep the result and one limitation.
Close your question · 10 minutes
Return to When should AI ask for help, and what should it bring? and your own opening question. Answered, partly answered or still open? Show the evidence. Save the updated AI Work Delegation Blueprint. A reasoned decision to keep a step human-led is useful work.
Common trap
Do not promote the task’s authority after merely rewriting an instruction. Do not make every novice execute three tools in four minutes.
What if every case asks a person? Revisit the job boundary; it may still provide useful preparation, but it is not autonomous completion.
Where this appears in real work
Leads and booking: use normal, ambiguous and refusal cases; preserve consent and appointment evidence. The class uses the fictional packet; selected project demonstrations and technical walkthroughs can add detail later.