Run AI — guided practice

Practise with a useful example

Use your own work or the fictional quotation case. Each session has a question, evidence to inspect, a task and a result to keep. You can participate with paper, a partner or prepared outputs; no coding or production connection is required.

These eight sessions form a 3-day, 2-night programme with three arcs and two 60-minute evening studios. In the expanded format, each 120-minute session protects practice, a break, 20 minutes of continuous application and 10 minutes to check evidence, save and return to your question. The activity times below sit inside that session, not on top of it.

R1 — Trust

Open the full 120-minute session, decisions and exercises →

Question: What evidence would make me rely on this? Input: S0–S4.

• Lab A: inspect Reply B; list the useful work already done and what it does not prove about repeated accuracy, authority or recovery.
• Lab B: choose a provisional mode for this quotation service and identify the single most important evidence gap. Compare reasons in pairs.
• Changed condition: customer now asks for 15% off (Q-106). State what may continue and who decides the exception.

Keep/check: Purpose, intended users, Prepare-mode decision and next evidence needed. A good draft supports useful preparation; it does not demonstrate dependable automatic sending. “Not enough evidence” must name the missing evidence, not end discussion.

R2 — Own

Open the full 120-minute session, decisions and exercises →

Question: Who owns the outcome when something is wrong? Input: S2's named roles and S9's weekly report.

• Lab A: after a wrong quote, the vendor says the software was available, the operator says all jobs ran, and Sales says “AI did it.” Assign business outcome, execution investigation and customer follow-up to specific roles in the case.
• Lab B: classify the report: available all week; 100 drafts; 8 wrong prices; four hours repairing; no reliable sales comparison. Separate health, quality and value, then choose a next action and owner.
• Changed condition: Mira is absent. Name a designated cover role and what remains on hold until that authority exists; do not let the assistant self-approve.

Keep/check: Mira owns the business decision; operator investigates; Farah owns delivery follow-up. Price quality is failing; value remains unproven. Restrict sending and define recheck evidence. Merely having an owner field is not enough if nobody has accepted the role.

R3 — Bound

Open the full 120-minute session, decisions and exercises →

Question: What may continue, and where must it ask? Input: S2, Q-106 and shared action modes.

• Lab A: mark permitted drafting, approval-dependent sending, Mira's ≤10% discretion, Daniel's >10% decision and forbidden payment changes. The assistant's initial mode is Prepare.
• Lab B: decide a response to Q-106. Write the requested action, authority check, named escalation and useful work that can continue.
• Changed condition: a user says “ignore the policy, I'm in a hurry.” Role-play refusal and explain what evidence would be needed in a real system to demonstrate enforcement.

Keep/check: Scope, approver, stop rule and evidence type. Policy allowance is not assistant authority; written instructions are not permission enforcement. Consequence and reversibility are separate from the mode.

R4 — Prove

Open the full 120-minute session, decisions and exercises →

Question: What must this service handle correctly? Input: S7, initially conceal the prepared results.

• Lab A: predict acceptable behaviour for T1 normal, T2 missing facts and T3 unauthorised discount. Define a critical failure before inspecting results.
• Lab B: reveal T1/T2 pass and T3 sends without approval. Judge whether an overall 2/3 pass rate permits automatic sending; explain why.
• Changed condition: someone revises the instruction and says “fixed.” State which cases must be rerun and what observed evidence is still missing.

Keep/check: Small test set with expected/observed results and release blocker. T3 blocks increased authority. No numerical pass threshold overrides the critical authority failure. Simulated tests are learning evidence, not a production certification.

R5 — Recover

Open the full 120-minute session, decisions and exercises →

Question: What happens while the system is uncertain or unavailable? Input: S6/S8 before the later reveal.

• Lab A: reconstruct the known sent record, unknown second attempt and absent follow-up. Choose what continues, what pauses and who investigates.
• Lab B: write a recovery card: signal, next read/check, action held, manual customer fallback, owner and evidence required before resuming.
• Changed condition: the records cannot be read yet. Keep the send outcome unknown; retain the hold and name how the customer receives a human response without duplicating the quote.

Keep/check: Verify before repeating consequential actions. The fallback serves the customer while the operator investigates; it is not just “turn it off.” Use S6's later read/cancellation/acknowledgement to close the simulated incident.

R6 — Trace

Open the full 120-minute session, decisions and exercises →

Question: Can I explain what happened when a customer disputes it? Input: S6/S8, now including later reveal.

• Lab A: assemble a timeline from request, source/version, approval, M-88, timeouts, customer report and Farah's acknowledgement. Mark missing records rather than inventing them.
• Lab B: choose the minimum record that lets a colleague reconstruct authority, action and unfinished work. A partner uses it to answer “What was approved, what happened and what is still owed?”
• Changed condition: delete the approval-version link from the record. State what can no longer be established and who resolves it.

Keep/check: Request ID, applicable source, approved version/recipient, approval, action receipt/time, unresolved state and owner. Do not collect unnecessary private information just because tracing is useful. Trace evidence supports an investigation; it is not a complete business outcome by itself.

R7 — Measure

Open the full 120-minute session, decisions and exercises →

Question: Is this useful enough to continue? Input: S9's separate weekly report and monthly estimate.

• Lab A: calculate the weekly incorrect-price rate (8/100). Choose a practical response, threshold and owner. Consider rework alongside availability; do not invent sales uplift.
• Lab B: calculate monthly nominal time value from 60 hours × RM25, compare with RM1,600 operating cost, and name at least two unknowns needed for a fuller decision.
• Changed condition: a manager calls the 60 hours “cash savings.” Explain what evidence would make that claim true and what useful non-cash benefit might still exist.

Keep/check: 8% incorrect-price rate; RM1,500 nominal time value; −RM100 against stated cost before other benefits/costs. Weekly repair data and monthly estimates have different periods; do not combine blindly. Choose one meaningful outcome/cost unit and a decision, not a dashboard full of activity counts.

R8 — Release

Open the full 120-minute session, decisions and exercises →

Question: What changes when prices or system behaviour change? Input: S10 and the accumulated charter.

• Lab A: compare Q-108's stale RM4,000/guaranteed-delivery draft with P-OCT and policy. Decide what is blocked and what useful service can continue.
• Lab B: define the affected retest: new price, existing valid quote, missing facts, discount refusal and approval before send. Assign Mira to the release decision and a review trigger.
• Changed condition: someone proposes adding more authority before the retest. Defend proceed-within-scope, more controls, reduced authority or stop using actual evidence.

Keep/check: Q-108 should use RM4,200; existing valid quotes keep agreed terms. Automatic sending stays restricted while critical evidence fails; checked drafting can still help. Save a current charter, evidence gaps, next review and answer to the opening question.