Give AI Work — guided practice
Eight practical sessions with supplied inputs, evidence checks and a useful takeaway.
Practise all eight Give AI Work decisions using supplied records or your own workflow without needing a live customer system.
Required: your AI Work Delegation Blueprint and either your own workflow or the starter case. AI/GPT is not required for the core simulation.
Work through the session decisions, inspect the evidence, label simulated versus observed results honestly and save each decision in the blueprint.
Eight worked decisions plus a tested delegation blueprint showing action, authority, evidence, escalation and handoff.
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.
G1 — Authorise
Open the full 120-minute session, decisions and exercises →
Useful question: What can AI actually do for me? Input: S0, S2 and S6.
Show the benefit first: Q-104 becomes a checked draft and an explicit request for approval instead of another loose message. In a five-minute role-play, move the training record from draft to approved to sent, using Mira's version-specific approval and M-88. Have a learner hold Farah's acknowledgement separately. Say “This is a simulation; nothing has been sent outside class.”
- Bite A: in 12 minutes identify draft creation, sending and follow-up ownership as separate changes. Mark which need external authority.
- Bite B: in 12 minutes choose Prepare or Act after approval for Q-104 and identify who approves what.
- Bite C: in 12 minutes repeat with Q-106's 15% request. Preserve preparation but escalate the discount to Daniel; the assistant cannot approve it.
Keep/check: Blueprint names useful action, target, mode and approver. Show which state changed, why it was allowed and what remains human-owned. Use designed/simulated/observed labels accurately.
G2 — Contract
Open the full 120-minute session, decisions and exercises →
Useful question: What exactly am I handing over? Input: Q-104/Q-106, S2 and S6. The proposed action is a contract field; approval must precede execution.
| Activity | Do this in 12 minutes | Keep and reveal |
|---|---|---|
| G2A — target and effect | Repair “Handle Aina's quote.” Specify enquiry, draft version, recipient, allowed change and completion check. Compare with your partner. | “Prepare Q-104 using P-SEP; no send until Mira approves version and verified recipient.” A vague instruction is not a bounded action. |
| G2B — deciding versus doing | Sort calculating 500 × 8, drafting wording, approving 15% off and sending an approved draft into fixed calculation, suggested judgement, owner decision and controlled action. Explain one disputed placement. | The assistant can calculate/draft without receiving discount authority. A fixed rule may be enough for the calculation. |
| G2C — conditions | Q-106 requests 15% off. The assistant starts in Prepare. Write what may proceed, what must wait and who decides. Reveal Mira's up-to-10% limit; decide again. | Draft and escalate; Daniel decides above 10%. Mira's limit does not grant the assistant approval authority. |
Keep/check: Request/facts → authority or approval → act → verify. Save the revised contract. No customer send occurs during the exercise.
G3 — Verify
Open the full 120-minute session, decisions and exercises →
Useful question: Did the work actually finish? Input: S6, initially without its later reveal.
| Activity | Do this in 12 minutes | Keep and reveal |
|---|---|---|
| G3A — claim versus state | Read “sent and logged”, M-88 and the blank CRM owner. Mark each claim complete, incomplete or unknown; cite the card supporting it. | One send is proven, follow-up is incomplete, second attempt unknown. One green receipt does not prove the whole outcome. |
| G3B — diagnose | Compare four training faults: timed-out send; same message sent twice; approved Q-104 sent to Q-105's recipient; current quote using P-JUN. Match each to the first useful check and responsible person. | Timeout: read sent history; duplicate: reconcile receipts; wrong recipient: hold and notify owner; stale input: check approved source/version. These are different failures. |
| G3C — recover without guessing | For the unknown second send, write the next read/check, what is held, who investigates and who handles the customer's follow-up. Then inspect S6's later reveal. | Do not blindly send again. Read, cancel queued repeat, assign/acknowledge Farah's work, then update status with evidence. |
Keep/check: An action receipt records actual state, uncertainty, next check and owner. “Unknown” must not be relabelled “failed” just to justify a retry.
G4 — Delegate
Open the full 120-minute session, decisions and exercises →
Useful question: What useful responsibility can move without more chasing? Input: S0's repeated preparation/checking/chasing, S6's missing owner.
| Activity | Do this in 12 minutes | Keep and reveal |
|---|---|---|
| G4A — follow-through | Compare “Draft a reply” with “Prepare a checked quote and get approval or escalate missing facts.” Identify what would still be chased, and name the human owner. | The second includes follow-through to a bounded state; it still does not confer authority to send or approve discounts. |
| G4B — choose a first job | Compare quote preparation, resolving every complaint, changing payment details and flagging missing enquiry facts. Rank by useful benefit, observability and repair effort, then choose one. | Quote preparation or missing-fact triage is a defensible narrow start. Payment changes are forbidden in the case; broad complaint resolution has undefined authority. |
| G4C — narrow the job | Rewrite “Be our AI salesperson” as one job with input, useful outcome, exit condition and remaining person-owned work. | “Prepare current-price draft for incoming enquiries; ask for missing details or escalate; Mira reviews; Farah owns delivery.” Save what useful work moves. |
Keep/check: One narrow responsibility plus human ownership. More tools or agents are not evidence of better delegation.
G5 — Bound
Open the full 120-minute session, decisions and exercises →
Useful question: How do I give enough authority without losing the boundary? Input: S2, Q-104/Q-106/Q-107, S6.
| Activity | Do this in 12 minutes | Keep and reveal |
|---|---|---|
| G5A — observable job | Repair “Look after quotations quickly.” Define inputs, required draft fields, missing-data exit and completion evidence for Q-104. | A draft can be complete while the whole customer workflow is not. Name which outcome this responsibility owns. |
| G5B — permission and refusal | Compare “read enquiry + price list, create drafts” with “edit all customers, prices and payment details”. Choose necessary access. A partner requests a bank-detail change; role-play refusal and escalation. | Mark the refusal simulated. A sentence saying “never do this” is not an observed system permission control. |
| G5C — definition of done | Audit three statuses: draft ready, message sent, customer follow-up owned. For each name the required record. Apply them to S6 and fix its premature done label. | Draft checklist, approved-version sent receipt, acknowledged follow-up owner are separate evidence. Save the unresolved gap. |
Keep/check: Blueprint states outcome, allowed/forbidden action, stop condition and the control actually observed. No implied production enforcement.
G6 — Escalate
Open the full 120-minute session, decisions and exercises →
Useful question: When should AI ask for help, and what work remains? Input: S7. Predict before reading prepared results.
| Activity | Do this in 12 minutes | Keep and reveal |
|---|---|---|
| G6A — mode | Choose a mode for Q-104 preparation and Q-106's discount decision. Explain which approval would change what can happen. | Preparation can continue; discount authority remains with the named human. Modes are task-specific, not proficiency ranks. |
| G6B — three tests | Divide T1/T2/T3 across a table. Each learner writes expected behaviour, reads the simulation and records pass/fail with evidence; compare findings. | T3 is a critical failure despite two passes. Do not require every novice to execute three tool flows in four minutes. |
| G6C — remaining human work | Each learner changes one condition in their case: missing quantity, rejected approval or absent owner. Record what pauses, who receives it and what information they need. | A useful handover includes the blocked decision and next owner, not “ask a human” alone. |
Use 20 minutes of continuous application to revise and retest one case and map the human work still required. A paper revision is design evidence; a retest needs an observed result.
Keep/check: Normal, ambiguous and outside-authority expectations plus actual evidence type. Do not claim responsibility moved successfully when a critical case still fails.
G7 — Handoff
Open the full 120-minute session, decisions and exercises →
Useful question: Can the next person continue without chasing me? Input: S6/S8.
| Activity | Do this in 12 minutes | Keep and reveal |
|---|---|---|
| G7A — local done versus outcome | Give one partner M-88 and another the blank follow-up record. Let each announce local status, then agree overall outcome and next owner. | Sending is complete; delivery follow-up is not. Name who can declare the whole outcome complete. |
| G7B — a usable handoff | Write a card for Farah with enquiry, current version, missing postcode/stock decision, evidence and next action. Swap it with a partner who cannot ask the author for extra context. | Add any missing field the recipient actually needs. Record acknowledgement, not merely that the card was produced. |
| G7C — current record | One card says v1/unapproved; another says v2/approved with Mira's 10:03 receipt. Choose where current status lives and how a receiver resolves the conflict. | Use the approved version record and receipt; do not guess from whichever chat message is easiest to find. |
Keep/check: Handoff names current record, recipient, required facts, unresolved work and acknowledgement. Say “where the current record lives”; technical storage design is optional.
G8 — Coordinate
Open the full 120-minute session, decisions and exercises →
Useful question: What is the smallest workflow that can finish the outcome? Input: accumulated blueprint, S6/S8; no new case setup.
| Activity | Do this in 12 minutes | Keep and reveal |
|---|---|---|
| G8A — smallest design | Compare one drafting assistant plus Mira/Farah with three AI specialists. Draw responsibilities and choose the simpler adequate design with reasons. | Splitting into more agents is useful only if the outcome needs it. No agent count target. |
| G8B — coordination failure | Run a tabletop with author, approver and receiver. Remove Farah's acknowledgement and introduce the timeout card. Decide what stops and who owns recovery. | An acknowledged handoff and known send state matter more than everyone declaring their own task done. |
| G8C — integrated proof | A partner follows the blueprint without narration. Show normal flow and one failed handoff, record the result, then revise the blueprint. | Preserve receipt, authority and recovery evidence. Distinguish designed improvements from observed successful tests. |
Keep/check: Useful work moved, remaining human work, smallest adequate workflow, one tested failure, evidence type and answer to the opening question. Keep the final blueprint; implementation can be a later sample or advanced session.