20–30 minutes + optional 10-minute message challenge · No account, code or API key · Original fictional case
01 / Put the human decision in the workflow
Keep the facts, the draft and the authority separate.
A model can turn supplied text into a plausible message. Fluency does not establish that a claim is true, current or yours to promise. You still need a source for the facts and an accountable person for a commitment.
In this exercise, supported means the selected source backs the claim. Revise means the source contradicts it or supports something narrower. Ask the owner means a required approval or fact is missing. There may be more than one sensible next action; explain yours.
The review boundary
01 / EvidenceRead the current sources
Check the version and scope.
02 / AssistanceInspect the draft
Separate each claim.
03 / Your judgmentCheck facts and permission
Correct, withhold or ask.
04 / Accountable actionGive the owner a next step
A draft cannot approve itself.
02 / Review the evidence
Would you send this?
Read the sources and their roles. Choose what each sentence needs, then compare your reasoning with the teaching notes.
Review a confident draft against the actual dispatch note and communication policy. Your work in each case stays here while this page is open.
AI draft / authored teaching example
01Your delivery is confirmed for Tuesday 3 November, 09:00–12:00.
02There is no new delivery charge on this order.
03We have waived all future delivery fees for you.
No score or automatic send decision. Unselected choices remain unrecorded.
03 / Keep work another person can receive
Write the message you can support.
Draft a short correction. Cite the source IDs, retain the uncertainty and name who must act next. Try the changed-source and customer-message cases before reading their notes.
Give a colleague the sources and your record. Ask them to locate one changed decision without your narration. Record their actual action and any help. Notes stay in this tab and disappear on reload; nothing is uploaded.
Copy the workbook
A common AI failure / Indirect prompt injection
A source can try to become an instruction.
An email, document or search result may contain instructions aimed at an AI assistant. If the assistant treats those instructions as governing the task, the external content can redirect its work. In the third case, M-9 tries to manufacture an approval and suppress the explanation.
Keep the legitimate request, identify the embedded instructions and use the actual policy and owner decision. Do not assume that a polite prompt, a citation or a human spotting this example makes the system secure. A real application also needs enforced permissions and action checks.
The facilitator plan follows all three cases, swaps author and receiver roles and includes a separate later task. Each person keeps their own explanation, assistance and receiving account. This is a ready-to-use practice plan; it does not announce a staffed class.
Use an approved environment and permitted data. Name the task, source owner, draft step, review decision and receiving action. Capture a baseline before changing the workflow, then try a new case with different facts. Ask your manager which actions you may take and which still require approval.
The practice above uses authored text and local notes. It does not run a model, verify a live policy or send a message. For real work, use your organization’s actual policies, source versions and permissions.