Practice exercise · 5 minutes to start · no account or email needed
Helpful, but Unsupported
Map each claim in an AI assistant’s answer to a fixed policy, catch what the policy contradicts or never says, and write a corrected answer that still helps.
What you will practice
Check an assistant’s answer claim by claim against the policy it should rely on, route what it cannot answer to the right person, and keep the parts that genuinely help.
An AI assistant drafted a reply to a real-sounding question. It is friendly, specific, and fast. Some of its specifics come from the policy. Some do not.
Assistants tend to fill gaps with plausible detail: a typical timeline, a usual fee, a reassuring promise. When a claim is not in the source, specificity is not evidence. The fix is not a vaguer answer; it is an answer that says what the policy says and sends the rest to someone who can answer it.
A quick example
Fictional store policy: “Returns are accepted within 30 days with a receipt.”
Before
“You can return it within 30 days, even without a receipt.”
After
“Returns are accepted within 30 days with a receipt. If you don’t have your receipt, our service desk can tell you what options are available.”
The first half was supported; the second half contradicted the policy. The revision keeps the useful part and routes the unanswered part to a person.
Your turn — about 5 minutes
Fictional community clinic. A message arrives: “I need my records sent to my new doctor before my appointment on Friday. Also, can I take ibuprofen before my blood draw?” An assistant drafted the reply below.
Read the policy excerpt and the question, then classify each claim in the assistant’s answer. A claim the policy does not allow staff to make counts as contradicted.
Appointment and Records Policy (fictional, version 3)
- §1 New-patient appointment requests made online are reviewed within two business days.
- §2 Please give at least 24 hours’ notice to cancel. The clinic does not charge a cancellation fee.
- §3 Records copies require a signed request form. Records requests take 10–15 business days to process. Electronic copies are free; ask the front desk for the current paper-copy fee.
- §4 Front-desk and messaging staff do not give medical advice. Clinical questions go to the nurse line.
Assistant’s answer
Choose one answer for each statement. 0 of 6 answered.
Administrative example only. It contains no clinical guidance and is not advice about medication. Follow your organization’s privacy policy; never enter patient information into an unapproved tool.
Review rubric
Use these questions on your own work, not just on the exercise.
- Mapped
- Is every claim tied to a policy section, or marked “not in the policy”?
- Contradictions
- Are contradicted claims removed or restated in the policy’s terms?
- Specifics
- Do numbers, fees, and timelines appear only where the policy states them?
- Commitments
- Are promises (expediting, approvals, reviews) made only where the policy grants them?
- Routing
- Are out-of-scope questions sent to the person or team who can answer them?
- Still useful
- Does the corrected answer still answer everything the policy does allow you to answer?
- Version
- Do you know which version and date of the policy you checked against?
Also check, for this kind of work
- No clinical advice appears in an administrative reply; clinical questions are routed to clinical staff.
- No patient-identifying details are pasted into an AI tool your organization has not approved for them.
Go deeper — 20–30 minutes with your own work
- Choose one policy, procedure, or guideline you answer questions about, and note its version and date.
- Write a realistic question someone might ask. Use no real person’s details.
- Ask an AI tool your organization approves to answer it, first without the policy and then with the relevant excerpt (if you are allowed to share it with that tool).
- Map every claim in each answer: supported, contradicted, or not in the policy.
- Write the answer you would actually send, and note who the unanswered parts should go to.
Stop or change the exercise if
- The policy is confidential and the tool is not approved for it.
- You find that two official sources disagree. Stop and ask which one is authoritative; do not pick one silently.
Keep this: Claim-to-source map
A table that shows where each claim came from and what you did about it.
What you type here stays in your browser and is lost when you leave the page. Keep real work details in the tools your organization approves.
Reflect
Which unsupported claim sounded most helpful, and what made it convincing?
Do it again
Try the other case, or apply the map to an assistant your team already uses. Then “One Mistake, One Checklist” shows how to turn a recurring catch into a standing check.
- Would You Send This? — Check an AI-written update against its source record and fix every statement the record does not support before it goes out.
- One Mistake, One Checklist — Turn one concrete lesson from your work into a short checklist someone else could use, with a clear trigger, specific checks, the judgment it cannot replace, and its limits.
- Could Someone Continue This Tomorrow? — Write a handoff note that lets a colleague, or a fresh AI session, pick up the work without you: goal, state, sources, decisions, open questions, next actions, and what to leave out.
A personal next step
Which capability should you practice first?
The free 10-minute assessment helps you see which tasks AI may affect, where your professional judgment matters, and which capability to strengthen first. You get a personal report, a learning path, and a four-week Development Plan. It is an educational starting point, not a job prediction or a credential.
Map My Work