AI becomes much more useful when it can work from material you provide: a policy, brief, meeting note, or spreadsheet. It can also become more dangerous. A tidy summary may omit the one condition that changes the decision. This lesson is about using documents as evidence, not as decoration.
Use only a document your organization permits in the chosen tool. For practice, the fictional source below is safer than a real contract or customer record. You will ask for a summary, then check every important line against the source. The assistant’s wording is not a substitute for the document.
Source note
“40 attendees. Venue undecided. Invitation needs approval.”
Give the source a narrow job
Imagine this fictional workshop note: “The workshop is two hours, capped at 40 attendees. The venue is undecided. A hands-on activity must take at least 30 minutes. We have not approved a price or refund policy. The team will send an invitation only after legal approves the registration wording.” That is the whole source for this exercise.
Ask for a decision brief, not a generic summary. A manager needs to know what is decided, what is pending, and who must act. Put the source between clear markers. Ask the assistant to say “not stated” when the note does not answer a question. This reduces the temptation to complete missing details from general knowledge.
Try this
Copy a source-bound prompt
Paste only the fictional note below, or replace it with an approved public document.
Use only the source note below. Make a two-column table: “Decided” and “Still needs a decision.” Then draft a six-sentence briefing for the event owner. After each material statement, quote the exact phrase from the note that supports it. If the note does not say, write “not stated”; do not fill the gap. SOURCE NOTE: The workshop is two hours, capped at 40 attendees. The venue is undecided. A hands-on activity must take at least 30 minutes. We have not approved a price or refund policy. The team will send an invitation only after legal approves the registration wording.
Why it works: The source is visible and the output asks for supporting phrases. You still compare the quotes with the note yourself.
Check omissions as carefully as inventions
A summary can be technically true and still misleading. “The workshop has a 40-person cap” is correct, but if the same briefing implies that invitations are ready to send, it has omitted the legal-approval condition. Check both directions: did it add a fact that is not there, and did it leave out a condition that matters?
For long documents, ask the assistant to identify the section or page behind each claim. Then open that location. If the source has conflicting versions, do not ask the model to quietly choose one. Ask it to show the conflict and the decision that a person must make.
Match the tool to the material
Some tools can read an uploaded file or connected workspace; others only see what you paste. Availability, retention, and access rules vary by product and account. Confirm your organization’s approved tool and data policy before uploading anything confidential. A personal account’s settings are not the same thing as a company approval.
For your first document task, choose something you can verify in ten minutes. A one-page public agenda, a published FAQ, or a fictional policy is ideal. Build trust by finding the assistant’s limits on simple material before asking it to navigate a dense contract or a live customer complaint.
Practice assignment
The source-check challenge
Use the fictional note or an approved public one-page document.
- 1Ask for decided facts, open questions, and a six-sentence brief.
- 2Underline each claim in the answer and point to its source phrase.
- 3Find one important condition the summary might omit.
- 4Revise the briefing so a manager could use it without mistaking a pending decision for a fact.
What to save
A briefing with a source map and a “not stated” list.
You are ready when
Every material claim has a matching source passage; missing and conflicting details stay visible.
Further reading
Sources and guidance behind this lesson, not required reading to finish it.
- OpenAI Academy: Prompting
Updated September 2026; clear task, context, output, and iteration.
- OpenAI Academy: Responsible and safe use of AI
Check workplace rules and use appropriate human review.
- OpenAI: Data controls in ChatGPT
Controls differ by plan and workspace; organizational policy still applies.
- NIST: Generative AI Profile
Explains risks including confident false statements and fabricated citations.