Choose where an AI answer should come from
The public web can't tell you whether your manager signed off the proposal. Pick the source that can answer the question before you ask an AI to sound certain about it.
What you'll make
a one-line source plan for a work question, and an answer that stays inside that source.
What you'll need
ChatGPT, Claude, Gemini or Copilot in your work account, one question, and the file or page that ought to answer it. About ten minutes.
Example result.
Question: Has the onboarding pilot been signed off? Source that can answer it: the steering group’s decision note of 2 October (internal file). Web search off. Answer: Not yet. Section 2 says budget approval is pending and no start date has been agreed.
1. Decide what kind of question it is
Write the question in one sentence, then decide which of three jobs it is. Each one needs a different source.
- A fact about the world now. “Can Canva export a presentation as PPTX?” The vendor’s current help page answers this. Canva’s file-type page lists PPTX for presentations, and warns that the file may look different in PowerPoint.
- What our own material says. “Has our onboarding pilot been signed off?” Only the decision record, or the person who owns the decision, can answer this. The public web can’t, however many articles it finds about onboarding pilots.
- What we should do. “Which onboarding format should we pilot?” This needs your goals, constraints and evidence for each option, then a human decision. The AI can help prepare that; it can’t make it.
Why this works: most wrong answers come from a sensible source used for the wrong job. A supplier’s marketing page can describe a product, but it can’t prove your licence includes it.
What you should see: one question, one job and the name of the source that has authority over it.
2. Ask for a source plan before the answer
Paste the question and list what you’re able to use. Ask for the plan first, not the answer.
Before answering, help me choose the evidence for this question: [question]
Sources I can use: [files, pages or workspace sources]
Say whether this is a factual lookup, a question about my own material, or a research and decision question. Name the source that could establish the answer and why. List any evidence still missing. Don't search the web or answer yet.[question] is your one sentence. [files, pages or workspace sources] is everything you’re allowed to use, named so the AI can’t treat everything it can reach as relevant.
Why this works: separating the plan from the answer makes the assistant commit to a source before it has a conclusion it wants to support.
What you should see: a short plan naming one main source, plus a “missing” list. If it names “general knowledge” for an internal question, push back now.
3. Switch to that source and ask
Set the tool so it uses the source the plan named, and only that.
| Tool | Where the source switch is |
|---|---|
| ChatGPT | + › Web search for the web; + › Add photos & files for a file; + › Deep research for a broad question |
| Claude | + › Web search (a checkmark means it’s on); attach a file with +. On Team and Enterprise an owner turns web search on in Organization settings › Capabilities |
| Gemini | Add files › Files to upload; Add files › Deep Research for a broad question |
| Copilot | The Work and Web toggle; see work and web sources in Copilot |
For an internal question, turn web search off where your tool lets you. Then ask:
Answer this question using only the source I've attached: [question]
Give the answer, the exact section or line that supports it, and any condition that changes its meaning. Keep your own suggestions separate from what the source says. If the source doesn't answer the question, say so and name the confirmation needed.Why this works: a source that isn’t switched on can’t leak into the answer. Asking for the exact line gives you something to check.
What you should see: an answer with a section or line reference you can find in the file.
Check the result
- Open the cited section and read it. The answer should say what it says, with the same conditions.
- Look for anything that came from somewhere else: a date, a “typically” or a figure the source doesn’t contain.
- If the question was internal, check that no web citation appears.
- Anything the source can’t establish is marked unresolved, not filled in.
The usual failure: the answer is right about the source, then adds something plausible from elsewhere. A pending decision becomes “expected to start in November”, taken from general knowledge or a web result. Spot it by asking where each sentence came from.
The sentence "[sentence]" doesn't come from the agreed source. Remove it and answer again from that source only. If the source doesn't settle the point, mark it unresolved and say who could confirm it.[sentence] is the line you couldn’t find in the source.
Worked example: a pilot that isn’t signed off yet
Fictional example.
Steering group decision note, 2 October
Section 2: The onboarding pilot is recommended for November. Budget approval is pending. A final start date has not been agreed.
Question asked (web search left on): Has the onboarding pilot been signed off, and when does it start?First answer:
The pilot was recommended in the 2 October note (section 2). Budget approval is pending, but pilots of this kind usually start at the beginning of the month, so you can plan for 3 November.
- The first sentence is supported. Section 2 says “recommended”, not “approved”, and the answer keeps that.
- “Plan for 3 November” isn’t in the note. It came from general patterns, and web search was left on. That’s the usual failure: a reasonable guess dressed as a date.
- After the repair prompt: “Not signed off yet. Section 2 says the pilot is recommended for November, budget approval is pending and no start date is agreed. Unresolved: the start date. The steering group would need to confirm it.” That answer is less helpful-sounding and more useful.
Task kit for this job: Download (ZIP) · Reusable instructions
Go deeper
Give research a scope and a stopping rule
Before a broad research task, agree the scope, the source standard, the time period and what result you expect. Decide in advance what would be enough to support a recommendation, and what stays with the decision owner. A longer source list isn’t automatically stronger evidence. Use the research-report audit to check the major claims.
Keep different source types labelled
When an answer mixes your files, public documentation and the assistant’s own suggestions, ask for each to be labelled. A proposal shouldn’t borrow the authority of a nearby citation that supports a different statement. For internal summaries, ask for conflicting versions to be listed separately, so an old policy and a new draft don’t merge into a rule neither one states.
Check coverage before conclusions
If the assistant read only part of a document, or only some of the files, keep that limit visible in the answer. Use the file-coverage guide to see what was actually available to it.