Writing a research report with AI — separate the research from the writing
The step everyone skips is the third one, and it is the one that makes the report defensible.
AI Tools Hub Editorial · published · updated
Four steps
| Step | Work | AI involvement |
|---|---|---|
| Define the question | One line stating what the report answers | None |
| Research | Find sources, extract points | High |
| Attribute | Tie each claim to a URL and date | Moderate, human verified |
| Write | Turn the material into prose | High |
The question has to have an answer
"Trends in the AI market" is not a report. "Should we enter this segment next year" is. Without it the research has no boundary.
Research output is notes, not prose
- Record the URL and publication date for every source
- Verify figures against the original, not a secondary citation
- Record what you could not find, so the gaps are visible
- Keep contradictory sources, both of them
Search-enabled tools speed this up. Check that the URLs they return exist; fabricated links still appear.
Label each claim
Market size was X in 2025 (source, published March 2026). It may double by 2027 (inference: extrapolated from the above growth rate). Our advantage would be in Y (our judgement).
Organised this way, the reader can weigh the evidence themselves.
Tell it not to fill gaps
When generating prose, state explicitly that nothing outside the supplied material may be written. Without that instruction the gaps get filled with generalities, and the attribution work is wasted.
FAQ
Is this slower?
Yes. For an external report or a decision document, being checkable matters more. Internal information sharing can skip it.