Section 1
Separate desk research from primary research
Desk research is reading what already exists: filings, pricing pages, job postings, industry reports, forums, review sites, regulatory registers. AI compresses this dramatically. It can read fifty competitor pricing pages and return a structured comparison in an hour, which used to be a week. Primary research is talking to buyers. AI cannot do this for you, and the current fashion for synthetic customers is worth resisting. A model asked to role-play your buyer will produce the average of what people write publicly about that role, which is precisely the information that is already priced in. Use the first to arrive at your interviews better prepared. Do not use it to skip them. The document-handling mechanics are covered in [AI-Powered Document Processing for Startups](/blog/ai-powered-document-processing-for-startups).
Section 2
The fabrication problem, and the rule that fixes it
Models produce fluent citations that do not exist. Report titles that sound right, page numbers that look right, percentages that are wrong by a factor nobody notices. One rule handles almost all of it: nothing enters a document unless you have opened the source yourself and can point to the sentence. If a claim survives that test, it is yours to use. If it does not, delete it rather than softening it. A hedged fabrication is still a fabrication. In practice this means running research with tools that retrieve and link, keeping the link next to the claim from the first draft onward, and treating any number without a live link as a placeholder rather than a finding.
Section 3
Sizing a market without inventing one
Top-down sizing, taking a large industry figure and claiming a percentage, is where fabricated numbers cluster. Bottom-up sizing is harder and defensible. Build it from countable things: registered businesses of a given type in a given country, average spend on the category you can evidence from a public price, replacement or renewal frequency. Each input needs a source. The output is a range, not a point. AI helps by finding and structuring those countable inputs quickly, and by stress-testing the arithmetic. It should never supply the inputs from memory. A market size a founder built this way survives due diligence, and more usefully, it tells you whether the business can reach the revenue it needs. The adjacent forecasting discipline is in [Leveraging AI Automation for Predictive Sales Analytics](/blog/leveraging-ai-automation-for-predictive-sales-analytics).
Section 4
Where AI genuinely improves interviews
The strongest use is before and after the conversation, not during it. Before: draft the guide, then have a model critique it for leading questions and for places where you asked about the future rather than the past. Buyers are unreliable about what they will do and reliable about what they did. After: transcribe every interview, code the transcripts against a consistent scheme, and look for the phrase repeated across people who have never met. That repeated phrase is usually your positioning, written by the market. Twenty interviews analysed this way beat two hundred survey responses, because you can ask why. Keep the raw transcripts. They become the evidence base for every claim you make later.
Section 5
Bias, consent, and the limits of scraped data
Public text is not a representative sample. Review sites over-represent extremes, forums over-represent technical users, and social platforms over-represent whoever posts. Research built only on scraped text will describe the loudest tenth of a market with great precision. Correct for it deliberately: name who is missing from your sample and go find three of them. Ask who would never appear in this data and why. On consent, record interviews only with permission, tell participants how the recording will be used, and remove identifying detail before any transcript goes near an external tool. Competitor research has a line too: public information is fair, and anything obtained through misrepresentation is not. Presenting the resulting insight is covered in [Storytelling in the Age of AI and Automation](/blog/storytelling-in-the-age-of-ai-and-automation).
Section 6
Know when the research is finished
The failure specific to AI-assisted research is that it never ends. Synthesis is cheap, so the deck keeps growing while the decision waits. Set the exit condition first. Write down the decision the research must inform, the two or three things you would need to believe to proceed, and the evidence that would change your mind. Stop when those are answered, even if the document looks thin. A useful signal: when your fifth consecutive interview produces nothing you have not heard, you have reached saturation on that segment. Move to a different segment or move to selling. Extending the research past that point is procrastination with citations.