How Many AI Interviews You Need: Sample Size for Qualitative Research
There is no single correct number of interviews. Size the study for each group you compare, launch in waves and stop when new interviews stop adding new themes.
Overview
AI interviews make it easy to run more sessions than a person could. More sessions are not always better. Each session adds reading and checking time, and a large study of the wrong people is still wrong.
Qualitative sample size is a judgment. Ask how many groups you compare and how different the people in each group are. Then watch when new sessions stop adding new themes. This is often called saturation.
In User Evaluation, the default target for an AI interview study is 12. You can pick from 3 to 40. You pay at launch and purchases are final, so a smaller first launch is often the safer plan.
Rules
- 01
Size the study for each segment, not for the whole study.
If you compare small and large accounts, each group needs enough sessions to show its own themes. A total of 12 split 10 and 2 tells you almost nothing about the smaller group.
Do notTarget: 12 customers.
DoTarget: 12 customers, with a quota of 6 small accounts and 6 large accounts.
- 02
Launch in waves.
A first wave shows if the guide works and which themes appear. You can then decide to run a second study with the same guide, or to stop.
- 03
Check for saturation after each wave.
Code each new session. Count the codes that are new for your main question. When a wave adds no new codes for that question, more sessions of the same group add little.
- 04
Set the target to the number of usable sessions you need.
Some sessions will be fake, careless or have bad audio. In User Evaluation, a flagged session is replaced and does not count toward the target. So the target is the number of good sessions.
- 05
Make the audience narrower before you make the sample larger.
A mixed audience needs many more sessions before themes repeat. Tighten the screener to the people who can answer the research question.
- 06
Do not report percentages from a qualitative sample.
A percentage suggests a measurement of the whole market. Report what most or some participants said, and how many, and show the clips.
Do not42% of users hate the export.
DoMost participants who export data described the same workaround with a spreadsheet.
- 07
Use a text survey when you must know how common something is.
Interviews tell you why and how. To learn how often a reason occurs, ask a closed question to a larger sample after the interviews.
- 08
Plan the analysis time with the sample.
An AI can write the first report, but you must still check the claims against the clips. More sessions mean more clips to check.
Checklist before launch
- You listed each segment that you will compare.
- Each segment has its own quota.
- The screener is as narrow as the research question allows.
- The first launch is a wave, not the full budget.
- You have a rule for when to stop: a wave with no new codes for the main question.
- You know how many clips you can check before the report is due.
- The report plan uses counts of participants, not percentages.
Questions
Enough that new interviews in each group stop adding new themes for your main question. Plan for each group you compare, and check after each wave.
Not by default. The same rules apply. AI interviews make a larger sample cheaper and faster, but checking the findings still takes time.
Saturation is the point where new sessions stop adding new themes or codes. It depends on the question and on how similar the participants are.
Ask real people your question
The AI moderator interviews screened people by voice or text. Each claim in the study report links to its clip. The first 3 interviews are free.