How Many AI Interviews Are Enough? Replace “It Feels Saturated” with a Recomputable Stopping Rule

Once AI interviews drive the cost of a single session close to zero, "let's run a few more" becomes a nearly frictionless choice, and "how many is enough" turns into a question every project has to answer on the spot. "It feels like themes are saturating" cannot answer it: give the same transcripts to two researchers and you will often get two different stopping points, with no way for either to show it is the better one. A recomputable alternative is to fix a stopping rule in advance, lock a theme pool from the first few interviews, recompute the share of newly added themes at fixed intervals, and stop when that share falls below a declared threshold. That converts a feeling into a calculation a colleague can redo and a methods section can report.

The stopping rule only answers when to stop; it does not answer where to start. The starting point follows from information power: the more focused the aim, the more homogeneous the sample, the more established the theory, the better the dialogue quality, and the more targeted the analysis strategy, the fewer sessions are needed. A workable order is three steps: judge information power to set an expected range, compute saturation batch by batch while collecting, then screen for false saturation, which AI interviews are especially prone to through scripted probing and summary-layer coding.

Separate Two Judgments: Where to Start and When to Stop

These are usually collapsed into "interview until saturation," but they rest on different evidence. The starting point is a design-time planning question driven by the aim, sample homogeneity, theoretical preparation, and the analysis plan. Stopping is an execution-time monitoring question driven by the observed growth curve of coded themes. Merging them has concrete costs: without planning, recruitment keeps escalating; without calculation, a claim of saturation cannot be checked.

A systematic analysis of interview-based health research over 15 years and three disciplines found that most papers either omitted a sample-size rationale or leaned on "saturation" and pragmatic constraints in general terms, while framing their sample size as "small" only after the fact. In practice, the word has become a reason not to explain, rather than a judgment that requires declared parameters and a visible calculation.

Putting both judgments in one table removes much of the mid-project disagreement:

Information Power: Five Parameters That Set the Starting Point

Information power reframes sample size from "how many bodies" to "how much study-relevant information this sample carries." Five parameters move it, and any one of them changes the number of sessions, so writing one sentence about each is more useful than arguing about whether 12 or 20 is right.

Empirical benchmarks inform the starting range but are not the answer. A systematic review of 23 empirical studies found that saturation was reached within 9 to 17 interviews, or 4 to 8 focus groups, mainly in studies with relatively homogeneous populations and narrowly defined objectives; the outliers, including multi-country studies, meta-themes, and "meaning" saturation, needed materially larger samples. An earlier stepwise thematic analysis of 60 in-depth interviews showed a similar shape: 70% of 114 themes appeared within the first six interviews and 92% by the twelfth. Those figures describe the curvature of theme accumulation, not a passing score of 12. Treating 12 as a fixed threshold is exactly what ignores the five parameters above.

Compute It Batch by Batch

Turning saturation into arithmetic requires three declared quantities: the base size (how many initial interviews define what is already known), the run length (how many consecutive interviews form one comparison window), and the new information threshold (the share of newly added themes, relative to the base pool, below which you accept saturation). All three must be fixed before coding begins; during fieldwork you only compute.

The table below is an illustrative example that demonstrates the calculation only; the numbers do not represent any real project. A brand studies what triggers second-week churn among urban white-collar users who registered and churned within three months. The aim is narrow, so information power suggests an expected range of 12 to 16 interviews. Base size is set at 4, run length at 2, and the threshold at 5%. Each run shifts forward by one interview, so consecutive windows overlap:

The calculation needs only two numbers: how many themes appeared for the first time inside the current window, and how many themes the base pool contains. Recording what each new theme actually was, rather than only counting it, pays off when someone asks why you stopped at nine: you can show whether that single new theme opens a new decision branch and judge whether to confirm it with two more interviews.

The Same Transcripts Stop Later Under Different Parameters

Stopping at nine is not the uniquely correct point; it is a product of the parameters. Change them and the conclusion moves. With a base size of 6 (26 themes) and a run length of 3, holding the threshold at 5%, interviews 7-9 add 3 themes (11.5%), interviews 8-10 add 2 (7.7%), and interviews 9-11 add 1 (3.8%), so the threshold is crossed only after interview 11. Tighten the threshold to zero new themes and you need consecutive runs with no additions: in this example the study runs to interview 12. Nine, eleven, and twelve are the same transcripts, distinguished only by which parameters were declared in advance.

That is precisely why the rule must be fixed before coding. If parameters can be chosen after seeing results, some combination will always look just adequate, and saturation shifts from a judgment into a narrative device. One technical detail is easy to miss: the denominator cannot be the full dataset collected so far. Using everything as the denominator drives new themes toward the denominator total, inflates saturation mechanically, and guarantees a "saturated" result, which makes it useless for decisions during fieldwork. The base pool exists to give the denominator a fixed reference point.

Three Kinds of False Saturation, More Common in AI Interviews

The stopping rule produces a number, and whether that number means anything depends on where themes came from. Each situation below lowers the count of new themes through the measurement method rather than through genuine exhaustion, and each is more likely when AI conducts the interview.

• Scripted probing that manufactures repetition. Generative models tend to ask similar questions about similar answers, so the same topic surfaces with every participant and the theme pool appears to stop growing early. A 2026 controlled evaluation of six leading models as interviewers found clear differences in whether a follow-up was warranted at all, and required experts to annotate items one by one on criteria such as necessity, context-awareness, and whether a skip was justified; the study also reported operational risks from unstable output formats. Probing quality therefore cannot be assumed: sample a fixed share of follow-ups and judge whether each one fills a gap in that participant's evidence or pulls the conversation back to the model's own script.
• Coding on summaries instead of source text. If theme coding rests on model-generated summaries, "no new themes" may only mean the summaries have hit their compression ceiling. Code from original passages, keep summaries as an index, and retain at least one verbatim quote per theme so a reviewer can see how strong the underlying evidence really is.
• Channel saturation mistaken for saturation. Twenty interviews drawn from a single recruitment channel show that this channel has no more to say. Write a coverage matrix before recruitment, stratify on the differences that matter (usage frequency, region, role), and track growth within each stratum rather than only in the total.

A closely related misreading treats "no new themes" as "no counterexamples." A theme list that stopped growing says nothing about whether the evidence behind each theme is concrete or whether cases contradict the main conclusion. After the stopping decision, run a reverse check: find at least one counterexample or boundary condition for the leading conclusion. Finding none usually means the sample is too homogeneous to test it.

Write the Stopping Rule into the Protocol and the Report

Parameters constrain only once they are written down. The protocol should state the base size, run length, and threshold with reasons, plus an expected range and a ceiling. The ceiling matters because it gives the project a recruitment endpoint if themes keep growing. During fieldwork, log new themes and their share per batch, and note who coded each batch; when humans and AI code together, log how disagreements were resolved.

The report should cover four things: sessions completed with recruitment channels, the new-theme count and share for every run, why fieldwork stopped where it did, and which conclusions hold regardless of the stopping point. Reporting standards for interview and focus-group research likewise require sampling approach, sample-size rationale, and refusal information, for a straightforward reason: readers cannot return to the field and must judge credibility from those parameters. Unfinished saturation does not need to be hidden. Stating that a particular group of participants is still missing and that the related conclusions remain provisional is more useful than a vague claim of saturation.

When This Calculation Is the Wrong Tool

Saturation is the most used and most abused concept in qualitative research, and it does not fit every tradition:

• Reflexive thematic analysis does not treat saturation as a validity criterion. Methodological commentary argues that using saturation as a sample-size rationale conflicts with how that approach understands researcher reflexivity and theme generation; a better rationale starts from the research question and the depth of analysis.
• Estimating population proportions requires probability sampling. Questions such as "what share of users hit this problem" are estimation problems answered by sampling design and sample-size calculation. Interview saturation cannot substitute; no number of interviews adds the precision a proportion needs.
• Studies of a few extreme cases. Their value lies in information density, not theme coverage. State the stopping condition as completeness of case information (key processes, timeline, consequences) rather than the absence of new themes.
• Theory-driven verification interviews. When conceptual categories come from existing theory, an initial analysis sample plus a stopping criterion works better, for example analyzing 10 interviews to confirm category coverage and then continuing until 3 consecutive interviews add nothing new.
• Interviews whose purpose is questionnaire design. If the goal is generating items and options, stop when the item pool covers every response branch and edge case, not when themes repeat; the checks that matter are whether options are exhaustive and whether wording is understood as intended.

Running This Inside WeJot

WeJot conducts AI interviews over text, voice, or video with automated probing and structured synthesis, so the rule above can sit directly in the workflow: write theme objectives into the interview design, export transcripts and new themes batch by batch, and compute the saturation share with the parameters declared in advance. Control probing quality with a fixed audit sample instead of assuming the machine asked well. When a study also needs to know how many people face the problem, build a questionnaire baseline first and recruit interview participants from it, rather than inferring population shares from interviews; see combining questionnaires and AI interviews. Session counts and sample sizes for the quantitative arm follow a different logic, covered in calculating sample size. Because the guide determines whether stable themes can be coded at all, start from writing an AI interview guide, or open the user research AI interview template to begin.

Method Sources and Recalculation Notes

• Guest, Namey & Chen (2020), PLOS ONE — A simple method to assess and report thematic saturation in qualitative research (base size, run length, new information threshold, and a step-by-step worked example)
• Hennink & Kaiser (2022), Social Science & Medicine — Sample sizes for saturation in qualitative research: A systematic review of empirical tests (9-17 interviews, 4-8 focus groups)
• Guest, Bunce & Johnson (2006), Field Methods — How Many Interviews Are Enough? (early concentration of theme discovery across 60 interviews)
• Malterud, Siersma & Guassora (2016), Qualitative Health Research — Sample Size in Qualitative Interview Studies: Guided by Information Power (the five parameters)
• Francis et al. (2010), Psychology & Health — What is an adequate sample size? Operationalising data saturation for theory-based interview studies (initial analysis sample and stopping criterion)
• Vasileiou et al. (2018), BMC Medical Research Methodology — Characterising and justifying sample size sufficiency in interview-based studies (how sample-size rationales are reported in practice)
• Braun & Clarke (2021), Qualitative Research in Sport, Exercise and Health — To saturate or not to saturate? (questioning saturation as a sample-size rationale for thematic analysis)
• Panfilova et al. (2026), Scientific Reports — The AI interviewer: multi-faceted evaluation of adaptive questioning by large language models (expert-annotated evaluation of follow-up behavior and format-stability risk)
• Tong, Sainsbury & Craig (2007), International Journal for Quality in Health Care — COREQ: 32-item checklist for interviews and focus groups (sampling and sample-size reporting items)