Skip to content
The method

Listening that scales without losing depth.

Every Tacita project starts from a method built over years of consulting and observing processes. We design AI interviewers that know what to ask, when to dig deeper and how to hand the material back to the people who have to decide.

The problem

The alternatives to Tacita.

Anyone gathering information from people (customers, employees, stakeholders) has always had to choose between speed and depth. Forms scale but stay shallow. Human interviews go deep but don't scale. Neither one, on its own, solves the problem.

📋

The multiple-choice form

Fast and scalable, but it only gives back what you thought to ask about, and never gets at the causes.

HighScalability
LowDepth
~€300Per form and end-to-end analysis
Schuman & Presser (1981) showed that the categories researchers anticipate often don't match what people would say spontaneously. You risk measuring the wrong answers.
🎤

The human interview

Deep and rich, but slow, expensive and hard to take to scale.

LowScalability
HighDepth
€80–300+Per interview
Hennink et al. (2017) show that 16–24 interviews are needed for full understanding. A typical project takes weeks of work and thousands of euros.

The interview with Tacita

The depth of an open interview, with the speed and scale of digital.

HighScalability
HighDepth
20%Of the human cost
Guest et al. (2006) show that 92% of themes emerge within 12 interviews. With Tacita you can run all of them on the same day, with personalized follow-ups and structured output.
How it works

No technical knowledge required.

We look together at your request and the context behind it. The process is designed to be readable even by people who aren't AI experts, which makes it easier to get started and easier to use the result.

1

We design the interviewer

Together we define objectives, themes and the criteria for digging deeper. We build the conversational agent on a framework calibrated on the specifics of the project.

2

We listen at scale

Tacita runs the interviews in parallel, with open questions and adaptive follow-ups. It gathers structured material while keeping depth and methodological consistency.

3

We deliver usable insight

The project closes with outputs that decision-makers can actually read: reports, segmentations, thematic maps, structured data. Not documents to interpret, but tools to use.

The configuration method

You don't improvise a good AI interviewer.

The quality of what you gather depends almost entirely on how well the interviewer was designed, before the first conversation even starts. Execution is scalable and repeatable. Design is not. And design requires something technology alone doesn't bring. It takes knowledge of organizations, an understanding of the psychological mechanics of an interview, and the ability to see which conversations certain structures simply can't have on their own.

🎯

Diagnosis before design

Every organization arrives with a stated problem that is almost always a simplification of the real one. Working out what is worth exploring (and with what kind of interviewer) is the first act of design, not a preliminary step to rush through.

🎤

The tone that determines what comes out

An interviewer perceived as judgmental shuts answers down. An excessively empathetic one produces comforting narratives that aren't much use. The optimal calibration point (curious enough to push for depth, neutral enough to lower defenses) doesn't come out of a generic prompt. It is designed.

🔓

The conversations that stay stuck

In every organization there are conversations that are structurally blocked (the feedback that never travels upward, the dissent nobody voices, the knowledge people guard closely). Knowing where they sit, and designing the interviewer to unblock them, is the least visible and most decisive part of the work.

Execution

Scale doesn't lower quality.

Gathering hundreds of conversations in parallel is the technical part. Making sure each conversation is purposeful, coherent and able to adapt to the person on the other side is the methodological part. The second shapes the first decisively. An agent designed with care carries the method into every session, with no need for constant supervision.

🌐

Every interview starts from a context

The agent doesn't come to the conversation unprepared. It knows the project objective, the profile of the person it is talking to and the questions still open. That makes every session purposeful rather than generic.

🔄

The conversation adapts

The script exists, but it isn't rigid. The agent knows when to dig deeper, when to change angle, when to leave more room. It isn't improvisation. It's the result of a design that anticipated the most common variations and calibrated the responses.

🚫

Difficult answers are expected

Not everyone cooperates straightforwardly. Some back away, rationalize, change the subject. The interviewer is designed to recognize these signals and respond without pushing, because an answer obtained under pressure is worth nothing.

The results

The output is designed like everything else.

Gathering conversations isn't enough. The value is produced in the distance between what people said and what the organization needs to understand before it can decide. You cross that distance with an analysis method that knows what to look for, because it already knows what the final result has to be good for.

📋

What was stated

Every explicitly expressed piece of information is extracted, categorized and mapped back to the project framework. The system knows what to look for because the framework was defined before any gathering began.

🔎

What was not stated

In conversations there is information that surfaces without the interviewee ever formulating it explicitly. Very specific examples, hesitations, exceptions mentioned as if obvious. These are often the most useful parts, and they require a level of analysis that goes beyond transcription.

Output built for decision-makers

The result is not a transcript to interpret. It is a document designed for people who have to do something with that information. Priorities, recommendations, insight maps. The form depends on who will read it and with what objective.

Sampling method

The right sample changes the quality of the whole project.

Interviews gather valuable knowledge. But if the sample of people to interview is random or skewed, even the best insight risks staying partial.

That's why our process starts with a step many skip. We work alongside you to build a representative sample, using a qualitative segmentation method consolidated across dozens of real projects.

Instead of selecting people by role or seniority, we distinguish them by their needs, the problems they face and the behaviors observable in their working context. Every segment that emerges has qualitative characteristics of its own; the sample is built to cover all of them proportionately and meaningfully.

The tool below helps calculate the size and composition of the ideal sample for your project.

Population 200
Homogeneity Medium · 75%
Insight level
Segments / strata
Distinct roles, sites, clusters
Interviews +30% versus human benchmarks (AI buffer) · Survey without design effect (DEFF = 1.0)
Validated
14
Interviews for 80%
Code saturation
Thematic saturation doesn't depend on the size of the population, but on how heterogeneous it is. Larger populations tend to have more internal segments, so the number rises slightly — but it stays orders of magnitude below surveys.
Indicative
155
Multiple-choice survey for 80%
Indicative estimate
11×
Multiplier
Additional multiple-choice surveys required

Qualitative model (interviews)

The number of interviews is calculated by combining homogeneity, thematic complexity and level of saturation. With more segments, saturation has to be reached in each group separately: the total is therefore saturation per segment × number of segments. In conservative mode a +30% buffer is applied to account for the difference between an AI interviewer and an experienced human one.

  • Guest, G., Bunce, A. & Johnson, L. (2006). How Many Interviews Are Enough? Field Methods, 18(1), 59–82.
  • Hennink, M.M., Kaiser, B.N. & Marconi, V.C. (2017). Code Saturation Versus Meaning Saturation. Qualitative Health Research, 27(4), 591–608.
  • Hagaman, A.K. & Wutich, A. (2017). How Many Interviews Are Enough to Identify Metathemes? Field Methods, 29(1), 23–41.
  • Hennink, M. & Kaiser, B.N. (2022). Sample Sizes for Saturation. Social Science & Medicine, 292, 114523.

Quantitative model (multiple-choice survey)

In optimistic mode a design effect (DEFF) is applied for heterogeneity; in conservative mode DEFF = 1.0. Segments compute Cochran per sub-population. Thematic complexity does not change the Cochran formula (more themes = more questions, not more respondents), but a multi-theme questionnaire increases the drop-out rate: the note below the result indicates how many additional invitations are needed to compensate for dropout (Krosnick & Presser, 2010).

  • Cochran, W.G. (1977). Sampling Techniques (3rd ed.). Wiley.
  • Krosnick, J.A. & Presser, S. (2010). Question and Questionnaire Design. Handbook of Survey Research (2nd ed.).
  • Schuman, H. & Presser, S. (1981). Questions and Answers in Attitude Surveys. Academic Press.

Note: these numbers are indicative estimates. The survey model assumes simple random sampling and p = 0.5 (maximum variance). The conservative estimate is designed to be defensible; the optimistic one reflects academic benchmarks for experienced human interviewers.

Want to go deeper into the research?

We've put together a complete document with all the academic sources, the formulas and the practical implications for designing research that actually works.

Start here

Tell us about a real case and we'll find the right framework. With Tacita, we solve it.

If you want to work out which use case fits your HR project best, the right next step is a short, concrete conversation. Chat with Lara to find out. Otherwise: