A real Themera analysis — 14 remote-work survey responses
This is a read-only preview of the Themera dashboard.
💡How to read this analysis
Themera automatically reads qualitative data (like interview transcripts or survey responses) and uses AI to identify common themes, build a structured codebook, and tag every excerpt.
Below is the output of unstructured survey responses about remote work. The AI generated a summary, built a Codebook, and organized Verbatim Evidence. The Coverage panel shows exactly what was (and wasn't) coded.
Manual tagging in legacy tools (NVivo, MAXQDA) takes days of work and costs €300–€1,200/year. Themera does the heavy lifting in minutes, saving immense time and money while keeping you in full control.
📚Identified Themes (Codebook)
💬Evidence by Theme
Before the analysis: pseudonymisation
Names, institutions and places are replaced before anything is analysed. Rules handle emails, phone numbers and IBANs; a language model handles names and places, because a rule cannot tell whether "Bergmann" is a surname or an occupation. German inflection is handled too — Müllers and Müllern are replaced along with Müller, which manual find-and-replace almost always misses.
I: Frau Tölle, wie erleben Sie die Arbeit am LKH Feldkirch? B: Seit ich 2019 dort angefangen habe, hat sich viel verändert. Dr. Bergmann hat die Station übernommen. Schreiben Sie mir an u.toelle@lkh-feldkirch.at. I: Und Tölles Einschätzung dazu?
I: Frau [Person 1], wie erleben Sie die Arbeit am [Organisation 1]? B: Seit ich [Datum 1] dort angefangen habe, hat sich viel verändert. Dr. [Person 2] hat die Station übernommen. Schreiben Sie mir an [Kontakt 1]. I: Und [Person 1] Einschätzung dazu?
You review every proposed replacement before it is applied — a technical term replaced by mistake damages your data as badly as a missed name exposes someone. Your real names are never stored: only a salted hash, which is enough to keep the same person as [Person 1] across every transcript in the study without us ever holding the key. Try it — free on every plan.
After the analysis: intercoder agreement
The answer to the question every supervisor asks. You re-code a random 20% sample yourself, blind to what Themera assigned, and get Cohen's kappa per category — plus a paragraph for your methods section. Note that running a model twice and comparing it with itself measures determinism, not codebook clarity; an independent human pass is what the criterion actually asks for.
| Category | κ | Used |
|---|---|---|
| Isolation and loneliness | 0.88 | 24 |
| Blurred boundaries | 0.81 | 19 |
| Autonomy over the day | 0.76 | 17 |
| Communication overhead | 0.41 | 12 |
| Home workspace quality | — | 0 |
The per-category table is the useful part. Communication overhead at 0.41 is a definition problem you can fix — sharpen it and re-run. Home workspace quality shows "—" because neither coder used it: with no variance there is nothing to correct against chance, and reporting 0 there would claim disagreement about a category both coders agreed on perfectly.
For your methods section
The codebook was generated computationally from the full dataset (n = 312 units of analysis) and then reviewed and revised by the researcher. The revised codebook of 9 categories was applied to the full dataset, with every assignment accompanied by a verbatim extract. To assess intercoder agreement, a random sample of 20% (n = 62) was independently re-coded blind to the computational assignments. Percentage agreement was 91.4% and Cohen's kappa, pooled across categories, was κ = 0.79 (substantial). Disagreements were reviewed and led to clarification of the affected category definitions.