06 · Freelance 2025–26 · White-label engagement

Scaling Qualitative Analysis with AI

An organization wanted to know how its people really communicate, and where the culture gets stuck. What can millions of messages show that a handful of interviews can't?
MethodsA statistical sweep of 2.8 million messages, then AI-assisted coding of the exchanges the numbers flagged, validated against my own hand-coding, and a hypothesis matrix rated by strength of evidence.
Key findingFull-dataset analysis surfaced tensions no interview sample would have caught, and located them in specific parts of the organization.
What changedLeadership workshopped the findings into a five-point action plan for the communication culture.

Understand + Align

Chart forms from the analysis

A · VOLUME PER PERSON AGAINST WHEN THEY ARRIVED the tail the numbers flag for close reading more none MESSAGES POSTED earliest accounts newest accounts ONE DOT IS ONE PERSON B · THE SAME PEOPLE GROUPED INTO ARRIVAL COHORTS MEAN MESSAGES first cohort most recent cohort ONE BAR IS ONE QUARTER OF NEW ACCOUNTS

ILLUSTRATIVE. THE FORMS ARE FROM THE STUDY; THE DATA DRAWN HERE IS INVENTED. CLIENT FIGURES AND FINDINGS REMAIN CONFIDENTIAL.

The two instruments the sweep ran on, drawn with invented numbers. The scatter finds the few people who carry most of the volume; the cohort bars turn individual noise into a comparison between groups.

The human–AI analysis loop

1 · FRAME Researcher sets thequestions, codebook 2 · EXTEND AI codes the wholeset, not a sample 3 · VALIDATE Every batch againsthand-coded samples 4 · INTERPRET Researcher reads:findings, tensions a batch that disagrees sends the frame back to step 1
The workflow I validated in my thesis and now use commercially. The return edge is what makes it a loop: a batch that disagrees with my hand-coding sends the frame back to step one.

The storyWorking white-label as the research engine behind a partner consultancy, I owned the analysis pipeline, synthesis and reporting. Three kinds of evidence met: platform statistics across thirteen months and roughly 750 people, the partner's employee interviews in Finnish and English, and close discourse analysis of seven episodes the numbers flagged, over 400 messages coded one by one. The coding ran on the human-AI workflow from my master's thesis: I set the frame, AI extends the coding across the full dataset, every batch is checked against hand-coded samples, and interpretation stays human.

← Back to the overview

Questions or thoughts? I'd love to hear from you!

Book a 30-min call