You have completed the interviews. You have 40 transcripts, or 40 sets of notes, sitting in a folder. The findings are in there somewhere. The problem is getting them out.
This guide walks through the complete workflow, from preparing your raw transcripts to producing a structured, evidence-based thematic analysis. It assumes you are using Skimle, but the preparatory steps apply to any systematic analysis approach.
Step 1: Prepare your transcripts
Before uploading anything, spend 30 minutes on preparation. The quality of your analysis depends significantly on the quality of what goes into it.
Format your files consistently. Each transcript should be a separate document. Accepted formats include PDF, Word (.docx), and plain text (.txt). If you have interview notes rather than full transcripts, format them as prose paragraphs rather than bullet lists where possible. Bullet notes can be processed, but connected prose produces richer thematic output.
Make each document self-contained. If you refer to the participant as "she" throughout the notes, consider adding a single-line header identifying who the interview was with (role, organisation, or demographic group, not name if anonymisation matters). The analysis will be more interpretable later.
Anonymise if needed. If your transcripts contain participant names that should not appear in analytical output, replace them before upload. Consistent labels ("Participant A," or role-based labels like "Head of Procurement") work better than random anonymisation that makes transcripts unreadable. See anonymising qualitative data for a detailed approach.
Decide on your metadata scheme. This is the most important preparatory decision. Metadata lets you cross-tabulate themes by participant characteristics later: which themes appear more among senior managers than junior staff, which appear in the European sample but not the US sample, which are specific to churned customers rather than active ones. Useful metadata fields for most research projects:
- Participant type or role
- Organisation type or size
- Geography
- Interview date (for longitudinal analysis)
- Any other grouping variable relevant to your research question
Write your metadata scheme down before you start uploading.
Step 2: Create a project in Skimle
In Skimle, every corpus of documents lives in a project. Create a new project with a name that reflects the scope of analysis, not just the client or topic (e.g., "Customer exit interviews Q1 2026" rather than just "Acme").
Set the project language if your transcripts are not in English. Skimle supports over 100 languages. For multilingual projects where interviews were conducted in different languages, upload each document in the language it was recorded in rather than translating first. Translation introduces distortion; cross-lingual analysis within Skimle preserves the original meaning more reliably.
Step 3: Upload and tag your documents
Upload your 40 transcripts. Skimle accepts bulk upload: you can drag all 40 files at once rather than uploading individually.
Once uploaded, add metadata to each document according to your scheme using the Metadata editor. This pays off significantly during analysis when you can filter themes by any metadata dimension to spot qualitative and quantitative differences (e.g., how much or how women vs. men talk about a topic).
If your interviews contain any clues, Skimle can extract some metadata automatically - simply the AI to look for e.g., "Home town" or "Age" from the interview text and it will do it's best to fill the field.
For projects where the transcripts were recorded and transcribed via Skimle's audio transcription, first convert the media files to text. See the practical setup for audio recording and transcription for that workflow.
Step 4: Run the analysis
Skimle's AI analysis reads every uploaded document and identifies the thematic patterns across the corpus. For a 40-transcript project, processing typically takes a few minutes.
You can choose either fully automated thematic analysis, where Skimle creates the full categorisation scheme from scratch based on e.g., the questions asked in the interviews, or give Skimle insights types you are interested in (e.g., "Expressions of emotion" or "Statements about competition").
The analysis takes you chosen scope and approach and produces:
- Insight types (higher-order patterns of meaning across multiple transcripts)
- Categories and sub-categories (the grouping layer that organises related themes)
- Insights and quotes (specific coded passages from the transcripts, each linked to a theme and traceable back to the source document and verbatim text)
The number of themes will vary with the diversity of your corpus. A focused study with a specific research question and a relatively homogeneous participant group will produce 8-15 themes. A broad exploratory study across diverse participants may produce 20-30 before consolidation.
Step 5: Review the categories in Category View
This is where your analytical work begins. The AI's output is a starting point, not a finished analysis.
Open the Category View, which shows you each theme with its supporting insights (coded passages from the transcripts)
Review each theme critically. For each theme the AI has identified, ask:
- Do the supporting insights actually share a meaningful pattern, or has the AI grouped superficially similar passages under one label?
- Is the theme label accurate? Often the AI's label captures part of what the theme is about but misses the more interesting interpretive dimension.
- Does the evidence support the theme as described? Read through 5-10 of the supporting quotes to check.
Merge, split, and rename. After reviewing, you will typically:
- Merge 2-3 themes that turn out to be dimensions of the same underlying pattern
- Split 1-2 themes that turn out to contain two distinct patterns the AI grouped together
- Rename most themes to something more analytically precise than Skimle's initial label
- Edit the categorisation logic to restructure a part of the hierarchy, e.g., instead of sorting topic you might want to structure it by type of actor, which can be done automatically
Add your own coding. If you notice a quote in the transcripts that the AI has not captured as a distinct insight, you can create code it manually. This is the key safeguard against the AI missing unexpected patterns. It is good practice to review the data both from the theme angle as well as the document angle, which is possible thanks to Skimle's two-way transparency.
For a detailed walkthrough of this review process, see thematic analysis in qualitative research.
Step 6: Cross-tabulate by metadata
Recommended reading
Discovering themes in the data using metadata variables - advanced analysis with Skimle
This is where the metadata you set up in Step 3 pays off.
Enter the Data View to see how different metadata dimensions explain the data:
- Which themes appear across all participant types, and which are specific to one group?
- How do the themes from interviews conducted early in the project compare to those from later interviews (relevant for tracking how a situation evolved)?
- Do European and US participants describe the same themes differently, or do entirely different themes dominate in each geography?
Cross-tabulation often produces the most valuable findings in a multi-segment study. An overall theme that appears across all 40 interviews may be interesting; the finding that it appears in 18 of 20 enterprise interviews but only 4 of 20 SME interviews is actionable.
In the Category View you can also compare the qualitative differences between segments - not just how much they mention specific themes but how they talk about them.
Step 7: Build the synthesis narrative
Themes are not findings. Findings are the interpretive statements about what the themes mean in relation to your research question.
Work through your confirmed themes and draft the synthesis:
What is the pattern? State the theme in one sentence that captures what the evidence actually shows, not just what the topic is. "Onboarding complexity" is a topic. "Participants expected self-service onboarding but encountered a process that required three separate introductory calls with the support team, creating friction at the moment of highest motivation" is a finding.
What is the evidence? Select 2-3 representative quotes that illustrate the theme. Avoid quotes that are so general they could support almost any claim; prioritise specific, concrete passages.
What does it mean? Connect the finding to the research question or the decision the research is informing. "The onboarding friction explains the pattern we observed in the product data: users who do not complete setup within 7 days have a 40% lower 6-month retention rate."
How consistent is the pattern? State which participant groups showed this theme and which did not. If the pattern is strong in one segment and absent in another, say so.
Write the synthesis as a short document, one section per major theme. This is the analytical output that will form the basis of your report or presentation. A great starting point for the report (be it Word or PowerPoint) is to export the project in the Export Center.
For guidance on presenting these findings to a leadership audience, see presenting qualitative research findings to executives.
Time benchmarks for a 40-transcript project
| Stage | Estimated time - example only |
|---|---|
| Transcript preparation and anonymisation | 1-2 hours |
| Upload, tagging, and metadata | 30-60 minutes |
| AI analysis | 5-15 minutes |
| Theme review and consolidation | 3-6 hours |
| Cross-tabulation and comparison | 1-2 hours |
| Synthesis narrative | 3-6 hours |
| Total analyst time | 8-17 hours |
Compare this to manual thematic analysis of the same corpus, which would typically require 80-150 hours of coding before synthesis begins. The analyst time saving is in the coding and initial theme-identification stages; the synthesis and interpretive work takes similar time regardless of method.
For some the time savings is used on something completely different, while in many cases it translates to conducting more interviews and spending more time with important stakeholders to ensure the findings translate to impact.
Frequently asked questions
What if my notes are not full transcripts?
Detailed notes work. Brief bullet points lose too much context to be reliably coded. If your notes are brief (3-5 bullets per interview), consider whether you can expand them from memory while it is still fresh. A 15-minute expansion session per interview, writing the notes into prose, significantly improves what the analysis can produce. Going forward, recording and AI transcription produces a full text with very little overhead. See the practical setup guide.
How do I handle very different interview lengths?
Consistency of coverage matters more than length. A 90-minute interview with 20 topics covered is not necessarily richer than a 45-minute interview with 8 topics explored deeply. When your corpus has significantly variable interview lengths, note which interviews are more or less detailed and factor that into how you weight the evidence. Very short interviews (under 20 minutes) from the same study as 90-minute interviews should be flagged: they are contributing much less evidence per document.
What if the AI groups together things that should be separate themes?
This is common and expected. The AI's initial theme structure is a draft, not a final analysis. The review step (Step 5) is specifically designed for this. Split the merged theme into its components, review the evidence for each component, and give each a distinct label. The AI gives you a starting structure; you give it analytical precision.
How do I know when I have done enough review?
You have reviewed enough when: you can describe what each theme means and why the evidence supports it; you have checked the underlying quotes for at least your 5-6 most important themes; and you can identify any themes in your final set that are not supported by the evidence and have removed or revised them. There is no formula. The standard is whether you could defend your analysis in front of a sceptical audience.
Can this workflow handle mixed data (interviews plus documents)?
Yes. Skimle processes interviews and documents in the same project. If your study includes both interview transcripts and, say, customer support tickets or open survey responses, upload all of them into the same project and tag each document type in the metadata. The analysis will identify themes across the full corpus, and you can filter by document type to see where each theme appears.
Ready to start with your own corpus? Try Skimle for free and run the full workflow on your first 40 transcripts.
Related reading: Manual interview coding is too slow: how to speed up theme analysis | Thematic analysis in qualitative research | Turning interview notes into insights at scale
About the authors
Henri Schildt is a Professor of Strategy at Aalto University School of Business and co-founder of Skimle. He has published over a dozen peer-reviewed articles using qualitative methods, including work in Academy of Management Journal, Organisation Science, and Strategic Management Journal. His research focuses on organisational strategy, innovation, and qualitative methodology. Google Scholar profile
Olli Salo is a former Partner at McKinsey & Company where he spent 18 years helping clients understand the markets and themselves, develop winning strategies and improve their operating models. He has done over 1000 client interviews and published over 10 articles on McKinsey.com and beyond. LinkedIn profile
