Understanding the principles behind good AI-assisted analysis is one thing. Putting them into practice is another.
This article walks through what a human-in-the-loop workflow actually looks like, from the moment data comes in to the point where an insight is ready to share. It also includes a real-life example of the workflow in practice.
The most useful way to think about AI in research analysis is not as automation, but as collaboration. AI handles the parts that are time-consuming and mechanical. Researchers handle the parts that require judgment, context, and interpretation.
This division of tasks matters because using AI to speed up the interpretive parts of research analysis without human oversight and accountability increases the risk of poorly-informed decisions and undesirable consequences. A transcript processed incorrectly can easily be caught and fixed. An insight interpreted without context, or a pattern accepted without scrutiny, can travel far before anyone questions it.
Keeping the researcher in the loop at every interpretive step is what makes the workflow trustworthy. Human judgment is crucial for interpretation. That has always been true, and AI does not change it.
Here’s an overview of the workflow:
Data Collection and Transcription
Session Summarization
Querying Your Data
Researcher Review and Refinement
Tagging Your Data
Pattern Detection and Clustering
Validation Against Raw Evidence
We’ll now go through a stage-by-stage walkthrough of the analysis workflow, from the moment a session is recorded to the point where a pattern becomes a finding you can stand behind. Each stage is both an opportunity for AI support and for you to judge what that support is actually worth.
Transcription is where AI is most reliable. Converting audio to text is straightforward, and if there are any typos or errors, you can check it against the recording. Many tools, including Condens, also support translation alongside transcription. This matters more than it might initially seem: it means you can work with research conducted in other languages, pull in sessions run by colleagues in other markets, and build on findings from studies you weren’t a part of without language being a barrier to what’s usable.
Accuracy does vary however. Poor audio, overlapping speakers, and unfamiliar accents can all affect transcript quality. So spot-checking against the original recording is a step worth building into your process, especially for quotes you intend to use.
However, transcripts are only part of the picture. AI can turn audio into words, but it can’t capture the researcher’s in-session thinking and observations: the moment a participant hesitated, the idea that surfaced mid-interview, etc.
Live note-taking during sessions helps enrich the transcript with context. In Condens, Live Notes lets you capture observations, reactions, flagged moments, and emerging ideas in real time. They’re also embedded directly alongside the transcript. So when AI summarizes or surfaces patterns later, it can work with data deepened by your input.
Think of AI summaries as a navigation tool that tells you where to look. They’re a starting point, not a substitution for the raw data or complete analysis outputs.
The single most important thing you can do to improve summary quality is to provide context. Without it, AI defaults to frequency: what was mentioned most often instead of what matters most to your study.
The prompt “summarize this interview” produces a topic overview. The prompt “summarize this interview with a focus on pain points, workarounds, and moments of frustration or surprise” produces something you can actually work with.
Look for tools with summaries that include citations and direct links back to the specific transcript moments they draw from. So that validating a claim takes seconds rather than requiring a full re-read.
The most natural way to start analysis is with questions. You just finished a session, and you already have things you want to know: did this participant mention the same friction point as the last one? What did they say about the onboarding flow? Was there a moment where their sentiment shifted?
In an analysis tool like Condens, you can ask those questions directly against your session in plain language. Just type your question and you’ll get a cited answer drawn from the transcript. In a way, it’s like interviewing your raw data, rather than searching for keywords or scrolling through timestamps.
This makes it a helpful first move. Patterns can begin to emerge before you have done any formal analysis, and specific answers surface in seconds. From there, you can follow up with a closer read of the raw transcript, move into clustering, or stop here if you only need specific answers rather than a full synthesis.
This is the step that gets skipped most often, usually because a solid AI summary makes it feel redundant. But it’s not.
Your job is to check what was filtered out, and to bring the kind of contextual reading that AI just can’t.
AI can identify that a participant used the word “frustrating” three times. But unless AI is trained on visual intelligence, it can’t know that they smirked when they said it, that they were describing a tool they’d already stopped using, or that their answer might have been shaped by how the question was framed. That kind of reading requires someone who was in the room, or close to it.
This is also where the most interesting analysis often happens. Where you notice things that do not fit the emerging pattern and point you in a direction you weren’t expecting.
Practically, this looks like:
Reviewing the AI summary and flagging anything that feels incomplete
Adding your own observations and interpretations as highlights
Marking moments that feel significant (even if you can’t articulate why yet)
Noting anything in the raw data that contradicts or complicates what the summary surfaced
Tags have always done two jobs: making data findable and making patterns visible. AI search changes the first of those significantly. Because when you can ask AI a question and get cited evidence from across your repository, data doesn’t need to be perfectly tagged to be retrievable. The pressure to build a complete taxonomy before your first session, and to apply it flawlessly throughout, is genuinely lower than it used to be.
What tags are still essential for is the second job. A tag applied to a highlight is a human judgment call, which can accumulate into a structured record of what your team has decided is meaningful, traceable across studies, and comparable over time.
This shifts tagging from a front-loaded obligation to something more like a byproduct of active analysis that’s applied as you go. The resulting taxonomy then tends to reflect what actually matters rather than what seemed important before the research started.
One thing that remains important regardless of how AI develops: metadata. Things like participant role, company size, or research phase are often not present in the content itself, and AI search can’t infer them reliably. “What did enterprise users say about onboarding?” is only answerable if that context about the users is already attached to the data.
For a closer look at how AI-powered search is changing the role of taxonomies, read Do We Still Need Taxonomies Now That We Have AI-Powered Search?
This is where AI’s ability to work across large volumes of data becomes most valuable. Rather than manually comparing notes across sessions, you can ask AI to find patterns in your dataset like what’s recurring and what’s diverging.
AI clusters around language patterns. So two participants can describe the same experience in different words and end up in separate clusters. On the other hand, two participants can use similar language for entirely different experiences and land in the same cluster. The highlights that do not cluster at all are worth examining carefully. That’s often where the outliers are, and a single data point that does not fit the pattern can sometimes be exactly what you are looking for.
The key distinction is between AI identifying patterns and AI interpreting them. The first is something AI does well. The second should still be left to a researcher.
Before a pattern becomes a validated insight, it needs to be traceable to specific moments in specific sessions, not just to an AI summary or a cluster label.
Validation means being able to answer:
Can you point to two or three moments in the raw data that support this insight?
Does it still hold up in context?
Is there any contradicting evidence?
Are there any participants whose experiences do not fit the pattern?
Research that can’t be traced back to evidence rarely survives the first hard question in a stakeholder readout. That’s why maintaining a clear evidence trail isn’t just good practice, it’s essential to upholding your research’s credibility.
This section shows what it looks like when the workflow detailed in the previous section is applied to a real project. Special attention is given to where AI changed the process and where it didn’t.
The example comes from Sina Richter, User Researcher at Condens. The study was a journey map project on how researchers decide on a tool stack. It was run twice, roughly eighteen months apart, with the same research question but a different workflow each time: first manually, then with AI-supported analysis in Condens.
The table below compares each stage of the process side by side. Two stages look identical in both approaches. The rest is where the difference shows up.
Stage | Without AI | With AI |
|---|---|---|
1. Framing the research | Collaborative alignment on assumptions, goals, and key questions across teams. | Same process. AI doesn’t change the framing conversation. |
2. Starting from existing evidence | Filter the repository by global tags, drag items onto a whiteboard, walk through what the team remembers from past sessions. Dependent on how well things were tagged at the time. | Plain-language search across the entire repository. Questions like “What do we know about how researchers discover new tools?” return cited answers, including highlights from raw data that never made it into a formal report. |
3. Preparing for sessions | Recruiting and interview guide preparation as usual. Discussion guide refined through manual review. | Same preparation, with AI helping fine-tune the discussion guide. Project questions set up in advance and run against each transcript as it comes in. Transcript generation starts during the debrief. |
4. During and after sessions | Record, take notes, debrief. Analysis happens separately, often with a significant time gap. | AI processes transcripts immediately, with Live Notes linking in-session observations to transcript moments. Prepared research questions run against the data right after each session, surfacing relevant moments without a full rewatch. AI-generated summaries provide a first overview. |
5. The deeper analysis | Filter highlights by tag, drag onto an affinity map, cluster manually, repeat. Joint session on the same board, then a longer individual finishing phase. Multiple passes to ensure nothing was missed. | The whiteboard already has structure when analysis begins. Tag filtering and AI-assisted clustering work in tandem, with manual refinement throughout. AI kickstarts the process, supports double-checking, and populates clusters with supporting evidence once the researcher has shaped them. |
6. The output | Journey map with key findings, highlights, and workshop-based synthesis. | Same type of output. No fundamental change in format or deliverable. |
7. The outcomes | Slower start, higher friction to begin analysis. Gaps between sessions and synthesis make it harder to re-engage with the data. | Faster entry into analysis. Immediate access to highlights, themes, and summaries lowers the barrier to start while context is still fresh. |
The parts AI changes the most are where it helps you get to the evidence faster, lowers the hurdle to start, reduces the time gap between session and analysis, and suggests a first structure you can work with and refine.
Worth noting is that the “framing the research” and “delivering the output” stages are identical in their approaches, with the thinking at the start and the deliverable at the end looking the same. The difference is everything in between: how quickly prior knowledge is surfaced, how sessions feed into analysis, and how much mechanical work the researcher has to do before the real interpretation can begin.
The judgment calls are still the researcher’s, like what a theme means, which observations matter to which stakeholder, how everything fits into something a team can act on. AI hasn’t changed that part. What’s changed is how much time and energy you have left for interpreting the data after the time-consuming and mechanical tasks are done.
AI can take on the time-consuming, mechanical parts of analysis, but the judgment calls, like what a theme means and what a team should act on, should remain with the researcher. A workflow that keeps you in the loop at every interpretive step is what lets you move faster without losing the rigor that makes your research credible.
If you’d like to dive deeper, check out our free guide. It explores what good analysis looks like with and without AI, how to ask better questions to get better outputs, and how to maintain rigor, traceability, and reproducibility once AI is part of your workflow. You’ll also find what 330+ research practitioners told us about how they use AI today, plus a case study on how UX researcher Michaël Dufranne uses AI to save three days per study.
Iva is the Senior Product Marketing Manager at Condens, where she focuses on positioning, messaging, and G2M. With eight years in B2B SaaS, she's developed a real appreciation for the messy, ambiguous early stages of building where there's no playbook yet and the work gets to shape itself. She's passionate about grounding marketing in real customer insight and translating that into work that resonates with the people it's meant to reach.
