Episode 4 is here
This episode is our answer to the question many SaaS tool companies are asking: are we afraid of Claude, GPT, Fable and the rest of them? Short version, we are not. We are heavy users of these advanced AI tools and reach for the newest ones the moment they land. The episode is about why a better model makes us more excited rather than more nervous, and what that says about how analysis tools should be built.
The engine and the car
The frame for the whole conversation is an analogy. A frontier model is an engine, and a very good one. Ten years ago the state of the art was closer to a water wheel: you could count word frequencies and see which ones co-occurred, which is roughly where NVivo, ATLAS.ti and MAXQDA started their mixed-methods features. GPT-3.5 was the steam engine. What we have now is closer to a petrol or electric engine.
But an engine is still an engine. If you want to get somewhere, you have to build the car around it: the transmission, the steering, the seats, the dashboard. That is the business Skimle is in. We take the best engines on the market and build the vehicle that professionals can actually drive to work. For the wider argument about how AI is reshaping the software layer underneath knowledge work, see the death of SaaS or a renaissance of better software.
Why the raw engine is dangerous for analysis
We spend a good part of the episode on an experiment we wrote up on the blog. Olli generated 650 short customer comments, sorted them randomly, and labelled every second row "Finnish customer" and every other row "American customer", completely at random. He then asked the raw model whether there were cultural differences in the data. It confidently found plenty: the Finns terse and reserved, the Americans expressive and emotional, complete with advice on how to run the advertising campaign differently.
Then he flipped the labels and asked a fresh instance the same question, this time telling it to act as a professional market researcher and not to hallucinate. It produced the same confident cultural story, except now, because it had been told to cite evidence, it went and found real quotes to support conclusions drawn from pure noise. Convincing, useless, and genuinely dangerous if a real advertising budget rode on it.
The lesson is not that the model is bad. It is that a bare model asked to "analyse the data" often just writes something in the genre of a synthesis. It produces an output that resembles what an analyst would give you without any analysis having taken place. We unpack the failure modes in hallucinations, context and the black box, and the same distinction runs through our guide on how to code qualitative interviews with AI: chatbots summarise text, they do not code it.
The interim artifact: transparency you can trace
The design principle that separates a car from a bare engine is transparency, and specifically the interim artifact. Between the raw data and the final summary there should be a structured, inspectable object: the coded data. Academic users of the legacy tools know this well as the coded transcript. Skimle produces it instantly and transparently, so you can trace any conclusion down to the source material and back up again, and it keeps an analytical log of what the AI actually did.
This is the same reason agentic coding works so well in software: between the request and the compiled program sits the codebase, a structured artifact that both humans and machines can read. A senior engineer does not read every line any more, but at the moment of truth they can ask exactly what was done and why. We want that same interim object for qualitative analysis. It matters even more when you work in a team, and doubly so for credibility: "the computer told me the Americans are more expressive" is not a finding. This is the heart of our argument for two-way transparency, and the architectural reasons behind it are in why RAG-to-riches does not work.
Predictability and genuine control
Two more design criteria follow from the same thinking. The first is predictability. Agentic AI effectively rebuilds the car every time you ask, and you do not quite know what you will get. We are aiming for something closer to Excel: a tool that takes the same shape every time, responds the same way, and can be mastered like a musical instrument. That is what lets you build routines, train a team, describe your method for an academic paper, and trust that a chart means the same thing this week as it did last week.
The second is genuine user control, which connects to what people call cognitive offloading. The harder it is to steer a tool, the more you simply accept whatever it hands you. We build for experts who should be more cognitively engaged, not less, so imposing your own framing on the data has to be easy: change the category structure yourself, test an alternative reading, dig deeper, without prompting your way back into a black box. That frees experts to spend their time in the interpretive stage rather than weeks in mechanical first-order coding. The risk of surrendering that judgement is the subject of is AI destroying the ability to think, and the wider stakes in quality as the differentiator in the era of AI.
From single-player to multiplayer
Most agentic tools today are single-player, and often single-session: a lot happens in the background of one chat window that you cannot pause, branch, or hand to a colleague. Real knowledge work is a team sport. When you have thousands of pages across several languages, the old fix was to split the reading, one person takes these interviews, another takes those. That is not the optimal way to use a team.
A shared, structured analysis lets everyone see the big picture and then dig into the themes they care about across the whole corpus, rather than being boxed into a slice of it. It also changes the team model. The traditional consulting "surgeon" pattern, one person operating while others pass the instruments, is limited by the surgeon's own attention and bias. We would rather elevate everyone in the team to a full collaborator who understands the whole context. This is the collaboration story we tell in how humans and agents collaborate in qualitative research, and it is why Skimle is also a database you can revisit: add a new interview and it is coded into the same scheme as the rest, ready to compare over time and share with colleagues. A car that stays where you park it, rather than driving itself back to the garage and looking different every time you open the door.
New in Skimle: Agentic analysis
We close on the updates. We have plugged in the latest frontier model, and the difference is immediately visible in better-written summaries and sharper chat responses. The bigger news is agentic analysis: you collaborate with the AI to define the research focus and how to code the data, then Skimle codes the interviews in its structured, predictable way, with direct quotes that are verified and organised into categories.
One feature in Agentic analysis we like is counter-evidence. The tool is forced to stay within the guard rails of the process, and for every theme it systematically searches the full corpus for cases that contradict it, which is exactly the discipline good qualitative research demands. Agentic analysis produces an executive summary, an extensive research report for anyone writing up or presenting, and a data table behind every claim showing which informants it rests on and whether the evidence is cross-case or within-case. One way to describe it: two decades of qualitative research craft distilled into something that helps every user, exposed transparently rather than hidden away.
It is a good time to be alive for anyone who learns to use these tools well. The two failure modes are equally wrong: reject AI as slop and refuse to touch it, or use it blindly and ship the slop yourself. The sophisticated middle, learning to use it properly, is where the value is. If you want to see how Skimle approaches it, how Skimle works is the place to start.
About Skimle and Skimlecast
Read more about Skimlecast and watch Episode 1 here! You can watch all our episodes on our Spotify or YouTube channels. If you have any comments, feedback, topics you would want us to cover or other things you want to share, please connect with us through our form or via email through olli@skimle.com.
If you are interested in using Skimle, check out how Skimle works. You can also try Skimle for free and see how AI-assisted qualitative analysis handles everything from academic interview data to consulting and due diligence at scale.
Meet the cast
Henri Schildt is a Professor of Strategy at Aalto University School of Business and co-founder of Skimle. He has published more than 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
