Nobody hands a finance director a single number and expects it to be accepted on faith. A revenue forecast comes with a model behind it: assumptions in one tab, calculations in the next, a formula bar that shows exactly how each cell got its value. If a number looks wrong, you don't argue about the number. You open the formula, find the input that's off, and fix it. That's not a nice-to-have feature of spreadsheets. It's the entire reason we trust them with decisions worth real money.
Now compare that to how a lot of AI-assisted qualitative analysis gets used today. Paste the interview notes in, get five themes back. No visible steps. No way to see what got read and what got skipped. No formula bar. Just an answer, delivered with the same confident tone whether it's right or invented.
Qualitative findings, from customer interviews, market research, and voice of customer feedback, routinely inform decisions worth as much as any number in a financial model. There's no good reason to accept less rigour for one than the other.
Why did we ever accept opacity in quantitative work?
We didn't, and that's the point. Until the 1950s, every calculation of consequence was done by hand, and the habit that came out of that era never left: show your work. When computers took over the arithmetic, the expectation didn't disappear, it moved into the tool. A spreadsheet is popular specifically because it keeps the manual-era transparency while removing the manual-era labour. Every cell traces to a formula. Every formula traces to an input. The computation got automated; the visibility didn't.
More sophisticated quantitative tools, statistical software, econometric models, keep the same bargain. You can always ask "how did you get that number" and get a real answer, not a shrug.
Qualitative analysis, historically, earned its rigour a different way: through sheer manual effort. A researcher who spent three weeks reading every transcript by hand had, by necessity, actually engaged with all of it. The rigour was real, but it was also slow enough that scaling it to hundreds of documents was never realistic for most commercial timelines.
AI removed the speed constraint. It has not, in most tools, replaced the transparency that manual reading used to guarantee as a side effect. That's the gap.
What does "showing the steps" actually mean for text?
Not literally exposing a large language model's internal weights. Nobody, including the model's own creators, can fully explain why a given output emerged from billions of parameters, and pretending otherwise would be its own kind of theatre. The useful analogy isn't "show me the neural network." It's "show me the same kind of audit trail a spreadsheet gives me."
For qualitative analysis, that means three specific things:
Every category traces to specific evidence. Not "customers mentioned pricing concerns," but the exact quotes, from the exact documents, that a category is built on. This is the equivalent of a formula referencing specific cells rather than a hardcoded number nobody can trace.
Every document shows what was and wasn't used. Open a source transcript and see which passages fed into the analysis, highlighted, and which didn't, on a plain background. This is closer to an audit trail than a formula bar: it's the check that catches what the analysis missed, not just what it found. We've written more on why this second direction of transparency matters more than most tools admit.
Every step is editable, not just visible. A spreadsheet formula you disagree with, you change, and the result updates. An analysis category that's mislabelled or wrongly grouped should work the same way: directly editable, not locked behind a black box that only takes prompts.
Isn't this just "human in the loop"?
The phrase gets used constantly and means almost nothing operationally. Every AI vendor claims a human stays in control. In practice, "human in the loop" often means: raw data goes in, a black box does something, an answer comes out, and the human's role is to click accept.
That's not a loop. That's a rubber stamp with a delay built in.
A genuine loop requires the same thing a spreadsheet requires: intermediate states you can actually see and change. Panko's decades of research into spreadsheet accuracy concluded, across fifteen years of studies, that spreadsheet errors are both common and non-trivial, meaning even a fully visible, fully editable tool doesn't eliminate mistakes on its own. What it does is make mistakes findable. An expert reviewing a spreadsheet can catch a wrong formula because the formula is right there to inspect. An expert reviewing an AI-generated theme summary with no visible working has nothing equivalent to inspect, so errors that would be obvious in a spreadsheet slip through undetected in a black-box qualitative tool.
According to KPMG's 2025 global study on trust in AI, 66% of people rely on AI output without evaluating its accuracy, and 56% report having made a mistake in their work as a result. That's not a criticism of the people; it's what happens when a tool gives you nothing to check against. Remove the formula bar from Excel and usage patterns would look identical.
What this looks like built into the tool, not bolted on
The point isn't that visible steps are a feature you add later. It's that the whole workflow needs to be structured around them from the start, the way a spreadsheet is.
Skimle's approach follows the same logic Excel does for numbers: bring in audio, video, text, or tabular data, run rigorous analytical workflows to identify categories, and keep every step traceable in both directions, from summary to source and from source back to what was and wasn't picked up. Categories and coding decisions stay manually editable throughout, the same way a formula stays editable after a spreadsheet calculates it. The output isn't just a report; it's charts, cross-tabs, and export formats built to communicate a story the same way a well-built spreadsheet model communicates one.
For market researchers and consultants used to defending a quantitative model line by line, this should feel familiar rather than novel. It's the same standard, applied to text.
The test to apply to any AI-assisted analysis
Before trusting a theme, a summary, or a set of findings an AI tool produced, ask the spreadsheet question: if I disagreed with this number, could I find out why it's what it is?
If the honest answer is "no, it just came out that way," you're looking at a black box wearing an analysis-shaped label, not an analysis. If the answer is "yes, here's the exact evidence, here's what else was in the data, and here's how to change it," you're looking at something closer to what a spreadsheet has trained every professional to expect by default.
Qualitative insights shape pricing decisions, product roadmaps, and market positioning worth as much as anything in a financial model. There's no reason the standard of proof should be lower just because the underlying data is a sentence instead of a number.
Try Skimle for visible-steps analysis
If you want your qualitative findings to hold up to the same "show me the formula" standard you'd apply to a spreadsheet, it's worth testing on a real project.
Try Skimle for free and trace a finding back to its source, then back further to what else was in the document.
Related reading:
- Two-way transparency: creating confidence in AI to make it useful for real work
- A citation is not an analysis
- 9 design criteria for AI qualitative analysis tools
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
