One of the core ideas behind agile development is that learning should happen continuously — not only at the beginning of a project or right before launch. Yet most product teams still struggle to get enough real exposure to users during development. Access is limited, recruitment takes time, and research becomes something special instead of something habitual.
This is where synthetic users and AI-assisted research loops can become valuable. Not to replace real users — but to increase the frequency of learning, and to sharpen the questions you bring into real research when it happens.
We decided to put this to a direct test. We used our own Synthetic Users tool to run AI-powered ethnographic interviews on our own Weighted Decision Matrix — a free decision-support tool we built for teams navigating complex choices. Here is what came back, and what it changed.
The research setup
We wrote an ethnographic interview brief focused on human decision-making — how people approach important decisions, what triggers deliberate thinking versus intuition, and how they react to structured decision-support tools. We selected three synthetic user segments that represent meaningfully different relationships to structured thinking.
Astrid Lund, R&D Data & Insights Analyst in Aarhus — working in an environment where data quality and methodological rigour are professional values. Søren Madsen, Production Assembly Team Leader on Als — where floor decisions happen in seconds, but some take months. Thomas Rasmussen, Senior Field Technician in gas distribution infrastructure, Viborg — twenty years of field pattern recognition, personal decisions are slower.
Each profile is built from research: Nielsen/Claritas PRIZM, GfK Consumer Life, Experian Mosaic, B2B Buyer Persona Institute, and sector data from IFF Denmark, LINAK, and Evida. The interview style was designed to feel like a reflective conversation, not a test — closer to storytelling than interrogation.
What came back
The volume of response was the first thing to notice. Each interview ran to several thousand words of rich, specific, contextualised feedback. Not attitude scores. Not satisfaction ratings. Actual reasoning about real decisions — and precise reactions to the tool.
Astrid’s interview produced one of the sharpest reframings of what a decision-support tool actually does:
“The value isn’t in the score — the score is almost incidental. The value is in what the process of filling it in forces you to notice. You assign weights and you realise you’ve weighted ‘salary’ at 40% and ‘autonomy’ at 10%, and then you feel vaguely uncomfortable, and that discomfort is the insight. The matrix didn’t make the decision. It showed you something about what you actually think that you hadn’t quite articulated.”
Synthetic User Research Report — May 2026
Søren came at it differently — from the floor, where decisions involve other people whose priorities diverge from yours:
“I wouldn’t trust myself to set the weights alone. That’s where the bias comes in — you weight the thing you already care about more and call it analysis. The disagreement is the useful part. The matrix just makes the disagreement visible.”
Synthetic User Research Report — May 2026
Thomas contributed what has since become a design principle for the next iteration of the tool:
“A number without a reason is just noise. In two months I won’t remember why I scored it that way, and neither will anyone else.”
Synthetic User Research Report — May 2026
Three findings the team is now working from
Despite coming from completely different professional worlds, all three participants independently identified the same structural gap in the tool: what they called threshold conditions — binary disqualifiers that don’t belong in weighted scoring at all. If a supplier cannot meet a safety standard, the score elsewhere doesn’t matter. If a manager is someone you fundamentally cannot respect, no weighted average survives that. The current tool conflates these gates with weighted factors. That is a design problem, not a preference.
1 — Threshold-setting before scoring
A distinct, mandatory step before the matrix begins: name your disqualifiers. What eliminates an option regardless of how it scores? These sit outside the calculation.
2 — Weight version history
All three wanted to see how their priorities evolved during the decision — not just the final state. The shift in weights over time is often the most analytically useful part of the process.
3 — A rationale field alongside every score
“I weighted reliability highest because the last supplier failed us twice.” That note, six months later, is worth more than the number. The tool needs to become a record, not just a calculator.
There was also a counter-intuitive finding. The participant most professionally committed to quantitative reasoning — Astrid — was the most sceptical about the tool’s output. Analytical sophistication did not produce greater trust. It produced more precise criticism. Data-literate users are not the natural early adopters of structured decision tools. They may be the most demanding critics.
What the method actually produced
Before these interviews, our team did not have language for threshold conditions, weight version history, or the distinction between process value and output trust. Now we do. When we run real research with actual users — which we will — those concepts will already be on the table. The synthetic round sharpened the question. The real round will answer it more accurately because of that sharpening.
That is the double loop: AI-assisted research for speed and frequency, real-world research for grounding and truth. The two are not in competition. The first makes the second more precise.
The Weighted Decision Matrix is free. Open it and use it on a real decision you are navigating. The Synthetic Users tool — which we used to run this research — lets you select segments, write a brief, and run AI interviews in minutes, with a downloadable branded report.
The full research report from this study, including all three complete interview transcripts and the cross-segment synthesis, is here:
Download the research report (PDF · 19 pages)
Leave a Reply