The data was immaculate. The truth was somewhere else.

When the analysis for our first report came back, it was a lovely thing to look at. Clean tables. Every base accounted for. Fourteen concerns ranked in a tidy descending order, each with a percentage attached. You could have lifted any chart straight into a board deck and nobody would have questioned it.

That neatness is exactly what I want to talk about. Because the most important thing in that report was not in any of the tidy tables. It was in the gap between two of them, and no amount of processing power would have found it without knowing something about how human beings, and specifically women, answer questions.

What the data faithfully recorded

Asked to rate fourteen areas of life as concerns, the women in our survey put personal safety dead last. Seven per cent called it a significant concern. Half said it worried them "not at all".

Let me be precise about that number, because precision matters here. The data is not wrong. The data is not bad. It is a faithful record of exactly what 107 women chose when shown that scale. Taken at face value, it supports one conclusion, cleanly: women are not worried about safety.

Face value is the problem. Ask a woman to rate her personal safety as a "concern" and something human happens in the seconds before she answers. She compares herself to women in genuine danger. She decides her own experience does not clear the bar. She is not reporting risk. She is reporting whether she feels entitled to complain. The scale captures her answer perfectly. It just cannot capture the negotiation she performed to arrive at it.

We knew that negotiation would happen before fieldwork opened, because we are women, we research women, and we have sat across from enough women to recognise the reflex. So we built for it. Later in the same survey we asked an open question with no mention of safety anywhere in its wording: what one thing would make life easier for women right now? Sixteen per cent raised safety, violence or misogyny unprompted. Six of those women had rated safety "not at all" a concern minutes earlier.

Same women. Same survey. Twenty minutes apart. Two different answers, and both of them true. The rated scale told us what women will claim for themselves. The open question told us what they carry. The insight lives in neither number. It lives in the distance between them, and that distance only exists because humans designed the instrument knowing it would.

What machines read, and what they miss

This is where I part company with the current excitement about automated insight. AI is extraordinary at the part of research that is arithmetic. It will count, cross-tab, rank and visualise faster than any team of humans, and without transcription errors. Anyone who tells you it adds nothing to research is selling you nostalgia.

But AI reads data at face value, because face value is all data has. It has never sat opposite a woman who says "I'm fine" in a tone that means the opposite. It does not know that people fill out surveys tired, guarded, generous, performing a little, minimising a lot. It has no feel for which throwaway line is actually the pressure point.

One woman in our survey, asked what would make life easier, wrote: "I honestly don't know. I have no headspace to think about ways to make life easier. My brain is full."

A machine codes that as a non-response and moves on. A human reads it and recognises the most complete description of the mental load in the entire dataset. She could not answer the question because answering the question was another task. The finding is not in what she said. It is in why she could not say it.

The unsaid works the same way. Not one woman in our survey asked for confidence coaching. That absence is a finding, because an entire industry is built on selling women confidence they never requested. But a machine cannot flag what is missing, because it has no expectation of what should be there. Absence only becomes signal when someone knows the territory well enough to notice the shape of the hole.

And when automation does misread people, it misreads them with total confidence. Our own survey platform's quality checker flagged the phrase "gender pay equality" as profanity, and a woman's suggestion for a women-run taxi service as gibberish. Real answers from real women, machine-labelled as junk. The system was not malicious. It was doing what automated systems do: applying a general model to a specific human being, and never once suspecting the model might not fit her.

The illusion of robustness

This matters more right now than ever, because businesses are wiring their data together at speed. Survey platforms feed dashboards, dashboards feed AI agents, agents pass findings to each other through connected pipelines. Every connection makes the output smoother. Numbers reconcile. Sources agree. It all looks so streamlined, so definitive, so robust.

Neat is not the same as true. A pipeline that faithfully carries the answer to a question no one interrogated does not strengthen the finding. It launders it. A single-scale safety number flows through the stack looking every bit as solid as a triangulated one, and by the time it reaches a board slide, nobody is asking how it was made. That is the illusion: the polish of the surface standing in for the quality of what is underneath. The smoother the dashboard, the less anyone checks.

Humans understand humans

So here is where we have landed, and it is our proposition, stated plainly.

The tools belong in research. We use them and will keep using them, for everything that is genuinely a counting problem. But insight is not a counting problem. It lives before the data, in designing questions that anticipate how real people, in all their muckiness, will actually answer them. And it lives after the data, in reading what people meant rather than what they typed. It lives in triangulation: quant against qual, the rated against the unprompted, the said against the unsaid, until the places where the sources disagree start telling you the truth the individual numbers could not.

That is judgement, not processing. It comes from knowing people, and in our case from knowing women, because we answer questions the same way ourselves. We know the reflex to minimise because we have felt it. We know what "my brain is full" means because we have written it.

The data in our report was immaculate, and I stand behind every number in it. But the numbers were the record, not the insight. The insight came from two humans who knew, before a single response arrived, where the truth would try to hide.

Trust the data. It is telling you exactly what people said. Then bring in the humans, because someone still has to work out what they meant.

If you would like the report, please contact us via our contact form or email sev@moodinsights.co.uk