The question no AI can ask for you

The decision that happens before the research starts

Every research project has a moment that determines whether the work will be useful. It is not the moment the data comes in. It is not the moment the analysis is complete. It is the moment someone decides what question the research is actually trying to answer.

That moment is easy to underestimate because it does not look like work in the way that data collection and analysis look like work. There is no output to show. No deliverable to review. Just a conversation, or a thinking process, that ends with a question written down somewhere. The brief gets issued, the project starts, and the question recedes into the background.

But the question is doing more structural work than anything else in the project. It determines what data is relevant and what is not. It determines what a meaningful finding looks like as opposed to an interesting one. It determines, above all, what the research could actually change: what decisions are downstream of it, and what evidence would be sufficient to move those decisions in a different direction.

Getting that question right is the most consequential research skill there is. It is also the one that AI cannot perform.

What question design actually involves

The surface version of question design is straightforward: define the research objective, translate it into a set of questions, issue the brief. Most research projects go through some version of this process. The brief exists. The objective is written down. The project proceeds.

The deeper version is harder and less visible. It involves interrogating the stated objective until you reach the real one. It involves asking who is going to use the findings and what they are actually trying to decide. It involves identifying the assumptions that are already baked into the way the question has been framed, and asking whether those assumptions should be tested rather than taken for granted.

And it involves asking the question that separates useful research from research that produces findings without consequence: what would change our minds?

That last question is the hardest one. It requires the researcher, and the client or stakeholder behind the brief, to specify in advance what evidence would be sufficient to alter a conclusion they may already be inclined toward. It requires intellectual honesty about the difference between research that is genuinely open to finding something unexpected and research that is looking for confirmation of a position already held.

Most research briefs do not answer that question. They describe what the research will cover. They do not specify what the research could overturn.

Why this cannot be automated

The case for AI assistance in research rests on genuine capability. Pattern detection across large datasets, thematic clustering, sentiment analysis at scale: these are tasks where AI adds real value and where the practitioner’s time is better spent elsewhere. The argument for using AI in research is not wrong. It applies to a specific part of the research process.

Question design is not that part.

The reason is structural. AI systems are optimised to work with what exists: data that has been collected, text that has been written, patterns that are present in a corpus. Question design is about deciding what should exist. It requires reasoning about what is not yet known, what has not yet been asked, and what the consequences of asking it in one way rather than another will be.

That reasoning depends on context that is not in the data. It depends on knowing the organisation that commissioned the research: its politics, its prior commitments, the decisions it is actually trying to make as opposed to the decisions it says it is trying to make. It depends on knowing the market or category well enough to recognise when a question is framed in a way that will produce a misleading answer. It depends on professional experience accumulated across many projects, which generates the intuition that something about this brief is not quite right, that the question as written will not get to what actually matters.

None of that is accessible to a system that has not been in the room, has not read the previous three research reports for this client, and has not watched a board debate the same strategic question for two years without resolving it.

The brief that looked complete

Consider a brief that arrives with a clear objective: understand consumer attitudes toward a new product category. The target audience is defined. The methodology is suggested. The timeline is set.

A researcher who takes that brief at face value and begins designing the research will produce work that answers the question as written. Consumer attitudes, understood. Report delivered.

A researcher who interrogates the brief will ask a different set of questions before any fieldwork begins. What decision is downstream of this research? If attitudes are positive, what changes? If attitudes are negative, what changes? Has the organisation already committed to entering this category, in which case the research is being used to validate a decision rather than inform one? Is the real question not about attitudes toward the category but about whether this particular organisation is credible in it?

Those questions cannot be answered by analysing data. They require a conversation with the people who commissioned the research, an understanding of the strategic context, and the professional confidence to push back on a brief that is framed in a way that will produce findings the organisation cannot use.

That is what experienced researchers do. It is the work that happens before the work. And it is the work that determines whether everything that follows is pointed in a useful direction.

The cost of the wrong question

Research conducted against the wrong question is not neutral. It consumes budget and time. It produces findings that look authoritative. And it answers something other than what the organisation needed to know.

The most common version of this failure is not dramatic. The research is not obviously wrong. The findings are defensible. The report is well-constructed. But the organisation makes the decision it was going to make anyway, because the research did not address the specific uncertainty that was driving the decision. The brief was too broad, or framed around the wrong variable, or asked about attitudes when the real question was about behaviour.

The less common but more damaging version is the brief that has a directional assumption embedded in it that goes unexamined. The question is designed in a way that makes a particular finding more likely. The research confirms what the organisation expected to find. The decision is made. The assumption was wrong.

Both failures share a root cause: the question design stage was treated as a formality rather than as the substantive intellectual work it is. The brief was taken as given rather than interrogated. The researcher’s job was understood as answering the question, not as testing whether the question was the right one.

What this means for AI-assisted research

The practical implication for practitioners using AI tools is about sequencing and scope. AI belongs in the research process at the stages where the question has already been established and the task is to gather and analyse evidence against it. It does not belong at the stage where the question itself is being determined, because that stage requires the kind of contextual, relational, and experiential reasoning that AI tools do not have.

This is not a limitation that will be resolved by more capable models. The question of what an organisation actually needs to know, as distinct from what it thinks it wants to find out, is not a data problem. It is a judgement problem. It requires a practitioner who can sit across from a client and recognise the gap between the question on the brief and the question that would actually be useful. That recognition comes from experience, from understanding of organisational behaviour, and from the professional relationship that makes it possible to say: I do not think this is the right question.

Those capabilities compound in the practitioner over time. They do not transfer to a model.

The research pipeline that uses AI well is one in which the practitioner has done the question design work first, established what would count as a meaningful finding, and then deployed AI assistance to gather and process evidence efficiently against that established frame. Mimir operates on this logic: continuous monitoring of unprompted consumer conversation provides the evidence base, but the practitioner defines what questions that evidence is being gathered to answer.

The sequence matters. AI after the question, not before it.

This is the third article in a series on the irreplaceable practitioner. The series begins with you are still the star. The previous article covers why the researcher has to be able to defend every finding. The next article addresses the gap between pattern recognition and insight: why surfacing a pattern in data is not the same as understanding what it means.

Mimir monitors the conversations your briefs are missing, continuously and without prompting. start for free.