Part 1 of a series on what research actually owns from here.
The line you hear at every panel is that AI democratized research, and in the narrow sense it did. Anyone can run a study now, with templated guides and unmoderated tests that spin up in an hour and survey tools that draft their own questions, so the barrier that used to sit in front of collection has mostly fallen away. That part is true, and it is also not the part that matters, because collecting data has been getting cheaper for a decade and the field survived all of it. SurveyMonkey did not need a language model, and panel vendors were selling two hundred responses in a week back when GPT-2 was still a curiosity. If cheap gathering were going to reduce research to clerical work, it had every chance and mostly didn't, because the value was never in the gathering. It sat in the two things a machine was not supposed to be able to do: design the instrument well, and read what came back. The update worth sitting with is that a machine now does the first of those genuinely well and produces a convincing version of the second, and those are not the same achievement even though they look alike from across the room.
The Instrument and the Reading #
Start with the instrument, because this is where researchers get defensive and lose the room before the argument has even started. Writing a survey that actually maps to your goals, drafting an interview guide that gets at the question instead of circling it, sequencing the probes so the early answers don't contaminate the later ones, all of that was real craft, and for years it was the thing you could point to when someone asked what the researcher added over a sharp PM with a Google Form. Hand a current model a clear set of goals now, Fable 5 or ChatGPT 5.5 on the max-thinking or pro modes especially, and it will map questions to those goals cleanly, strip the double-barreled ones, catch the leading phrasing, and hand you back a guide a competent researcher would mostly sign off on. That is not a hollow imitation of research. It is a good instrument, often a genuinely good one, and insisting otherwise mostly signals that you last tried this a year or two ago, which the people you are arguing with can tell immediately. So I concede it plainly, and on the record: the skilled front end of collection has commoditized, and it commoditized because the tools got good at the work, not because anyone's standards slipped.
The reading is where the same tools do something that looks identical from a distance and is not. A PM drops the readout into a model and asks for the takeaways, and a clean interpretation comes back a few minutes later, fluent and confident and already shaped for a slide, and the decision gets made that afternoon on the strength of it. The output is not clumsy, and that is exactly the trap, because it reads as well as the interview guide did and so it earns the same trust the interview guide earned. But writing questions from goals is a task that lives entirely inside the prompt, while reading a finding is a task that mostly lives outside it. The model that wrote you a good guide had everything it needed sitting in the request. The model reading your results does not, and the worse part is that it has no way of noticing what it is missing.
Do that across a few hundred decisions a quarter and the shape of the whole function changes without a memo ever going out. The studies still get commissioned, the instruments still get written, increasingly by the tools themselves, and somebody still gathers the data. But the step where research actually turns into a decision, the reading, has quietly migrated to whoever happens to be holding the model, and it is not coming back through the front door you have been standing in front of.
The Instinct Is to Lock the Door #
The reflex, when you feel that happening, is to protect the data. You lock down the repository, you put research in the path of anything that ships, you try to make sure nothing goes out the door without passing through you first. It feels like taking control, and it is aimed at the wrong door entirely. Collection is the part that is commoditizing, the cheap part that scales, the part vendors want and PM tooling wants and the org is quietly delighted to let spread because none of it costs a headcount. Plant your flag on data gathering and you have planted it on the one piece nobody was trying to take from you in the first place, which means you can win that fight completely and still lose, because while you were standing guard over the warehouse the expensive work, the reading, got routed to whatever tool sat closest to the person actually making the call.
That routing is not hypothetical, and in plenty of organizations it is being built on purpose. Something gets stood up that ingests every readout the research team has ever produced, every deck and every topline, and generates its own synthesis on demand behind a clean, fast interface that the PMs understandably love. Nobody who built it can tell you how it decides which of two contradicting studies to trust, or whether it has any idea that the segment in one of them was redesigned out of existence a year ago, because it doesn't. It reads the documents in front of it and produces something that reads like research. That is a research-shaped object, a phrase I have used before and intend to keep using because nothing better has come along, and the trouble with it is precise rather than vague: it has the full shape of analysis while missing the only thing that makes analysis worth paying for, which is the context that is not written down anywhere in the document it is reading.
A Reading That Knows Where It Sits #
The difference, then, is not skill but reach, and that is the part people miss when they want to be reassured that the model simply isn't good enough yet. It is good enough. The tool interprets one report in a vacuum because the deck you hand it is the entire world it has access to, and inside that world it performs well. What it does not know is what the organization already believes about these users, or that this finding cuts hard against a study from last spring, or that the flow in question hasn't been looked at since a redesign changed everything underneath it. It cannot tell you how confident anyone should be, and it never will from the document alone, because confidence is not a property of the document. It is a property of everything standing around the document.
You read the same deck against all of that, whether or not you narrate it to yourself while you do it. You read it against coverage, which asks whether you have even looked at this or are quietly guessing. You read it against freshness, which asks when you last looked and whether the ground has moved since. You read it against confidence, which asks how much this should actually move you given everything else you already know. The reading you produce that way knows where it sits in the larger picture, while the tool's reading floats, perfectly fluent and anchored to nothing.
That larger picture is the Frame, and I am not going to re-explain it here because I have done that already: the linked piece is the short version, where the concept first took shape, and the long version is in the book. What matters for this piece is what the Frame carries that the document itself doesn't: coverage, freshness, confidence, and an owner. Three of those tell you how good the model of your users currently is, and the fourth is the one this whole piece has been circling from the start.
I should be honest about the limit of my own argument here, because there is one, and it follows directly from everything above. If the problem is that the context lives outside the prompt, then a good enough tool wired into a genuinely maintained repository closes some of the gap, and probably more of it than I am fully comfortable admitting in print, because you would be handing the model the very thing it was missing. But a pile of old readouts is not a maintained model of your users, and most repositories are the pile rather than the model, which means the tool inherits whatever state you actually kept and almost nobody kept state. So the gap holds for now, and it holds mostly because of a failure on our side rather than because the tool couldn't in principle do the job once we had done ours. I am not sure that is a comforting reason for it to hold, and I would rather say so than pretend the moat runs deeper than it does.
And there is a harder question waiting behind that admission, which is what happens on the day somebody actually keeps state, wires good retrieval into a Frame that is genuinely maintained, and asks why a human should own the assessment at all rather than supervise the system producing it. I do not have a complete answer, and I am suspicious of anyone who claims one this early; it is probably its own piece. What I can see is that my argument, pushed that far, stops being about capability and turns into an argument about accountability, and I would rather make that turn out loud than let it happen quietly.
Control Is a Tell #
The accountability version starts with retiring the word control, because the moment research says that word out loud the argument is already lost. A team that says it needs to control the data sounds exactly like a guild protecting its turf, and every PM in the room hears it that way and quietly arrives at the opposite of the conclusion you wanted. They are not wrong to hear it that way either, because "let us gatekeep the inputs" is a genuinely bad pitch and it deserves to lose on its merits.
Strip the territorial language off it, though, and the thing actually being asked for is accountability rather than control, which is both harder to say and less flattering to want. Somebody in the organization has to be accountable for what the company believes about its users, and has to answer for it when the belief turns out to be wrong, and that is a real job rather than a turf claim. Hand interpretation to everyone with a login and that accountability does not get distributed across the org the way the optimistic version promises. It simply evaporates, and what you are left with is not better understanding but more of it, more confident and contradicting itself across four teams at once, with nobody in a position to say which version is the one the company actually holds. An intelligence function, the thing I keep saying UXR has to become, is not the team that hoards the inputs but the team that owns the assessment. I made the structural version of that case already, and this is the same argument standing one layer down, at the level of who gets to say what a finding means.
Where I Keep Landing #
Where all of this keeps taking me is somewhere close to the opposite of the reflex. Not rebranding the team an intelligence function while the same request queue runs underneath the new name, and not fighting harder for control of the warehouse either. Letting collection come from wherever it wants to come from, from PMs and vendors and panels and the synthetic stuff, and letting the model draft the instruments too, since we have already established that it can. If the value was never in the gathering, then defending the gathering is precisely what makes research read like a vendor with a territory problem, and the thing worth owning has to sit somewhere else. As far as I can tell it is the gate, the point at which an input, wherever it originated, only becomes part of what the organization believes by passing through the Frame and getting reconciled against everything already there.
The picture I keep drawing for myself is an ordinary Tuesday. A PM runs a self-serve study on the checkout flow and the tool hands back a finding, users want the order summary step gone, confident and bulleted. At the gate, that finding meets the rest of the picture: a study from the spring with twice the sample that found the opposite, a redesign in between that makes both readings suspect, a segment nobody has properly looked at in eight months. What goes back to the PM is a position rather than a summary of their deck: the finding is real but it is the weakest of three signals on this flow, the strongest one points the other way, nobody has looked properly since the redesign, and here is the two-day study that settles it. No tool wired to that one deck writes that paragraph, because the paragraph is mostly made of things the deck does not contain. Whether most research teams are actually in a position to stand at that gate today is a separate question, and I go back and forth on it.
Part of why I go back and forth is velocity, because in this version it stops being optional. The two-day study has to actually take two days, because if the reading lives with you and the answers take three weeks, the organization will route around you to the tool that answers in seconds, and from where they sit that will feel like an entirely reasonable thing to do rather than a betrayal. So speed and ownership travel together, and I have come to think of them as one problem wearing two faces rather than two, which is most of what the book is about. I will leave the how in the book rather than pretend it compresses into a paragraph without losing the parts that make it work.
Which Half You Keep #
From here I can only see two ways this goes, and I am suspicious of clean binaries, mine included, but I keep failing to find the third one. In the first, you own the interpretation layer and you let collection commoditize all around you, because the collecting was never where the value lived, and the organization finds it cannot make a confident call about its own users without the thing you maintain. You end up in the room, but not because you asked to be there or made a case for a seat. You are in the room because the decision is simply wrong without you, and everyone involved can feel it. In the second, you keep the data. You win the access fight and you guard the repository and you wrap a tidy process around all the inputs, and it looks for a while like you protected something. Then one quarter a system gets switched on that does the reading well enough, and you find out what is actually left for the team whose only remaining job is to go collect more of the thing that was already cheap to begin with. It is a vendor relationship at that point, no different in kind from the panel vendors you used to manage, except that it is happening inside your own company where it is harder to see clearly and much harder to leave. I do not know which of those two most teams are going to walk into. I have a guess, and I do not love it.
🎯 If this one landed, subscribe**to The Voice of User. Essays on UXR, the organizational dysfunction underneath it, and the parts of the AI conversation the vendors are too invested to say out loud.
📖 The full version of this argument, velocity and the Frame and how to actually hold both at once, is the book. It is up for preorder now.