Some Thoughts on Always-On Participant Access

You can make research fast. My book is largely about how: scope tight, route to the right depth, run the mode the question actually calls for. Micro, sprint, deep. All of that compresses the time between a question and an answer.

None of it touches the one part that stays slow. Getting the right people in front of you.

Participant access is the tax nobody put in the budget. You can cut a study down to a single sharp question you could answer in an afternoon, and then wait three weeks for a panel. The speed you built on the method side gets eaten whole on the supply side. The faster the rest of the process gets, the more absurd the recruiting bottleneck looks sitting in the middle of it.

So the constraint was never really the research. It was the supply.

That is what sent me back to something I had read a while ago on the (now inactive) Dave Hora's blog, Dave's Research Co. He calls it always-on user access. He arrives at it through a Wardley map of how research is unbundling, and the phrase is his shorthand for one specific evolution on that map: specialized recruiting turning into always-on user access. The idea is that a stack of ordinary technologies maturing at once (video calls, scheduling tools, customer contact databases, and Research Operations as a function) has quietly changed the dominant mode of reaching users. Recruiting used to be a specialist activity you commissioned per study. It is becoming a standing capability that is simply on, the way electricity is on. The pool exists, the contact data is maintained, the scheduling runs itself, and reaching a user stops being a project and becomes a background utility. His argument is that once access is always on like that, the case for bundling research into a dedicated function weakens, and the work starts redistributing toward product managers, designers, and tooling.

The piece is from 2024, and in this field that is a long time ago. A lot of the scaffolding around it has dated. The map was drawn before AI moderation was real, before synthetic participants were a serious conversation, before half the tooling that now sits on top of access existed. Read it today and some of the surrounding assumptions feel like they belong to an earlier phase of the discipline.

The concept itself does not date. That is the tell of a good one. Always-on access as a description of where recruiting is heading is as true now as it was then, arguably truer, because the enabling technologies he pointed at have only matured further. The frame holds even as the world he drew it in has moved on.

What the article could not have accounted for is the thing that makes the concept finally useful rather than merely correct. When he wrote it, an open pipe did not help you much, because the person meant to walk through it could run five sessions in a day and then had to sleep. Access was solved on paper and throttled in practice. So the concept sat there, accurate and inert.

AI-moderated interviewing is what changed. Sessions in parallel, at the participant's hour, in the participant's language, with structured output instead of a pile of recordings. That is a different operating speed, not a better interview, and it is exactly the piece the 2024 version was missing. Fold the durable concept together with the new modality and you get something neither one is on its own.

Which is the whole point I want to make here. Getting users into the room was only ever half the problem. Hora named that half and named it well. The other half is what you do once they are in the room, and until very recently there was no good way to do it at any scale, so access on its own just meant more conversations with the same reachable people. Pair the open pipe with a real way to convert access into evidence and you finally have the whole thing. Leave them unpaired and you have volume, dressed up as insight.

Everything below is an argument for taking the pairing seriously. It is also thinking in progress, not a finished position. I know which way I am leaning, I have not earned all of it yet, and I will flag the places where the argument runs ahead of the evidence.

The concept is innocent

I said the concept stops at access. That is its honest scope, and the scope is where the trouble starts. The damage sits in the gap between what an open pipe enables and what teams do with it.

Here is what happens in practice.

When contact was expensive, it came bundled. You wrote a plan, because you were not going to waste a slot that took three weeks to get. You had a second person on the call, because the slot was scarce enough to justify two salaries. You wrote up what you learned, because the effort of getting there demanded a receipt. And you could remember which study a claim came from, because there had been six studies that year.

Cheap contact strips the bundle. What you get instead is more contact events with less rigor per event, no verification step, and no artifact at the end except a recording nobody will watch and a Slack message that says "talked to three users, they all hated it."

Then the accounting error arrives. Teams start reading contact volume as evidence volume. Forty conversations this quarter sounds like a lot of knowing. It is a lot of contact. Whether any of it changed what the organization actually believes about its users is a separate question, and nobody is asking it, because the number is right there and the number is big.

Contact is not coverage

An organization's picture of its users has properties worth tracking separately. How much of the user base it accounts for. How recently it was checked. How well supported each part of it is. And whether anyone is responsible for keeping it current.

Always-on access moves exactly one of those. Recency. It moves it beautifully. Your last contact with a user is measured in days instead of quarters and the chart goes up and to the right and everybody feels excellent.

Coverage does not move. In most implementations it quietly degrades, because standing panels are made of the reachable. Opted-in users. Active users. People who answer emails, who tolerate a video call, who have not churned, who like you enough to keep saying yes. The cheaper contact gets, the harder teams lean on that pool, because leaning on that pool is what makes it cheap.

So you end up extremely current about the twelve percent of your users who enjoy talking to you, and increasingly stale about everyone else. Recency hides the coverage failure and, worse, launders it. A picture checked recently feels like a picture checked well.

Nobody notices, because the users who would have contradicted you were never in the room. They were never in the room before either. The difference is that before, recruiting friction was a standing reminder that your sample was a compromise. Frictionless recruiting removes the reminder and keeps the compromise.

I suspect this generalizes past panels, and I want to leave it here as a note to myself more than a finished argument. Every instrument a team adopts serves some of those properties and is silent on the others. Analytics dashboards are strong on recency and weak on why. Surveys buy coverage and lose depth. Nothing serves all four, and almost nobody asks which ones a new tool is actually moving before they buy it. Worth working out properly at some point.

The other half of the loop

Back to the fold, because this is where the concept gets completed rather than corrected.

The objection to AI moderation is that its ceiling is lower than a good human moderator on a hard question. True, and it matters. It is also the wrong comparison. The relevant comparison is against the study nobody ran, because there was no capacity to run it, which is most of them.

Paired with micro research, it becomes a working two-stage process.

Stage one is access. The pipe is open, the pool is standing, contact costs nothing. This is Hora's contribution and it holds.

Stage two is conversion. Small, tightly scoped studies, one question each, moderated at machine cadence, producing structured claims instead of transcripts. Not a quarterly deep dive chopped into pieces. Genuinely small units of inquiry, cheap enough to run continuously and specific enough to be verified afterward. The machinery of that verification (what counts as a claim, how it earns a confidence level, when it expires) is the back half of the book, and I am not rebuilding it here.

What I cannot hand off to the book is the objection to my own comparison, because the comparison is new. The study nobody ran produces no evidence, which means it produces no bad evidence. A cheap, badly run study produces confident bad evidence at machine cadence, and confident bad evidence compounds in a way that admitted ignorance does not. So the pairing only beats the unrun study if the verification step holds at the new speed. On paper it holds. Whether it holds in practice, at volume, with real teams cutting real corners, is the thing in this piece I am least sure about, and it is the first thing I am setting up to test. If it does not hold, the loop is just volume with better paperwork, and I will report that too.

The two stages need each other. Always-on access without a conversion mechanism produces conversation volume and calls it insight. AI-moderated micro research without always-on access is an excellent instrument with nobody to point it at. Together they amount to something the discipline has been claiming to want for a decade and has never actually had: continuous engagement with users that yields evidence rather than anecdote.

Routing when the cost goes away

One consequence worth being honest about.

I route work across modes, micro through deep, and cost used to do part of that routing for me. Deep work was expensive so you saved it for stakes that justified it, micro was cheap so you ran it constantly. A crude prioritization function, but a function. The loop collapses the cost of micro and sprint work toward zero and the crude function stops working, so routing has to be justified on what you are trying to learn rather than what you can afford. A gap in recency routes differently than a gap in coverage, which routes differently than a gap in support. You cannot fix a coverage gap by talking to your standing panel forty more times, and you cannot fix a support gap by talking to anyone quickly. That is the note-to-self from the coverage section refusing to stay a note. I suspect the real routing logic lives inside it, and I have not worked it out yet.

There is a specific failure mode that comes with the speed. When talking to a user costs nothing, it becomes trivially cheap to confirm what the team already believes. Somebody has a hunch on Tuesday, six sessions by Friday, hunch validated, in the sense that six people selected by a person holding the hunch said things compatible with the hunch. That is not new. What is new is the throughput, and you can now generate confirmation faster than anyone can audit it.

So the allocation rule has to be structural, and this is an experiment I am committing to in public. Reserve a fixed share of the loop for the regions nobody is currently curious about: the segments untouched in eight months, the places where support is thinnest, the users who churned and will not answer. I do not know the right share. I am starting at a fifth, mostly because a fifth is a number, and I will adjust it when I find out what it costs. Curiosity is a terrible sampling frame. It goes exactly where the team already suspects something is there.

Why this raises the value of the layer above

Hora frames the shift in evolutionary terms, with access moving from specialized toward industrialized and commodity. Take that seriously and it produces a conclusion the map does not draw.

Commoditizing an input raises the value of the layer directly above it. Take his electricity image all the way: cheap electricity did not make electricians rich, it made everything that runs on electricity valuable. If access to users becomes a utility, and moderation and analysis become throughput, then the scarce thing is the maintained, owned, versioned account of who those people are and what they do. The interpretation layer.

Unbundling is real and the interviewing is going to keep leaking out. That leak is the argument for owning the layer above it.

What this means for ResOps

This is the part I have been circling for a while without writing down, and I think the exposure here is larger than the function realizes.

ResOps today runs on a service model, and UXR is the client. I go to ResOps and say I need eleven people who ordered from a merchant they had never used before in the last thirty days. ResOps writes the screener, sources the pool, handles incentives, books the calendar, and hands the sessions back to me. Intake, queue, delivery.

That relationship is where most of the friction in the function lives. Requests arrive vague and get sharpened by someone who was not in the planning conversation. Queues force prioritization arguments between researchers who all think their study is the urgent one. Lead times get quoted, missed, and resented. The researcher thinks operations is slow. The operator thinks the request was underspecified and arrived on a Thursday. Both are usually right, and neither problem is fixable inside the model, because the model puts a service desk between a researcher and the people they are trying to understand.

Always-on access removes the brokerage entirely. The researcher reaches the user directly. Which means the question is no longer what ResOps does differently. It is what ResOps is.

If ResOps is defined as the function that brokers contact, then direct access deletes the job, slowly, in a way that will be visible in headcount before it is visible in strategy. If ResOps is defined as the function that governs access, then direct access is the best thing that has ever happened to it, because governing an open pipe is a substantially harder and more valuable job than filling a closed one.

Governance is not a softer word for gatekeeping. Nobody needs a slower service desk. It means owning the conditions under which access happens and the standards the output has to meet.

I have not run any of what follows, so read it as bets, not a playbook. Six of them. I expect at least one to be wrong and I cannot tell you which one. That is what makes them bets.

Own panel composition as a coverage instrument. Most panels are currently optimized for responsiveness, which is a throughput property. The panel should be optimized for representativeness against the actual user base, with the gaps documented and visible. Publish the coverage map. Make it uncomfortable.

Own the recency clock. Someone has to know which segments have not been touched in six months and say so without waiting to be asked. No other part of the org is positioned to monitor this.

Govern claims, not studies. Study-level governance made sense when studies were the unit of work. In a continuous loop the unit is the claim, and claims need provenance, a confidence level, and a decay date. A repository full of reports is an archive. A repository of verified claims with expiry is infrastructure.

Take the instrument layer before someone in product tooling does. Somebody has to hold quality standards for AI-moderated studies: what a good protocol looks like, what probing depth is acceptable, which questions should never go to an automated moderator, how output gets validated before it enters the record. That job does not exist yet in most organizations. It is also the bet I am most convinced by and can prove the least, which tells you what stage this thinking is at.

Meter access against the picture, not against demand. The thankless one. If everyone can run studies, someone has to manage participant load, oversampling, and fatigue, and that means telling teams no when they have burned through their share of a shared resource. It will be unpopular and it is the job.

Refuse the volume metrics. Sessions run, time to recruit, cost per participant. Fine when contact was the constraint. Actively misleading now, because they will show the function performing beautifully right up until somebody notices the organization's picture of its users is broad, shallow, and confidently wrong. Get the metrics changed before that happens.

Hora described how users get into the room, and that description was accurate before there was any good way to handle what came next. I think there is one now. I am not certain, which is why my own loop is currently set up as an experiment rather than a rollout. The bet is that the organizations pairing the open pipe with a real conversion mechanism will spend the next two years building the layer that sits on top of cheap contact, and the ones that do not will spend the same two years having an enormous number of conversations with the same eleven hundred people who like them.

Ask me in six months how much of this survived contact.

🎯 Insight, not volume. That's the whole deal. Subscribe to The Voice of User.