AI isn't a variable, it's a constituency. And understanding this difference changes if you're studying effects of AI in a snapshot vs. studying impact on AI as an evolving party over time.
There's a specific kind of vertigo going around UX research right now. The methods most of us trained on assume a stable object: you study a system, it behaves the same way next week, and your findings hold long enough to ship against them. AI breaks that assumption, and a growing body of writing has started naming the break out loud — talk of "UX 3.0," for autonomous, self-learning systems that don't sit still the way the deterministic software our methods were built for did, treating AI as a research participant. All of it makes sense, but the subtext under most of it is anxiety: the ground is moving and the playbook might be obsolete.
I felt that vertigo too. Then something settled it, and this piece is really just me thinking that through out loud. I realized I'd been researching a non-human, semi-autonomous thing in the middle of my studies for years — I just called it a marketplace. Not identical to AI. But close enough that the muscle transfers, and close enough that the panic is, I think, a little overblown. So this isn't a new framework. It's a reflection: an argument that a lot of us are better prepared for this moment than the newness makes us feel, and that the preparation is hiding in our own past work — and I'm hoping this piece will help you, or urge you, to look a little closer.
The reframe
The Heuristic
The B2B/B2C split is really about decision context
Every UX researcher learns the B2B/B2C distinction early. It's a genuinely useful heuristic. In B2C, the user and buyer are usually the same person. Populations are huge, decision cycles are short, and you can lean quantitative — big-N surveys, unmoderated usability, A/B tests at volume. In B2B, the model inverts. The user, buyer, and budget-holder are often three different people. Populations are small and hard to reach, so you go deep: contextual inquiry, longitudinal studies, embedding in workflows. Decision cycles are long, stakes are high, and success is measured in workflow fit and integration, not engagement.
I started my career researching logistics software for planners and dispatchers — deeply technical, domain-heavy B2B work where every interview required me to understand freight operations, route constraints, and fleet windows before I could ask a single useful question. Unlike a consumer app, a bad UX moment costs real spend in real time — so behavioral and diary data matter more than stated preference. Then I moved to Meta and consumer scale flipped the whole playbook toward B2C. But the problems that taught me the most were neither — they were B2B2C marketplaces, a creator economy in the Metaverse and then advertiser tooling on Instagram, where the platform only works if both sides win and something automated in the middle decides how the two sides even find each other.
The B2B/B2C split presents itself as a taxonomy of users. It's actually a taxonomy of decision context.
How complex the decision is, how much domain knowledge you need before you can ask one good question, and how many parties' incentives you have to reconcile. And in those systems I learned the thing this whole essay is about, which is that the platform is not the business. My own path taught me that the slow way.
The Realization
The platform has a health, and it isn't the same as the objective
Three studies, one lesson arriving three times. Each time, the object I was really studying wasn't a user — it was the automated thing standing between two sides.
The thing in the middle
Logistics and the Avatars SDK taught me the thing in the middle is itself a party. At Optym I wasn't really researching how dispatchers used software. I was researching how they trusted an optimization engine to make calls they used to make by hand — which driver takes which route, when to cede control to the machine. The interface was the surface; the real question was what makes a human comfortable letting an algorithm run their operation. I hit the same dynamic with the Avatars SDK — game studios weren't evaluating our UI, they were deciding whether ceding avatar rendering to Meta's code would break their pipeline. In both, the intermediary had a logic of its own that determined everything, and studying it as inert plumbing would have missed the entire question.
At Instagram Ads, advertisers wanted verified phone numbers for their leads, and the fix on the table was an OTP flow — a user sees an ad, submits their number, gets a verification code. From the advertiser's side, pure upside: cleaner data, higher contact rates, better ROI. By every business measure, the customer was happy. Then I tested it with consumers and the floor dropped out. Nobody expects an OTP from Instagram inside an ad. They read it as phishing — the brand overreaching, a trust violation — and it broke their entire model of what an ad interaction is. Here's the thing that reorganized my thinking. The advertiser pays Meta. The consumer pays nothing, never signed up to be a customer, and their trust is the single thing the whole marketplace is built on — lose it and inventory quality degrades and the system caves from the bottom up. So there were two different things I could advocate for in that room, and they were not the same thing. I could advocate for the advertiser's data quality. Or I could advocate for consumer trust. But I advocated for the health of the platform — what every future dollar quietly depended on.
The advertiser gradient taught me the platform shapes what its users become. "Advertisers" aren't one audience; they're a spectrum of sophistication, and the platform sits at every point on it. At one end, the solo business owner on an automated campaign — time-poor, no marketing expertise, attached to every dollar, and perfectly happy to let the machine handle it because they have no expertise to defend. At the other, the agency trading desk — deep experts, embedded workflows, buying committees, people who'll clock in five seconds whether you understand viewability and incrementality. Same platform, opposite relationships to the automation in the middle. And here's the part that matters later: the automation can meet the novice two ways. It can scaffold them toward competence, or it can do the thinking so completely they never build any. The platform is not neutral about what its users become. Its health includes their growth — another thing no quarterly objective was measuring.
The advertiser gradient
That's what I mean by a constituency. Not a party with feelings — the optimization engine didn't have any, and the model won't either. A party with a role, a logic, and a health that no one at the table is paid to protect, so it goes unspoken unless a researcher speaks for it. And there's a new one in the room now.
The Jump
AI is the newest thing in the middle
The jump is short. AI is an evolving, semi-autonomous intermediary with a health of its own that is distinct from the business's objectives for it — and most of the field is still treating it as a variable, the knob you toggle to see what happens to your real subjects. "Does the AI suggestion improve task completion? Do users trust AI results more than human ones?" AI as the independent variable, humans as the dependent one.
Run the marketplace lens over that and it collapses exactly the way the two-party playbook did. The business objective for an AI system is adoption, engagement, reliance, deflection — the machine gets used, and used more. The health of the AI as a constituency is a different thing: whether it's legible, whether its confidence is calibrated, whether users can correctly tell when to trust it and when to push back, and — the marketplace's parting gift — whether it's building its users or deskilling them.
Where objective and health diverge
Those two diverge at a precise and dangerous point. The business reads high reliance as a win. Read against the system's health, high reliance is a question — because a user who over-relies has stopped being able to catch the machine when it's wrong. It's the ad-load problem again: the metric climbs while the health erodes, and the side holding the target can't see it.
Where This Leaves Us
The panic is a little overblown — the muscle already transfers.
This is also where the "AI breaks our methods" anxiety turns out to be half-right in a way we already know how to handle. The genuinely new wrinkle is that this intermediary evolves — the model you studied isn't the model in production next month — so you can't hold it constant, and a one-shot study can't capture a relationship that's still forming. That's real. But researching for a non-human party with a health of its own is a muscle a lot of us have already built. We just built it on things that sat still.