A dispatch from 2036

In 2036 there is an agency that keeps a synthetic panel running that nobody has briefed in four years. It answers anyway. Every morning it files opinions on products that were never launched, for clients who never asked.

The agency keeps it on as a reminder, the way a lighthouse keeps a foghorn: not because it says anything true, but because it says something, endlessly, at a price, and for a few strange years the entire industry mistook that for research.

This is an account of how we got from there to here. It runs through a reckoning with a stupid name and a |serious| invoice.

I don’t care < Someone else was paying

For three years, roughly 2023 to 2026, market research ran its AI transformation on somebody else’s money. Every synthetic panel, every auto-coded transcript, every agentic pipeline was priced against tokens subsidised by venture capital. The machine seemed cheap because the machine was a loss leader. Agencies rebuilt their margins on top of it and called the rebuild innovation.

Then June 2026. The industry called it the Tokenpocalypse. The numbers were these: flat subscriptions flipped to meters overnight, bills jumped twentyfold, and companies that had spent a year telling staff to use AI as much as possible blew through annual budgets by spring and were rationing by summer. The detail that should have chilled every researcher came out of a leaked audio clip: it was not the engineers burning the tokens. It was the ordinary knowledge workers, turning documents into decks, summarising, drafting, asking. Us. We were the consumption.

The paradox made it worse. Token prices fell throughout, halving in a year, but consumption grew four and a half times over, because cheaper units invite heavier use. The unit cost of intelligence collapsed and the bill went up anyway. The industry had built its methods during a | pricing fiction |, and when the fiction ended, the methods were wearing price tags nobody had ever read.

Two rackets, it turned out, on one invoice. Research had always sold certainty it did not possess. It had lately been selling it at a price that did not exist.

F**k > The machine keeps sending bills

The volume work went first, as everyone had predicted. Trackers, concept screens, the endless U&As, all swallowed by automation during the subsidy years. Agencies shed the humans who used to do the work and pointed proudly at the margin.

Then the meter started, and the automated layer that had replaced those humans began charging like a law firm. Agencies discovered they had | sacked their own capability | and were renting it back by the token. The ones that had gone furthest, fastest were the most exposed.

Corvid Insight went down in March 2028 and became the case study. A mid-size London shop, well liked, early to everything. It had cut its field and coding teams to eleven per cent of 2024 headcount and moved the entire quant operation onto agentic pipelines during the cheap years. When its platform contracts repriced, the cost of running its flagship tracker rose above what it charged the client for it. Corvid spent eight months paying to deliver its own product before the bank called time. The final all-staff email, later posted to a forum, contained the line that stuck to the era: we automated the work and kept the overheads.

The survivors learned it the hard way. Never build a methodology on a subsidy.

The chart is calm > The chart is lying

Right, so did token create or kill triangulation? Does that make any sense and do we even care about the answer?

When tokens were near-free, agencies ran everything through the machine and smoothed the outputs into one clean number. Certainty theatre was cheap to stage. When every token cost, redundancy had to justify itself, and every method had to argue for its presence on the job. Which forced the question the industry had spent decades avoiding: what was each method actually for?

The answer that emerged: methods were for disagreeing with each other. A human survey said one thing. A synthetic panel said something adjacent but warped. Behavioural data said a third thing. The | old agency | reconciled and smoothed. The post-meter agency presented all three and said the finding was the gap. Where synthetic and human split, something was culturally unstable, and unstable was where the money was.

Research had always been directional. A 43% was never a fact about the world, it was a reading of a thing that kept moving after it was measured. The Tokenpocalypse made pretending otherwise unaffordable. Conflicting methods stopped being a quality problem and became the product, not through philosophical maturity but because nobody could pay for the pretence anymore.

A thousand ghosts > A hundred people

A thousand synthetic respondents at true inference cost, against a hundred humans on a panel the agency already owned. By 2028 the spreadsheet answer was | the humans |.

The cost crisis settled the quality debate by accident. The industry went back to people not out of principle but because the machines started invoicing properly, and the principle turned out to have been right all along. Synthetic respondents were trained on the said. They could tell you what people like your sample would plausibly say. They could not surprise you with the unsaid. They had no subconscious to slip. They never contradicted themselves in the third hour and then defended the contradiction with a story about their gran.

Left alone > The machine went feral

Qual did not shrink. It became the scarcest asset in the stack.

The mechanical layer went into tooling early: transcription, coding, first-pass themes. What remained was presence, and presence could not be token-optimised. More than that, every serious synthetic operation needed | human depth | work as its ground truth, the reference against which the artificial panel was checked for drift. Depth interviews became calibration data, which meant every machine-heavy agency needed a serious qual operation feeding it or its synthetic assets slowly went feral.

The machine made the human interview more valuable, and then the meter made it cheaper by comparison. Almost nobody in 2026 had priced that in.

Attack attack! > Confirm! Conform!

The agency that came out the other side was small. Eight people, sometimes twelve. No field department, because fieldwork was orchestrated. No processing floor. A | methods contrarian | whose job was breaking the house methodology. A synthetic wrangler who audited artificial respondents for warp. An epistemics lead, a real title by 2030, who decided how much any finding deserved to be believed. And a role that would have been laughed out of the room in 2025: a token economist, who decided which questions were worth asking a machine at all.

Revenue came from three places. Disagreement audits, because every client ran AI research in-house and it was cheap, fast, and quietly wrong in ways they could not detect, so auditing the machine paid better than being the machine. Judgement retainers, a standing subscription to scepticism, priced like legal counsel because that was roughly what it was. And instrument licensing: proprietary synthetic panels tuned to a category, rented out with the astronomer attached.

Custom work came back angry. Not the 45-minute questionnaire revived as craft, but adversarial design. Synthetic respondents pre-lived a launch a thousand times and humans were interviewed about the runs that diverged most violently. Bespoke persona populations were built from a client’s own community, then stress-tested against the real community until the agency knew exactly where the mirror distorted. Research was | commissioned to attack the strategy | rather than confirm it, and whatever survived shipped.

Ten thousand lanyards >Twelve necks

Scarcity did what no conference keynote ever managed. It made the industry deliberate.

The tokenmaxxing era had mistaken consumption for capability. Usage dashboards were status symbols. Asking the machine everything felt like rigour and was its opposite, because a question you have not costed is a question you have not thought about. What came after was leaner by necessity and better by accident. Smaller models where | small was enough |. Prompts written like telegrams. Studies scoped like expeditions, because expeditions have supply constraints, and constraints force you to decide what you actually need to know.

The agencies that died kept laminating, bolting synthetic panels onto old decks and presenting the output with the old false certainty at the new true price. The ones that survived had already made peace with the directional nature of the work. They treated methodology as an argument, not a liturgy. They said we do not know, but we know where to look, and they said it before the invoice forced them to.

Somewhere the empty panel is still responding. Filing opinions into the void, confident, fluent, billed monthly. In four years it has never once mentioned its gran. It is the most honest artefact the industry owns, because it holds the question every researcher should be asking about every method on every job.

Is anyone actually in there? And what does it cost to find out?