Twenty-three centuries ago, a Chinese wheelwright named Bian interrupted a nobleman who was reading the words of dead sages. The knack of making a wheel, Bian said, could not be put into words - he couldn't teach it even to his own son. What can be written down, he told the nobleman, is only the leftovers of knowing.Every debate about AI in the insights industry is a rerun of that conversation, though almost nobody in the industry seems to have noticed.The current argument runs on Western management vocabulary: augmentation versus replacement, human-in-the-loop, efficiency gains. It's a language built for describing workflows, not for describing knowledge — which is why it keeps failing to explain what, exactly, a trained researcher does that a model cannot. Older traditions asked better questions, and they asked them long before anyone had data to be drowned in.

Start with the pause. Japanese aesthetics has a word - ma (間) — for the interval that carries meaning: the silence in conversation, the empty space in a room, the beat before an answer that changes what the answer means. A transcript deletes ma by design. A language model, trained entirely on what was said, cannot in principle recover what happened between. When a Mumbai homemaker pauses before telling you she finds a retail experience "okay," the pause is the finding. Everything after it is the cover story.

Jain philosophy offers a second correction. Anekāntavāda - many-sidedness - holds that reality never yields to a single standpoint; every truth is true only "in some respect." The blind men and the elephant is a Jain parable before it is a boardroom cliché. Now notice what a large language model is built to do: collapse many sides into one fluent, confident answer. That's not a bug of current systems; it's the shape of the tool. But in cultural research the contradiction is the data. The consumer who distrusts algorithms and delegates her shopping to one isn't confused - she's a finding that only a many-sided frame can hold.

Clifford Geertz, closer to our own trade, made the same cut with his wink and twitch: physically identical movements, separated only by cultural reading. AI gives us thin description at infinite scale - every twitch in ten thousand transcripts, tagged and clustered by breakfast. The thick part, deciding which twitches were winks, remains stubbornly on our side of the desk.And Bian's wheel explains why it will stay there. The wheelwright's point was not that his craft was mysterious, but that it transferred only through practice - never through text. Text is all a model has. The moderator's read of a room, the ethnographer's sense that a kitchen is staged, the analyst's instinct that a theme is too neat to be true: these are knacks, apprenticed rather than downloaded. What can be written down is the leftovers - and the leftovers are precisely what the models are trained on.

None of this is an argument against AI in research. Our firm uses it daily, and gladly - to scan categories at speed, to sort mountains of transcript, to pressure-test drafts. Bian was not against books; he was against confusing books with knowing. The distinction the old traditions insist on is not human versus machine but articulable versus tacit - and it tells you exactly where to draw the line in your workflow, with more precision than any efficiency framework can.

The insight industry's future belongs to firms that automate the leftovers and protect the knack. The wheelwright knew which was which. So did the Jain logicians, and the tea masters, and a certain anthropologist watching for winks in a busy market. The answer to the AI question is not new. We just stopped reading the people who already gave it.