Borrowed Certainty

THE AUGMENTED PRACTITIONER — INTERLUDE NO. 2

A colleague on my team hit a wall this week. Not a hard task — an unfamiliar one. That conversation clarified something I've been circling for months.

People use the word “insight” for two different things. AI is very good at one and structurally incapable of the other.

The first is retrieval. Ask a model what Indian consumers believe about premiumisation, and it gives you something fluent and structured — because thousands of articles and reports already contain opinions on the subject. It isn't discovering anything. It's recombining what's already been said, weighted by what gets said most. That's useful. It's also memory with better packaging.

The second is discovery. Ask the same model what a specific distributor in Coimbatore will actually pay for a new SKU next quarter, given his cash cycle and his relationship with three competing principals. There's no text anywhere that contains that answer. Nobody's written it down. It exists — if it exists at all — inside one person's head, and the only way to get it out is to go and ask him.

That's the line: retrieval versus discovery. AI has no ceiling on the first and no floor under the second. The failure isn't that the model says “I don't know.” It doesn't know that it doesn't know. It answers the Coimbatore question with the same confidence as the premiumisation question, because both come out of the same machinery. That confidence is borrowed — from the shape of a good answer, not from any contact with the thing being asked about.

Borrowed Certainty: insight that sounds earned but was never checked against a real person, in a real market, at a real moment.

Why this bites harder in India than the pitch decks admit

Most of the “AI will replace analysts” conversation runs on a US-and-Europe training corpus — enormous, dense, recent. Ask a US-trained model about American consumer behaviour and it's interpolating between millions of closely related data points. The certainty is still borrowed, but it's borrowing from something adjacent.

The numbers on this are worse than most people assume. English makes up roughly 90 percent of the text most large models are trained on. Hindi, Tamil, Bengali, Telugu, Marathi, and Gujarati — combined — make up about half a percent of Common Crawl, the dataset behind much of that training. Hindi, the third most spoken language on earth, doesn't rank in the top twenty languages by volume in that corpus. Tier 2 and Tier 3 buying behaviour, SME purchasing psychology, distributor economics in a specific state — almost none of it exists in a form a model can retrieve. The model doesn't get quieter when you ask about rural Karnataka. It gets louder. It fills the gap with the nearest available pattern — usually a US or urban-India proxy — and delivers it with the same fluency as everything else. The gap doesn't make these systems humble. It makes them confident about the wrong thing.

This is the terrain we work in at ResearchFox. Foreign companies evaluating India entry aren't looking for what's already written about India — if it were already written down, they wouldn't need us. They need someone to ask the distributor, the regulator, the category head who hasn't given an interview, the plant manager who'll only speak off the record. That's discovery work. You can't retrieve it. You have to go and get it.

What doesn't change, and what does

I'm not saying AI is useless here. That would fail my own test for writing this. AI compresses everything downstream of discovery — transcription, coding, first-pass synthesis, cross-referencing forty interviews for contradictions that would take a person a week to spot. We've built our own workflow around exactly this. The compression is real and it matters.

What doesn't compress is going and asking, and knowing which answer to trust when three respondents contradict each other. That judgment isn't a lookup. It's built from having been wrong before, in this market, with real money on the line — a different kind of learning than being trained on text about being wrong.

So the position isn't “AI can't do market research.” It's narrower, and I believe it holds up longer: AI cannot manufacture ground truth. It can only work with ground truth someone already went and got. The moment a business treats retrieval as if it were discovery, they're not getting insight. They're getting borrowed certainty. They won't find out it was borrowed until it's expensive to find out.

Where I'd want to be challenged

Isn't this just what every profession says right before it gets automated?

Maybe. But look at the shape of the argument. Most “AI won't replace us” pieces argue from a current capability gap — nuance, context, cultural sensitivity — and gaps close as models improve. I'm arguing something narrower: this is a data-existence problem, not a skill problem. No amount of scale retrieves an answer that was never recorded anywhere. That's a different claim, and it should be judged differently.

Won't synthetic panels just train on more India-specific data over time?

Some of the gap closes. But the opinion that matters most in a fast-moving market is, by definition, the one nobody's written down yet. The gap doesn't disappear — it moves forward in time, permanently one step behind whoever goes and asks first.

Can't an AI agent just go and ask, the way a human would — a voice bot calling the distributor directly?

This is the strongest version of the counterargument, and it deserves a direct answer, not a dodge. Agents that conduct interviews and generate synthetic personas mirroring real respondents already exist and are improving fast. But notice what the agent still can't do: decide which of three contradictory answers to trust, or know that this particular distributor always underquotes in the first call. That judgment isn't built from access to the person. It's built from having been wrong about this market before, with money on the line. An agent can close the access gap. It cannot close the accountability gap — and that's the harder one.

Does this apply outside market research?

I believe it applies to any judgment that depends on unrecorded human intent — hiring, negotiation, pricing a specific deal, reading a specific room. Same pattern: AI accelerates everything around the judgment. It doesn't replace finding out.

I want to know where this breaks. If you find I've got the ceiling wrong, tell me.

Ashwin · freshwin.in