The Job Split in Two

Going into last year's launch I had a view, and I held it fairly confidently. If the AI could handle discovery, eligibility, availability, documentation and follow-up — which it could — then the channel partner was a cost we'd been paying for an information asymmetry that no longer existed.

That's a clean argument. It's also the argument that most people in enterprise technology are making about intermediaries in general, in most industries, right now.

It was wrong, and the data that corrected me came from our own launch.

What we ran

A residential launch, roughly five weeks, mid-August to late September 2025. No sample flat. No site office. 8,838 registrations against 1,419 homes released. Six times oversubscribed, ₹640 crore allotted. Around three lakh unique visitors, 17,000 OTP-verified logins, about 45,000 conversations handled by AI.

Eight percent of buyers completed the whole thing without speaking to a human being.

That eight percent is the number people repeat back to me, and I understand why — it's the futuristic one. It isn't the interesting one.

The interesting one is sixty-eight. Sixty-eight percent of registrations still arrived through channel partners, on a launch where the AI was doing all of the work I'd assumed the channel partner was there to do.

I'd removed the reason for the intermediary to exist, by my own theory, and two-thirds of the business still came through them.

What I found when I looked properly

So I went and worked out what a channel partner actually does across a transaction. Not what the org chart says. What the hours go to.

Checking whether a unit is available. Explaining who qualifies for what. Chasing a document that was promised on Tuesday. Answering a question that's been answered three times, because the person asking is about to spend fifty lakh and is frightened. Following up on a Sunday evening because that's when the family is together and can talk about it.

That's most of his hours. Nearly all of it moved to the machine, and moved cleanly, because it's information work with a right answer.

Then the other side. Sitting in a living room with a family that is deciding whether to commit a decade of savings. Reading a room well enough to know that the wife has a concern she hasn't voiced and that the deal dies if it isn't addressed. Being reachable in eighteen months when possession is delayed and somebody is angry. Having a name in that neighbourhood that's worth something, and therefore something to lose.

That's almost all of his value. None of it moved.

The job split in two. Only one half went.

Why this isn't a comfortable finding

I could tell this as a warm story about the irreplaceable human element, and it would do well on LinkedIn, and it would be dishonest.

The uncomfortable part is that the half that moved was most of the hours. If a channel partner's income was calibrated to a job where seventy percent of the time went on information work and thirty percent on judgment, and the information work is now free, then the economics of that role have changed permanently even though the role survives.

What survives is the part of the job that was always underpriced, now carrying the whole weight.

That has consequences I don't think the industry has thought about. The people who are good at the judgment half will do better than before, because the tedious half no longer rations their capacity. The people who were competent at the information half and mediocre at the judgment half have a problem. And the entry route into the profession — which was, universally, doing the information work for a few years until you learned the judgment — has quietly closed.

That last point is the one I'd watch. It shows up in the macro data too. Stanford's Digital Economy Lab, working off ADP payroll records through June this year, finds employment for 22-to-25-year-olds in AI-exposed occupations running about nineteen percent below where it would otherwise be, and the gap has widened from fifteen percent a year earlier. The mechanism they identify isn't firing. It's reduced hiring.

You don't lose your job. You just never get the first one.

The distinction that actually predicts things

The most useful thing in the Stanford work, for an operator, is a split they make explicitly: where AI is used in an automation mode, employment declines. Where it's used in an augmentative mode, employment is flat or rising.

That matches what we saw. We didn't set out to automate the channel partner, mostly because we couldn't have — they own the relationships. What we did was remove the information work from around them. The effect was augmentative by accident, and the 68% is what augmentation looks like in a number.

Anthropic's economic index, published in June off around 9,700 linked survey responses, adds a detail that lines up: early-career workers report that AI can handle roughly ten percentage points more of their tasks than workers with fifteen-plus years of experience. Experienced people consistently name judgment and context as the part that doesn't transfer.

Three different instruments — our transaction data, payroll microdata, usage telemetry — pointing the same way. The split runs through the middle of individual jobs rather than between them, and it separates codified work from tacit work.

What I'd tell someone running an intermediated business

Don't model this as headcount reduction. You'll get the number wrong and you'll damage the thing that actually produces revenue.

Work out which half of each role is information work with a right answer, and move that. Then look hard at what's left, because what's left is what you're actually paying for, and in most intermediated businesses it has been priced as though it came free with the information work.

Expect the intermediary to survive and the intermediary's economics to change. Expect your best people to get better and your average people to struggle. And expect a hiring problem in about three years, when you discover that the training ground you used to have was the work you automated.

One more caution about my own data. This is five weeks, one launch, one micro-market, and it's internal. It tells you what happened; it doesn't establish a law. I'd be careful of anyone stretching a single launch into a theory of labour — including me, in the paragraphs above.

The 68% is the number I trust most, because it's the one I'd have bet against.