Sixteen Percent

A World Development Report background paper, published at the end of July, went looking at 1,563 companies across 83 countries, all of them drawn from IFC's private-equity and venture portfolio — the kind of firm you'd expect to be ahead, and explicitly not representative of the typical emerging-market business. About 40% had adopted AI in some form.

Then they split the sample by something else entirely. Not sector. Not size. Not country income. They split it by whether the firm already had digital foundations — systems, data, processes that ran on software rather than on paper and phone calls.

Sixty-three percent of the digitised firms had adopted AI. Sixteen percent of the traditional ones had.

That's close to a four-fold gap, and digitalisation is the strongest correlate of adoption the paper finds. Their conclusion is worth reading twice: the risk is not lack of access to AI. It is that firms without digital foundations cannot integrate it.

I've been sitting with that number for a while, because it describes my industry precisely and nobody in my industry is talking about it.

The absorption problem

There's an assumption buried in almost every conversation about AI in India, and it goes roughly like this. The models are getting cheaper and better. Access is effectively universal — the same frontier model costs the same per token in Mumbai as it does in San Francisco. Therefore the playing field is level, and possibly even tilted our way, because our labour-cost base makes automation less urgent and our market size makes the prize large.

Every part of that is true. The conclusion still doesn't follow.

Access isn't the constraint. Absorption is. A model you can call from anywhere is worth nothing to a business whose inventory lives in a WhatsApp group, whose pricing sits in one person's head, and whose customer history is a stack of forms in a site office cupboard. You can't put an agent on top of that. There's nothing for it to stand on.

This is the part that gets missed when people compare adoption rates between countries. The gap is structural, and it sits underneath the AI question entirely. It isn't a talent story either, at least not straightforwardly. Nasscom's AI-Native Talent Index, published in July, scores roughly nine in ten of India's early-career technology workforce as AI-proficient or AI-native — though the same index puts deep engineering capability considerably lower, and the assessment is self-reported.

What the macro data says, and what it doesn't

The Bank for International Settlements published a bulletin in February that puts numbers on the consequence. A standardised improvement in AI preparedness lifts growth in an advanced economy by about 0.6 percentage points. In an emerging market, the same improvement delivers 0.45. Compound that across a decade and advanced economies pull ahead by more than two points of real GDP.

The mechanism isn't mysterious. BIS points out that agriculture, transport and construction have the lowest measured AI exposure of any sectors — and those three dominate output across emerging markets. The technology is least applicable exactly where the economy is heaviest.

I want to be careful here, because this is the point where macro findings usually get stretched. This is a modelling exercise, not an observation. It assumes a standardised improvement that nobody has actually made. And there's a genuine counter-current in the World Bank data that cuts the other way: in emerging markets, small and micro firms are adopting more than medium and large ones, which inverts the advanced-economy pattern completely. Small firms have less legacy to unwind. They can start digital.

So the picture is more specific than "emerging markets fall behind." The firms that fall behind are the ones with enough scale to have built physical process infrastructure, and not enough digitisation to replace it. That's a very specific kind of company, and in Indian real estate it's most of us.

What this looks like from inside

Here's the concrete version, from a business I run.

Selling a home in India has historically required a set of physical assets whose only job is to make a stranger trustworthy enough to take fifty lakh from. A sample flat, so the buyer can stand inside something real. A site office, so there's a person to ask. A channel partner with a reputation in that neighbourhood. None of this is operations. All of it is trust infrastructure, and all of it is on the balance sheet as though it were operations.

Now: could AI replace some of that? Yes, and I'll write about what happened when we tried. But notice what had to be true first. We needed inventory in a system rather than a spreadsheet. We needed pricing that could be queried. We needed a customer record that survived a conversation ending. We needed the eligibility rules written down somewhere other than in a manager's memory.

None of that is AI. All of it is the boring digitisation work of the previous decade, and it's the reason the AI worked at all.

The firms that will get nothing from AI over the next three years are the ones that skipped that decade and are now trying to buy their way past it. The model choice has almost nothing to do with it.

The uncomfortable part

There's a version of this argument that's comfortable, and I don't want to make it. The comfortable version says: do your digitisation first, then adopt AI, and all will be well. Sequential, manageable, and a consultant can sell you both phases.

I don't think it works like that, for a simple reason. The cost of the digitisation work hasn't fallen. The cost of the AI has. So firms are looking at a cheap, exciting, visible technology sitting on top of an expensive, dull, invisible prerequisite, and they're doing exactly what you'd expect — buying the top and skipping the bottom.

Then they get nothing, and conclude AI doesn't work in their industry.

McKinsey's numbers this year show something consistent with that. Eighty-nine percent of organisations use AI in at least one function. Thirty-seven percent can attribute any EBIT impact to it. About six percent report it mattering — and that figure has been flat for a year. What separates the six percent, on McKinsey's own reading, is that they redesigned how work happens rather than layering AI on top of what was already there. Nearly three-quarters of them did. A quarter of everyone else did.

Redesigning the workflow is the digitisation work. It's the same task, arriving with a better reason.

What I'd actually do

If you're running a business with physical process infrastructure in a market where trust is expensive, I'd suggest three things, and none of them is about choosing a model.

Find the one workflow that generates the revenue, and ask what it would take to run it without anybody physically present. Not to actually do that — just to find out what's missing. The list you get back is your real project.

Then look at what's on your balance sheet that exists only to make you trustworthy. Sample flats, branch networks, relationship managers, the call centre. Price it. That number is what AI is competing against, and in a low-trust market it's usually much larger than the efficiency savings everyone is chasing.

And be honest about the sixteen percent. If your inventory lives in a WhatsApp group, no model is going to help you this year. That's worth knowing before you spend, because it tells you what you're actually buying.