Something interesting has happened in our house over the last few weeks. For a long time, if we suddenly discovered that we were out of milk, needed something for dinner, or the children wanted something that was not available at home, the answer was usually Blinkit.

Recently, Sushmita has started ordering more often from Amazon Now. There is no great technological insight behind this change in consumer behaviour. Amazon has been offering cashbacks.

I had observed this shift, and it would probably have remained just another observation about competitive forces in India’s quick-commerce market. But something clicked while I was reading Prediction Machines by Ajay Agrawal, Joshua Gans and Avi Goldfarb.

The book was published in 2018, and its central argument is surprisingly simple. Much of what we call artificial intelligence can be understood economically as prediction. AI makes prediction cheaper, and when the price of something falls dramatically, we start using much more of it.

The interesting question, then, is not merely what AI can predict. It is what happens to businesses when prediction becomes sufficiently cheap and sufficiently good.

The authors use Amazon to construct one of the more memorable examples in the book.

Shop, Then Ship

The traditional model of online retail is straightforward. You shop. Amazon ships.

I search for a book, decide that I want it, place an order and Amazon sends it to me. There is very little uncertainty about my demand by the time the product leaves its warehouse. I have already told Amazon exactly what I want.

But Amazon also knows quite a lot about me. It knows what I have bought before, what I search for, and which products I looked at but didn’t buy. With enough data and sufficiently good prediction, perhaps Amazon could predict what I am going to buy even before I buy it.

And if its prediction became good enough, why wait for me?

Amazon could simply send the product. The model would change from Shop → Ship to Ship → Shop.

Instead of waiting for me to order something, Amazon predicts what I am likely to want and sends it to my house. I keep what I want and return the rest.

The authors are not necessarily predicting that Amazon will adopt this model. It is a thought experiment about what happens when the economics of prediction change enough to alter business strategy. Interestingly, Amazon had already obtained a patent relating to anticipatory shipping years before the book was published.

There is, however, an obvious problem with shipping before shopping. What happens when the prediction is wrong?

The Problem of Returns

Suppose Amazon predicts that I want five things and sends all five to my house. If I want four of them, the system works beautifully. If I want only one, Amazon now has four returns to process. And if I want none, Amazon has paid to move five products to my house and must then deal with moving all five back.

The better Amazon becomes at predicting my individual demand, the more attractive Ship then Shop becomes. But until prediction becomes sufficiently accurate, the cost and inconvenience of incorrect predictions can overwhelm the benefit.

That trade-off is the point of the example in Prediction Machines.

But there is another interesting possibility. Suppose the predictions become so good that Amazon is regularly sending me things I would otherwise have ordered. I no longer need to search for them, compare prices or even remember to buy them. They simply arrive.

That could make me a remarkably loyal Amazon customer. Not necessarily because I have consciously decided that Amazon is the best retailer, but because I am no longer making a fresh decision about where to shop. If I am not shopping, another retailer has fewer opportunities to persuade me to switch. The opportunity arises mainly when I want something Amazon hasn’t anticipated, or when its predictions disappoint me.

In that sense, Ship then Shop could change much more than the movement of goods. It could change the relationship between the retailer and the customer.

And now, reading the book eight years after its publication, I find myself looking at India’s quick-commerce industry and wondering whether it offers another way to use prediction. Not Ship then Shop. Something in between.

Predict, Then Position

Blinkit, Zepto and Swiggy Instamart don’t need to predict that I am going to order a packet of Amul butter at 7:42 this evening. They need to predict that people living around a particular dark store are likely to order enough packets of Amul butter for it to make sense to keep them there.

Instead of predicting individual demand, they can predict aggregated hyperlocal demand. And that changes where the inventory goes before the order exists.

The butter moves before I shop. But it doesn’t move all the way to my house. It moves from a larger supply network to a small dark store close to where I live. The store carries inventory selected for the demand expected in its neighbourhood. Only after I press the order button does the butter make the final journey to my house.

Of course, retailers have been forecasting demand and positioning inventory for a very long time. There is nothing fundamentally new about trying to stock goods where customers are likely to buy them. What makes the quick-commerce model interesting in this context is the granularity of the prediction and the proximity of the inventory to the customer.

So perhaps the model looks something like this:

Predict → Position → Shop → Ship

In the original Shop then Ship model, the customer orders before the retailer commits to sending a particular product to that customer. In Ship then Shop, the retailer makes that commitment based on prediction. Quick commerce occupies a space between the two.

It uses prediction to move inventory towards demand, but stops before assigning that inventory to a particular customer. If I don’t order the butter, it doesn’t have to be returned from my house. Someone else in my neighbourhood might order it.

The prediction doesn’t have to answer, Will Amit buy this packet of butter tonight? It needs to answer something closer to, How many packets of butter are customers around this dark store likely to buy before we replenish it again?

And changing the unit of prediction changes the economics of being wrong.

The problem of inaccurate prediction hasn’t disappeared. A dark store can still stock too much butter, or too little. But the uncertainty is now spread across a neighbourhood rather than concentrated in an individual customer’s predicted purchase. That, to me, is the fascinating part. Instead of improving prediction until Amazon can confidently send something all the way to my house, a retailer can change what it is trying to predict and stop the inventory a little earlier.

And there is another consequence of stopping there.

The Shopping Decision Survives

In the Ship then Shop world, the better Amazon becomes at predicting what I want, the less I need to shop for those things. The shopping decision could gradually disappear, taking with it many of the opportunities for another retailer to win my business.

But in Predict then Position, the shopping decision remains with me. The dark store may already have the butter waiting, but I still have to decide which app to open. That decision could depend on who has the product, who can deliver it fastest, who offers the better price, and which app is easiest to use. Or who happens to be offering cashback.

Which brings me back to Sushmita. Amazon Now has managed to persuade her to move some of our orders away from Blinkit. That is possible because the shopping decision still exists. Amazon is competing for an order that hasn’t yet been placed.

And this is where the two models could produce very different kinds of customer loyalty.

Ship then Shop, if it worked consistently well, could make a customer increasingly loyal to a particular retailer. Predict then Position doesn’t necessarily create that same loyalty. It can leave retailers competing with one another for each purchase.

But it may create another kind of loyalty altogether. Not loyalty to the retailer. Loyalty to the way of shopping.

When Prediction Machines was published in 2018, India’s quick-commerce industry had not yet developed into the business we know today. Now dark stores sit close to consumers in many Indian cities, allowing inventory to be positioned before demand becomes an actual order. Blinkit, Zepto and Swiggy Instamart have helped establish this model, and Flipkart and Amazon have entered the competition too.

And while the shopping decision remains, the expectations around shopping have changed enormously. The convenience of Blinkit, or any other quick-commerce service, is becoming rather difficult to give up. Once you become accustomed to having most everyday things available within ten or fifteen minutes at the click of a button, waiting until tomorrow begins to seem strangely unnecessary.

I may move from Blinkit to Amazon Now, or back again. But I am much less likely to give up quick commerce itself. That is a different kind of stickiness.

And What Happens Tomorrow?

There is, however, one more possibility. The fact that Predict then Position leaves the shopping decision with the customer doesn’t mean it will always remain that way.

Every time I open Blinkit or Amazon Now and place an order, I am telling the retailer something about myself. What I buy, how often I buy it, what I buy together, and how my purchases change over time. Quick commerce is not merely using prediction to position inventory. It is also creating opportunities to improve predictions about individual customers.

And as those predictions become cheaper and more accurate, perhaps the economics will change again. Today, Amazon Now needs to predict how much butter my neighbourhood will consume. Tomorrow, it may become sufficiently confident about how much butter I will consume, and when, to offer to replenish it automatically. Eventually, for sufficiently predictable purchases, it might make sense to send the butter before I have even thought about ordering it.

We would be back at Ship then Shop. And the advantage that quick commerce currently leaves open to competing retailers (the opportunity to win my business every time I shop) could begin to disappear.

Perhaps the dark store is not the final answer to the thought experiment in Prediction Machines. Maybe it is an intermediate arrangement that makes prediction useful today while generating the information that could make another arrangement economical tomorrow.

For now, Sushmita can move between Blinkit and Amazon Now depending on who is offering cashback. We have become accustomed to the convenience of quick commerce without becoming permanently attached to any one retailer. But who knows what happens when the retailer knows our shopping habits better than we do?

The book imagined a world in which prediction becomes so good that shopping might disappear. Eight years later, we are still shopping. We are simply doing it much faster. And maybe, without quite realising it, we are teaching the machines how to make even that unnecessary.

You may also like:

The Age of AI – Notes That Stayed – A discussion on what the future might be.

Co-Intelligence by Ethan Mollick: Why AI May Make Expertise More Important, Not Less – A view on expertise in the Age of AI

You Look Like a Thing and I Love You by Janelle Shane: Of Giraffes, Glitches, and Gentle Warnings – An introduction to AI, AI weirdness style