[This post originally appeared in HBR online on 3rd October, 2017]
by Ajay Agrawal, Joshua Gans and Avi Goldfarb
How will AI change strategy? That’s the single most common question the three of us are asked from corporate executives, and it’s not trivial to answer. AI is fundamentally a prediction technology. As advances in AI make prediction cheaper, economic theory dictates that we’ll use prediction more frequently and widely, and the value of complements to prediction – like human judgment – will rise. But what does all this mean for strategy?
Here’s a thought experiment we’ve been using to answer that question. Most people are familiar with shopping at Amazon. Like with most online retailers, you visit their website, shop for items, place them in your “basket,” pay for them, and then Amazon ships them to you. Right now, Amazon’s business model is shopping-then-shipping.
Most shoppers have noticed Amazon’s recommendation engine while they shop — it offers suggestions of items that their AI predicts you will want to buy. At present, Amazon’s AI does a reasonable job, considering the millions of items on offer. However, they are far from perfect. In our case, the AI accurately predicts what we want to buy about 5% of the time. In other words, we actually purchase about one out of every 20 items it recommends. Not bad!
Now for the thought experiment. Imagine the Amazon AI collects more information about us: in addition to our searching and purchasing behavior on their website, it also collects other data it finds online, including social media, as well as offline, such as our shopping behavior at Whole Foods. It knows not only what we buy, but also what time we go to the store, which location we shop at, how we pay, and more.
Now, imagine the AI uses that data to improve its predictions. We think of this sort of improvement as akin to turning up the volume knob on a speaker dial. But rather than volume, you’re turning up the AI’s prediction accuracy. What happens to Amazon’s strategy as their data scientists, engineers, and machine learning experts work tirelessly to dial up the accuracy on the prediction machine?
At some point, as they turn the knob, the AI’s prediction accuracy crosses a threshold, such that it becomes in Amazon’s interest to change its business model. The prediction becomes sufficiently accurate that it becomes more profitable for Amazon to ship you the goods that it predicts you will want rather than wait for you to order them. Every week, Amazon ships you boxes of items it predicts you will want, and then you shop in the comfort and convenience of your own home by choosing the items you wish to keep from the boxes they delivered.
This approach offers two benefits to Amazon. First, the convenience of predictive shipping makes it much less likely that you purchase the items from a competing retailer as the products are conveniently delivered to your home before you buy them elsewhere. Second, predictive shipping nudges you to buy items that you were considering purchasing but might not have gotten around to. In both cases, Amazon gains a higher share-of-wallet. Turning the prediction dial up far enough changes Amazon’s business model from shopping-then-shipping to shipping-then-shopping.
Of course, shoppers would not want to deal with the hassle of returning all the items they don’t want. So, Amazon would invest in infrastructure for the product returns — perhaps a fleet of delivery-style trucks that do pick-ups once a week, conveniently collecting items that customers don’t want.
If this is a better business model, then why hasn’t Amazon done it already? Because if implemented today, the cost of collecting and handling returned items would outweigh the increase in revenue from a greater share-of-wallet. For example, today we would return 95% of the items it ships to us. That is annoying for us and costly for Amazon. The prediction isn’t good enough for Amazon to adopt the new model.
That said, one can imagine a scenario where Amazon adopts the new strategy even before the prediction accuracy is good enough to make it profitable because the company anticipates that at some point it will be profitable. By launching sooner, Amazon’s AI will get more data sooner, and improve faster. Amazon realizes that the sooner it gets started, the harder it will be for competitors to catch up. Better predictions will attract more shoppers, more shoppers will generate more data to train the AI, more data will lead to better predictions, and so on, creating a virtuous circle. In other words, there are increasing returns to AI, and thus the timing of adopting this kind of strategy matters. Adopting too early could be costly, but adopting too late could be fatal.
The key insight here is that turning the dial on the prediction machine has a significant impact on strategy. In this example, it shifts Amazon’s business model from shopping-then-shipping to shipping-then-shopping, generates the incentive to vertically integrate into operating a product-returns service (including a fleet of trucks), and accelerates the timing of investment due to first-mover advantage from increasing returns. All this is due to the single act of turning the dial on the prediction machine.
Most readers will be familiar with the outcome of companies like Blockbuster and Borders that underestimated how quickly the online consumer behavior dial would turn in the context of online shopping and the digital distribution of goods and services. Perhaps they were lulled into complacency by the initially slow adoption rate of this technology in the early days of the commercial internet (1995-1998).
Today, in the case of AI, some companies are making early bets anticipating that the dial on the prediction machine will start turning faster once it gains momentum. Most people are familiar with Google’s 2014 acquisition of DeepMind – over $500M for a company that had generated negligible revenue, but had developed an AI that learned to play certain Atari games at a super human performance level. Perhaps fewer readers are aware that more traditional companies are also making bets on the pace the dial will turn. In 2016, GM paid over $1B to acquire AI startup Cruise Automation, and in 2017, Ford invested $1B in AI startup Argo AI, and John Deere paid over $300M to acquire AI startup Blue River Technology – all three startups had generated negligible revenue relative to the price at the time of purchase. GM, Ford, and John Deere are each betting on an exponential speed up of AI performance and, at those prices, anticipating a significant impact on their business strategies.
Strategists face two questions in light of all of this. First, they must invest in developing a better understanding of how fast and how far the dial on their prediction machines will turn for their sector and applications. Second, they must invest in developing a thesis about the strategy options created by the shifting economics of their business that result from turning the dial, similar to the thought experiment we considered for Amazon.
So, the overarching theme for initiating an AI strategy? Close your eyes, imagine putting your fingers on the dial of your prediction machine, and, in the immortal words of Spinal Tap, turn it to eleven.
The ideas here are adapted from our forthcoming book “Prediction Machines: The Simple Economics of Artificial Intelligence.” (Harvard Business School Press, April 2018)