From Prediction to Transformation

Start-up founder: “It will give them insights.”

We wish we had a dime for every time an entrepreneur gave that answer to the mentors and investors at the Creative Destruction Lab, a global seed-stage start-up program we created at the University of Toronto.

Though it’s the stock response, “insights” is precisely the wrong way to think about how an advance in AI will create value. In fact, we feel that “insights” often is code for “We don’t know what to do with our AI’s predictions.”

A much better answer would be to describe the decisions that the predictions will improve, because AI has value only if it leads to better decision-making.

The good news for entrepreneurs is that the opportunities for AI to do that are countless. The number of decisions businesses make has been rising, and the need to make the right ones—in every area of operations—has never been greater. Consider that in 1960 just 6% of jobs required core decision-making skills such as problem-solving, diagnosing, strategizing, and prioritizing, according to research by David Deming of the Harvard Kennedy School. By 2018 that number had reached 34%.

But as we’ll show in the following pages, implementing AI isn’t just about improving specific decisions. Decisions in one area of an organization usually have an impact on decisions made in others, so introducing AI often entails revisiting and redesigning whole systems of decision-making. Let’s begin by looking at a specific example of an initiative where that was the case and how AI ended up completely changing the way the system involved created value.

How New Zealand Won the America’s Cup

Sailboat makers and sailors have been refining their techniques for 5,000 years. Even though commercial shipping no longer relies on wind for propulsion, innovations in sailing have never stopped.

The top prize in sailing (and the oldest trophy in international sports) is the America’s Cup. Today the race is as much about technology as about the skills of the crew. Millions of dollars go into boat design. Since the physics of wind, water, and ships are well understood, competitors use simulators to identify the most effective designs and to test boats without actually building them. The team with the best simulator gains a big advantage, as Emirates Team New Zealand discovered in 2017, when it won the cup.

Two decades after Edison switched on the light bulb, only 3% of U.S. businesses used electricity.

As the members of the team planned for the 2021 race, they wondered if they could speed up the design process. Partnering with McKinsey, the global consultancy, they identified the main bottleneck to innovation: human sailors. It takes time for a human crew to sail a boat in the simulator; there’s no way to increase the pace at which its members react to conditions and maneuver the boat in response. The sailors work on a human time scale, and that isn’t fast enough.

Using technology similar to the AI that beat the world’s top players of the popular strategy board game Go, the team taught an AI program to sail. The bot didn’t need to sleep or eat, and it could run thousands of simulations in the same time that it took the human crew to run just a handful. After eight weeks the AI started to beat the sailors in the simulator.

That’s when things got interesting. The AI began teaching the human sailors new tricks. As a member of the development team told Wired magazine, “The bot was actually doing things that felt counterintuitive to the sailors, but they’d try them out on the water and they’d actually work.” Previously, the boat designers had needed humans to test out any innovation. Figuring out the best way to use a newly designed boat could take weeks.

The AI, in contrast, could experiment with multiple variations of the boat simultaneously, 24 hours a day. It could try different racing tactics. It sped up the cycle of design itera-tion and the development of new maneuvers. Once the AI figured out a superior solution, the human sailors could copy it. As one team member put it, “Accelerating the learning process is extremely valuable, both in terms of allowing the design team to explore as much of the design space as possible and the sailors to maximize performance for a given design.” That year Emirates Team New Zealand claimed the trophy, winning seven races to three.

Why was this use of AI so novel? Setting aside the impressive technology that allowed for simulations in complex environments, the key impact was at the system level. The AI wasn’t handing Emirates Team New Zealand some insights. Instead, it was being built into a system of decision-making.

Race preparation involves two types of decisions: those about boat design and those about sailing maneuvers. While simulators had long been used for boat design, the maneuvers had always been worked out by humans. The AI didn’t actually pilot the boat in the race—the rules still require that real boats be piloted by real people—but it sped up the innovation process and allowed better coordination between boat design and sailing maneuvers. The complete system of the simulated boat and the AI sailor enabled improvements to both kinds of decisions.

Why System Change Takes Time

It can take a while for the systemic impact of a new technology to become apparent. When a technology emerges, people initially apply it narrowly. When electric power was invented as a substitute for steam power, for instance, businesses used it where the water needed for steam was hard to come by. Two decades after Edison switched on the light bulb, only 3% of U.S. businesses used electricity. Similarly, in 1987, decades after the introduction of computers into businesses, the economist Robert Solow noted, “You can see the computer age everywhere but in the productivity statistics.” The potential of computing was clear, but the impact remained muted.

The same thing has happened with AI. Despite some alternative branding, new AI technologies are basically advances in statistics. They make it possible to predict more-multifaceted outcomes and, in doing so, take advantage of data that otherwise might be unexploited. And their initial applications were focused on what they could immediately deliver: better and cheaper predictions than humans were making.

Translation software, an early application of AI, is a good example. It predicts how people would translate a given text from one language to another on the basis of how real humans have translated previous texts. AI classifying medical images, another early application, predicted what expert radiologists would say that scans showed. Both applications leverage the wisdom of the crowd, which can often make far more accurate predictions than one person can. Applications like these can have enormous commercial value. Take the Canadian company Verafin, which was acquired by Nasdaq for $2.75 billion. Why? Because its AI-driven technologies for identifying financial fraud were being used by hundreds of financial institutions as a substitute for the security teams that used to perform that function.

These new applications may drive some important advances, but they’re hardly transformational. They slot into existing businesses without much fuss, precisely replacing the humans who traditionally made predictions. In all other respects, the businesses are unchanged.

The introduction of AI into your company’s decision-making doesn’t affect just you. It also affects your partners in the value chain and the ecosystem you operate in.

But when we consider the impact of electricity and computerization, we don’t think of narrow applications; we think of transformation. Thanks to electricity, factories no longer had to be located near water and have multiple stories to optimize the use of steam. They could be located hundreds of miles away from a water source and spread out on the same floor, making a new type of mass-production system feasible. Computers had the same impact. They evolved from being glorified calculating machines into what Steve Jobs described as “bicycles for the mind”—not substitutes for it.

And that’s the real lesson from the America’s Cup. Emirates Team New Zealand didn’t take the people out of the process. Yes, it is possible to imagine a fully automated solution that makes every decision. But that approach is surely a rarity. AI prediction provides information that improves decisions, which are made by people. Interestingly, with AI the difference is not so much whether machines do more but who the best people to make decisions are.

How AI Is Changing Decision-Making

When Apple launched the smartphone revolution, no one thought, It’s curtains for the taxi industry. But ride-sharing was possible only because internet-connected mobile phones allowed people to hail rides through an app and get navigation information cheaply. In London, for instance, it takes three to four years for cabdrivers to learn all the city’s streets and the best routes through them. Today AI on smartphones allows anyone to predict the best routes there, taking into account traffic conditions. If smartphones didn’t exist, the taxi business might still be thriving.

Most decisions require two things of the decision-maker: the ability to predict the possible outcomes of a decision, and judgment. Prediction is largely based on data. (Given the available routes and the traffic conditions, how long is the trip likely to take?) Judgment is basically a subjective assessment of contextual factors that are not easily reduced to data. (Will this customer prefer a quick trip or the scenic route?)

Taxi drivers have both skills. Ordinary drivers are more limited; they can gauge passenger preferences (judgment) but are less adept at navigation (prediction). But pair ordinary drivers with navigation software, and they match taxi drivers. Add a platform that eliminates the need for a mileage meter, a method of taking payment, and a central dispatcher who assigns drivers to fill pickup requests, and any driver with access to the platform can offer rides.

The platforms and their AI had two important effects. First, vastly more people could now be involved in making decisions about rides, and second, drivers’ control over decisions decreased. Because the ride-share platform could match drivers and passengers and identify the best routes, all the drivers had to do was focus on providing comfortable and pleasant rides that satisfied the customers assigned to them. Those two effects weakened the power of traditional taxi drivers and transformed the industry.

In some cases, AI simply concentrates decision-making without changing who has control. Look at the hiring process, which in most large organizations is managed by the human resources department. Traditionally, hiring has involved a great many HR people who make a lot of small decisions, especially about screening applicants, which can require teams of people looking through hundreds of résumés in order to identify promising candidates to interview. Thanks to AI, one HR executive can decide what criteria to use to decide who gets an interview. The basic process and the key decision-maker remain the same, but fewer people are needed.

In other cases, AI radically centralizes decision-making, completely changing how and where it happens. Credit card verification is a case in point. Before the rollout of connected devices that automatically validate cards, merchants would make their own judgments about whether to accept someone’s card. They could reject it if they suspected fraud—for instance, if someone’s signature didn’t match the one on the card or a customer didn’t have supporting ID. And they could readily accept cards from regular customers. But systems driven first by crude database checks and now by AI prediction have automated the process. Credit card purchases are approved according to rules set by a small group of people, most likely a committee, which creates the risk parameters embedded in the programs that run verification devices.

In still other cases the introduction of AI not only leaves decisions in the hands of existing decision-makers but makes their (more decentralized) judgment more important. The use of AI in medical imagery is a case in point.

Treatment decisions arising from a diagnosis are and always have been made by the patient’s physician. But before the advent of AI prediction, a physician would often call in an expert radiologist, who would perform a medical imaging procedure like an MRI, ultrasound, or X-ray and use his or her judgment to make a diagnosis. In effect, the decisions of the radiologists were needed for the physicians to make their decisions. With AI-enabled diagnosis replacing the radiologist’s judgment, the only judgment now involved in treatment decisions is that of the patient’s physician. That consequently makes the physician more important and powerful and the radiologist less so.

In all these cases the application of AI has changed how and by whom decisions are made. But the introduction of AI into your company’s decision-making doesn’t affect just you. It also affects your partners in the value chain and the ecosystem you operate in. What works for you may create problems for them. Let’s look now at how that can happen.

How AI Shifts Uncertainty

Imagine you’re running a restaurant. Diners come in and order meals. The cooks then make them. At any given time there are constraints on what dishes they can make, which are driven by the skill of the chefs, the total number of orders, and the availability of ingredients and equipment. If you allow your customers to order any dish they might fancy, there will be problems.

What you do, therefore, is set a menu. You limit the choices of your customers so that you can actually make what they order. From the perspective of the kitchen, the menu creates reliability and prevents unexpected surprises. Every week you need to order ingredients, which are based on the menu. If guacamole is on the menu, you need avocados. You order 100 pounds every week. Sometimes that’s too much, and you throw out the excess. At other times 100 pounds is too little, and you miss out on sales.

Let’s say you adopt AI for demand forecasting (what customers will choose), and you find that it works. Now some weeks you order as little as 30 pounds. Other weeks you need 300 pounds. You waste less and sell more. Profitability rises.

The adoption of AI will often involve a system that finds an optimal balance of modularity and coordination.

But your local supplier has been used to buying 100 pounds for you each week. Now it faces more unpredictability because of you. Its other customers are also using AI for demand forecasting, and demand starts to fluctuate wildly. So the supplier decides to adopt AI for its own demand forecasting. It used to order 25,000 pounds of avocados a week. Now its order varies from 5,000 pounds to 50,000 pounds. Your supplier’s source of the fruit, in turn, needs to develop AI, and its orders begin to fluctuate too. And so it goes all the way to the growers who need to make crop-size decisions a year or more in advance.

What this shows is that while AI can be used to resolve one person’s uncertainty, that effect doesn’t spread to decisions throughout a system. The fundamental problem—that demand needs to be aligned with supply—hasn’t really been solved. Like a stone thrown into a pond, your own AI solution has ripple effects on other decisions in the system.

That leaves us with something of a paradox. The value of AI comes from improving decisions by predicting what will happen with factors that might otherwise be uncertain. But a consequence is that your own decisions become less reliable for others. Introducing AI into the value chain means that your partners in it will have to coordinate a lot more to absorb that uncertainty.

Coordinating Systems to Align Effort and Resources

The restaurant manager has to make several other decisions besides predicting demand—for example, what to offer on the menu. If the AI ripple effect means that the grower can’t supply enough avocados, then the restaurant needs to change the menu. It probably won’t do so unless it knows the avocados aren’t available, which requires coordination across decision-makers. That coordination has two aspects:

Synchronizing the work. Consider the operation of a crew team of eight rowers. Two things determine how it will perform in a race: whether its members are rowing in unison, and how they adjust rowing speed as the race progresses to ensure that no one on the team runs out of energy before the finish. The coxswain, who sits at the back of the boat, is essential for the second but not the first function. That might seem surprising, since the coxswain is coordinating the rowers to keep the same time by calling out, “Stroke! Stroke! Stroke!” But that task doesn’t require a separate person; one of the rowers could do it, and in fact, this occurs in races in which crew boats don’t have coxswains. But when it comes to monitoring strategy in a race and obtaining cues about the status of individual rowers—that is, gathering information and aggregating it—the coxswain is critical. The coxswain can assess the need for changes in the team’s rhythm and adjust the message to rowers accordingly. The coxswain is there because the team needs to ensure that adaptation is made in a synchronized fashion.

Assigning resources. The coordination challenge also involves a class of problems that Paul Milgrom and John Roberts call assignment problems—situations in which you need to assign resources to an activity but you know that only a certain amount of them will be used. Any more would be wasted; any less would be insufficient. Consider ambulance dispatching. If all the ambulances in a network received an emergency message and then chose individually whether to respond, you would often end up with no responders or with too many. To ensure that only one responds, you need a central dispatcher, whether human or software, that receives calls (that is, information) regarding an emergency and then assigns one ambulance to respond. In this case sending the “wrong” ambulance (one that is perhaps too far away or doesn’t have the right equipment) is far less of a problem than sending none or sending too many.

Both coxswains and dispatchers are communication systems that ensure that the bad outcomes that could arise from a lack of synchronization or poor resource assignments don’t occur. Similarly, when AI causes coordination problems, new communication systems may be required to overcome them. It is through smart investment in coordination that organizations will be able to fully realize the promise of AI.

So what does “smart” look like in this context?

Combining Coordination with Modularity

Ideally, a system would be able to coordinate entirely through communication, as crew coxswains and ambulance dispatchers do. But communication isn’t always enough. A restaurant can’t create alignment along the supply chain through communication alone because its chain spans thousands of kilometers and many months. The investment would be prohibitively expensive and time-consuming.

What’s the solution? Let’s consider the operations of Amazon. It supplies millions of products all over the world. That involves procuring them, storing them in warehouses, capturing customer orders, and shipping items to those customers. But it also involves helping the customers work out what to purchase in the first place—that is, providing them with recommendations. Amazon faces the same problem our restaurant does. It wants to supply customers with what they want when they want it, but products don’t magically appear, because their supply chains are complex.

Let’s say that Amazon’s AI-based recommendation engine predicts that the best product to suggest to a customer is probably unavailable. What should Amazon do?

It’s tempting to think that if you don’t have a product available, you shouldn’t recommend it to a customer. The problem is, how do you know whether the AI’s prediction was correct, and the customer really wanted it? If you recommend only what you have, you miss opportunities to learn and grow.

That’s precisely why Amazon includes recommendations for products that are out of stock and will take longer to reach its customers. The decisions are coordinated in the sense that Amazon communicates the likely delay to the customers. The customers may well choose products that are available, but occasionally they won’t. Amazon then learns how much effort it needs to make to carry inventory for the out-of-stock items.

Achieving this balance requires careful design. Amazon has a modular organization that has allowed it to slot better AI predictions into recommendations that minimize the impact on the rest of the organization. But the inventory and ordering decisions it makes cannot be fully independent from the AI recommendation system precisely because customers’ choices and reactions give rise to information that needs to be acted on by the logistics department.

The adoption of AI will often involve a system that finds an optimal balance of modularity and coordination. Modularity insulates decisions in one part of the organization from the variability—the ripple effects—that AI creates in others. It reduces the need for reliability. Coordination, in contrast, counters the lack of reliability that comes alongside AI adoption. Successful AI systems enable coordination where possible, and modularity where necessary.

. . .

As we hope will be clear by now, the promise of AI’s prediction technology is similar to that of electricity and personal computing. Like them, AI began by resolving a few immediate problems, creating value in isolated, tightly bounded applications. But as people engage with AI, they will spot new opportunities for creating solutions or improving efficiency and productivity. Restaurants, for example, will most likely become more deeply embedded in their own supply chains and perhaps more flexible in their menu offerings. As these opportunities are realized, they will create new challenges that in turn provide more opportunities. So as AI spreads across supply chains and ecosystems, we will find that all the processes and practices we took for granted are being transformed—not by the technology itself but by the creativity of the people who are using it.

Copyright 2022 Harvard Business School Publishing Corporation. Distributed by The New York Times Syndicate.