Understanding the fundamental economics of AI

Strategy & Leadership

ISSN: 1087-8572

Article publication date: 21 December 2022

69

Abstract

Purpose

Despite the hype about transformative technology, the authors of “Power and Prediction: The Disruptive Economics of Artificial Intelligence” see the recent AI advances as all basically ‘better statistical techniques’ that allow us to take really big data sets and come up with more refined predictions.

Design/Methodology/Approach

University of Toronto experts, Ajay Agrawal, Joshua Gans and Avi Goldfarb explain why transformation of business model by AI will be some time in the future when we move beyond simply substituting the new technology into existing systems and start to leverage its potential to enable the reimagining of old system solutions and innovate radically new value propositions.

Findings

What economic history tells us is that technology-driven transformation does not come easy and real adoption only occurs when new systems are created.

Practical Implications

As there are likely many decisions in your organization that have been codified into rules, AI offers the potential to turn them to dynamic decisions.

Originality/Value

To realize the full potential of AI, companies need to adopt a “system mind-set,” in contrast to the “task-level thinking” that still predominates in most companies.

Citation

Leavy, B. (2022), "Understanding the fundamental economics of AI", Strategy & Leadership, Vol. ahead-of-print No. ahead-of-print. https://doi.org/10.1108/SL-11-2022-0111

Publisher

:

Emerald Publishing Limited

Copyright © 2022, Emerald Publishing Limited


In 1987, Nobel-winning economist, Robot Solow famously remarked, “You can see the computer age everywhere but in the productivity statistics.”[1] Today, the same might be said of the artificial intelligence (AI) age. While the technology continues to advance relentlessly, its anticipated transformational impact on the economy as a whole has yet to occur. Why so, and what can be done to accelerate productivity? These are the questions that University of Toronto experts, Ajay Agrawal, Joshua Gans and Avi Goldfarb set out to answer in their latest book: Power and Prediction: The Disruptive Economics of Artificial Intelligence.[2]

In their previous book Prediction Machines: The Simple Economics of Artificial Intelligence, the authors thought they had said “all we needed to on the economics of AI.”[3] However, they have since come to realize that while “we had been focused on the economics of AI itself – lowering the cost of prediction – we underestimated the economics of building the new systems in which AIs must be embedded.” The latter hold the key to understanding more fully the pace of AI adoption, and how it will “eventually sweep across industries, entrenching some incumbents and disrupting others.”

Ajay Agrawal and Joshua Gans are both Professors of Strategic Management and Avi Goldfarb is Professor of Marketing at the Rotman School of Management, University of Toronto, where Professor Agrawal is also the founder of the Creative Destruction Lab and Professors Gans and Goldfarb are the lab’s Chief Economist and Chief Data Scientist respectively. Their interviewer, Brian Leavy, is Emeritus Professor of Strategy at Dublin City University Business School and a Strategy & Leadership contributing editor (brian.leavy@dcu.ie).

Strategy & Leadership: In your earlier book, Prediction Machines, you set out the fundamental economic properties of the new technology. Please remind us.

Joshua Gans: When most people think about AI, they imagine intelligent machines like the helpful robots featured in movies, such as R2-D2 and WALL-E, or brilliant teammates such as Data from StarTrek or J.A.R.V.I.S. from Iron Man. Such images of AI in the popular culture have one thing in common: they can think, reason and have agency. However, this is still far from the technology we have today. The recent AI advances are all basically “better statistical techniques” that allow us to take really big data sets and come up with more refined predictions. This advance is still very significant, and as it reaches its potential it will dramatically reduce the cost of prediction.

S&L: Given the growing excitement around AI as a potentially transformative technology, why are we seeing so little impact in terms of overall productivity growth in the economy so far?

Gans: Taking any one decision that is made under uncertainty and adding better prediction in the form of AI machine learning can be very important. The problem with doing that is that you are confining your AI adoption to what we call “point solutions.”

When we looked back in history to other general purpose technologies that were transformative – such as electricity or information technology – we actually found a similar pattern. Initially, the technical achievements outstripped adoption in practice. Take electricity. The electric light bulb was first switched on by Edison in 1879 but it was another forty years before electricity really began transforming things. We refer to this interval as “the Between Times.”

In industrial settings, the immediate opportunity for electricity was to provide an alternative source to steam power. The value the point solution entrepreneurs promised was lower cost and other benefits specific to particular factory types. However, only so much of a power bill can be sliced off by changing a power source. What a point solution did not offer was a reason to use more power. In contrast, system solution entrepreneurs ask what would a factory look like if you designed it from scratch, given what you know about the novel properties of electricity. Electric power, in effect, equalized the economic value of space within a factory whereas steam power diminished with distance from the source. Henry Ford was a system solution entrepreneur. He could not have invented the production line for the Model T car with steam power. Only electricity could achieve that, and these system changes altered the industrial landscape. Only then did electricity show up in the productivity statistics in a big way.

This is why we call today “the Between Times” for AI. We have just been exploiting point solutions. Transformation will be some time in the future. when we move beyond simply substituting the new technology into existing systems and start to leverage its potential to enable the reimagining of old system solutions and innovate radically new value propositions.

S&L: Where do you hope to see your recent book, Power and Prediction: The Disruptive Economics of Artificial Intelligence helping to accelerate more widespread adoption of prediction technology?

Gans: The new book, like our first one, is all about calibrating expectations. We find that when people are too pessimistic about a technology’s potential they miss opportunities. But when they are too optimistic, they waste expenditures by pursuing applications that are not yet fit for purpose.

We want people to understand that the real returns from AI are from transformation and are in the future. But just because it will take time to work out these systems solution innovations, doesn’t mean you should wait. Instead, the book offers ways of thinking about transformation and getting businesses started on finding precisely what AI avenue they need to explore.

A basic economic framework for assessing transformative AI

S&L: Assessing the transformative potential of AI-based system solutions starts with understanding the main economic trade-offs in choosing between following a rule or using AI prediction to turn it into a decision in any given situation. What are these trade-offs?

Gans: When people form habits or keep to rules and routines, they are acknowledging that the costs of trying to optimize are too high and are seeking ways to reduce their cognitive load. The economy pretty much runs on this premise, which Nobel Laureate Herbert Simon labelled “satisficing.”

Two broad considerations drive decisions: high versus low consequence and cheap versus expensive information. As AIs become more powerful they lower the cost of information and prediction. They also increase the relative returns to decision making compared to using rules by reducing the costs of recalibrating your assumptions about the world. As there are likely many decisions in your organization that have been codified into rules, AI offers the potential to turn them to dynamic decisions. In so doing, you no longer have to accept the costs associated with the inflexibility of rules and, by changing your actions as predictions change, you can deliver better outcomes.

It must be noted, however, that rules not only lower cognitive costs but also increase reliability when embedded in systems characterized by interdependencies in choices and outcomes. For example, most organizations rely on Standard Operating Procedures (SOPs), which are rules. SOPs reduce cognitive load while also increasing reliability. If you are going to use AI prediction to turn rules into decisions, then you may need to redesign the system to account for the reduced reliability.

S&L: Why do you see “the primary benefit of AI” as its ability to “decouple prediction from the rest of the decision-making process?”

Gans: To make a decision first requires knowing what your options are – we call that your action space. Second, when there is uncertainty you want to be as informed as possible. This is what we call prediction and it is where AI comes in. Finally, you have to know what the overall return or payoff is from taking different actions as a result of predictions that are being made. In particular, precisely because any AI prediction is likely to be useable but imperfect, you will want to know what the costs are to you of “errors” – that is, doing the wrong thing because the prediction was incorrect. To ascertain payoffs or costs of errors, you need to have “judgment.”

Judgment is all about outcomes. Sometimes, judgment is something you can work out just as you calculate different rates of return in a spreadsheet. Other times, judgment comes from experience with what the outcomes were for similar decisions in the past. Finally, judgment may well have an emotional or personal component. Regardless of the source, you need judgment to make a decision.

For many of us, when we make decisions we don’t neatly separate out the prediction part from the judgment part. But when you are handed an AI prediction machine, in using it you are forced to consider judgment more clearly. What is more, it may well be that someone who was skilled in decision-making prior to AI, may not be the right person to make that decision with AI. This is because the potential for the application of “better” judgment may rest with someone who wasn’t previously very good at decision-making because they weren’t good at prediction.

Our “go to” example on this is what mobile navigation tools did to driving. Taxi drivers were good at their job because they had learned the fastest route between two places. That helped them be better drivers when thinking about its whole purpose as transportation. Hand any other driver a mobile phone with the app Waze on it, and they have the prediction capabilities of the best taxi drivers. Those people have similar judgment – how to drive a car safely – as taxi drivers but couldn’t compete on prediction (see Figure 13.3). Now they can, and the AI’s ability to decouple prediction from judgement has meant that a vastly greater pool of people can make good pick-up to destination route decisions, causing a transformative impact on the taxi services industry.

S&L: For every rule or standard operating procedure in an organization, “there is the uncertainty that led to it.” How does such “hidden uncertainty” represent major economic opportunity for rule substitution by AI prediction?

Gans: The famous economist George Stigler once remarked, “If you never miss the plane, you're spending too much time in airports.” The architects who designed the new Terminal 2 of South Korea’s Incheon airport have provided an alternate experience model. Arrive early for a flight there and you can visit a spa, gamble at a casino, take in an art exhibition, watch a dance performance or go ice skating. You can also shop to your heart’s content, grab a meal or catch some sleep at the “NAP zone.” For new airport terminals this is becoming the norm. Singapore recently installed a garden with a five-story waterfall. Doha offers a swimming pool and kid’s entertainment center. Vancouver has an aquarium.

People are choosing to spend more time at airports because getting to your flight has become more uncertain. There’s traffic, parking and security lines to contend with. And the consequences of failing to get to your flight on time have risen.

Modern airports are a monument to what we call “hidden uncertainty.” When people do not have the information they need to make optimal decisions – say, regarding when to leave for the airport – they will adopt rules. If you don’t feel the cost of a rule, the uncertainty behind its adoption moves to the background while the consequences of that entire system of choices can be seen in the form of gleaming new buildings with amazing five-story waterfalls.

So it is not simply the rules themselves that represent the target of opportunity for AI enabled new decisions. It’s also the edifices and scaffolding that have been built up to hide the uncertainty that leads to feelings of waste and inefficiency in the rules we have adopted. Not only are they a sign that there is an opportunity for AI, they also represent the magnitude of that opportunity. Just comparing the world on the other side of the arrival gates to those of departure, where these two areas are separate, helps us visualize how big this hidden opportunity might be. Indeed, for airports, some very simple applications of AI represent a threat to all they currently stand for.

Recognizing the disruptive nature of high-impact AI

S&L: A key consequence of the adoption of any new AI system solution is “its implications for power.” What are these likely to be?

Gans: What economic history tells us is that technology-driven transformation does not come easy and real adoption only occurs when new systems are created. It’s important to understand that this process will be disruptive at both industry and firm levels for at least three reasons: First, as we’ve already discussed, the opportunities for the application of AI can be hidden from view, thus existing industries are vulnerable to blind spots. Second, the challenges and trade-offs in taking down existing systems and building new ones are part and parcel of the process of creative destruction that accompanies transformational technological change. Third, as old systems are replaced with new, there is necessarily a shift in economic power that makes the accumulation of power the reward for system innovation and potential disruption something to fear and resist.

So what managers need to understand is that within existing organizations there will be winners and losers in terms of their change in economic power before and after the transformation. Thus, a key challenge for managers will be how to bring everyone along so that during the change process, which may take some time, key shared assets like brand and reputation are not lost. Furthermore, to realize the full potential of AI, companies need to adopt a “system mind-set,” in contrast to the “task-level thinking” which still predominates, and which has disruptive implications for talent recruitment and development. Executives will need to become alert to the potential interdependencies involved in innovating new high impact system solutions, both within and across organizational boundaries.

S&L: What area of business management offers the biggest potential for AI to raise the productivity of the economy and living standards as a whole?

Gans: AI is hard to adopt when it means changing things continually. It is easiest to adopt when you are using AI as part of a process that is already about change. Using AI to help you invent new things, such as new stable molecules or new stronger material designs, will allow you to build new products and those products, rather than AI themselves, are what will be brought to business activity. So the good news is that scientists are embracing AI for innovation. AI is going to have its biggest impact on overall productivity where it supercharges research and development processes.

S&L: Finally, what is the thinking behind the “AI Systems Discovery Canvas” tool you offer business leaders to help them uncover potential opportunities to design innovative systems solutions.

Gans: Most companies have created systems comprising so many interdependent rules, along with so much associated scaffolding to manage uncertainty, that it’s difficult to think about how to undo parts of it and contemplate the new system design possibilities AI predictions afford. So, we suggest starting from scratch and trying to understand your industry at its economic essence, which is what the AI System Discovery Canvas aims to help do. Constructing it involves three steps: (1) articulate the mission; (2) reduce the business to the fewest possible decisions required to achieve the mission if you had super powerful high-fidelity AIs; and (3) specify the prediction and judgement associated with each of the primary decisions. Figure 17.2 shows what a completed canvas might look like for the home insurance industry.

There are two reasons why such an exercise is valuable. One is that it requires you to go back to first principles and consider anew the primary decisions that go into fulfilling your organization’s mission. Some of those decisions may already exist as rules, and may offer opportunities to adopt prediction that will turn those rules into decisions. A second reason is that you can use it to evaluate the system implications of particular AI solutions.[4]

Figures

AI prediction causes decoupling

Figure 13.3

AI prediction causes decoupling

AI systems discovery canvas: home insurance

Figure 17.2

AI systems discovery canvas: home insurance

Notes

1.

Solow, R.M. (1987), Review of Stephen S. Cohen and John Zysman, The Myth of the Post-Industrial Economy, New York: Basic Books (New York Times, July 12th).

2.

Agrawal, A., Gans, J. and Goldfarb, A. (2018), Prediction Machines: The Simple Economics of Artificial Intelligence, Boston: Harvard Business Review Press.

3.

Agrawal, A., Gans, J. and Goldfarb, A. (2022), Power and Prediction: The Disruptive Economics of Artificial Intelligence, Boston: Harvard Business Review Press.

4.

For a more detailed illustration of the use of the “AI Strategy Discovery Canvas” see Agrawal et al. (2022) cited above, particularly Chapters 17 and 18.

Corresponding author

Brian Leavy can be contacted at: brian.leavy@dcu.ie

Related articles