Artificial Intelligence (AI) offers the prospect of transformational changes but history tells us that this is likely to take longer than AI enthusiasts are currently suggesting. That’s because radical change will only occur when there is a systems-level approach to adopting AI. That’s the view of Joshua Gans, co-author with Ajay Agrawal and Avi Goldfarb, of Power and Prediction, the disruptive economics of artificial intelligence.
Gans, a professor of Strategic Management at Toronto’s Rotman School of Management, is one of the most highly regarded academic experts on AI in the world and a cofounder of a leading global programme for the commercialisation of AI.
“It’s easy to be seduced by the technical achievements of AI but existing systems are slowing things down and businesses are telling us that they are not adopting AI as much as they thought they would. We should look at history to see the compelling reasons why it shouldn’t be any different,” Gans tells The Irish Times.
The real breakthrough came when entrepreneurs started to rethink the factory
Gans and his colleagues note the similarities between AI adoption and the early roll-out of electricity. Thomas Edison famously demonstrated the electric light bulb in 1879 yet 20 years later, only 3 per cent of US households had electricity and factories had barely more take-up. Twenty years later, however, that number had accelerated to over half the population.
There was plenty of hype around electricity in the early years but not much to show for it.
Electricity replaced steam in factories. Steam was wasteful and electricity’s immediate opportunity was to provide a much more efficient source of power at the same point in factories where steam was used. This is what the authors refer to as a point solution.
The next development was to mount an electric drive on a single machine, in what would now be referred to as an application solution. The challenge was that individual machine tools such as drills and metal cutters had to be totally redesigned to take advantage of having an individual unit electrical engine.
Nineteenth century factories, however, were designed to leverage steam, with power distributed to individual machines through a central shaft upon which belts and pulleys were hung. The real breakthrough came when entrepreneurs started to rethink the factory. The question was: what would a factory look like if you designed it from scratch, given what you now know about electricity?
Electricity revolutionised factory design. As the authors note: “Henry Ford could not have invented the production line for the Model T car with steam power. Only electricity, decades after its commercial promise was shown, could achieve that. Yes, Ford was a care entrepreneur, but he was largely a system solution entrepreneur. These system changes altered the industrial landscape. Only then did electrification show up in the productivity statistics.”
Roll forward to modern times.
A 2020 study by Sloan Management Review and global consultancy group BCG found that just 11 per cent of organisations had seen significant financial benefits from AI. This wasn’t for lack of trying as 57 per cent had deployed or piloted an AI strategy.
Like electricity, the authors argue, we are currently in “The Between Times”.
AI can do so much more but changing the status quo involves shifting power and those who have it are naturally reluctant to cede it
Good point solutions now abound but for now they are largely in the low hanging fruit territory.
Financial services is an obvious case. For years, firms in this sector have employed teams of data scientists to predict fraud and money laundering, and noncompliance in financial transactions. They are a model for AI, with its capacity for enhanced prediction.
No surprise then that Nasdaq, a New York-based stock market, paid $2.75 billion (€2.6 billion) in 2020 to acquire Canadian AI firm Verafin, which had invested heavily and built leading edge tools to identify fraud and validate the identity of customers. Verafin allows Nasdaq to do what it has long done, except better. It fits within an existing systems architecture.
AI can do so much more but changing the status quo involves shifting power and those who have it are naturally reluctant to cede it. In the early 2000s, for example, video rental firm Blockbuster could see where the industry was going at the time as much as its rival Netflix but its franchisees were disadvantaged by the changes it tried to make it and wielded their power to support the status quo. Blockbuster collapsed.
Consider insurance. The business model here concentrates on calculating likely losses across the portfolio insured and then building in a margin to create profit. Premiums spread risk among the pool but in a relatively crude way. AI could in theory upend this model by incentivising risk reduction behaviours in a way not done now.
If a home insurance company used AI to predict that a particular homeowner had an especially high risk of electrical fire or flooding, rather than charging them a high premium, the company could share that information so that customers could take action to lower their risk, such as investing in low-cost devices for early detection of heightened fire or flood risk.
One of the major problems about clinging to the status quo is that it risks disruption
While a few insurance companies have dabbled in this area – think tracking devices in cars for young learner drivers – few have gone down this path with any scale. Changing the model from insuring risk to mitigating risk could make good business sense in the long-term, but it would lower premiums and change the business model, upsetting certain stakeholders in the short-term.
Understanding power dynamics in organisations is key as Gans notes:
“The challenges that existing businesses have in adopting new systems is that they are doing some of the current stuff well. Moreover, you will find a load of people who are benefiting from the current system. Then you’ll find people advocating for AI on the basis of the exciting new possibilities it offers. What that tells you is that the AI will redistribute economic power among those people – and those people know it. The change in economic power is the challenge, that’s the hard bit for the organisation’s leaders to navigate.”
One of the major problems about clinging to the status quo is that it risks disruption. The advantage moves to the nimble, often a lean start-up with no organisational drag. Gans notes that start-ups are constantly testing, running hypotheses and refining their business propositions based on feedback, something that on the surface appears much harder for legacy businesses to do.
One solution is to borrow from the toolkit of the lean movement and introduce a blank canvas approach to business modelling, posing a lot of “what if?” type questions. Simulation can help here. It is possible to use digital assets, including digital twins of physical environments, to simulate different options and use AI to predict the outcome of each.
Disrupters have one advantage in deploying AI, as long as they employ speed. The key is capturing feedback data, which enables prediction performance to improve constantly. In this way, AI systems display a very human trait – they can learn from outcomes as the authors explain:
“For data to generate a long-term advantage, early movers need to harness feedback data. Operating in the field, they can collect powerful feedback data that they can use to improve their predictions, making it harder for others to catch up. The advantage isn’t in launching when others can’t. The advantage is that launching enables the collection of feedback data.”
Take Google, for example. Because of the massive investment it has made in talent as well as data, it makes very accurate predictions about who wants what and when, which makes it an excellent targeted advertising platform. It has built-in feedback loops enabling it to see whether its predictions are accurate or not. If it finds it has made an error, it corrects the model for next time, which makes it difficult for competitors to compete.
Prediction machines have a long way to go before they replace humans
Fast feedback loops lead to races to market. Imagine the first AI that is able to safely navigate a car through New York City. Once it receives regulatory approval, the AI will continue to collect data and get better and better. When a second AI is approved it won’t have the same quantity and quality of data, lowering its appeal.
As the authors note, for an AI to be accurate, it needs enough data. Gathering the data to achieve this minimum scale takes time and effort. The advantage to launching first depends on how much effort is required to have a commercially viable prediction.
Sometimes not much effort is needed. Early internet search engines were not especially accurate as there was a high tolerance for error. This encouraged fierce competition. Contrast that with autonomous vehicles where the tolerance for error is extremely low. AI’s efficacy needs to be measurably better than humans if people’s lives are at stake. The cost of this is high, which discourages competition.
Gans and his co-authors write thoughtfully as well as insightfully about the possibilities of AI, but also see the challenges and maintain a healthy scepticism throughout the book.
Sometimes the hype simply doesn’t stand up to scrutiny. Some have suggested that AI could replace radiologists and it’s understandable why – the machines are demonstrably better at predicting the right diagnosis from an image. But that fails to understand the other important things radiologists do. The authors list 30.
Prediction machines have a long way to go before they replace humans. The key is knowing when and how to use them.
Power and Prediction, the disruptive economics of artificial intelligence, by Ajay Agrawal, Joshua Gans and Avi Goldfarb is published by Harvard Business Review Press.