AI and Hype: An Assessment of McKinsey’s Forecast for Artificial Intelligence

  This report assesses an analysis by the McKinsey Global Institute on the future of artificial intelligence (AI), with a focus on the impact of AI on productivity and economic growth. It focuses on productivity and economic growth because thought leaders such as Erik Brynjolfsson, Andrew McAfee, Martin Ford, Peter Diamandis, and Ray Kurzweil claim that there are endless opportunities arising from AI and related technologies such as robots and drones. Many of them argue that the main challenge for policy makers is to prevent mass unemployment in the face of rapid improvements in productivity that will also disrupt many incumbents.

   Others have made bolder claims. For example, Accenture, Frontier Economics and Allianz Global Artificial Intelligence claim that artificial intelligence-enabled technologies could double the economic growth rates of many advanced countries by 2035[i]. PricewaterhouseCoopers predicts that AI will drive global Gross Domestic Product (GDP) gains of US$15.7 trillion by 2030[ii]. But these reports provide few details[iii] and optimistic predictions continue to be made even though productivity improvements remain flat and even as two precursors to AI, the Internet of Things and Big Data, have grown rapidly over the last 10 years with combined markets of over $300 Billion[iv].  

     This report uses a June 2017 report by the McKinsey Global Institute to assess these bold forecasts, as it also critiques the report from the McKinsey Global Institute. It focuses on the McKinsey Global Institute because it is a think tank for McKinsey and Co., the world’s leading managerial consulting firm, and because the report provides many more details than do other reports such as one from PwC cited in the previous paragraph. McKinsey’s report (Artificial Intelligence: The New Digital Frontier?) assesses AI’s impact on four categories of activities and for five industries. The four activities are: 1) enabling companies to better project and forecast to anticipate demand, optimize R&D, and improve sourcing; 2) increasing companies’ ability to produce goods and services at lower cost and higher quality; 3) helping promote offerings at the right price, with the right message, and the right target customers; and 4) allowing them to provide rich, personal, and convenient user experiences. The five industries are retail, electric utilities, manufacturing, health care, and education.

     My assessment relies on a 40 year-career in academic and professional positions related to new technologies, on information available from the online media, and on several books about AI. Many searches were done to check the veracity of specific events or forecasts in the McKinsey Report, often checking whether a success story has been replicated in other companies. The books include the AI Advantage by Thomas Davenport, Prediction Machines by Ajay Agrawal, Avi Goldfarb, and Joshua Gans, The Rise of the Robots by Martin Ford, The Second Machine Age and Machine Platform Crowd by Erik Brynjolfsson and Andrew McAfee, The Future of the Professions by Daniel and Richard Susskind, and AI Super-Powers by Kai-Fu Lee. Davenport’s book is referenced the most partly because it is the most recent, published in September, 2018.

 Assessing the Categories and Industries

   I begin with an assessment of the four categories and five industries. Clearly numbers two and four will have more impact on productivity than will the other two. The lower cost and higher quality of number 2 and the rich, personal, and convenient user experiences of number 4 are much more relevant to productivity than are better forecasting or better promotions (which are emphasized by many books including Prediction Machines). Although categories 2 and 4 will likely have a large impact on competition between firms and the best strategies for them, they will likely have a small impact on cost and quality in most industries (see next section for most details). Exceptions include industries where fixed assets have a huge impact on costs and thus increased asset utilization through better price promotions and targeting can lead to dramatically lower costs, perhaps for example in hotels and airplanes. As for the five industries chosen to analyze, they represent a good mix of economic sectors, although the largest consumer expenditure (housing) is not considered.

    Exhibit 1 is taken directly from the McKinsey report (Exhibit 6 in the report). It summarizes the impact of AI on the four categories and five industries. This includes the general impact of AI on better forecasts, more efficient operations, better price offerings, and richer and more personal experiences for each of the five industries. The first question concerns the magnitude: how much will AI improve forecasts, automate operations, optimize pricing, or personalize products and services? It is easy to argue that AI will have some impact on all four categories in each of the five industries, but the real question is how much of an impact and when.

    A second question to ask is: how important are these four categories to the cost and quality of products and services, particularly from the five industries listed in the exhibit? We have already noted that categories 2 and 4 probably have more impact on cost and quality than the other two categories. Nevertheless, we would like to know more than details than this; for example, what percent of costs are represented by those that will be impacted by AI? What aspects of quality are important to customers and which of them can be improved by AI? Without knowing these types of answers, it is difficult to reach conclusions about AI’s impact on productivity.

   A third question, one related to the second question, is what are the current levels of inefficiencies in specific industries? Some of AI’s purported benefits will come from lower inventories, less equipment downtime, and higher quality and yields. Thus, a key question is what are these levels in specific industries? Are there currently high inventories, large downtimes, and low yields that can be improved by AI? Similarly, other benefits include better back-office operations including development processes. What percent of costs do these operations represent?

Exhibit 1: Artificial intelligence can create value across the value chain in four ways

     A fourth question to ask is: what are the historical trends in each of the four categories and five industries? Improvements have been occurring in inventories, downtime, and yields for many decades, largely driven by improvements in previous generations of information technology (IT), and thus we would expect some further improvements even without AI. Knowing the historical trends would help us estimate the magnitude of the likely effect from AI, also in comparison to past technologies. For example, will AI have a greater impact on retail than did bar-code scanners, which reduced the cost of groceries for consumers by an estimated 1.4% according to a McKinsey study[v]. Or will the impact of AI be another example of Solow’s paradox, "You can see the computer age everywhere but in the productivity statistics.”

    Exhibit 2 is also taken directly from the McKinsey report (Exhibit 7). Based on analyses of the five industries, it summarizes the likely impact of AI on the four categories for each of the five industries. These summaries reference claims made by incumbents and startups on what they believe their AI products and services can do. A first comment is that firms often exaggerate the benefits of their products and services and furthermore only a small number of startups succeed, so we must be careful about believing their claims and extrapolating from them. History is littered with failed startups and with over-hyped technologies such as nuclear fusion, solar water heaters, synthetic fuels, hydrogen vehicles, cellulosic ethanol, the Strategic Defense Initiative (a.k.a., Star Wars), microfinance, the aerospace hysteria of the 1960s, and magnetic levitating trains.

   These historical examples should be kept in mind as we look for evidence for the claims made by AI startups and incumbents. Are the claims consistent with claims made by other firms and by users of these technologies? Are they consistent with answers to the four questions raised above? The larger the number of startups and incumbents making similar claims and the larger the number of users confirming these claims, the more believable the claims become. The subsequent sections focus on each box in Exhibit 2, beginning with those for retail, and ending with back-office operations, a category not emphasized by the McKinsey and not in Exhibit 2, but one receiving much emphasis by others.

3. Retail

    Exhibit 2 says that better projections through AI will lead to lower inventories (20% reduction), fewer product returns (2 million), and higher profits (1-2% EBIT). But why 20% and 20% of what, floor space or warehouse inventory? Also, how high are inventories as a percent of sales; how much has this dropped over the last 40 years; and how much further can we expect them to drop? We can also ask similar questions about product returns. Two million is what percentage of purchases, how has this percentage increased or fallen over the last 40 years and how might AI impact on this in the future?

Exhibit 2. AI can help capture significant gains, across the value chain

    Exhibit 2 says that more efficient production through AI-based AVs will lead to a “30% reduction of stocking time” in warehouses. But is stocking time in warehouses important? For example, how much retail costs are represented by stocking time in warehouses, or even in stores? Or are other things such as cost of floor or warehouse space, damaged goods, or delivery activities more important drivers of cost and how might AI impact on them?

   The text in the report provides more details than do the two exhibits, and sometimes they address the estimates made in the exhibits for retail. For instance, the text says these types of things: “AI-based approaches to demand forecasting are expected to reduce forecasting errors by 30 to 50 percent” and reduce “lost sales due to product unavailability by up to 65 percent.” The report continues: “Costs related to transportation and warehousing and to supply chain administration are expected to decrease by 5 to 10 percent and 25 to 40 percent, respectively. With AI, overall inventory reductions of 20 to 50 percent are feasible.” These estimates are plausible, although more details would be nice. Furthermore, without a discussion of cost structure or references to historical trends (as noted in Section 2), it is hard to understand the plausibility of these estimates or their importance to productivity.

    The text in the report also emphasizes increasing online sales, consistent with other analyses and with all historical trends. It mentions the use of digital assistants and voice recognition for placing orders, and drones for delivery. While digital assistants and voice recognition may be plausible, the report does not mention the problems with drones making landings at their final destinations or even the need for high energy density batteries to carry objects for any distance and instead emphasizes the impact of deep learning on the feasibility of drones. This is unfortunate because better batteries are needed to carry objects any reasonable distance and even to enable accurate and stable landings. Given the slow diffusion of electric vehicles (and thus poor cost-performance of batteries), it is likely that drones will be limited to high value items. And even for these applications, it will probably take many years before batteries have sufficiently high energy densities to enable long-range services or precision landings in crowded areas.  

    Neither the text nor the table mention another likely outcome for retail, the elimination of check-out. Many stores have experimented with self-checkout and some are now experimenting with automated checkouts. For example, Thomas Davenport describes Amazon’s pilot Go stores in Seattle in his book The AI Advantage. Amazon’s “Just Walk Out Technology automatically detects when products are taken from or returned to the shelves and keeps track of them in a virtual cart. When you’re done shopping, you can just leave the store. Shortly after, we’ll charge your Amazon account and send you a receipt.” Nevertheless, the impact on costs will likely be small because humans will still work in these stores, accepting deliveries, stocking shelves, and changing prices.

4. Electric Utilities

    Exhibit 2 says that deep learning “to predict power demand and supply” can “cut national electricity usage by 10%,” an ambitious goal. But how does a better match between supply and demand lead to less energy usage? A cut in peak energy usage would be more believable, but the exhibit and the text in the report emphasize the ability of deep learning to cut energy usage.

    Exhibit 2 also says that machine learning and smart sensors can increase energy production by 20% and improve profits (EBIT) by 10-20%. The former purportedly comes from better optimization of assets and the latter from enhanced predictive maintenance, automated fault prediction, and increased capital productivity. But the 20% figure for increased energy production assumes that assets are down for at least 20% of the time; is this true? Base load electricity plants (coal and nuclear) are running a high percentage of the time due to the high costs of shutting them down while gas, oil, and solar are used to meet peak demand. And the shutdowns are largely planned, such as for reloading nuclear fuel. Better data on the amount of unplanned downtime would be useful to understand the potential impact of AI on energy costs. 

  Exhibit 2 also says that machine learning can help users automatically switch between electricity providers thus giving them $10-$30 savings on monthly bills. Do these different prices reflect dramatically different supplier costs or just that different suppliers will offer different time-of day prices because they have different cost structures? If it is the latter, then the impact on productivity will be small.

    The text in the report provides details, but little of it provides justification for the estimates in Exhibit 2. For example, the text cites an academic paper and the success of DeepMind, a Google owned-startup, to justify the impact of AI-based forecasting to reduce energy usage. The academic paper (Jaime Buitrago and Shihab Asfour, Short-term forecasting of electric loads using nonlinear autoregressive artificial neural networks with exogenous vector inputs, Energies, vol 10, no 1, 2017) only discusses forecasting techniques and does not say anything about the possible cost reductions from better forecasts.

    The success of DeepMind might be a bigger reason for optimism about AI. It purportedly reduced energy usage at a Google data center by 20% and Brynjolfsson’s and McAfee’s 2017 book, Machine, Platform, Crowd: Harnessing Our Digital Future, provides more details on this example, apparently one of many successes that were achieved by Google’s DeepMind by 2015. DeepMind’s algorithms used many years of historical data from thermometers (inside and outside), pressure gauges, hydrometers, and other sensors to suggest better settings for pumps, coolers, cooling towers, and other equipment, admittedly a complex task that might benefit from AI. On the other hand, McKinsey’s extrapolation from data centers to DeepMind’s claims are less believable. Can DeepMind really reduce the UK’s energy usage by 10% and maximize renewable energy? The claims would be more believable if information about other data centers and other energy applications existed. My online search for such examples found very little evidence of this[i], certainly not proof that they don’t exist, but evidence that it might it might not occur, or it might take a while before large economy wide energy reductions (and thus improvements in energy productivity) are achieved through AI.

   The report also emphasizes smart meters and thus presumably some of the estimates in Exhibit 2 assume that smart meters will provide some if not most of the purported reductions in energy usage. However, the text mostly emphasizes the large number of smart meters that have been installed, driven largely by government subsidies, and does not mention, much less demonstrate falling energy usage from the installation of smart meters. My search for independent analyses of smart meters also found no evidence of a linkage between them and lower energy usage. Instead my search found that the advertised potential 2020 energy savings figure for the UK has been more than halved since the campaign began, dropping from ￡26 to just ￡11 a year for duel-fuel bills[i] and the cost of smart meters and their installation has risen, not good news for smart metes[ii]. Google’s financial reports also do not support the idea that smart meters will have a large impact on energy efficiency. Google reports revenues of $726 million with losses of $621 million in 2017, results that do not suggest smart meters are reducing energy usage[iii].

     The text uses a claim from GE’s wind turbine business and a startup to justify more energy output through AI. For example, GE says AI could boost a wind farm’s energy production by as much as 20 percent and create $100 million in extra value over the lifetime of a 100-megawatt farm. Upside Energy, a startup that received a UK government grant to use AI to manage a portfolio of batteries and other storage assets, estimates that machine learning could be used to help unlock up to six gigawatts of demand-side flexibility that can be shifted during the evening peak without affecting end users. The report uses these examples to “estimate that optimizing preventive maintenance, automating fault prediction, and increasing capital productivity through AI applications could increase power generation earnings before interest, taxes, depreciation, and amortization (EBITDA) by 10 to 20 percent,” estimates that assume much larger reductions in cost than 10 to 20%. But the text provides few details, and thus how generalizable are the claims made by GE and Upside Energy? My independent search for similar examples did not find anything noteworthy.

 For the rest of the report, see my papers on Research Gate: https://www.researchgate.net/profile/Jeffrey_Funk/research

[i] Energy firms are running out of time to meet 2020 smart meter rollout deadline, Which?, Alex Geraghty, 19 November 2018.

[ii] Smart meter benefits cut by old technology and rising costs: Energy companies may end up spending billions unless they can install smart meters more cheaply, Cliff Saran, 23 Nov 2018, ComputerWeekly.com.

[iii] https://www.recode.net/2018/4/23/17272756/google-alphabet-nest-q1-earnings-2018-revenue-operating-loss

[i] The title of this June 11, 2018 FastCompany article (Microsoft is using AI to cut the cloud’s electric bill) suggests that AI is having an impact on energy costs but the text of the article focuses more on the benefits of outsourcing data processing to the cloud than to AI, which can also be seen in the sub-title (Microsoft’s cloud is far more energy-efficient and carbon-efficient than traditional on-site data centers).

[i] https://www.accenture.com/us-en/insight-ai-industry-growth  https://hk.allianzgi.com/zh-cn/retail/our-products/fund-in-focus-landing/allianz-global-artificial-intelligence

[ii] www.pwc.com/gx/en/issues/data-and-analytics/publications/artificial-intelligence-study.html

[iii] www.pwc.com/gx/en/issues/analytics/assets/pwc-ai-analysis-sizing-the-prize-report.pdf

[iv] Bain argues the market for IoT will grow to about $520B in 2021, more than double the $235B spent in 2017. Discussed in Forbes by Louis Columbus,  August 16, 2018. IDC predicts revenue from the sales of big data and business analytics applications, tools, and services will increase more than 50%, from nearly $122 billion in 2015 to more than $187 billion in 2019. Jessica Davis, Information Week, Big Data, Analytics Sales Will Reach $187 Billion By 2019, May 24, 2016

[v] https://www.mckinsey.com/~/media/McKinsey/Business%20Functions/McKinsey%20Digital/Our%20Insights/Where%20machines%20could%20replace%20humans%20and%20where%20they%20cant/SVGZ-Sector-Automation-ex3.ashx