Machine, Platform, Crowd: Harnessing Our Digital Future is a 2017 book by Andrew McAfee and Erik Brynjolfsson (MIT), published in June 2017 by W. W. Norton. It is a sequel to the same authors' The Second Machine Age (The Second Machine Age — Brynjolfsson and McAfee (2014)). The book argues that three technology-driven trends (the expanding capabilities of machines, the rise of digital platforms, and the emergence of the online crowd) require companies to rethink three long-standing balances: between human minds and machines, between products and platforms, and between a firm's internal core and the external crowd. It draws mainly on economics, alongside computer science, psychology, sociology, history and management science, and closes each chapter with a summary and a set of questions for managers rather than a prescriptive playbook.
Summary of the argument
The authors hold that minds, products and the core are not obsolete; human abilities, good products and strong organizational capabilities remain necessary to business success. Their claim is that the second element of each pair (machine, platform, crowd) became much more capable within a few years before publication and therefore needs to be reconsidered. Each of the book's three parts opens with a chapter describing the "standard partnership" between the two elements that took shape about twenty years earlier and the pressures on it, then proceeds through chapters arranged, in the authors' phrase, on a "science fiction gradient" of increasingly unusual developments. The final chapter of each part asks whether computers can be creative, whether the whole economy will become an on-demand economy, and whether companies are an endangered species; the authors' short answers are yes, kind of, and no.
Chapter 1: The triple revolution
The book opens with three examples, one for each trend.
- Machine. Go has roughly 2 × 10^170 possible board positions, more than the number of atoms in the observable universe (about 10^82), and top players cannot fully explain their own strategies, an instance of what the authors call Polanyi's Paradox ("we know more than we can tell"). Commentators in 2014 and December 2015 still expected computer Go champions to be a decade or more away. In January 2016 Google DeepMind published AlphaGo in Nature. Rather than being programmed with human strategy, AlphaGo learned from 30 million board positions from an online repository plus another 30 million generated by self-play. It beat European champion Fan Hui 5–0 in secret in October 2015; after Gary Marcus noted that Fan Hui was ranked only #633 in the world, it beat Lee Sedol 4–1 in Seoul in March 2016, Sedol winning the fourth game.
- Platform. The authors cite strategist Tom Goodwin's March 2015 observation that Uber owns no vehicles, Facebook creates no content, Alibaba holds no inventory and Airbnb owns no real estate. They report that Uber then carried over a million riders a day in 300 cities in 60 countries and Airbnb listed 640,000 lodgings in 191 countries; that Walmart, with more than 150 distribution centers, 6,000 trucks and $180 billion of property and equipment, had a lower market value than Alibaba, which enabled over half a trillion dollars of sales in 2016; that Alibaba's marketplaces enabled $17.8 billion of sales on Singles Day (November 11, 2016), three times the combined US Black Friday and Cyber Monday total; and that Facebook, visited daily by 936 million people, earned $6.4 billion in revenue and $2 billion in profit in the second quarter of 2016 without producing its own content.
- Crowd. General Electric, with a $5.2 billion annual R&D budget, used FirstBuild, a co-creation community launched with the University of Louisville in 2014, to crowdsource the design of a home nugget-ice maker. A 2015 online competition was won by Ismael Ramos of Guadalajara, and a July 2015 Indiegogo campaign for the resulting Opal, priced at $399 (later $499) against a $150,000 goal, raised more than $2.7 million, shipping to more than 5,000 preorder customers in late 2016. The authors state that GE wanted the market intelligence more than the money.
The authors date a first phase of the second machine age to the mid-1990s, when computers took over routine work and productivity growth accelerated (after Robert Solow's 1987 remark that computers were visible everywhere but in the productivity statistics). They place the start of a second phase in the 2010s, marked by Google's autonomous cars (2010), IBM Watson's Jeopardy! win (2011), more than a billion smartphone users by 2012, and machines doing nonroutine work by learning rather than following codified instructions.
As a historical caution, they describe factory electrification. Manufacturing accounted for about 50% of US GDP in the 1910s, yet the large industrial trusts largely failed to exploit electricity: a 1935 survey by Shaw Livermore found that over 40% of trusts formed between 1888 and 1905 had failed by the early 1930s, and a study by Richard Caves and colleagues found that 42 firms dominant in 1905 saw average market share fall from 69% to 45% by 1929. The authors, citing Paul David and Gavin Wright, argue that the gains came from redesigning production (unit drive, assembly lines) rather than swapping motors for steam engines, and that incumbents' expertise and status-quo bias kept them from seeing this. They argue that companies that keep today's technological and organizational status quo are making the equivalent of the choice to keep steam power or group drive.
Part 1: Mind and machine
Chapter 2: The hardest thing to accept about ourselves
The authors describe a "standard partnership" formed in the 1990s through business process reengineering (Hammer and Champy's 1993 Reengineering the Corporation), enterprise systems (adopted by over 60% of the Fortune 1000 by 1999) and the web: computers handled routine record-keeping while people exercised judgment. Drawing on Daniel Kahneman's System 1 and System 2, they argue human judgment is subject to many biases and that fast intuition hijacks deliberate reasoning to justify snap decisions. They adopt the acronym HiPPO ("highest-paid person's opinion"), coined in 2006 by Avinash Kaushik and Ronny Kohavi, for opinion-driven decision-making.
They argue that, where the option exists, data and algorithms usually produce better decisions and forecasts than experienced human experts, and that many decisions should be turned over to algorithms, with people sometimes kept in the loop for commonsense checks and sometimes removed. Examples include FICO credit scoring (Fair Isaac was founded in 1956; by 1999 loan requests of $50,000 or less were decided entirely by computer, and a 2007 Federal Reserve report said scoring reduced opportunities for illegal discrimination), Amazon recommendations, and Google's structured approach to hiring. They also discuss algorithmic bias, including Harvard professor Latanya Sweeney's finding that Google searches for names more common among African Americans were more likely to show ads suggesting an arrest record, and state that the test for an algorithm is not whether it is flawless but whether it outperforms available alternatives and can be improved. In an "inverted partnership," subjective human judgments are quantified and fed into data-driven analyses. They recommend predicting less and experimenting more, and they cite the US Office of Personnel Management's underground "Paperwork Mine," where retirement paperwork took sixty-one days in 1977 and still did at the time of writing, against two days in Texas's digitized process.
Chapter 3: Our most mind-like machines
The authors trace two approaches to AI: rule-based or symbolic AI, which they describe as dormant and unlikely to revive outside narrow domains, and machine learning, which they say is finally doing useful work. They attribute its progress to larger networks, faster and more specialized hardware (including GPUs developed for gaming), more data and improved algorithms. They report Elliot Turner's estimate that the computing needed for a cutting-edge machine learning project could be rented from Amazon Web Services for about $13,000 by fall 2016, and quote Geoffrey Hinton's retrospective view that success was largely a matter of data and computation. They state that neural networks have succeeded mainly at supervised learning and made little progress on unsupervised learning; that supervised learning suits pattern matching, diagnosis, classification, prediction and recommendation; that machine learning is early in its spread and now available on demand in the cloud; and that AI systems still lack common sense.
Chapter 4: Hi, robot
The authors argue that many business processes are "virtualizing," moving to digital channels with fewer people and often only the customer involved (airline kiosks, Global Entry, ATMs and mobile check deposit), and that this is a long-term trend rather than only customer self-selection. Following Gill Pratt of the Toyota Research Institute, they describe a "Cambrian Explosion" in robotics as machines learn to see, and attribute it to five drivers they abbreviate as DANCE: data, algorithms, networks, the cloud and exponentially improving hardware. They report IBM's estimate that 90% of the world's digital data had been created in the previous twenty-four months. They expect robots to be used where work is dull, dirty, dangerous and dear, while arguing that human agility and dexterity will keep people working alongside robots for some time. They treat 3D printing as an example of digital tools spreading into manufacturing and raising the rate of innovation.
Chapter 5: Where technology and industry still need humanity
The authors argue that computers can now meet most definitions of creativity, citing generative design (software that produces complete designs, such as a 3D-printed heat exchanger, without a human drafting them) and a Baylor College of Medicine and IBM study in which Watson generated useful scientific hypotheses about kinases that activate the protein p53. They argue that machine designs often differ from human ones, which they regard as beneficial for diversity of viewpoints, and suggest machines handle "busywork" and first drafts that people extend. They hold that computers do not understand the human condition and do not expect a decent machine-written novel soon. Following MIT colleague Deb Roy, they argue that work drawing on social drives (compassion, pride, embarrassment, envy, justice, solidarity), such as coaching, leadership, teamwork and empathy, will remain with people for some time, and that high-level social skills may become more valuable than quantitative ones, with the combination paying most.
Part 2: Product and platform
Chapter 6: The toll of a new machine
The authors describe how the internet disrupted retailing, journalism, photography and recorded music over twenty years. In the mid-1990s, US newspapers (2,400 of them) generated $46 billion in annual revenue, 30% of it from classified ads in 1995. They attribute the disruption to the "free, perfect, and instant" economics of digital information goods on networks: near-zero marginal cost of copies (a gigabyte of storage cost $0.02 in 2016 against $11 in 2000), perfect replication and instant global distribution. Networked goods gain value as users grow, producing demand-side economies of scale. They define a platform as a digital environment with near-zero marginal cost of access, reproduction and distribution, and argue that platform economics, Moore's law and combinatorial innovation continue to reshape industries from computer hardware to music.
Chapter 7: Paying complements, and other smart strategies
Using the iPhone, which sold over 6 million handsets in its first year, the authors recount how Steve Jobs initially resisted outside apps before Apple opened the App Store. They explain complements (goods for which a price drop in one raises demand for the other) and argue that opening a platform to outside contributors raises demand for the owner's product, with digital complements often being free, perfect and instant. Platform owners must curate contributions to maintain quality, compete on attracting and curating contributions, and find it much harder to succeed once two platforms are established, particularly when consumers will not "multihome." They stress user interface and experience and describe many platforms as two-sided.
Chapter 8: The match game: why platforms excel
Platforms are spreading into physical goods and services, which the authors call O2O (online-to-offline), using ClassPass as an example: founder Payal Kadakia ended its ClassPass Unlimited plan (launched May 2014) in November 2016 because heavy use made it unsustainable. Because perishable inventory such as a class slot or a seat has low marginal cost but fixed capacity, O2O platforms use revenue-management techniques (the authors recount American Airlines' use of revenue management against People Express in the 1980s) to match supply and demand. They report that China has been a center of O2O innovation and that O2O platforms can scale quickly because they add members fast, control the customer experience, use existing capital and labor, and apply data and algorithms to matching, with investors funding aggressive expansion.
Chapter 9: Do products have a prayer?
The authors argue that platforms can capture much, most or all of an industry's value. Uber, founded after Travis Kalanick and Garrett Camp's 2008 difficulty hailing a cab in Paris, reportedly generated $20 billion in gross bookings in 2016 and by June 2016 was valued at $68 billion after raising $15 billion. Los Angeles taxi trips fell almost 30% within three years of Uber and Lyft's arrival from 8.4 million in 2012; San Francisco's Yellow Cab Cooperative filed for bankruptcy in January 2016; New York medallion prices, above $1.3 million in 2013, fell by as much as half. The authors attribute platform success to reducing information asymmetries, to two-sided and multisided networks with cross-elasticities, and to switching costs that reward early investment in market share. Incumbent product makers risk looking like utilities with limited profit and growth, and platforms may try to erode incumbents' brands. They argue the disruptive potential of O2O platforms is more limited where physical goods are differentiated and customers can be locked in, explaining why they expect Airbnb not to displace hotels.
Part 3: Core and crowd
Chapter 10: That escalated quickly: the emergence of the crowd
The authors define the crowd as the opposite of the core: huge, diverse, largely uncontrollable and messy, with an emergent structure that stock markets, prediction markets and search engines extract information from. Drawing on Friedrich Hayek and Polanyi's Paradox, they argue that overcentralization fails because people cannot always articulate what they have, know, want or can do. They present Linux as a crowd product built under "geeky leadership" following principles of openness, noncredentialism, self-selection, verifiability and clarity of goals. As a contrast they describe Nupedia, the expert-only encyclopedia Jimmy Wales and Larry Sanger launched in 1999, which after eighteen months and $250,000 had twelve completed articles; its wiki-based successor Wikipedia, launched January 15, 2001, had 617 articles by the end of that month, 19,000 by year end, and 36 million articles in 291 languages by 2016. They conclude that following only some of the principles works poorly and that finding the right balance takes trial, error and luck.
Chapter 11: Why the expert you know is not the expert you need
The authors argue that credentialed experts are repeatedly outperformed by uncredentialed crowd members, because the knowledge needed to solve a problem often lies in a distant domain and its location is hard to predict. In a study by Karim Lakhani, Kevin Boudreau and colleagues, a white-blood-cell gene annotation problem was reframed as a general algorithmic task and posted to Topcoder, a platform with about 400,000 developers; over fourteen days, 122 people or teams from sixty-nine countries made 654 submissions. Most were less accurate than the benchmarks, the NIH's MegaBLAST (1 million sequences in about 4.5 hours at 72% accuracy) and the idAb algorithm (under forty-eight minutes at 77% accuracy), but thirty beat MegaBLAST's accuracy, sixteen beat idAb's, eight reached the estimated 80% maximum, and the three fastest ran in sixteen seconds; total prize money was $6,000. About half the entrants were still students, and none were academic or industrial computational biologists. They describe ways for the core to tap the crowd, crowd efforts that need little core at all, and crowd-based startups challenging incumbents.
Chapter 12: The dream of decentralizing all the things
The authors present Bitcoin, proposed in Satoshi Nakamoto's October 31, 2008 paper, as a demonstration that fully decentralized communities combining cryptography, economics, code and networks can create money. They note its exchange-rate volatility (over $1,100 in November 2013, less than $250 in January 2015, more than $830 two years later) and argue that the blockchain, an open, transparent, global, flexible and immutable ledger, may matter more than the currency, especially with smart contracts. They present the view that large organizations have become too powerful and that extreme decentralization offers an alternative as a position held by some, and say early initiatives indicate strong demand for new ledger technologies, while the eventual roles of Bitcoin and blockchain remain unclear.
Chapter 13: Are companies passé? (Hint: no)
The authors describe The DAO, a decentralized autonomous organization built on Ethereum that raised $162 million in twenty-eight days in May 2016 before an attacker drained roughly one-third of its funds by exploiting its code. Using transaction cost economics (Ronald Coase), incomplete-contracts theory (Sanford Grossman and Oliver Hart) and residual rights of control, they argue that firms exist because contracts cannot specify every contingency, and that managers also coordinate people, articulate goals and strategy and shape culture. They note that although technology has lowered transaction costs, economic activity in most industries and geographies is concentrating in fewer companies. They conclude that companies will remain part of the economic landscape for a long time, and that the leading companies of the second machine age, while operating differently, will almost all be recognizable as companies.
Key claims
- Machine learning has overtaken symbolic AI and, by getting past Polanyi's Paradox, lets machines do nonroutine work; the authors urge defaulting to data and algorithms over HiPPO judgment where the option exists, while noting algorithmic bias.
- The "free, perfect, and instant" economics of digital goods and network effects let platforms capture much of an industry's value, with complements, curation, two-sided networks and early scale as the main mechanisms; O2O platforms extend this to physical goods, though less so where products are differentiated.
- The crowd can outperform the core because relevant knowledge is dispersed and hard to locate, but crowds need structure and "geeky leadership."
- Radical decentralization has limits: incomplete contracts and management's other functions mean companies persist.
Reception and later assessments
The framework was written before the transformer architecture and the generative-AI wave, but several of its threads connect to AI dynamics of 2024–2026. On the Machine axis, the HiPPO-versus-algorithm question recurs in discussions of agentic deployment (Agentic AI, AI and Productivity, Generative AI and the Nature of Work). The Platform axis is reflected in characterizations of frontier labs (Anthropic, OpenAI, Google DeepMind) as platforms rather than single-product firms, with their API and Cowork ecosystems described as two-sided markets. The Crowd axis is reflected in open-weight models (Llama, Qwen, DeepSeek, Kimi K2) positioned as crowd responses to the closed-frontier-lab core, in the open-source versus closed-model debate, and in Hugging Face as a distribution crowd.
The authors' specific forecasts have aged unevenly. Their claims about machine-over-HiPPO decision-making in structured-data domains, platform dominance in consumer markets, and crowd disruption in software (Linux and open source) have largely held. Other forecasts require updating: the authors were optimistic about platform and API openness, whereas the 2020s saw retrenchment in API access (for example Twitter/X and Reddit); they were bullish on blockchain as a crowd substitute for core institutions, a position complicated by the 2018–2024 crypto cycles; and their 2017 framing of AI predates the transformer, leaving most of the book's AI predictions conservative by 2026 standards. These retrospective assessments are not drawn from the book itself and are not attributed to a named commentator.
Provenance
Andrew McAfee and Erik Brynjolfsson are both affiliated with MIT. Erik Brynjolfsson's later work with Lindsey Raymond on generative AI builds on this framework (Generative AI and the Nature of Work, Generative AI at Work). The raw file is a full-text PDF extraction of the book.
Relationships
- supports: Generative AI and the Nature of Work — Brynjolfsson's more recent work with Raymond on generative AI builds on this framework.
- supports: Generative AI at Work — customer-service RCT evidence for Machine-axis claims.
- related: The Digitalist Papers (Stanford, Volumes 1–2) — Brynjolfsson is Vol. 1 co-editor; framework informs his editorial perspective.
- related: The Industrial Explosion (Davidson, Hadshar) / Three Types of Intelligence Explosion (Davidson, Hadshar, MacAskill) — concepts that extend Machine-axis to AI-on-AI recursion.
- related: Open-Source AI / Open-Weight Models — crowd-axis applied to frontier models.
- predecessor: The Second Machine Age — Brynjolfsson and McAfee (2014) — same authors' prior book.