Sources for Book 2: Business in the Age of AI Agents
Status: this English source list is a working version, not the final edition. It gives readers the current named links for the selected English manuscript; final English source activation and release verification remain separate.
This page brings together bibliographic cards and verified direct links for Book 2. Cards follow the book's reading order, and their numbers match the source footnotes in the manuscript. Printed page numbers are intentionally omitted because they change with format and typesetting.
How to read the cards
- The description identifies the source and explains which fact, quotation, or qualification it supports.
- Open sources lists the named original publication, official document, independent confirmation, bibliographic record, or useful context.
- Gaps in numbering are expected: some numbers belong to glossary or explanatory notes that are not repeated here.
A linked source may open in the language in which it was originally published.
From the Author: When the World Around the Individual Is Rebuilt
Quotes and epigraphs
::: {.source-note} 1. The line “The future is already here—it's just not evenly distributed” is commonly attributed to William Gibson, with references to a 1993 Fresh Air interview and a 2003 Economist article. :::
Books and frameworks
::: {.source-note} 7. Fiduciary duty is the obligation to act in the client's interest. AI can provide a first analysis, but licenses, conflict disclosures, and legal responsibility remain with a person or organization. :::
Markets, companies, and products
::: {.source-note} 2. Agentic commerce is a layer of protocols and rules that lets programs search, compare, negotiate, and pay for a person or company. Here it is a framework for a transition already under way, not a numerical forecast.
Open sources: Stripe — Agentic Commerce Protocol; Google Cloud — Agent Payments Protocol; Google for Developers — Universal Commerce Protocol. :::
::: {.source-note} 3. VkusVill's public MCP server gives an AI agent access to the product catalog, ingredients, prices, ratings, and a checkout link. The user still takes the final purchase action.
Open sources: Source 1: habr.com. :::
::: {.source-note} 4. SparkToro, using Similarweb data, estimated that 68.01 percent of US Google searches from January through April 2026 ended without an external click. Semrush found a 92–94 percent zero-click share for Google AI Mode in an early sample of almost 69 million US desktop sessions from May through July 2025. The products, periods, and methods differ, so the figures should not be compared directly.
Open sources: Source 1: blog.google; Source 2: sparktoro.com; Source 3: semrush.com. :::
::: {.source-note} 6. McKinsey's lower headcount does not show that AI “cut the jobs.” Public reporting attributes the decline partly to the consulting slowdown and the reversal of pandemic-era expansion. McKinsey leadership has also said that roughly a quarter of the firm's work uses outcome-based arrangements. The figures support a shift in the commercial model, not a simple automation story. :::
Concepts used across the books
::: {.source-note} 5. Outcome-based payment shifts the unit of value from hours, slides, or analysis to a measurable outcome: what changed in revenue, risk, speed, or decision quality. :::
Chapter 1. AI Is Not a Plug-In—It Is a Reconstruction
Quotes and epigraphs
::: {.source-note} 8. Robert M. Solow, review of The Myth of the Post-Industrial Economy, “We'd Better Watch Out,” 1987. His line about the “computer age” gave the productivity paradox its memorable formulation: the technology was already visible everywhere, while the gain had not yet appeared in the productivity statistics.
Open sources: standupeconomist.com. :::
Research, reports, and data
::: {.source-note} 9. MIT NANDA, The GenAI Divide, 2025. This chapter treats it as a practitioner report about the scale of the problem: under the report's methodology, most generative-AI pilots had not produced measurable P&L impact at the time studied. It is not presented as a peer-reviewed economic law.
Open sources: fortune.com; virtualizationreview.com. :::
::: {.source-note} 12. Deloitte, State of AI in the Enterprise 2026. The relevant signal is the gap between employee access to sanctioned AI and mature governance of autonomous agents.
Open sources: Deloitte. :::
::: {.source-note} 13. McKinsey & Company / QuantumBlack, The State of AI (2025–2026). This chapter uses the research as context: adoption is growing faster than strategy, governance, and enterprise-wide scaling.
Open sources: McKinsey. :::
History and institutional context
::: {.source-note} 14. Paul A. David, “The Dynamo and the Computer,” 1990. The historical framework is that electricity created its larger effect not when a factory merely changed its motor, but when the floor plan and production flow were reorganized around distributed electric drive.
Open sources: researchgate.net; aei.org. :::
::: {.source-note} 16. Erik Brynjolfsson, Daniel Rock, and Chad Syverson, “The Productivity J-Curve,” 2021. A major technology first requires investments in processes, skills, and organizational capital that conventional measurement may not fully capture; the measured return can appear later.
Open sources: American Economic Association; ideas.repec.org. :::
Books and frameworks
::: {.source-note} 17. Ajay Agrawal, Joshua Gans, and Avi Goldfarb, Power and Prediction, 2022. Their framework distinguishes point, application, and system solutions and explains why a point use of AI rarely changes the economics of an entire process.
Open sources: avigoldfarb.com; shortform.com. :::
Author cases
::: {.source-note} 22. This is the author's contract-review case from Book 1. AI did not act as an autonomous lawyer or merely as a comment-writing assistant. It served as internal memory and changed the route taken by routine questions before external legal review. Final legal judgment remained with a person. :::
Synthesis notes
::: {.source-note} 20. MIT NANDA provides a signal about pilots without measured financial impact; Deloitte describes the distance between technology adoption and governance readiness; Brynjolfsson, Rock, and Syverson provide the productivity J-curve; Agrawal, Gans, and Goldfarb distinguish point, application, and system solutions. Together, they support the chapter's synthesis: meaningful AI value requires changes to both process and management. :::
::: {.source-note} 23. Solow and David explain why the impact of a major technology can appear after a delay. Brynjolfsson, Rock, and Syverson describe a related measurement and investment pattern as the productivity J-curve: organizations first rebuild processes, skills, and organizational capital; the measured return may follow later.
Open sources: Paul A. David — The Dynamo and the Computer; NBER — The Productivity J-Curve. :::
Glossary notes
::: {.source-note} 10. Margin is the difference between a selling price and the costs directly associated with that sale. In rough terms, it is what remains after the product, delivery, and other direct costs. See the Reader Glossary. :::
::: {.source-note} 11. P&L means profit and loss: revenue minus expenses. It is where a business can see whether it made more money, not merely whether a task became faster. See the Reader Glossary. :::
::: {.source-note} 15. The J-curve describes a pattern in which investment and disruption come first and growth follows later—if the process has in fact been redesigned. See the Reader Glossary. :::
::: {.source-note} 18. Output is the artifact produced. Outcome is the useful result the work was meant to create. See the Reader Glossary. :::
::: {.source-note} 19. AI-native describes a process or business rebuilt around AI, rather than an old arrangement decorated with AI. See the Reader Glossary. :::
::: {.source-note} 21. A reliability threshold is the point at which an AI-enabled workflow has become dependable enough in a specific task to justify changing the process around it. See the Reader Glossary. :::
Chapter 2. From Output to Outcome
Quotes and epigraphs
::: {.source-note} 24. Peter F. Drucker, Management: Tasks, Responsibilities, Practices (1974). Drucker writes, “Efficiency is concerned with doing things right. Effectiveness is doing the right things.” His 1963 Harvard Business Review article “Managing for Business Effectiveness” develops the same distinction.
Open sources: Harvard Business Review — Managing for Business Effectiveness. :::
Research, reports, and data
::: {.source-note} 26. Microsoft, 2026 Work Trend Index. The study surveyed 20,000 people who use AI in ten countries. Microsoft calls the gap the Transformation Paradox: employees are ready to reinvent work while organizational metrics, incentives, and norms reinforce the old way.
Open sources: Microsoft — Work Trend Index 2026. :::
::: {.source-note} 28. Deloitte, State of AI in the Enterprise 2026. The relevant signal is the gap between plans for autonomous agents and mature guardrails around them.
Open sources: Deloitte — State of AI in the Enterprise 2026. :::
::: {.source-note} 29. McKinsey & Company / QuantumBlack, The State of AI (2025) and AI trust research (2026). This chapter uses the work as context: scaling is growing, but governance and strategy lag.
Open sources: McKinsey / QuantumBlack — The state of AI in 2025. :::
::: {.source-note} 30. Gartner, 2025 forecast: more than 40% of agentic-AI projects may be canceled by the end of 2027 because of escalating costs, unclear business value, or inadequate risk controls. This is a vendor forecast, not an observed failure rate.
Open sources: Source 1: gartner.com. :::
::: {.source-note} 31. OpenAI, GDPval, 2025. The full benchmark contains 1,320 real-world deliverable tasks across 44 occupations. This chapter uses it as a benchmark of tasks and artifacts, not as a field measurement of business outcome.
Open sources: OpenAI — GDPval. :::
::: {.source-note} 32. OpenAI's GDPval supports the roughly 100× model-only time and API-cost comparison and explicitly excludes human oversight, iteration, and integration. The human-plus-AI comparison in the prose comes from Shakked Noy and Whitney Zhang, “Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence,” Science, 2023: average completion time fell 40% and evaluated quality rose 18% in midlevel professional writing tasks.
Open sources: OpenAI — GDPval. :::
::: {.source-note} 33. Erik Brynjolfsson, Danielle Li, and Lindsey R. Raymond, “Generative AI at Work,” Quarterly Journal of Economics, 2025. The study covered more than 5,000 customer-support agents and found an approximately 15% rise in issues resolved per hour, with the largest gains among less-skilled and less-experienced workers.
Open sources: Oxford Academic — Generative AI at Work; NBER — Working Paper 31161. :::
::: {.source-note} 41. U.S. Consumer Financial Protection Bureau, Circular 2022-03: a complex algorithm does not release a creditor from the duty to provide a specific and accurate reason for an adverse credit action, including refusal to increase a credit limit.
Open sources: Source 1: consumerfinance.gov. :::
History and institutional context
::: {.source-note} 42. Wells Fargo, 2018 disclosure: an error in a mortgage-modification eligibility tool led to incorrect denials and, in hundreds of cases, foreclosure. This is an example of the cost of an uncontrolled automated calculation, not a generative-AI case.
Open sources: Source 1: sec.gov; Source 2: schatz.senate.gov. :::
Books and frameworks
::: {.source-note} 34. Goodhart's law in Marilyn Strathern's 1997 formulation: when a measure becomes a target, it ceases to be a good measure.
Open sources: EconBiz — Goodhart, Problems of Monetary Management. :::
::: {.source-note} 36. Mik Kersten, Project to Product and the Flow Framework: measure the flow of value from idea to customer rather than resource activity.
Open sources: Simon & Schuster — Project to Product; Planview — Output to Outcome and outcome trees. :::
::: {.source-note} 37. Han Lee, Evaluation and Alignment. Verification must be designed into the process as a condition, not attached at the end.
Open sources: Manning — Evaluation and Alignment: The Seminal Papers. :::
::: {.source-note} 40. The Federal Reserve and OCC model-risk framework covers development and use, independent validation, monitoring, governance, and controls. The OCC's 2026 revision explicitly excludes generative and agentic AI from the guidance's formal scope, while stating that general risk-management and governance practices should still guide tools outside it.
Open sources: Source 1: federalreserve.gov; Source 2: occ.treas.gov. :::
::: {.source-note} 44. The World Economic Forum and work by Daron Acemoglu and coauthors on pro-worker AI treat the future of work as a design choice rather than a predetermined technological outcome. :::
::: {.source-note} 45. Andrej Karpathy, “Software 3.0,” 2025. As generation becomes cheap, verification becomes a bottleneck; narrow tasks, checkable steps, and a short leash become more valuable.
Open sources: Latent Space — transcript of Karpathy's Software 3.0 talk. :::
Synthesis notes
::: {.source-note} 39. GDPval shows the distance between producing a deliverable and putting it to use; Microsoft and Deloitte show the distance between individual readiness and organizational design; Mik Kersten contributes the value-flow metric; Han Lee treats evaluation as a condition of a mature process. Together they support the chapter's synthesis: more produced material does not by itself create more value. :::
Glossary notes
::: {.source-note} 25. Output is the artifact produced. Outcome is the useful result the work was meant to create. See the Reader Glossary. :::
::: {.source-note} 35. Value flow means how much value actually reaches the customer from an initial idea, and how long that journey takes, rather than how much material was produced. See the Reader Glossary. :::
::: {.source-note} 38. Evaluation is regular, built-in verification of AI output: a method defined in advance for deciding whether the result can be trusted. See the Reader Glossary. :::
::: {.source-note} 43. A human in the loop is the critical point at which a person, rather than AI, approves, signs, or stops a decision. See the Reader Glossary. :::
Page notes
::: {.source-note} 27. A prompt is an instruction or request given to AI. A good prompt can accelerate one step. A process is the full route, including verification and accountability. See the Reader Glossary. :::
Chapter 3. A Team with Agents
Quotes and epigraphs
::: {.source-note} 46. Andy Grove, High Output Management (1983). Grove defines managerial output as the output of the organizational units under a manager's supervision or influence.
Open sources: Penguin Random House — High Output Management. :::
::: {.source-note} 59. “The Buck Stops Here” appeared on a sign on U.S. President Harry S. Truman's desk. In his 1953 farewell address, Truman expanded the idea that the person making a decision cannot pass its responsibility to someone else.
Open sources: trumanlibrary.gov. :::
Research, reports, and data
::: {.source-note} 55. Anthropic, “Building Effective AI Agents” (December 19, 2024), and “How We Built Our Multi-Agent Research System” (June 13, 2025). The first describes evaluator–optimizer and orchestrator–workers workflows and recommends adding complexity only when it improves outcomes. The second documents an orchestrator–worker production architecture. “Persistent roles” is the author's organizational pattern, not an Anthropic category.
Open sources: claude.com. :::
::: {.source-note} 56. Gartner, “Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027,” June 25, 2025. The release attributes the forecast to escalating costs, unclear business value, and inadequate risk controls; it also defines agent washing and estimates that only about 130 of thousands of claimed agentic vendors have substantive agentic technology. This is a forecast, not an observed failure rate.
Open sources: Gartner — over 40% of agentic AI projects may be canceled by 2027. :::
::: {.source-note} 60. Jessica Lyons, “Why Amazon Hates ‘Human-in-the-Loop’ AI Governance,” The Register, June 20, 2026, reporting a direct interview with Eric Brandwine, vice president and distinguished engineer at Amazon Security. Brandwine argues that repeated approval decisions degrade human attention and advocates end-to-end accountability, distinct agent identities, permissions, and auditability. This is a practitioner counterpoint, not an independent controlled study.
Open sources: theregister.com; thenextweb.com. :::
Books and frameworks
::: {.source-note} 54. “Productive people do not automatically create a productive organization” is the author's management proposition. Without common management, multiple agents create coordination costs, so each needs an owner, boundaries, and a shared control environment.
Open sources: Source 1: vikulin.ai. :::
::: {.source-note} 58. Paul R. Daugherty and H. James Wilson, Human + Machine (2018). The authors use “the missing middle” and “fusion skills” for the space in which people and machines strengthen one another. :::
Markets, companies, and products
::: {.source-note} 48. Public agent catalogs and repositories contain hundreds of templates across functions such as finance, support, legal work, recruiting, and analytics. “Hundreds” is an order-of-magnitude signal, not evidence of mass production deployment. :::
::: {.source-note} 51. Anthropic, “Introducing Claude Tag,” June 23, 2026. The official announcement describes one Claude in a selected Slack channel, shared visibility, channel context, permissioned access to other sources, and connected tools. It does not by itself prove broad enterprise adoption.
Open sources: Source 1: anthropic.com. :::
::: {.source-note} 62. Microsoft, “Least Privilege for AI Agents (Agentic Identities + RBAC),” Microsoft Learn, July 2026. The official guidance covers excessive permissions, prompt injection through over-broad tools, weak audit trails, incomplete revocation, dedicated identities, named owners, lifecycle controls, logging, and tested disablement. :::
Concepts used across the books
::: {.source-note} 50. The harness is the author's organizing frame. The result depends not only on the model but on the system around it: context, tools, memory, permissions, and verification.
Open sources: Source 1: latent.space; OpenAI — harness engineering. :::
::: {.source-note} 53. The author's AI-native Operating Model is Work → Agents → Context → Tools → Control → Evidence. The Agent Contract defines duties, boundaries, and the person who owns the result. :::
Glossary notes
::: {.source-note} 47. An AI agent is a program that does more than answer. It can take actions: search, calculate, write, send, and use other systems. See the Reader Glossary. :::
::: {.source-note} 49. A harness is everything around the model that makes it useful: context, tools, memory, verification, and constraints. The model is the engine; the harness is the rest of the car. See the Reader Glossary. :::
::: {.source-note} 52. An agent contract is a short, one-page description of an agent's role, access, boundaries, escalation, verification, and owner. It turns another bot into a manageable participant. See the Reader Glossary. :::
::: {.source-note} 57. Fusion skills are capabilities at the boundary between person and machine: frame a task for AI, check machine work, handle exceptions, and retain judgment and responsibility. See the Reader Glossary. :::
::: {.source-note} 61. Least privilege means giving an agent only the access required for its task: closed by default, with only the necessary permissions opened. See the Reader Glossary. :::
Chapter 4. A One-Person Company
Quotes and epigraphs
::: {.source-note} 63. Sam Altman, interview with Alexis Ohanian at a JPMorgan investor event, January 2024. The established English wording describes a private betting pool about the first year in which a one-person company reaches a billion-dollar valuation. In this chapter it is a forecast the author tests against the distinction between a function and a firm.
Open sources: every.to — the one person billion dollar company. :::
::: {.source-note} 73. Naval Ravikant, “Product and Media Are New Leverage,” from the How to Get Rich series (2018–2019). Ravikant describes code and media as permissionless leverage. The chapter extends the mechanism to agents without claiming that he wrote about current agent systems.
Open sources: x.com; nav.al. :::
Research, reports, and data
::: {.source-note} 67. Polsia's direct founder interview and Listen Labs, “Our Intern Built the First Zero-Person Company” (2026). Listen Labs reports a closed-loop agent experiment with 2,000 interviews, 100 concepts, and more than 400 paying users, while also disclosing a loss, human identity and credential steps, and weak research memory. These are vendor and founder experiments, not proof of a fully autonomous firm.
Open sources: Polsia — AI That Runs Your Company; Paperclip — Product Definition; Felix Craft — AI That Works. :::
Books and frameworks
::: {.source-note} 69. The direct Polsia founder interview describes one person directing a system of specialized research, growth, product, engineering, QA, and deployment agents. “One person is a system” remains the author's organizing proposition, not a claim that Polsia proves a general market rule.
Open sources: Microsoft Asia — One Person Is a System. :::
::: {.source-note} 75. Tiago Forte, “Building a Second Brain: The Definitive Introductory Guide.” Forte defines a Second Brain as an external, centralized digital repository for what a person learns and the sources behind it.
Open sources: litres.ru. :::
::: {.source-note} 77. Melissa Valentine and Michael S. Bernstein, Flash Teams: Leading the Future of AI-Enhanced, On-Demand Work (MIT Press, 2025), building on their research into dynamically assembled expert teams. The chapter extends the organizational pattern to mixed teams of agents and people.
Open sources: mitpress.mit.edu. :::
::: {.source-note} 79. Sharon Gai, How to Do More with Less: Future-Proofing Yourself in an AI-Driven Economy (Wiley, 2026). The book supports the chapter's qualitative shift in how individuals and lean teams use AI; this chapter does not import a numerical effect from it.
Open sources: Wiley — How to Do More with Less. :::
Markets, companies, and products
::: {.source-note} 66. Kevin Rose, direct interview with Ben Cera, solo founder of Polsia, “The Solopreneur Revolution Is Here,” March 31, 2026. The interview describes market research, landing pages, Meta ads, cold outreach, a night cycle, scaling failures, support agents, and a one-person operating model. All scale and business claims remain founder-reported, not independently audited. :::
::: {.source-note} 70. Klarna, “AI Helps Klarna Cut Marketing Agency Spend by 25% and Run More Campaigns,” May 28, 2024, plus CEO Sebastian Siemiatkowski's 2025 statements on restoring human service after cost-first automation reduced quality. The marketing figures are company-reported; the reversal is a leadership account, not an independent causal audit.
Open sources: klarna.com; customerexperiencedive.com. :::
::: {.source-note} 78. Listen Labs, “Our Intern Built the First Zero-Person Company” (2026). The company reports 2,000 interviews over two weeks, 100 concepts tested through parallel subagents, more than 400 paying users, $1,293 in revenue, and $2,000 in ad spend. It is a disclosed company experiment, not an independent study. :::
::: {.source-note} 80. The direct Polsia founder interview documents specialized agents for research, growth, product, engineering, QA, and deployment, with the founder retaining strategy, taste, and ambiguous decisions. “Conductor” remains the author's management metaphor.
Open sources: Anthropic — Claude for Small Business. :::
::: {.source-note} 81. The direct Polsia interview describes Ben Cera building alone with AI for sixteen hours a day and discusses scaling pain and infrastructure failures. The chapter uses it as a warning, not a model to imitate. :::
::: {.source-note} 83. Direct Polsia accounts describe a repair loop: product-management triage, engineering fix, QA verification, and deployment, plus scaling failures and support agents built to repair them. This is a founder-reported engineering pattern, not an independent reliability audit. :::
Author cases
::: {.source-note} 64. The author's personal system triages email, identifies important messages, monitors deals, and assembles a morning picture from dozens of sources. It is used as the smallest layer of the “supervised function” model. :::
::: {.source-note} 65. Author case from Book 1, Chapter 10: the initial design work for complex corporate B2B deals, previously performed by analysts and architects, was carried largely by the author and one colleague with intensive AI support. :::
Glossary notes
::: {.source-note} 68. An orchestrator directs agents and combines their results instead of performing every task by hand: the conductor, not the performer. See the Reader Glossary. :::
::: {.source-note} 72. Leverage is a multiplier of effort: something that turns one action into a much larger result. Capital and labor are leverage; code and agents can be leverage too. See the Reader Glossary. :::
::: {.source-note} 74. A harness is everything around the model that makes it useful: context, tools, memory, verification, and constraints. The model is the engine; the harness is the rest of the car. See the Reader Glossary. :::
::: {.source-note} 76. A flash team is a temporary configuration of agents and, when needed, people for a specific goal: assemble, deliver, disband. See the Reader Glossary. :::
::: {.source-note} 82. A human in the loop is a critical process point where a person confirms, stops, or signs a decision rather than an agent. In a one-person company, it is the final support for accountability. See the Reader Glossary. :::
Page notes
::: {.source-note} 71. A CRM system records customers and deals: who the customer is, the current stage, promises made, and the next contact. See the Reader Glossary. :::
Chapter 5. Small and Midsize Businesses: AI as an Affordable Staff
Quotes and epigraphs
::: {.source-note} 84. Theodore Levitt, The Marketing Mode: Pathways to Corporate Growth (1969), where he credits the line to the advertising man Leo McGivena. The shortened “people want a hole, not a drill” version now circulates under Levitt's name; the wording here is his. The metaphor supports the chapter's practical rule: begin AI adoption with the customer's pain and required outcome.
Open sources: esade.edu — competing customer outcomes; library.hbs.edu — what customers want from your products. :::
Research, reports, and data
::: {.source-note} 95. Nextiva, “The Secret Cost of Missed Calls for Businesses” (2025), estimates that small businesses can miss about one quarter of calls during after-hours, weekends, busy periods, and staff breaks. It is a commercial illustrative estimate, not a neutral market-wide statistic.
Open sources: nextiva.com — whats the cost missed calls. :::
Books and frameworks
::: {.source-note} 86. The three value levers—new money, lower cost, and customer attention or trust—are the author's framework. Felix Oberholzer-Gee's Better, Simpler Strategy (2021) provides the broader economic foundation: value is created by increasing willingness to pay or reducing the cost required to serve. :::
::: {.source-note} 87. Julien Bek, “Services: The New Software,” Sequoia Capital, March 5, 2026. The essay distinguishes a copilot that sells a tool from an autopilot that sells completed work, and describes already-outsourced work as a practical entry point because scope and budget already exist.
Open sources: sequoiacap.com — services the new software. :::
::: {.source-note} 89. Sequoia's “Services: The New Software” argues that the budget for work is much larger than the budget for the tools used to perform it. An autopilot therefore competes on the cost of completed work, not only on software subscription price.
Open sources: sequoiacap.com — services the new software. :::
::: {.source-note} 94. David Yang's “three crystals” frame is illness, pill, and delivery: know the customer's pain, build the product that treats it, and reach the customer. The chapter uses it as an entrepreneurial operating metaphor.
Open sources: Source 1: forbes.ru. :::
Markets, companies, and products
::: {.source-note} 90. Tobi Lütke, Shopify memo, “AI Usage Is Now a Baseline Expectation,” April 2025. Before asking for additional headcount and resources, teams were expected to explain why the work could not be done with AI. This is a company policy, not evidence of a measured employment effect.
Open sources: theverge.com; businessinsider.com; firstround.com. :::
::: {.source-note} 91. Banco Santander, “Santander's Data & AI-First Strategy Accelerates Through OpenAI Collaboration,” 2025, and subsequent 2026 updates. The bank announced mandatory AI training for all employees beginning in 2026, including responsible-AI training, and later broadened AI access. This is a capability and HR-policy signal, not proof of a productivity result for every role.
Open sources: Santander — Data & AI-First Strategy. :::
::: {.source-note} 92. Jensen Huang's 2025 NVIDIA employee remarks urged broad AI use and rejected managers telling employees to use less of it. In later interviews he distinguished a job's purpose from the tasks and tools inside it. The chapter uses this as a management counterweight, not as a forecast that jobs cannot disappear.
Open sources: Fortune — Jensen Huang on using AI for every possible task; Business Insider — Huang on AI layoffs as a lazy narrative; Fast Company — Huang calls AI a lazy excuse for layoffs. :::
::: {.source-note} 93. Best Buy's official 2024 material describes a generative-AI self-service assistant for troubleshooting, delivery and scheduling changes, subscriptions, and memberships, plus tools that summarize calls and assist live customer-care employees. Klarna's public customer-service case and field research on AI assistance support the same division between safe first-line automation and prepared human work. No performance figure from one company is attributed to another.
Open sources: Best Buy — generative AI for customer support; Accenture — Best Buy humanizes customer experience with GenAI; OpenAI — Klarna customer-service case; CX Dive — Klarna reinvests in human customer service; Harvard Business School — when chatbots help people be more human. :::
::: {.source-note} 99. This is the chapter's macroeconomic frame: agents provide affordable staff, while durable advantage moves toward customer access, first-party data, and relationships. No universal platform-fee percentage is asserted. :::
Glossary notes
::: {.source-note} 88. An autopilot, or service-as-software, sells a completed function within narrow boundaries rather than a tool: a prepared application, a sorted request flow, or a contract draft for professional review. See the Reader Glossary. :::
::: {.source-note} 97. A platform's take rate is the share of a transaction it keeps for connecting the business with a customer. The deeper the dependency on one platform, the more margin and customer access the platform may control. See the Reader Glossary. :::
::: {.source-note} 98. A moat is a durable advantage that makes a customer harder to lose and a competitor's result harder to copy. Here, the moat is direct customer contact, first-party data, and a relationship the platform does not own. See the Reader Glossary. :::
::: {.source-note} 100. A human in the loop is a point where a person confirms, stops, or signs a decision. For an SMB, this is not bureaucracy. It protects customer trust and the owner's money. See the Reader Glossary. :::
Page notes
::: {.source-note} 85. SMB means small or midsize business: a company without a corporate budget or a separate department for every function. See the Reader Glossary. :::
::: {.source-note} 96. KYC, or know your customer, is the identity-verification process performed by banks and payment providers. A small business can use that infrastructure without maintaining its own compliance department. See the Reader Glossary. :::
Chapter 6. When the Customer Is Not Human
Quotes and epigraphs
::: {.source-note} 101. The line about a brand being what people say when you are not in the room is widely attributed to Jeff Bezos, but its exact original source has not been established. The wording remains subject to quotation clearance. :::
::: {.source-note} 112. In a reported May 2023 Goldman Sachs and SV Angel AI event, Bill Gates argued that the winner in personal digital agents could become the gateway that replaces many direct visits to search and shopping sites. It is an industry forecast, not a measured market result.
Open sources: Source 1: cnbc.com. :::
Research, reports, and data
::: {.source-note} 103. Gartner forecast in 2024 that traditional search-engine volume could fall 25% by 2026 as AI chatbots and virtual agents gain share. Generative-engine-optimization research separately tests how source material becomes visible inside generated answers. Forecast and research are evidence of direction, not a guaranteed traffic outcome for every business.
Open sources: Source 1: gartner.com; arXiv — GEO. :::
::: {.source-note} 104. Adobe Digital Insights reported rapid growth in generative-AI referrals to U.S. retail sites. In July 2025, those visitors were still 23% less likely to convert than non-AI traffic, although the gap had narrowed. Results vary by period, market, and category; the chapter does not claim a universal conversion advantage.
Open sources: Source 1: business.adobe.com. :::
::: {.source-note} 115. Pranjal Aggarwal and coauthors, “GEO: Generative Engine Optimization” (2023–2024). The study tested methods such as citations, statistics, quotations, and authoritative language and found that some increased source visibility in generated answers. Effects varied by method and domain.
Open sources: insightpartners.com — agent led growth. :::
Markets, companies, and products
::: {.source-note} 107. Official product material from Perplexity, OpenAI, Opera, and Google documents the move from browsing assistance toward agents that can act across pages. OpenAI describes Atlas agent mode as a preview that may make mistakes and pauses on certain sensitive sites; Google's 2025 Chrome announcement described agentic browsing as an upcoming capability. Product scope and availability change quickly. :::
::: {.source-note} 109. The Model Context Protocol (MCP) is an open way to connect an AI application to data and tools. Other protocols address agent-to-agent tasks, commerce, identity, authorization, or payment. They solve different layers and should not be treated as one standard.
Open sources: Anthropic — Model Context Protocol; Stripe — Agentic Commerce Protocol; Google Cloud — Agent Payments Protocol; Google for Developers — Universal Commerce Protocol; Source 5: yandex.ru; Source 6: banks.cnews.ru. :::
::: {.source-note} 110. Official 2025 announcements describe Mastercard Agent Pay, Visa Intelligent Commerce, OpenAI/Stripe checkout, and Google's Agent Payments Protocol. Their common elements include authenticated agents or credentials, user controls, permissions or mandates, and transaction records. Details and availability continue to change.
Open sources: stripe.com; cloud.google.com; digitalcommerce360.com; OpenAI. :::
::: {.source-note} 114. Amazon's 2025 Form 10-K reports $68.635 billion in advertising-services net sales. The number establishes the scale of paid visibility in one global marketplace; it is not a forecast that agent commerce will remove that revenue or a proxy for every platform.
Open sources: Amazon — 2025 Form 10-K. :::
::: {.source-note} 117. Agent-led growth is an emerging practitioner frame for B2B discovery before a sales conversation. The chapter uses the mechanism—machine-readable offers are easier for agents to include in a shortlist—without importing unstable survey percentages as permanent facts.
Open sources: Insight Partners — Agent-led growth. :::
::: {.source-note} 119. Resend and Supabase illustrate products with strong documentation, transparent self-service paths, and agent-friendly developer interfaces. Their cases are examples, not an independent head-to-head study, and the chapter imports no promotional growth figure.
Open sources: Insight Partners — Agent-led growth. :::
::: {.source-note} 120. WebMCP research prototypes test the benefit of exposing structured actions to browser agents instead of forcing them to infer every operation from a visual interface. The chapter uses this to illustrate an interface tax, not to claim universal cost or success rates.
Open sources: arXiv — webMCP. :::
::: {.source-note} 124. Industry leaders including Box CEO Aaron Levie have described agent work as a possible new business expense category. This is an industry view about the cost structure of work, not a measured effect or a universal accounting rule.
Open sources: Source 1: linkedin.com. :::
::: {.source-note} 125. Mastercard's official 2026 releases describe authenticated agentic transactions using Agent Pay and B2B agentic-commerce work around sourcing and payment. These are early program examples, not proof of a fully autonomous or universal agent-to-agent market.
Open sources: Mastercard — authenticated Agent Pay transactions. :::
::: {.source-note} 126. The toll-road conclusion is the author's fintech analysis: identity, authorization, payment, settlement, and model execution each have an economic owner. No universal agentic-commerce commission or fee is asserted.
Open sources: Mastercard — Agent Pay. :::
Concepts used across the books
::: {.source-note} 105. The three value levers—new money, lower cost, and customer attention or trust—are the author's framework introduced earlier in Book 2. This chapter focuses on the third: attention and trust must increasingly be earned from machines as well as people. :::
Glossary notes
::: {.source-note} 102. An AI agent is a program that can do more than answer: it can search, compare, fill out a form, or sometimes initiate a payment under defined authority. See the Reader Glossary. :::
::: {.source-note} 106. An agentic browser can navigate sites and take actions for a user instead of only displaying pages. Sensible implementations expose their work and pause before sensitive or irreversible steps. See the Reader Glossary. :::
::: {.source-note} 108. Prompt injection is an attack in which a site, document, or hidden text tries to give the agent instructions that conflict with its owner's intent. OpenAI identifies hidden malicious instructions on webpages or in email as an ongoing risk that safeguards cannot eliminate completely. See the Reader Glossary. :::
::: {.source-note} 111. Agent-led growth is the chapter's term for growth that occurs because a buyer's agent can find, evaluate, and select the product. Promotion shifts from persuasion aimed only at people toward evidence and interfaces a machine can use. See the Reader Glossary. :::
::: {.source-note} 113. Retail media is advertising sold by a commerce platform, including sponsored search positions and promoted product listings. Agents may respond differently from people, but paid or partner data can still influence what an agent sees. See the Reader Glossary. :::
::: {.source-note} 116. Machine trust is the chapter's working term for an offer an agent can recommend with less risk because its data, prices, conditions, reviews, and outside evidence are clear and consistent. It is a practical heuristic, not a settled technical standard. See the Reader Glossary. :::
::: {.source-note} 118. Token-to-value is an informal way to describe how much machine effort is required to understand and use an offer. It is analogous to time-to-value for a person. The term is an operating heuristic, not a standardized financial metric. See the Reader Glossary. :::
::: {.source-note} 121. Agent Experience (AX) means designing the information and transaction path for a machine user: explicit prices, terms, availability, evidence, and executable actions. It complements rather than replaces human user experience. See the Reader Glossary. :::
::: {.source-note} 122. Trust rails are the payment, identity, authorization, audit, and accountability systems needed for agents to transact safely. Whoever controls the rails can capture part of the value moving across them. See the Reader Glossary. :::
::: {.source-note} 123. KYC, or know your customer, is identity verification performed by banks and payment services. Agent commerce adds the need to authenticate the agent and connect its authority to a real person or organization. See the Reader Glossary. :::
Chapter 7. Children: What to Teach and Where to Guide Them
Quotes and epigraphs
::: {.source-note} 127. “Help me do it myself” is strongly associated with Montessori pedagogy, although the exact primary wording is difficult to establish. The epigraph remains subject to quotation clearance.
Open sources: myschoolticino.ch — help myself; yourtherapysource.com — help myself. :::
::: {.source-note} 139. Geoffrey Hinton's June 2025 Diary of a CEO interview used plumbing to illustrate his view that routine cognitive work may automate faster than skilled physical work in unpredictable environments. It is a forecast and career opinion, not labor-market proof. :::
::: {.source-note} 140. Jensen Huang has argued that AI infrastructure will require large numbers of electricians, plumbers, carpenters, and builders. The chapter's conclusion is not “everyone enter a trade,” but hands + mind + AI. :::
::: {.source-note} 143. Demis Hassabis described learning how to learn as a key skill for the next generation in 2025. This is an expert view, not a statistical prediction.
Open sources: greekcitytimes.com — demis hassabis learning how learn skills; hpcwire.com — googles nobel winning scientist says learning. :::
Research, reports, and data
::: {.source-note} 128. Erik Brynjolfsson, Bharat Chandar, and Ruyu Chen, Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence, Stanford Digital Economy Lab. The November 2025 revision, using ADP payroll records through September 2025, reports a 16% relative employment decline for workers ages 22–25 in the most AI-exposed occupations; the August 26, 2025 version reported 13%. It is an empirical association within the study design, not an explanation of the whole youth labor market.
Open sources: digitaleconomy.stanford.edu — canaries the coal mine six facts. :::
::: {.source-note} 129. World Economic Forum, Future of Jobs Report 2025. Employers surveyed expected 92 million displaced roles and 170 million new roles by 2030, and about 39% of key skills to change. These are employer expectations, not guaranteed outcomes.
Open sources: World Economic Forum — Future of Jobs Report 2025. :::
::: {.source-note} 131. Brookings Institution, A New Direction for Students in an AI World: Prosper, Prepare, Protect (2026). The report recommends cautious integration with safeguards rather than simple prohibition and documents benefits as well as risks in different educational settings.
Open sources: brookings.edu. :::
::: {.source-note} 132. Michael Gerlich's study reports an association between more frequent cognitive offloading to AI and lower critical-thinking scores. It is correlational and does not establish that AI use caused the lower scores.
Open sources: MDPI / Societies — «AI Tools in Society»; Source 2: ssrn.com. :::
::: {.source-note} 133. Oxford University Press, Teaching the AI-Native Generation (2025): about eight in ten surveyed UK students ages 13–18 used digital AI tools for schoolwork, 32% reported difficulty judging whether AI content was true, and nearly half wanted support from teachers.
Open sources: corp.oup.com — teaching the native generation; global.oup.com — oup native generation research report. :::
::: {.source-note} 135. Common Sense Media, Talk, Trust, and Trade-Offs (July 16, 2025), surveyed 1,060 U.S. teens ages 13–17. It reported 72% had used AI companions, 52% were regular users, about one third had chosen an AI companion instead of a person for a serious conversation, and 24% had shared personal information.
Open sources: Common Sense Media — «Talk, Trust, and Trade-Offs». :::
::: {.source-note} 136. Public reporting says Character.AI and Google agreed to settle several family lawsuits alleging harm to teenagers. Undisclosed settlement terms do not establish liability, causation, or a general effect of AI companions.
Open sources: cnbc.com. :::
::: {.source-note} 137. Italy's data-protection authority, Garante, fined Luka Inc., operator of Replika, €5 million in May 2025 for data-processing violations that included inadequate age-verification measures. The company challenged the decision.
Open sources: Source 1: garanteprivacy.it; Source 2: dataguidance.com. :::
Books and frameworks
::: {.source-note} 130. Matt Beane, The Skill Code (Harper Business, 2024). Beane frames capability development through challenge, complexity, and connection and uses robotic surgery to show how technology can improve work while weakening the novice's learning path.
Open sources: Google Books — The Skill Code; Harvard Business Review — Learning to Work with Intelligent Machines. :::
::: {.source-note} 141. Vivienne Ming's work, including Robot-Proof, emphasizes questions, human potential, and the capacity to develop people as answers become cheaper. The chapter's dinner-table exercises are the author's application. :::
::: {.source-note} 142. Daniel Susskind, What Should My Children Do? How to Flourish in the Age of AI (Penguin, announced for September 10, 2026; ISBN 9781837314591). Before publication, this chapter relies only on the publisher's description, not the unreleased text.
Open sources: Penguin — What Should My Children Do?. :::
Markets, companies, and products
::: {.source-note} 134. Google officially offered temporary student access to AI Premium; media reported a temporary ChatGPT Plus student offer, while universities adopted ChatGPT Edu and Gemini for Education. The chapter treats this as competition for early habit, not a claim about permanent pricing or intent inside every organization. :::
Glossary notes
::: {.source-note} 138. Shadow learning is Beane's term for informal workarounds used when the official training path is weak: observing recorded work, using simulations, and seeking real practice outside the standard route. See the Reader Glossary. :::
::: {.source-note} 144. A T-shaped profile combines broad literacy across fields with depth in one area. The horizontal stroke represents range; the vertical stroke represents the ability to go deep and produce serious work. See the Reader Glossary. :::
Chapter 8. The New Economy of Usefulness: Labor Markets and Platforms
Quotes and epigraphs
::: {.source-note} 145. John Maynard Keynes, “Economic Possibilities for Our Grandchildren” (1930). The ellipses mark two elisions: Keynes introduces the disease as one “of which some readers may not yet have heard the name,” and a sentence defining the term stands between the two halves quoted here. He coined technological unemployment for labor-saving discoveries outrunning our ability to find new uses for labor.
Open sources: marxists.org — our grandchildren. :::
Research, reports, and data
::: {.source-note} 146. World Economic Forum, Future of Jobs Report 2025: employers expected 170 million new roles and 92 million displaced by 2030, a net increase of 78 million. It is a survey forecast, not a promise.
Open sources: World Economic Forum — the future jobs report; World Economic Forum — future jobs report million new job 2. :::
::: {.source-note} 147. Erik Brynjolfsson, Bharat Chandar, and Ruyu Chen, Canaries in the Coal Mine?, Stanford Digital Economy Lab. The November 2025 revision, using ADP payroll records through September 2025, reports a 16% relative employment decline among workers ages 22–25 in the most AI-exposed occupations; the August 26, 2025 version reported 13%.
Open sources: digitaleconomy.stanford.edu — canaries the coal mine six facts; siepr.stanford.edu — canaries coal mine six facts about. :::
::: {.source-note} 148. A Stanford/SIEPR update using newer data reported that the young-worker decline in AI-exposed occupations had persisted and deepened. It remains an early empirical signal, not a final verdict on the labor market.
Open sources: digitaleconomy.stanford.edu — canaries the coal mine six facts; siepr.stanford.edu — canaries coal mine six facts about. :::
::: {.source-note} 150. JobCloud, jobs.ch AI Report 2026, analyzed 7.3 million postings placed on jobs.ch, jobup.ch, and JobScout24.ch between 2019 and 2025. The share aimed at career starters fell 32% in 2025 against the 2019–2022 average; within the roles the study classes as AI-exposed, the junior share fell 16% while the senior share rose 26%. It is a Swiss posting analysis, not a global measure or proof of AI causation.
Open sources: JobCloud — jobs.ch KI Report 2026; Source 2: swissinfo.ch. :::
::: {.source-note} 155. Official Chinese 2026 material on “AI + human resources and social security” describes applied employment-service scenarios. It supports one example of platforms and AI as employment infrastructure, not a universal law about all Chinese platforms.
Open sources: Source 1: nda.gov.cn; CyberPeace Institute — China AI and employment policy overview. :::
::: {.source-note} 158. Joint Chinese policy material on AI, human resources, and social security is one example of employment platforms developing beside public governance. It does not describe the entire Chinese platform economy.
Open sources: Source 1: nda.gov.cn. :::
::: {.source-note} 160. Skills England, Annual Skills Report 2026, reports that 27% of UK vacancies in 2024 were skills-shortage vacancies. The measure describes difficulty finding required skills, not a shortage in every occupation or experience level.
Open sources: Skills England — Annual Skills Report 2026. :::
::: {.source-note} 162. The chapter's UK window is a synthesis of two official datasets with different scopes. It supports the coexistence of higher-level skills shortages and a narrowing entrance; it does not claim AI is the sole cause.
Open sources: Skills England — Annual Skills Report 2026; UK Government — Entry-Level Hiring Snapshot. :::
History and institutional context
::: {.source-note} 153. Robert C. Allen's work documents the historical pattern known as Engels' Pause: productivity and output rose before real wages began catching up later. Periodization and estimates depend on the historical series used.
Open sources: nuff.ox.ac.uk. :::
Books and frameworks
::: {.source-note} 151. Cai Fang uses structural employment mismatch to describe old work disappearing faster than accessible new roles appear, with new roles often going to different people. The chapter uses the frame, not a universal quantitative model. :::
::: {.source-note} 154. Daron Acemoglu, David Autor, and Simon Johnson describe worker-enhancing AI as technology designed to increase the value of human skills rather than primarily substitute for labor. This is a normative design framework.
Open sources: NBER — Building Pro-Worker Artificial Intelligence. :::
::: {.source-note} 156. World Economic Forum, Four Futures for Jobs in the New Economy: AI and Talent in 2030 (January 7, 2026), presents scenarios rather than a forecast. The chapter uses the scenario in which AI intermediaries and platforms reshape access to work.
Open sources: World Economic Forum — Four Futures for Jobs in the New Economy; World Economic Forum — Four Futures report PDF. :::
::: {.source-note} 159. Wang Jiangping's 上善AI presents human-centered AI governance and the metaphor of guardrails rather than brakes. It is an author's policy framework, not statistical evidence.
Open sources: kpzg.people.com.cn — c404214. :::
::: {.source-note} 164. Kai-Fu Lee proposed a Social Investment Stipend in 2018 as payment for care, community service, and education rather than unconditional cash alone. It is an author's proposal, not an implemented national program.
Open sources: kaifulee.medium.com — how can coexist with humans bff091451442. :::
Markets, companies, and products
::: {.source-note} 149. SignalFire, State of Tech Talent Report 2025, reports from its platform data that recent-graduate hiring in Big Tech fell more than 50% from 2019 levels and more than 30% in startups. It is a private platform dataset, not a complete labor-market census.
Open sources: signalfire.com — signalfire state talent report; signalfire.com — signalfire state talent report 3; cnbc.com — entry level jobs hiring careers. :::
::: {.source-note} 157. Tobi Lütke, Shopify, “AI Usage Is Now a Baseline Expectation” (2025). The headcount test is a company policy, not independent evidence of an employment or productivity effect.
Open sources: Source 1: shopify.com; Source 2: theverge.com. :::
::: {.source-note} 161. UK government, A Snapshot of Entry-Level Hiring in the UK (April 2026), reports declining demand in 30 of 38 tracked entry-level occupations and distinguishes surplus entry-level skills from shortages that skew toward experience.
Open sources: UK Government — Entry-Level Hiring Snapshot. :::
Glossary notes
::: {.source-note} 152. Engels' Pause is the historical lag between productivity growth and workers' real wage growth. Technology creates more output before labor captures a comparable share. See the Reader Glossary. :::
::: {.source-note} 163. Universal basic income (UBI) is a recurring fixed payment made without employment, income, or work requirements. See the Reader Glossary. :::
Chapter 9. Physical AI and Who Owns the Rails
Quotes and epigraphs
::: {.source-note} 165. Vaclav Smil, Energy and Civilization (2017). The chapter uses energy as the universal physical requirement for action. Exact epigraph wording remains subject to quotation clearance.
Open sources: mitpress.mit.edu — energy and civilization. :::
Research, reports, and data
::: {.source-note} 166. Goldman Sachs Research gives a base-case sensitivity estimate of roughly $7.6 trillion in global AI-infrastructure investment during 2026–2031. It is a scenario estimate, not market consensus or a guaranteed build plan.
Open sources: Goldman Sachs — Tracking Trillions: AI Build-Out. :::
::: {.source-note} 168. Microsoft and Alphabet each planned roughly $175–190 billion of 2026 capex; Amazon reported about $132 billion in 2025 property/equipment cash capex. The text therefore uses the conservative group phrase “hundreds of billions a year.”
Open sources: Source 1: microsoft.com; Source 2: abc.xyz; Source 3: ir.aboutamazon.com. :::
::: {.source-note} 169. JLL projects nearly 100 gigawatts of new data-center capacity and up to $3 trillion of related investment during 2026–2030. It is an industry forecast, not a guaranteed construction schedule.
Open sources: jll.com — data center outlook. :::
::: {.source-note} 170. International Energy Agency, Energy and AI: data centers used about 415 TWh in 2024 and could use about 945 TWh in 2030 in the base case, with AI the main driver of growth.
Open sources: iea.org — executive summary. :::
::: {.source-note} 171. IEA and JLL identify grid connection and critical power equipment as data-center constraints. Lead times vary sharply by region and component; the chapter asserts no universal four-year wait.
Open sources: iea.org — executive summary. :::
::: {.source-note} 175. U.S. Bureau of Labor Statistics: electrician employment projected +9% during 2024–2034, about 81,000 openings annually, and median pay of $62,350 in May 2024.
Open sources: bls.gov — electricians. :::
::: {.source-note} 182. Morgan Stanley and Goldman Sachs publish widely different humanoid-robot forecasts using different horizons and methods. The chapter treats them as direction indicators, not a schedule.
Open sources: Source 1: morganstanley.com; Source 2: goldmansachs.com. :::
History and institutional context
::: {.source-note} 172. William Stanley Jevons, The Coal Question (1865), argued that greater efficiency in resource use can increase rather than reduce total consumption.
Open sources: oll.libertyfund.org — jevons the coal question. :::
::: {.source-note} 180. Moravec's paradox describes the asymmetry in which tasks easy for people can be difficult for machines while some abstract tasks difficult for people become tractable for computers.
Open sources: Google Books — Mind Children. :::
Books and frameworks
::: {.source-note} 179. Michiaki Tanaka describes physical AI as movement from screen output toward a perceive–act–verify loop in the physical world.
Open sources: publications.asahi.com. :::
::: {.source-note} 183. Chris Miller, Chip War, documents the concentration and vulnerability of the semiconductor supply chain. It supports the physical-fragility thesis, not a claim that only one firm controls every stage.
Open sources: Simon & Schuster — Chip War. :::
::: {.source-note} 184. Stephen Witt, The Thinking Machine (Viking, 2025), is used as a journalistic history of NVIDIA and the physical infrastructure behind generative AI.
Open sources: penguinrandomhouse.com — the thinking machine stephen witt. :::
::: {.source-note} 185. Karen Hao, Empire of AI (Penguin Press, 2025), documents labor, energy, water, and resource costs around AI development. The chapter uses the cost boundary without adopting every political conclusion.
Open sources: books.google.com — Empire. :::
::: {.source-note} 187. Yanis Varoufakis uses cloud rent inside his technofeudalism argument for payments captured through controlled digital infrastructure and market access. The chapter adopts the mechanism, not the full political thesis.
Open sources: mhpbooks.com — technofeudalism. :::
Markets, companies, and products
::: {.source-note} 173. Jensen Huang uses radiology to illustrate Jevons effects: cheaper analysis can create more scans and related work. It is an industry analogy from NVIDIA's CEO, not employment research. :::
::: {.source-note} 174. Microsoft estimates that the United States may need to train roughly 500,000 additional electricians over a decade for rising power and infrastructure demand. It is a corporate estimate.
Open sources: blogs.microsoft.com — winning the race. :::
::: {.source-note} 176. CNBC reported individual electricians at data-center construction sites earning $240,000–$280,000 with overtime, per diem, and premium rates. These are exceptional cases, not median earnings.
Open sources: CNBC — AI data-center buildout and skilled-trades pay. :::
::: {.source-note} 177. Jensen Huang's 2025–2026 remarks forecast rising demand for electricians, plumbers, carpenters, and other trades as AI infrastructure expands. This is an industry forecast. :::
::: {.source-note} 181. Unitree's official G1 base price was listed from approximately $13,500 as of July 20, 2026, before tax and delivery. It is the price of a robotics platform, not a ready autonomous employee. The RU body retained the earlier approximately $16,000 launch price.
Open sources: Source 1: unitree.com. :::
::: {.source-note} 186. Satya Nadella has argued that data centers need social license for energy use and must show measurable benefits to local communities. It is an industry position. :::
Glossary notes
::: {.source-note} 167. A hyperscaler builds and rents computing capacity at planetary scale: data centers, clouds, networks, and the surrounding infrastructure. See the Reader Glossary. :::
::: {.source-note} 178. Physical AI perceives and acts in the physical world through a robot, vehicle, machine, or another body rather than only producing an answer on a screen. See the Reader Glossary. :::
::: {.source-note} 188. Platform rent is recurring payment to an infrastructure owner for access to customers, tools, payments, data, or compute. See the Reader Glossary. :::
::: {.source-note} 189. Inference is running an already trained model on an input to calculate and return an output. Local inference runs that process on infrastructure under the user's or organization's control. See the Reader Glossary. :::
::: {.source-note} 190. Engels' Pause is the historical lag between productivity growth and workers' real-wage growth. See Chapter 8 and the Reader Glossary. :::
Chapter 10. The Future in 2030: Four Scenarios
Quotes and epigraphs
::: {.source-note} 191. Dennis Gabor, Inventing the Future (1963): “The future cannot be predicted, but futures can be invented.”
Open sources: quoteinvestigator.com — invent the future. :::
Research, reports, and data
::: {.source-note} 195. World Economic Forum, Four Futures for Jobs in the New Economy: AI and Talent in 2030. The WEF exercise uses AI development and workforce readiness as its axes; the matrix in this chapter is the author's own.
Open sources: World Economic Forum — here are four ways ais impact. :::
::: {.source-note} 202. Daron Acemoglu, David Autor, and Simon Johnson, “Building Pro-Worker Artificial Intelligence,” NBER Working Paper 34854 (2026). The authors distinguish five types of technological change; creating new tasks is clearly pro-worker.
Open sources: NBER — Building Pro-Worker Artificial Intelligence. :::
::: {.source-note} 203. ILO and NASK, global index of occupational exposure to generative AI. One in four workers is in an occupation with some exposure; task transformation is more likely than the complete automation of jobs.
Open sources: ILO — Generative AI and Jobs: A Refined Global Index. :::
::: {.source-note} 204. Stanford HAI, AI Index Report 2026. The chapter uses its broader finding that capabilities, investment, and adoption are advancing faster than transparency, evaluation, and governance, without relying on a single statistic.
Open sources: Stanford HAI — AI Index Report 2026. :::
History and institutional context
::: {.source-note} 192. Pierre Wack described Shell's scenario practice in Harvard Business Review in 1985. The 1973 oil shock is used as the classic illustration of the method.
Open sources: Harvard Business Review — scenarios uncharted waters ahead. :::
Books and frameworks
::: {.source-note} 194. Kai-Fu Lee and Chen Qiufan, AI 2041 (2021). The book appeared more than a year before ChatGPT's public launch. Its fictional scenes are followed by analytical commentary.
Open sources: Penguin Random House — AI 2041. :::
::: {.source-note} 200. The “Levels of AGI” framework uses performance, generality, and autonomy rather than a binary “AGI exists/does not exist” test.
Open sources: arXiv — 2311.02462. :::
Markets, companies, and products
::: {.source-note} 201. The chapter uses 2026 reporting on political resistance to data centers as an observation about public response, not as a statistical series.
Open sources: FT Alphaville — The AI backlash has already begun. :::
Glossary notes
::: {.source-note} 193. Scenario planning is a way to prepare for the future without predicting it: describe several plausible developments and test your plans in each. See the Reader Glossary. :::
::: {.source-note} 196. Multiplication World is this chapter's term for a scenario in which AI increases the value created by human judgment, responsibility, and customer relationships instead of removing the person from the work. See the Reader Glossary. :::
::: {.source-note} 197. Human Premium World, or Premium World, is this chapter's term for a scenario in which an abundance of machine generation increases the value of human trust, presence, signature, and responsibility. See the Reader Glossary. :::
::: {.source-note} 199. AGI, or artificial general intelligence, is a hypothetical AI that can handle a broad range of tasks at a human level rather than one narrow domain. See the Reader Glossary. :::
::: {.source-note} 206. No-regret moves are actions that pay off under every plausible future, so they can be made before uncertainty disappears. See the Reader Glossary. :::
Page notes
::: {.source-note} 198. The matrix is a map of tensions, not a strict prediction table. Human Premium World can become a broad norm beside Multiplication World or a premium segment layered over Fractured World. :::
::: {.source-note} 205. API stands for application programming interface: the rules through which one program, service, or agent communicates with another. Technical access can pass between systems; trust and relationships cannot. :::
Chapter 11. The Change Compass
Quotes and epigraphs
::: {.source-note} 207. Niccolò Machiavelli, The Prince, Chapter VI. The epigraph uses a public-domain English rendering of the passage about introducing a new order; final quotation clearance remains open. :::
Research, reports, and data
::: {.source-note} 211. Erik Brynjolfsson, Danielle Li, and Lindsey R. Raymond, “Generative AI at Work.” In customer support, the assistant produced the largest gains for less experienced workers; it was embedded in the work process.
Open sources: Oxford Academic — Generative AI at Work; NBER — Working Paper 31161. :::
::: {.source-note} 214. Deloitte, State of AI in the Enterprise 2026. The chapter uses company self-reporting: roughly one-fifth of respondents describe governance for autonomous agents as mature. :::
Books and frameworks
::: {.source-note} 215. Anthropic, “The Founder's Playbook” (2026). The chapter uses the principle that durable advantage comes from accumulated domain depth, data, and context—not access to the model alone. :::
Markets, companies, and products
::: {.source-note} 210. Salesforce and Heathrow, Hallie customer story and announcement (2025). The chapter uses it as a public illustration of choosing a narrow process, not as an independent effectiveness study. :::
::: {.source-note} 213. Julien Bek, Sequoia Capital, “Services: The New Software” (2026). The chapter uses the venture argument that budgets for performing work are larger than budgets for the tools used to perform it, not an empirical measurement. :::
Glossary notes
::: {.source-note} 208. The change compass is the author's final Book 2 framework: four points—where, people, effect, and control—inside the magnetic field of accumulated company context. See the Reader Glossary. :::
::: {.source-note} 209. Scoring evaluates a process through the same set of questions. The decision follows observable criteria rather than enthusiasm for the task. See the Reader Glossary. :::
::: {.source-note} 212. A baseline metric records the before state: what a process costs, how long it takes, how much work must be redone, and the volume before AI enters it. See the Reader Glossary. :::
::: {.source-note} 216. A gate is a control question before the next step. Until the answer is an honest yes, the next step does not begin. See the Reader Glossary. :::
Epilogue. Who Holds the Wheel
Quotes and epigraphs
::: {.source-note} 217. Lucius Annaeus Seneca, Moral Letters to Lucilius, Letter 71, section 3. The Latin maxim means that no wind can be favorable to a person who does not know the port toward which they are sailing; final quotation clearance remains open.
Open sources: en.wikisource.org — Letter. :::
Appendix 2. Will It Replace Us? AGI Without the Fog
Quotes and epigraphs
::: {.source-note} 218. A. M. Turing, “Computing Machinery and Intelligence” (1950). Turing replaces “Can machines think?” with the observable imitation game.
Open sources: Oxford Academic — Computing Machinery and Intelligence. :::
::: {.source-note} 219. On move 37 of game two, AlphaGo played where a professional was highly unlikely to play. The move first looked mistaken and later revealed a strong strategy. Fan Hui described it as outside the human repertoire. :::
Research, reports, and data
::: {.source-note} 221. Shin, Kim, and coauthors, PNAS (2023). Under the study's measures, professional Go players' move quality and novelty increased after superhuman AI appeared.
Open sources: PNAS — Superhuman AI and human decision-making. :::
::: {.source-note} 224. OECD and UNESCO AI-literacy frameworks emphasize critical evaluation, understanding AI's limits and effects, and remaining the author of one's work.
Open sources: OECD — Empowering Learners for the Age of AI; UNESCO — AI Competency Framework for Teachers. :::
History and institutional context
::: {.source-note} 220. Lee Sedol retired from professional Go in 2019. Yonhap and international reporting carried his explanation that an entity existed that could not be defeated. :::
::: {.source-note} 225. David Ricardo's principle of comparative advantage (1817): absolute superiority does not make it rational for one resource to perform every task.
Open sources: Library of Economics and Liberty — Ricardo's Principles. :::
Books and frameworks
::: {.source-note} 222. The “Levels of AGI” framework evaluates task breadth, performance, and autonomy instead of a binary AGI/no-AGI label.
Open sources: arXiv — Levels of AGI. :::
::: {.source-note} 223. METR develops a task-horizon measure: the length of human tasks an AI agent can complete reliably. The chapter uses the principle, not a current doubling-time figure.
Open sources: METR — Task-Completion Time Horizons. :::